linux/mm/swapfile.c

3692 lines
92 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* linux/mm/swapfile.c
*
* Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
* Swap reorganised 29.12.95, Stephen Tweedie
*/
#include <linux/blkdev.h>
#include <linux/mm.h>
#include <linux/sched/mm.h>
#include <linux/sched/task.h>
#include <linux/hugetlb.h>
#include <linux/mman.h>
#include <linux/slab.h>
#include <linux/kernel_stat.h>
#include <linux/swap.h>
#include <linux/vmalloc.h>
#include <linux/pagemap.h>
#include <linux/namei.h>
#include <linux/shmem_fs.h>
#include <linux/blk-cgroup.h>
#include <linux/random.h>
#include <linux/writeback.h>
#include <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/init.h>
ksm: let shared pages be swappable Initial implementation for swapping out KSM's shared pages: add page_referenced_ksm() and try_to_unmap_ksm(), which rmap.c calls when faced with a PageKsm page. Most of what's needed can be got from the rmap_items listed from the stable_node of the ksm page, without discovering the actual vma: so in this patch just fake up a struct vma for page_referenced_one() or try_to_unmap_one(), then refine that in the next patch. Add VM_NONLINEAR to ksm_madvise()'s list of exclusions: it has always been implicit there (being only set with VM_SHARED, already excluded), but let's make it explicit, to help justify the lack of nonlinear unmap. Rely on the page lock to protect against concurrent modifications to that page's node of the stable tree. The awkward part is not swapout but swapin: do_swap_page() and page_add_anon_rmap() now have to allow for new possibilities - perhaps a ksm page still in swapcache, perhaps a swapcache page associated with one location in one anon_vma now needed for another location or anon_vma. (And the vma might even be no longer VM_MERGEABLE when that happens.) ksm_might_need_to_copy() checks for that case, and supplies a duplicate page when necessary, simply leaving it to a subsequent pass of ksmd to rediscover the identity and merge them back into one ksm page. Disappointingly primitive: but the alternative would have to accumulate unswappable info about the swapped out ksm pages, limiting swappability. Remove page_add_ksm_rmap(): page_add_anon_rmap() now has to allow for the particular case it was handling, so just use it instead. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: Izik Eidus <ieidus@redhat.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Chris Wright <chrisw@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:59:24 +00:00
#include <linux/ksm.h>
#include <linux/rmap.h>
#include <linux/security.h>
#include <linux/backing-dev.h>
#include <linux/mutex.h>
#include <linux/capability.h>
#include <linux/syscalls.h>
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
#include <linux/memcontrol.h>
#include <linux/poll.h>
#include <linux/oom.h>
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
#include <linux/frontswap.h>
#include <linux/swapfile.h>
#include <linux/export.h>
mm/swap: add cache for swap slots allocation We add per cpu caches for swap slots that can be allocated and freed quickly without the need to touch the swap info lock. Two separate caches are maintained for swap slots allocated and swap slots returned. This is to allow the swap slots to be returned to the global pool in a batch so they will have a chance to be coaelesced with other slots in a cluster. We do not reuse the slots that are returned right away, as it may increase fragmentation of the slots. The swap allocation cache is protected by a mutex as we may sleep when searching for empty slots in cache. The swap free cache is protected by a spin lock as we cannot sleep in the free path. We refill the swap slots cache when we run out of slots, and we disable the swap slots cache and drain the slots if the global number of slots fall below a low watermark threshold. We re-enable the cache agian when the slots available are above a high watermark. [ying.huang@intel.com: use raw_cpu_ptr over this_cpu_ptr for swap slots access] [tim.c.chen@linux.intel.com: add comments on locks in swap_slots.h] Link: http://lkml.kernel.org/r/20170118180327.GA24225@linux.intel.com Link: http://lkml.kernel.org/r/35de301a4eaa8daa2977de6e987f2c154385eb66.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:39 +00:00
#include <linux/swap_slots.h>
mm/swapfile.c: sort swap entries before free To reduce the lock contention of swap_info_struct->lock when freeing swap entry. The freed swap entries will be collected in a per-CPU buffer firstly, and be really freed later in batch. During the batch freeing, if the consecutive swap entries in the per-CPU buffer belongs to same swap device, the swap_info_struct->lock needs to be acquired/released only once, so that the lock contention could be reduced greatly. But if there are multiple swap devices, it is possible that the lock may be unnecessarily released/acquired because the swap entries belong to the same swap device are non-consecutive in the per-CPU buffer. To solve the issue, the per-CPU buffer is sorted according to the swap device before freeing the swap entries. With the patch, the memory (some swapped out) free time reduced 11.6% (from 2.65s to 2.35s) in the vm-scalability swap-w-rand test case with 16 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test swapping, the test case creates 16 processes, which allocate and write to the anonymous pages until the RAM and part of the swap device is used up, finally the memory (some swapped out) is freed before exit. [akpm@linux-foundation.org: tweak comment] Link: http://lkml.kernel.org/r/20170525005916.25249-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:40:31 +00:00
#include <linux/sort.h>
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
#include <linux/completion.h>
#include <asm/tlbflush.h>
#include <linux/swapops.h>
#include <linux/swap_cgroup.h>
mm: create new mm/swap.h header file Patch series "MM changes to improve swap-over-NFS support". Assorted improvements for swap-via-filesystem. This is a resend of these patches, rebased on current HEAD. The only substantial changes is that swap_dirty_folio has replaced swap_set_page_dirty. Currently swap-via-fs (SWP_FS_OPS) doesn't work for any filesystem. It has previously worked for NFS but that broke a few releases back. This series changes to use a new ->swap_rw rather than ->readpage and ->direct_IO. It also makes other improvements. There is a companion series already in linux-next which fixes various issues with NFS. Once both series land, a final patch is needed which changes NFS over to use ->swap_rw. This patch (of 10): Many functions declared in include/linux/swap.h are only used within mm/ Create a new "mm/swap.h" and move some of these declarations there. Remove the redundant 'extern' from the function declarations. [akpm@linux-foundation.org: mm/memory-failure.c needs mm/swap.h] Link: https://lkml.kernel.org/r/164859751830.29473.5309689752169286816.stgit@noble.brown Link: https://lkml.kernel.org/r/164859778120.29473.11725907882296224053.stgit@noble.brown Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Geert Uytterhoeven <geert+renesas@glider.be> Cc: Trond Myklebust <trond.myklebust@hammerspace.com> Cc: Hugh Dickins <hughd@google.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Miaohe Lin <linmiaohe@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 01:20:47 +00:00
#include "swap.h"
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
unsigned char);
static void free_swap_count_continuations(struct swap_info_struct *);
static DEFINE_SPINLOCK(swap_lock);
static unsigned int nr_swapfiles;
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
atomic_long_t nr_swap_pages;
/*
* Some modules use swappable objects and may try to swap them out under
* memory pressure (via the shrinker). Before doing so, they may wish to
* check to see if any swap space is available.
*/
EXPORT_SYMBOL_GPL(nr_swap_pages);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
long total_swap_pages;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
static int least_priority = -1;
mm/swap: cache maximum swapfile size when init swap We used to have swapfile_maximum_size() fetching a maximum value of swapfile size per-arch. As the caller of max_swapfile_size() grows, this patch introduce a variable "swapfile_maximum_size" and cache the value of old max_swapfile_size(), so that we don't need to calculate the value every time. Caching the value in swapfile_init() is safe because when reaching the phase we should have initialized all the relevant information. Here the major arch to take care of is x86, which defines the max swapfile size based on L1TF mitigation. Here both X86_BUG_L1TF or l1tf_mitigation should have been setup properly when reaching swapfile_init(). As a reference, the code path looks like this for x86: - start_kernel - setup_arch - early_cpu_init - early_identify_cpu --> setup X86_BUG_L1TF - parse_early_param - l1tf_cmdline --> set l1tf_mitigation - check_bugs - l1tf_select_mitigation --> set l1tf_mitigation - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) The swapfile size only depends on swp pte format on non-x86 archs, so caching it is safe too. Since at it, rename max_swapfile_size() to arch_max_swapfile_size() because arch can define its own function, so it's more straightforward to have "arch_" as its prefix. At the meantime, export swapfile_maximum_size to replace the old usages of max_swapfile_size(). [peterx@redhat.com: declare arch_max_swapfile_size) in swapfile.h] Link: https://lkml.kernel.org/r/YxTh1GuC6ro5fKL5@xz-m1.local Link: https://lkml.kernel.org/r/20220811161331.37055-7-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:30 +00:00
unsigned long swapfile_maximum_size;
mm/swap: cache swap migration A/D bits support Introduce a variable swap_migration_ad_supported to cache whether the arch supports swap migration A/D bits. Here one thing to mention is that SWP_MIG_TOTAL_BITS will internally reference the other macro MAX_PHYSMEM_BITS, which is a function call on x86 (constant on all the rest of archs). It's safe to reference it in swapfile_init() because when reaching here we're already during initcalls level 4 so we must have initialized 5-level pgtable for x86_64 (right after early_identify_cpu() finishes). - start_kernel - setup_arch - early_cpu_init - get_cpu_cap --> fetch from CPUID (including X86_FEATURE_LA57) - early_identify_cpu --> clear X86_FEATURE_LA57 (if early lvl5 not enabled (USE_EARLY_PGTABLE_L5)) - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) This should slightly speed up the migration swap entry handlings. Link: https://lkml.kernel.org/r/20220811161331.37055-8-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:31 +00:00
#ifdef CONFIG_MIGRATION
bool swap_migration_ad_supported;
#endif /* CONFIG_MIGRATION */
static const char Bad_file[] = "Bad swap file entry ";
static const char Unused_file[] = "Unused swap file entry ";
static const char Bad_offset[] = "Bad swap offset entry ";
static const char Unused_offset[] = "Unused swap offset entry ";
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
/*
* all active swap_info_structs
* protected with swap_lock, and ordered by priority.
*/
static PLIST_HEAD(swap_active_head);
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
/*
* all available (active, not full) swap_info_structs
* protected with swap_avail_lock, ordered by priority.
* This is used by folio_alloc_swap() instead of swap_active_head
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
* because swap_active_head includes all swap_info_structs,
* but folio_alloc_swap() doesn't need to look at full ones.
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
* This uses its own lock instead of swap_lock because when a
* swap_info_struct changes between not-full/full, it needs to
* add/remove itself to/from this list, but the swap_info_struct->lock
* is held and the locking order requires swap_lock to be taken
* before any swap_info_struct->lock.
*/
static struct plist_head *swap_avail_heads;
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
static DEFINE_SPINLOCK(swap_avail_lock);
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
struct swap_info_struct *swap_info[MAX_SWAPFILES];
static DEFINE_MUTEX(swapon_mutex);
static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
/* Activity counter to indicate that a swapon or swapoff has occurred */
static atomic_t proc_poll_event = ATOMIC_INIT(0);
atomic_t nr_rotate_swap = ATOMIC_INIT(0);
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
static struct swap_info_struct *swap_type_to_swap_info(int type)
{
mm, swap: remove unnecessary smp_rmb() in swap_type_to_swap_info() Before commit c10d38cc8d3e ("mm, swap: bounds check swap_info array accesses to avoid NULL derefs"), the typical code to reference the swap_info[] is as follows, type = swp_type(swp_entry); if (type >= nr_swapfiles) /* handle invalid swp_entry */; p = swap_info[type]; /* access fields of *p. OOPS! p may be NULL! */ Because the ordering isn't guaranteed, it's possible that swap_info[type] is read before "nr_swapfiles". And that may result in NULL pointer dereference. So after commit c10d38cc8d3e, the code becomes, struct swap_info_struct *swap_type_to_swap_info(int type) { if (type >= READ_ONCE(nr_swapfiles)) return NULL; smp_rmb(); return READ_ONCE(swap_info[type]); } /* users */ type = swp_type(swp_entry); p = swap_type_to_swap_info(type); if (!p) /* handle invalid swp_entry */; /* dereference p */ Where the value of swap_info[type] (that is, "p") is checked to be non-zero before being dereferenced. So, the NULL deferencing becomes impossible even if "nr_swapfiles" is read after swap_info[type]. Therefore, the "smp_rmb()" becomes unnecessary. And, we don't even need to read "nr_swapfiles" here. Because the non-zero checking for "p" is sufficient. We just need to make sure we will not access out of the boundary of the array. With the change, nr_swapfiles will only be accessed with swap_lock held, except in swapcache_free_entries(). Where the absolute correctness of the value isn't needed, as described in the comments. We still need to guarantee swap_info[type] is read before being dereferenced. That can be satisfied via the data dependency ordering enforced by READ_ONCE(swap_info[type]). This needs to be paired with proper write barriers. So smp_store_release() is used in alloc_swap_info() to guarantee the fields of *swap_info[type] is initialized before swap_info[type] itself being written. Note that the fields of *swap_info[type] is initialized to be 0 via kvzalloc() firstly. The assignment and deferencing of swap_info[type] is like rcu_assign_pointer() and rcu_dereference(). Link: https://lkml.kernel.org/r/20210520073301.1676294-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Dan Carpenter <dan.carpenter@oracle.com> Cc: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Cc: Miaohe Lin <linmiaohe@huawei.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:37:09 +00:00
if (type >= MAX_SWAPFILES)
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
return NULL;
mm, swap: remove unnecessary smp_rmb() in swap_type_to_swap_info() Before commit c10d38cc8d3e ("mm, swap: bounds check swap_info array accesses to avoid NULL derefs"), the typical code to reference the swap_info[] is as follows, type = swp_type(swp_entry); if (type >= nr_swapfiles) /* handle invalid swp_entry */; p = swap_info[type]; /* access fields of *p. OOPS! p may be NULL! */ Because the ordering isn't guaranteed, it's possible that swap_info[type] is read before "nr_swapfiles". And that may result in NULL pointer dereference. So after commit c10d38cc8d3e, the code becomes, struct swap_info_struct *swap_type_to_swap_info(int type) { if (type >= READ_ONCE(nr_swapfiles)) return NULL; smp_rmb(); return READ_ONCE(swap_info[type]); } /* users */ type = swp_type(swp_entry); p = swap_type_to_swap_info(type); if (!p) /* handle invalid swp_entry */; /* dereference p */ Where the value of swap_info[type] (that is, "p") is checked to be non-zero before being dereferenced. So, the NULL deferencing becomes impossible even if "nr_swapfiles" is read after swap_info[type]. Therefore, the "smp_rmb()" becomes unnecessary. And, we don't even need to read "nr_swapfiles" here. Because the non-zero checking for "p" is sufficient. We just need to make sure we will not access out of the boundary of the array. With the change, nr_swapfiles will only be accessed with swap_lock held, except in swapcache_free_entries(). Where the absolute correctness of the value isn't needed, as described in the comments. We still need to guarantee swap_info[type] is read before being dereferenced. That can be satisfied via the data dependency ordering enforced by READ_ONCE(swap_info[type]). This needs to be paired with proper write barriers. So smp_store_release() is used in alloc_swap_info() to guarantee the fields of *swap_info[type] is initialized before swap_info[type] itself being written. Note that the fields of *swap_info[type] is initialized to be 0 via kvzalloc() firstly. The assignment and deferencing of swap_info[type] is like rcu_assign_pointer() and rcu_dereference(). Link: https://lkml.kernel.org/r/20210520073301.1676294-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Dan Carpenter <dan.carpenter@oracle.com> Cc: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Cc: Miaohe Lin <linmiaohe@huawei.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:37:09 +00:00
return READ_ONCE(swap_info[type]); /* rcu_dereference() */
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
}
static inline unsigned char swap_count(unsigned char ent)
{
return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
}
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
/* Reclaim the swap entry anyway if possible */
#define TTRS_ANYWAY 0x1
/*
* Reclaim the swap entry if there are no more mappings of the
* corresponding page
*/
#define TTRS_UNMAPPED 0x2
/* Reclaim the swap entry if swap is getting full*/
#define TTRS_FULL 0x4
/* returns 1 if swap entry is freed */
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
static int __try_to_reclaim_swap(struct swap_info_struct *si,
unsigned long offset, unsigned long flags)
{
swp_entry_t entry = swp_entry(si->type, offset);
struct folio *folio;
int ret = 0;
folio = filemap_get_folio(swap_address_space(entry), offset);
if (!folio)
return 0;
/*
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
* When this function is called from scan_swap_map_slots() and it's
* called by vmscan.c at reclaiming folios. So we hold a folio lock
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
* here. We have to use trylock for avoiding deadlock. This is a special
* case and you should use folio_free_swap() with explicit folio_lock()
* in usual operations.
*/
if (folio_trylock(folio)) {
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
if ((flags & TTRS_ANYWAY) ||
((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)))
ret = folio_free_swap(folio);
folio_unlock(folio);
}
folio_put(folio);
return ret;
}
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
static inline struct swap_extent *first_se(struct swap_info_struct *sis)
{
struct rb_node *rb = rb_first(&sis->swap_extent_root);
return rb_entry(rb, struct swap_extent, rb_node);
}
static inline struct swap_extent *next_se(struct swap_extent *se)
{
struct rb_node *rb = rb_next(&se->rb_node);
return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
}
/*
* swapon tell device that all the old swap contents can be discarded,
* to allow the swap device to optimize its wear-levelling.
*/
static int discard_swap(struct swap_info_struct *si)
{
struct swap_extent *se;
sector_t start_block;
sector_t nr_blocks;
int err = 0;
/* Do not discard the swap header page! */
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
se = first_se(si);
start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
if (nr_blocks) {
err = blkdev_issue_discard(si->bdev, start_block,
nr_blocks, GFP_KERNEL);
if (err)
return err;
cond_resched();
}
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
for (se = next_se(se); se; se = next_se(se)) {
start_block = se->start_block << (PAGE_SHIFT - 9);
nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
err = blkdev_issue_discard(si->bdev, start_block,
nr_blocks, GFP_KERNEL);
if (err)
break;
cond_resched();
}
return err; /* That will often be -EOPNOTSUPP */
}
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
static struct swap_extent *
offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
{
struct swap_extent *se;
struct rb_node *rb;
rb = sis->swap_extent_root.rb_node;
while (rb) {
se = rb_entry(rb, struct swap_extent, rb_node);
if (offset < se->start_page)
rb = rb->rb_left;
else if (offset >= se->start_page + se->nr_pages)
rb = rb->rb_right;
else
return se;
}
/* It *must* be present */
BUG();
}
sector_t swap_page_sector(struct page *page)
{
struct swap_info_struct *sis = page_swap_info(page);
struct swap_extent *se;
sector_t sector;
pgoff_t offset;
offset = __page_file_index(page);
se = offset_to_swap_extent(sis, offset);
sector = se->start_block + (offset - se->start_page);
return sector << (PAGE_SHIFT - 9);
}
/*
* swap allocation tell device that a cluster of swap can now be discarded,
* to allow the swap device to optimize its wear-levelling.
*/
static void discard_swap_cluster(struct swap_info_struct *si,
pgoff_t start_page, pgoff_t nr_pages)
{
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
struct swap_extent *se = offset_to_swap_extent(si, start_page);
while (nr_pages) {
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
pgoff_t offset = start_page - se->start_page;
sector_t start_block = se->start_block + offset;
sector_t nr_blocks = se->nr_pages - offset;
if (nr_blocks > nr_pages)
nr_blocks = nr_pages;
start_page += nr_blocks;
nr_pages -= nr_blocks;
start_block <<= PAGE_SHIFT - 9;
nr_blocks <<= PAGE_SHIFT - 9;
if (blkdev_issue_discard(si->bdev, start_block,
nr_blocks, GFP_NOIO))
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
break;
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
se = next_se(se);
}
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
#ifdef CONFIG_THP_SWAP
#define SWAPFILE_CLUSTER HPAGE_PMD_NR
#define swap_entry_size(size) (size)
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
#else
#define SWAPFILE_CLUSTER 256
/*
* Define swap_entry_size() as constant to let compiler to optimize
* out some code if !CONFIG_THP_SWAP
*/
#define swap_entry_size(size) 1
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
#endif
#define LATENCY_LIMIT 256
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
static inline void cluster_set_flag(struct swap_cluster_info *info,
unsigned int flag)
{
info->flags = flag;
}
static inline unsigned int cluster_count(struct swap_cluster_info *info)
{
return info->data;
}
static inline void cluster_set_count(struct swap_cluster_info *info,
unsigned int c)
{
info->data = c;
}
static inline void cluster_set_count_flag(struct swap_cluster_info *info,
unsigned int c, unsigned int f)
{
info->flags = f;
info->data = c;
}
static inline unsigned int cluster_next(struct swap_cluster_info *info)
{
return info->data;
}
static inline void cluster_set_next(struct swap_cluster_info *info,
unsigned int n)
{
info->data = n;
}
static inline void cluster_set_next_flag(struct swap_cluster_info *info,
unsigned int n, unsigned int f)
{
info->flags = f;
info->data = n;
}
static inline bool cluster_is_free(struct swap_cluster_info *info)
{
return info->flags & CLUSTER_FLAG_FREE;
}
static inline bool cluster_is_null(struct swap_cluster_info *info)
{
return info->flags & CLUSTER_FLAG_NEXT_NULL;
}
static inline void cluster_set_null(struct swap_cluster_info *info)
{
info->flags = CLUSTER_FLAG_NEXT_NULL;
info->data = 0;
}
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
static inline bool cluster_is_huge(struct swap_cluster_info *info)
{
if (IS_ENABLED(CONFIG_THP_SWAP))
return info->flags & CLUSTER_FLAG_HUGE;
return false;
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
}
static inline void cluster_clear_huge(struct swap_cluster_info *info)
{
info->flags &= ~CLUSTER_FLAG_HUGE;
}
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
unsigned long offset)
{
struct swap_cluster_info *ci;
ci = si->cluster_info;
if (ci) {
ci += offset / SWAPFILE_CLUSTER;
spin_lock(&ci->lock);
}
return ci;
}
static inline void unlock_cluster(struct swap_cluster_info *ci)
{
if (ci)
spin_unlock(&ci->lock);
}
mm: swap: add comments to lock_cluster_or_swap_info() Patch series "swap: THP optimizing refactoring", v4. Now the THP (Transparent Huge Page) swap optimizing is implemented in the way like below, #ifdef CONFIG_THP_SWAP huge_function(...) { } #else normal_function(...) { } #endif general_function(...) { if (huge) return thp_function(...); else return normal_function(...); } As pointed out by Dave Hansen, this will, 1. Create a new, wholly untested code path for huge page 2. Create two places to patch bugs 3. Are not reusing code when possible This patchset is to address these problems via merging huge/normal code path/functions if possible. One concern is that this may cause code size to dilate when !CONFIG_TRANSPARENT_HUGEPAGE. The data shows that most refactoring will only cause quite slight code size increase. This patch (of 8): To improve code readability. Link: http://lkml.kernel.org/r/20180720071845.17920-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:01 +00:00
/*
* Determine the locking method in use for this device. Return
* swap_cluster_info if SSD-style cluster-based locking is in place.
*/
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
static inline struct swap_cluster_info *lock_cluster_or_swap_info(
mm: swap: add comments to lock_cluster_or_swap_info() Patch series "swap: THP optimizing refactoring", v4. Now the THP (Transparent Huge Page) swap optimizing is implemented in the way like below, #ifdef CONFIG_THP_SWAP huge_function(...) { } #else normal_function(...) { } #endif general_function(...) { if (huge) return thp_function(...); else return normal_function(...); } As pointed out by Dave Hansen, this will, 1. Create a new, wholly untested code path for huge page 2. Create two places to patch bugs 3. Are not reusing code when possible This patchset is to address these problems via merging huge/normal code path/functions if possible. One concern is that this may cause code size to dilate when !CONFIG_TRANSPARENT_HUGEPAGE. The data shows that most refactoring will only cause quite slight code size increase. This patch (of 8): To improve code readability. Link: http://lkml.kernel.org/r/20180720071845.17920-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:01 +00:00
struct swap_info_struct *si, unsigned long offset)
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
{
struct swap_cluster_info *ci;
mm: swap: add comments to lock_cluster_or_swap_info() Patch series "swap: THP optimizing refactoring", v4. Now the THP (Transparent Huge Page) swap optimizing is implemented in the way like below, #ifdef CONFIG_THP_SWAP huge_function(...) { } #else normal_function(...) { } #endif general_function(...) { if (huge) return thp_function(...); else return normal_function(...); } As pointed out by Dave Hansen, this will, 1. Create a new, wholly untested code path for huge page 2. Create two places to patch bugs 3. Are not reusing code when possible This patchset is to address these problems via merging huge/normal code path/functions if possible. One concern is that this may cause code size to dilate when !CONFIG_TRANSPARENT_HUGEPAGE. The data shows that most refactoring will only cause quite slight code size increase. This patch (of 8): To improve code readability. Link: http://lkml.kernel.org/r/20180720071845.17920-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:01 +00:00
/* Try to use fine-grained SSD-style locking if available: */
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
ci = lock_cluster(si, offset);
mm: swap: add comments to lock_cluster_or_swap_info() Patch series "swap: THP optimizing refactoring", v4. Now the THP (Transparent Huge Page) swap optimizing is implemented in the way like below, #ifdef CONFIG_THP_SWAP huge_function(...) { } #else normal_function(...) { } #endif general_function(...) { if (huge) return thp_function(...); else return normal_function(...); } As pointed out by Dave Hansen, this will, 1. Create a new, wholly untested code path for huge page 2. Create two places to patch bugs 3. Are not reusing code when possible This patchset is to address these problems via merging huge/normal code path/functions if possible. One concern is that this may cause code size to dilate when !CONFIG_TRANSPARENT_HUGEPAGE. The data shows that most refactoring will only cause quite slight code size increase. This patch (of 8): To improve code readability. Link: http://lkml.kernel.org/r/20180720071845.17920-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:01 +00:00
/* Otherwise, fall back to traditional, coarse locking: */
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
if (!ci)
spin_lock(&si->lock);
return ci;
}
static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
struct swap_cluster_info *ci)
{
if (ci)
unlock_cluster(ci);
else
spin_unlock(&si->lock);
}
static inline bool cluster_list_empty(struct swap_cluster_list *list)
{
return cluster_is_null(&list->head);
}
static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
{
return cluster_next(&list->head);
}
static void cluster_list_init(struct swap_cluster_list *list)
{
cluster_set_null(&list->head);
cluster_set_null(&list->tail);
}
static void cluster_list_add_tail(struct swap_cluster_list *list,
struct swap_cluster_info *ci,
unsigned int idx)
{
if (cluster_list_empty(list)) {
cluster_set_next_flag(&list->head, idx, 0);
cluster_set_next_flag(&list->tail, idx, 0);
} else {
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci_tail;
unsigned int tail = cluster_next(&list->tail);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
/*
* Nested cluster lock, but both cluster locks are
* only acquired when we held swap_info_struct->lock
*/
ci_tail = ci + tail;
spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
cluster_set_next(ci_tail, idx);
spin_unlock(&ci_tail->lock);
cluster_set_next_flag(&list->tail, idx, 0);
}
}
static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
struct swap_cluster_info *ci)
{
unsigned int idx;
idx = cluster_next(&list->head);
if (cluster_next(&list->tail) == idx) {
cluster_set_null(&list->head);
cluster_set_null(&list->tail);
} else
cluster_set_next_flag(&list->head,
cluster_next(&ci[idx]), 0);
return idx;
}
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
/* Add a cluster to discard list and schedule it to do discard */
static void swap_cluster_schedule_discard(struct swap_info_struct *si,
unsigned int idx)
{
/*
* If scan_swap_map_slots() can't find a free cluster, it will check
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
* si->swap_map directly. To make sure the discarding cluster isn't
* taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
* It will be cleared after discard
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
*/
memset(si->swap_map + idx * SWAPFILE_CLUSTER,
SWAP_MAP_BAD, SWAPFILE_CLUSTER);
cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
schedule_work(&si->discard_work);
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
{
struct swap_cluster_info *ci = si->cluster_info;
cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
cluster_list_add_tail(&si->free_clusters, ci, idx);
}
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
/*
* Doing discard actually. After a cluster discard is finished, the cluster
* will be added to free cluster list. caller should hold si->lock.
*/
static void swap_do_scheduled_discard(struct swap_info_struct *si)
{
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *info, *ci;
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
unsigned int idx;
info = si->cluster_info;
while (!cluster_list_empty(&si->discard_clusters)) {
idx = cluster_list_del_first(&si->discard_clusters, info);
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
spin_unlock(&si->lock);
discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
SWAPFILE_CLUSTER);
spin_lock(&si->lock);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
__free_cluster(si, idx);
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
memset(si->swap_map + idx * SWAPFILE_CLUSTER,
0, SWAPFILE_CLUSTER);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster(ci);
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
}
}
static void swap_discard_work(struct work_struct *work)
{
struct swap_info_struct *si;
si = container_of(work, struct swap_info_struct, discard_work);
spin_lock(&si->lock);
swap_do_scheduled_discard(si);
spin_unlock(&si->lock);
}
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
static void swap_users_ref_free(struct percpu_ref *ref)
{
struct swap_info_struct *si;
si = container_of(ref, struct swap_info_struct, users);
complete(&si->comp);
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
{
struct swap_cluster_info *ci = si->cluster_info;
VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
cluster_list_del_first(&si->free_clusters, ci);
cluster_set_count_flag(ci + idx, 0, 0);
}
static void free_cluster(struct swap_info_struct *si, unsigned long idx)
{
struct swap_cluster_info *ci = si->cluster_info + idx;
VM_BUG_ON(cluster_count(ci) != 0);
/*
* If the swap is discardable, prepare discard the cluster
* instead of free it immediately. The cluster will be freed
* after discard.
*/
if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
(SWP_WRITEOK | SWP_PAGE_DISCARD)) {
swap_cluster_schedule_discard(si, idx);
return;
}
__free_cluster(si, idx);
}
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* The cluster corresponding to page_nr will be used. The cluster will be
* removed from free cluster list and its usage counter will be increased.
*/
static void inc_cluster_info_page(struct swap_info_struct *p,
struct swap_cluster_info *cluster_info, unsigned long page_nr)
{
unsigned long idx = page_nr / SWAPFILE_CLUSTER;
if (!cluster_info)
return;
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (cluster_is_free(&cluster_info[idx]))
alloc_cluster(p, idx);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
cluster_set_count(&cluster_info[idx],
cluster_count(&cluster_info[idx]) + 1);
}
/*
* The cluster corresponding to page_nr decreases one usage. If the usage
* counter becomes 0, which means no page in the cluster is in using, we can
* optionally discard the cluster and add it to free cluster list.
*/
static void dec_cluster_info_page(struct swap_info_struct *p,
struct swap_cluster_info *cluster_info, unsigned long page_nr)
{
unsigned long idx = page_nr / SWAPFILE_CLUSTER;
if (!cluster_info)
return;
VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
cluster_set_count(&cluster_info[idx],
cluster_count(&cluster_info[idx]) - 1);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (cluster_count(&cluster_info[idx]) == 0)
free_cluster(p, idx);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
}
/*
* It's possible scan_swap_map_slots() uses a free cluster in the middle of free
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
* cluster list. Avoiding such abuse to avoid list corruption.
*/
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
static bool
scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
unsigned long offset)
{
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
struct percpu_cluster *percpu_cluster;
bool conflict;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
offset /= SWAPFILE_CLUSTER;
conflict = !cluster_list_empty(&si->free_clusters) &&
offset != cluster_list_first(&si->free_clusters) &&
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
cluster_is_free(&si->cluster_info[offset]);
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
if (!conflict)
return false;
percpu_cluster = this_cpu_ptr(si->percpu_cluster);
cluster_set_null(&percpu_cluster->index);
return true;
}
/*
* Try to get a swap entry from current cpu's swap entry pool (a cluster). This
* might involve allocating a new cluster for current CPU too.
*/
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
unsigned long *offset, unsigned long *scan_base)
{
struct percpu_cluster *cluster;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci;
unsigned long tmp, max;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
new_cluster:
cluster = this_cpu_ptr(si->percpu_cluster);
if (cluster_is_null(&cluster->index)) {
if (!cluster_list_empty(&si->free_clusters)) {
cluster->index = si->free_clusters.head;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
cluster->next = cluster_next(&cluster->index) *
SWAPFILE_CLUSTER;
} else if (!cluster_list_empty(&si->discard_clusters)) {
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
/*
* we don't have free cluster but have some clusters in
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
* discarding, do discard now and reclaim them, then
* reread cluster_next_cpu since we dropped si->lock
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
*/
swap_do_scheduled_discard(si);
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
*scan_base = this_cpu_read(*si->cluster_next_cpu);
*offset = *scan_base;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
goto new_cluster;
} else
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
return false;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
}
/*
* Other CPUs can use our cluster if they can't find a free cluster,
* check if there is still free entry in the cluster
*/
tmp = cluster->next;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
max = min_t(unsigned long, si->max,
(cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
if (tmp < max) {
ci = lock_cluster(si, tmp);
while (tmp < max) {
if (!si->swap_map[tmp])
break;
tmp++;
}
unlock_cluster(ci);
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
}
if (tmp >= max) {
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
cluster_set_null(&cluster->index);
goto new_cluster;
}
cluster->next = tmp + 1;
*offset = tmp;
*scan_base = tmp;
return true;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
}
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
static void __del_from_avail_list(struct swap_info_struct *p)
{
int nid;
for_each_node(nid)
plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
}
static void del_from_avail_list(struct swap_info_struct *p)
{
spin_lock(&swap_avail_lock);
__del_from_avail_list(p);
spin_unlock(&swap_avail_lock);
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
unsigned int nr_entries)
{
unsigned int end = offset + nr_entries - 1;
if (offset == si->lowest_bit)
si->lowest_bit += nr_entries;
if (end == si->highest_bit)
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (si->inuse_pages == si->pages) {
si->lowest_bit = si->max;
si->highest_bit = 0;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
del_from_avail_list(si);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
}
}
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
static void add_to_avail_list(struct swap_info_struct *p)
{
int nid;
spin_lock(&swap_avail_lock);
for_each_node(nid) {
WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
}
spin_unlock(&swap_avail_lock);
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
unsigned int nr_entries)
{
unsigned long begin = offset;
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
unsigned long end = offset + nr_entries - 1;
void (*swap_slot_free_notify)(struct block_device *, unsigned long);
if (offset < si->lowest_bit)
si->lowest_bit = offset;
if (end > si->highest_bit) {
bool was_full = !si->highest_bit;
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
WRITE_ONCE(si->highest_bit, end);
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
if (was_full && (si->flags & SWP_WRITEOK))
add_to_avail_list(si);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
}
atomic_long_add(nr_entries, &nr_swap_pages);
WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (si->flags & SWP_BLKDEV)
swap_slot_free_notify =
si->bdev->bd_disk->fops->swap_slot_free_notify;
else
swap_slot_free_notify = NULL;
while (offset <= end) {
arch_swap_invalidate_page(si->type, offset);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
frontswap_invalidate_page(si->type, offset);
if (swap_slot_free_notify)
swap_slot_free_notify(si->bdev, offset);
offset++;
}
clear_shadow_from_swap_cache(si->type, begin, end);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
}
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
{
unsigned long prev;
if (!(si->flags & SWP_SOLIDSTATE)) {
si->cluster_next = next;
return;
}
prev = this_cpu_read(*si->cluster_next_cpu);
/*
* Cross the swap address space size aligned trunk, choose
* another trunk randomly to avoid lock contention on swap
* address space if possible.
*/
if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
(next >> SWAP_ADDRESS_SPACE_SHIFT)) {
/* No free swap slots available */
if (si->highest_bit <= si->lowest_bit)
return;
next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
next = max_t(unsigned int, next, si->lowest_bit);
}
this_cpu_write(*si->cluster_next_cpu, next);
}
static bool swap_offset_available_and_locked(struct swap_info_struct *si,
unsigned long offset)
{
if (data_race(!si->swap_map[offset])) {
spin_lock(&si->lock);
return true;
}
if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
spin_lock(&si->lock);
return true;
}
return false;
}
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
static int scan_swap_map_slots(struct swap_info_struct *si,
unsigned char usage, int nr,
swp_entry_t slots[])
{
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci;
unsigned long offset;
unsigned long scan_base;
unsigned long last_in_cluster = 0;
int latency_ration = LATENCY_LIMIT;
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
int n_ret = 0;
swap: try to scan more free slots even when fragmented Now, the scalability of swap code will drop much when the swap device becomes fragmented, because the swap slots allocation batching stops working. To solve the problem, in this patch, we will try to scan a little more swap slots with restricted effort to batch the swap slots allocation even if the swap device is fragmented. Test shows that the benchmark score can increase up to 37.1% with the patch. Details are as follows. The swap code has a per-cpu cache of swap slots. These batch swap space allocations to improve swap subsystem scaling. In the following code path, add_to_swap() get_swap_page() refill_swap_slots_cache() get_swap_pages() scan_swap_map_slots() scan_swap_map_slots() and get_swap_pages() can return multiple swap slots for each call. These slots will be cached in the per-CPU swap slots cache, so that several following swap slot requests will be fulfilled there to avoid the lock contention in the lower level swap space allocation/freeing code path. But this only works when there are free swap clusters. If a swap device becomes so fragmented that there's no free swap clusters, scan_swap_map_slots() and get_swap_pages() will return only one swap slot for each call in the above code path. Effectively, this falls back to the situation before the swap slots cache was introduced, the heavy lock contention on the swap related locks kills the scalability. Why does it work in this way? Because the swap device could be large, and the free swap slot scanning could be quite time consuming, to avoid taking too much time to scanning free swap slots, the conservative method was used. In fact, this can be improved via scanning a little more free slots with strictly restricted effort. Which is implemented in this patch. In scan_swap_map_slots(), after the first free swap slot is gotten, we will try to scan a little more, but only if we haven't scanned too many slots (< LATENCY_LIMIT). That is, the added scanning latency is strictly restricted. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. Multiple ram disks are configured as the swap devices. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random, so the swap space becomes highly fragmented during the test. In the original implementation, the lock contention on swap related locks is very heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap: 21.03 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.92 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.72 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 0.69 While after applying this patch, it becomes, _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 4.89 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 3.85 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.1 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88 That is, the lock contention on the swap locks is eliminated. And the pmbench score increases 37.1%. The swapin throughput increases 45.7% from 2.02 GB/s to 2.94 GB/s. While the swapout throughput increases 45.3% from 2.04 GB/s to 2.97 GB/s. Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:10 +00:00
bool scanned_many = false;
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
/*
* We try to cluster swap pages by allocating them sequentially
* in swap. Once we've allocated SWAPFILE_CLUSTER pages this
* way, however, we resort to first-free allocation, starting
* a new cluster. This prevents us from scattering swap pages
* all over the entire swap partition, so that we reduce
* overall disk seek times between swap pages. -- sct
* But we do now try to find an empty cluster. -Andrea
* And we let swap pages go all over an SSD partition. Hugh
*/
si->flags += SWP_SCANNING;
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
/*
* Use percpu scan base for SSD to reduce lock contention on
* cluster and swap cache. For HDD, sequential access is more
* important.
*/
if (si->flags & SWP_SOLIDSTATE)
scan_base = this_cpu_read(*si->cluster_next_cpu);
else
scan_base = si->cluster_next;
offset = scan_base;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
/* SSD algorithm */
if (si->cluster_info) {
if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
goto scan;
} else if (unlikely(!si->cluster_nr--)) {
if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
si->cluster_nr = SWAPFILE_CLUSTER - 1;
goto checks;
}
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
/*
* If seek is expensive, start searching for new cluster from
* start of partition, to minimize the span of allocated swap.
* If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
* case, just handled by scan_swap_map_try_ssd_cluster() above.
*/
scan_base = offset = si->lowest_bit;
last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
/* Locate the first empty (unaligned) cluster */
for (; last_in_cluster <= si->highest_bit; offset++) {
if (si->swap_map[offset])
last_in_cluster = offset + SWAPFILE_CLUSTER;
else if (offset == last_in_cluster) {
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&si->lock);
offset -= SWAPFILE_CLUSTER - 1;
si->cluster_next = offset;
si->cluster_nr = SWAPFILE_CLUSTER - 1;
goto checks;
}
if (unlikely(--latency_ration < 0)) {
cond_resched();
latency_ration = LATENCY_LIMIT;
}
}
offset = scan_base;
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&si->lock);
si->cluster_nr = SWAPFILE_CLUSTER - 1;
}
checks:
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
if (si->cluster_info) {
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
/* take a break if we already got some slots */
if (n_ret)
goto done;
if (!scan_swap_map_try_ssd_cluster(si, &offset,
&scan_base))
goto scan;
}
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
}
if (!(si->flags & SWP_WRITEOK))
goto no_page;
if (!si->highest_bit)
goto no_page;
if (offset > si->highest_bit)
scan_base = offset = si->lowest_bit;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
ci = lock_cluster(si, offset);
/* reuse swap entry of cache-only swap if not busy. */
if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
int swap_was_freed;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster(ci);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&si->lock);
/* entry was freed successfully, try to use this again */
if (swap_was_freed)
goto checks;
goto scan; /* check next one */
}
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
if (si->swap_map[offset]) {
unlock_cluster(ci);
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
if (!n_ret)
goto scan;
else
goto done;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
}
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
WRITE_ONCE(si->swap_map[offset], usage);
inc_cluster_info_page(si, si->cluster_info, offset);
unlock_cluster(ci);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
swap_range_alloc(si, offset, 1);
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
slots[n_ret++] = swp_entry(si->type, offset);
/* got enough slots or reach max slots? */
if ((n_ret == nr) || (offset >= si->highest_bit))
goto done;
/* search for next available slot */
/* time to take a break? */
if (unlikely(--latency_ration < 0)) {
if (n_ret)
goto done;
spin_unlock(&si->lock);
cond_resched();
spin_lock(&si->lock);
latency_ration = LATENCY_LIMIT;
}
/* try to get more slots in cluster */
if (si->cluster_info) {
if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
goto checks;
} else if (si->cluster_nr && !si->swap_map[++offset]) {
/* non-ssd case, still more slots in cluster? */
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
--si->cluster_nr;
goto checks;
}
swap: try to scan more free slots even when fragmented Now, the scalability of swap code will drop much when the swap device becomes fragmented, because the swap slots allocation batching stops working. To solve the problem, in this patch, we will try to scan a little more swap slots with restricted effort to batch the swap slots allocation even if the swap device is fragmented. Test shows that the benchmark score can increase up to 37.1% with the patch. Details are as follows. The swap code has a per-cpu cache of swap slots. These batch swap space allocations to improve swap subsystem scaling. In the following code path, add_to_swap() get_swap_page() refill_swap_slots_cache() get_swap_pages() scan_swap_map_slots() scan_swap_map_slots() and get_swap_pages() can return multiple swap slots for each call. These slots will be cached in the per-CPU swap slots cache, so that several following swap slot requests will be fulfilled there to avoid the lock contention in the lower level swap space allocation/freeing code path. But this only works when there are free swap clusters. If a swap device becomes so fragmented that there's no free swap clusters, scan_swap_map_slots() and get_swap_pages() will return only one swap slot for each call in the above code path. Effectively, this falls back to the situation before the swap slots cache was introduced, the heavy lock contention on the swap related locks kills the scalability. Why does it work in this way? Because the swap device could be large, and the free swap slot scanning could be quite time consuming, to avoid taking too much time to scanning free swap slots, the conservative method was used. In fact, this can be improved via scanning a little more free slots with strictly restricted effort. Which is implemented in this patch. In scan_swap_map_slots(), after the first free swap slot is gotten, we will try to scan a little more, but only if we haven't scanned too many slots (< LATENCY_LIMIT). That is, the added scanning latency is strictly restricted. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. Multiple ram disks are configured as the swap devices. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random, so the swap space becomes highly fragmented during the test. In the original implementation, the lock contention on swap related locks is very heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap: 21.03 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.92 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.72 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 0.69 While after applying this patch, it becomes, _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 4.89 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 3.85 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.1 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88 That is, the lock contention on the swap locks is eliminated. And the pmbench score increases 37.1%. The swapin throughput increases 45.7% from 2.02 GB/s to 2.94 GB/s. While the swapout throughput increases 45.3% from 2.04 GB/s to 2.97 GB/s. Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:10 +00:00
/*
* Even if there's no free clusters available (fragmented),
* try to scan a little more quickly with lock held unless we
* have scanned too many slots already.
*/
if (!scanned_many) {
unsigned long scan_limit;
if (offset < scan_base)
scan_limit = scan_base;
else
scan_limit = si->highest_bit;
for (; offset <= scan_limit && --latency_ration > 0;
offset++) {
if (!si->swap_map[offset])
goto checks;
}
}
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
done:
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
set_cluster_next(si, offset + 1);
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
si->flags -= SWP_SCANNING;
return n_ret;
scan:
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
while (++offset <= READ_ONCE(si->highest_bit)) {
if (unlikely(--latency_ration < 0)) {
cond_resched();
latency_ration = LATENCY_LIMIT;
swap: try to scan more free slots even when fragmented Now, the scalability of swap code will drop much when the swap device becomes fragmented, because the swap slots allocation batching stops working. To solve the problem, in this patch, we will try to scan a little more swap slots with restricted effort to batch the swap slots allocation even if the swap device is fragmented. Test shows that the benchmark score can increase up to 37.1% with the patch. Details are as follows. The swap code has a per-cpu cache of swap slots. These batch swap space allocations to improve swap subsystem scaling. In the following code path, add_to_swap() get_swap_page() refill_swap_slots_cache() get_swap_pages() scan_swap_map_slots() scan_swap_map_slots() and get_swap_pages() can return multiple swap slots for each call. These slots will be cached in the per-CPU swap slots cache, so that several following swap slot requests will be fulfilled there to avoid the lock contention in the lower level swap space allocation/freeing code path. But this only works when there are free swap clusters. If a swap device becomes so fragmented that there's no free swap clusters, scan_swap_map_slots() and get_swap_pages() will return only one swap slot for each call in the above code path. Effectively, this falls back to the situation before the swap slots cache was introduced, the heavy lock contention on the swap related locks kills the scalability. Why does it work in this way? Because the swap device could be large, and the free swap slot scanning could be quite time consuming, to avoid taking too much time to scanning free swap slots, the conservative method was used. In fact, this can be improved via scanning a little more free slots with strictly restricted effort. Which is implemented in this patch. In scan_swap_map_slots(), after the first free swap slot is gotten, we will try to scan a little more, but only if we haven't scanned too many slots (< LATENCY_LIMIT). That is, the added scanning latency is strictly restricted. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. Multiple ram disks are configured as the swap devices. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random, so the swap space becomes highly fragmented during the test. In the original implementation, the lock contention on swap related locks is very heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap: 21.03 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.92 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.72 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 0.69 While after applying this patch, it becomes, _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 4.89 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 3.85 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.1 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88 That is, the lock contention on the swap locks is eliminated. And the pmbench score increases 37.1%. The swapin throughput increases 45.7% from 2.02 GB/s to 2.94 GB/s. While the swapout throughput increases 45.3% from 2.04 GB/s to 2.97 GB/s. Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:10 +00:00
scanned_many = true;
}
swapfile: fix soft lockup in scan_swap_map_slots A softlockup occurs in scan free swap slot under huge memory pressure. The test scenario is: 64 CPU cores, 64GB memory, and 28 zram devices, the disksize of each zram device is 50MB. LATENCY_LIMIT is used to prevent softlockups in scan_swap_map_slots(), but the real loop number would more than LATENCY_LIMIT because of "goto checks and goto scan" repeatly without decreasing latency limit. In order to fix it, decrease latency_ration in advance. There is also a suspicious place that will cause softlockups in get_swap_pages(). In this function, the "goto start_over" may result in continuous scanning of the swap partition. If there is no cond_sched in scan_swap_map_slots(), it would cause a softlockup (I am not sure about this). WARN: soft lockup - CPU#11 stuck for 11s! [kswapd0:466] CPU: 11 PID: 466 Comm: kswapd@ Kdump: loaded Tainted: G dump backtrace+0x0/0x1le4 show stack+0x20/@x2c dump_stack+0xd8/0x140 watchdog print_info+0x48/0x54 watchdog_process_before_softlockup+0x98/0xa0 watchdog_timer_fn+0xlac/0x2d0 hrtimer_rum_queues+0xb0/0x130 hrtimer_interrupt+0x13c/0x3c0 arch_timer_handler_virt+0x3c/0x50 handLe_percpu_devid_irq+0x90/0x1f4 handle domain irq+0x84/0x100 gic_handle_irq+0x88/0x2b0 e11 ira+0xhB/Bx140 scan_swap_map_slots+0x678/0x890 get_swap_pages+0x29c/0x440 get_swap_page+0x120/0x2e0 add_to_swap+UX2U/0XyC shrink_page_list+0x5d0/0x152c shrink_inactive_list+0xl6c/Bx500 shrink_lruvec+0x270/0x304 WARN: soft lockup - CPU#32 stuck for 11s! [stress-ng:309915] watchdog_timer_fn+0x1ac/0x2d0 __run_hrtimer+0x98/0x2a0 __hrtimer_run_queues+0xb0/0x130 hrtimer_interrupt+0x13c/0x3c0 arch_timer_handler_virt+0x3c/0x50 handle_percpu_devid_irq+0x90/0x1f4 __handle_domain_irq+0x84/0x100 gic_handle_irq+0x88/0x2b0 el1_irq+0xb8/0x140 get_swap_pages+0x1e8/0x440 get_swap_page+0x1c8/0x2e0 add_to_swap+0x20/0x9c shrink_page_list+0x5d0/0x152c reclaim_pages+0x160/0x310 madvise_cold_or_pageout_pte_range+0x7bc/0xe3c walk_pmd_range.isra.0+0xac/0x22c walk_pud_range+0xfc/0x1c0 walk_pgd_range+0x158/0x1b0 __walk_page_range+0x64/0x100 walk_page_range+0x104/0x150 Link: https://lkml.kernel.org/r/20221118133850.3360369-1-chenwandun@huawei.com Fixes: 048c27fd7281 ("[PATCH] swap: scan_swap_map latency breaks") Signed-off-by: Chen Wandun <chenwandun@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Hugh Dickins <hugh@veritas.com> Cc: Kefeng Wang <wangkefeng.wang@huawei.com> Cc: Nanyong Sun <sunnanyong@huawei.com> Cc: <xialonglong1@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-18 13:38:50 +00:00
if (swap_offset_available_and_locked(si, offset))
goto checks;
}
offset = si->lowest_bit;
while (offset < scan_base) {
if (unlikely(--latency_ration < 0)) {
cond_resched();
latency_ration = LATENCY_LIMIT;
swap: try to scan more free slots even when fragmented Now, the scalability of swap code will drop much when the swap device becomes fragmented, because the swap slots allocation batching stops working. To solve the problem, in this patch, we will try to scan a little more swap slots with restricted effort to batch the swap slots allocation even if the swap device is fragmented. Test shows that the benchmark score can increase up to 37.1% with the patch. Details are as follows. The swap code has a per-cpu cache of swap slots. These batch swap space allocations to improve swap subsystem scaling. In the following code path, add_to_swap() get_swap_page() refill_swap_slots_cache() get_swap_pages() scan_swap_map_slots() scan_swap_map_slots() and get_swap_pages() can return multiple swap slots for each call. These slots will be cached in the per-CPU swap slots cache, so that several following swap slot requests will be fulfilled there to avoid the lock contention in the lower level swap space allocation/freeing code path. But this only works when there are free swap clusters. If a swap device becomes so fragmented that there's no free swap clusters, scan_swap_map_slots() and get_swap_pages() will return only one swap slot for each call in the above code path. Effectively, this falls back to the situation before the swap slots cache was introduced, the heavy lock contention on the swap related locks kills the scalability. Why does it work in this way? Because the swap device could be large, and the free swap slot scanning could be quite time consuming, to avoid taking too much time to scanning free swap slots, the conservative method was used. In fact, this can be improved via scanning a little more free slots with strictly restricted effort. Which is implemented in this patch. In scan_swap_map_slots(), after the first free swap slot is gotten, we will try to scan a little more, but only if we haven't scanned too many slots (< LATENCY_LIMIT). That is, the added scanning latency is strictly restricted. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. Multiple ram disks are configured as the swap devices. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random, so the swap space becomes highly fragmented during the test. In the original implementation, the lock contention on swap related locks is very heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock.get_swap_pages.get_swap_page.add_to_swap: 21.03 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.92 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.72 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 0.69 While after applying this patch, it becomes, _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 4.89 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 3.85 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.1 _raw_spin_lock_irqsave.pagevec_lru_move_fn.__lru_cache_add.do_swap_page: 0.88 That is, the lock contention on the swap locks is eliminated. And the pmbench score increases 37.1%. The swapin throughput increases 45.7% from 2.02 GB/s to 2.94 GB/s. While the swapout throughput increases 45.3% from 2.04 GB/s to 2.97 GB/s. Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200427030023.264780-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:10 +00:00
scanned_many = true;
}
swapfile: fix soft lockup in scan_swap_map_slots A softlockup occurs in scan free swap slot under huge memory pressure. The test scenario is: 64 CPU cores, 64GB memory, and 28 zram devices, the disksize of each zram device is 50MB. LATENCY_LIMIT is used to prevent softlockups in scan_swap_map_slots(), but the real loop number would more than LATENCY_LIMIT because of "goto checks and goto scan" repeatly without decreasing latency limit. In order to fix it, decrease latency_ration in advance. There is also a suspicious place that will cause softlockups in get_swap_pages(). In this function, the "goto start_over" may result in continuous scanning of the swap partition. If there is no cond_sched in scan_swap_map_slots(), it would cause a softlockup (I am not sure about this). WARN: soft lockup - CPU#11 stuck for 11s! [kswapd0:466] CPU: 11 PID: 466 Comm: kswapd@ Kdump: loaded Tainted: G dump backtrace+0x0/0x1le4 show stack+0x20/@x2c dump_stack+0xd8/0x140 watchdog print_info+0x48/0x54 watchdog_process_before_softlockup+0x98/0xa0 watchdog_timer_fn+0xlac/0x2d0 hrtimer_rum_queues+0xb0/0x130 hrtimer_interrupt+0x13c/0x3c0 arch_timer_handler_virt+0x3c/0x50 handLe_percpu_devid_irq+0x90/0x1f4 handle domain irq+0x84/0x100 gic_handle_irq+0x88/0x2b0 e11 ira+0xhB/Bx140 scan_swap_map_slots+0x678/0x890 get_swap_pages+0x29c/0x440 get_swap_page+0x120/0x2e0 add_to_swap+UX2U/0XyC shrink_page_list+0x5d0/0x152c shrink_inactive_list+0xl6c/Bx500 shrink_lruvec+0x270/0x304 WARN: soft lockup - CPU#32 stuck for 11s! [stress-ng:309915] watchdog_timer_fn+0x1ac/0x2d0 __run_hrtimer+0x98/0x2a0 __hrtimer_run_queues+0xb0/0x130 hrtimer_interrupt+0x13c/0x3c0 arch_timer_handler_virt+0x3c/0x50 handle_percpu_devid_irq+0x90/0x1f4 __handle_domain_irq+0x84/0x100 gic_handle_irq+0x88/0x2b0 el1_irq+0xb8/0x140 get_swap_pages+0x1e8/0x440 get_swap_page+0x1c8/0x2e0 add_to_swap+0x20/0x9c shrink_page_list+0x5d0/0x152c reclaim_pages+0x160/0x310 madvise_cold_or_pageout_pte_range+0x7bc/0xe3c walk_pmd_range.isra.0+0xac/0x22c walk_pud_range+0xfc/0x1c0 walk_pgd_range+0x158/0x1b0 __walk_page_range+0x64/0x100 walk_page_range+0x104/0x150 Link: https://lkml.kernel.org/r/20221118133850.3360369-1-chenwandun@huawei.com Fixes: 048c27fd7281 ("[PATCH] swap: scan_swap_map latency breaks") Signed-off-by: Chen Wandun <chenwandun@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Hugh Dickins <hugh@veritas.com> Cc: Kefeng Wang <wangkefeng.wang@huawei.com> Cc: Nanyong Sun <sunnanyong@huawei.com> Cc: <xialonglong1@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-18 13:38:50 +00:00
if (swap_offset_available_and_locked(si, offset))
goto checks;
offset++;
}
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&si->lock);
no_page:
si->flags -= SWP_SCANNING;
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
return n_ret;
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
{
unsigned long idx;
struct swap_cluster_info *ci;
unsigned long offset;
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
mm/swapfile.c: replace some #ifdef with IS_ENABLED() In mm/swapfile.c, THP (Transparent Huge Page) swap specific code is enclosed by #ifdef CONFIG_THP_SWAP/#endif to avoid code dilating when THP isn't enabled. But #ifdef/#endif in .c file hurt the code readability, so Dave suggested to use IS_ENABLED(CONFIG_THP_SWAP) instead and let compiler to do the dirty job for us. This has potential to remove some duplicated code too. From output of `size`, text data bss dec hex filename THP=y: 26269 2076 340 28685 700d mm/swapfile.o ifdef/endif: 24115 2028 340 26483 6773 mm/swapfile.o IS_ENABLED: 24179 2028 340 26547 67b3 mm/swapfile.o IS_ENABLED() based solution works quite well, almost as good as that of #ifdef/#endif. And from the diffstat, the removed lines are more than added lines. One #ifdef for split_swap_cluster() is kept. Because it is a public function with a stub implementation for CONFIG_THP_SWAP=n in swap.h. Link: http://lkml.kernel.org/r/20180720071845.17920-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:05 +00:00
/*
* Should not even be attempting cluster allocations when huge
* page swap is disabled. Warn and fail the allocation.
*/
if (!IS_ENABLED(CONFIG_THP_SWAP)) {
VM_WARN_ON_ONCE(1);
return 0;
}
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (cluster_list_empty(&si->free_clusters))
return 0;
idx = cluster_list_first(&si->free_clusters);
offset = idx * SWAPFILE_CLUSTER;
ci = lock_cluster(si, offset);
alloc_cluster(si, idx);
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
unlock_cluster(ci);
swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
*slot = swp_entry(si->type, offset);
return 1;
}
static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
{
unsigned long offset = idx * SWAPFILE_CLUSTER;
struct swap_cluster_info *ci;
ci = lock_cluster(si, offset);
memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
cluster_set_count_flag(ci, 0, 0);
free_cluster(si, idx);
unlock_cluster(ci);
swap_range_free(si, offset, SWAPFILE_CLUSTER);
}
int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
{
unsigned long size = swap_entry_size(entry_size);
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
struct swap_info_struct *si, *next;
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
long avail_pgs;
int n_ret = 0;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
int node;
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
/* Only single cluster request supported */
WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
spin_lock(&swap_avail_lock);
avail_pgs = atomic_long_read(&nr_swap_pages) / size;
if (avail_pgs <= 0) {
spin_unlock(&swap_avail_lock);
goto noswap;
}
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
atomic_long_sub(n_goal * size, &nr_swap_pages);
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
start_over:
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
node = numa_node_id();
plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
/* requeue si to after same-priority siblings */
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
spin_unlock(&swap_avail_lock);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&si->lock);
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
spin_lock(&swap_avail_lock);
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
if (plist_node_empty(&si->avail_lists[node])) {
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
spin_unlock(&si->lock);
goto nextsi;
}
WARN(!si->highest_bit,
"swap_info %d in list but !highest_bit\n",
si->type);
WARN(!(si->flags & SWP_WRITEOK),
"swap_info %d in list but !SWP_WRITEOK\n",
si->type);
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
__del_from_avail_list(si);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
goto nextsi;
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
}
if (size == SWAPFILE_CLUSTER) {
mm, THP, swap: fix allocating cluster for swapfile by mistake SWP_FS is used to make swap_{read,write}page() go through the filesystem, and it's only used for swap files over NFS. So, !SWP_FS means non NFS for now, it could be either file backed or device backed. Something similar goes with legacy SWP_FILE. So in order to achieve the goal of the original patch, SWP_BLKDEV should be used instead. FS corruption can be observed with SSD device + XFS + fragmented swapfile due to CONFIG_THP_SWAP=y. I reproduced the issue with the following details: Environment: QEMU + upstream kernel + buildroot + NVMe (2 GB) Kernel config: CONFIG_BLK_DEV_NVME=y CONFIG_THP_SWAP=y Some reproducible steps: mkfs.xfs -f /dev/nvme0n1 mkdir /tmp/mnt mount /dev/nvme0n1 /tmp/mnt bs="32k" sz="1024m" # doesn't matter too much, I also tried 16m xfs_io -f -c "pwrite -R -b $bs 0 $sz" -c "fdatasync" /tmp/mnt/sw xfs_io -f -c "pwrite -R -b $bs 0 $sz" -c "fdatasync" /tmp/mnt/sw xfs_io -f -c "pwrite -R -b $bs 0 $sz" -c "fdatasync" /tmp/mnt/sw xfs_io -f -c "pwrite -F -S 0 -b $bs 0 $sz" -c "fdatasync" /tmp/mnt/sw xfs_io -f -c "pwrite -R -b $bs 0 $sz" -c "fsync" /tmp/mnt/sw mkswap /tmp/mnt/sw swapon /tmp/mnt/sw stress --vm 2 --vm-bytes 600M # doesn't matter too much as well Symptoms: - FS corruption (e.g. checksum failure) - memory corruption at: 0xd2808010 - segfault Fixes: f0eea189e8e9 ("mm, THP, swap: Don't allocate huge cluster for file backed swap device") Fixes: 38d8b4e6bdc8 ("mm, THP, swap: delay splitting THP during swap out") Signed-off-by: Gao Xiang <hsiangkao@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Yang Shi <shy828301@gmail.com> Acked-by: Rafael Aquini <aquini@redhat.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Carlos Maiolino <cmaiolino@redhat.com> Cc: Eric Sandeen <esandeen@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: <stable@vger.kernel.org> Link: https://lkml.kernel.org/r/20200820045323.7809-1-hsiangkao@redhat.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-09-26 04:19:01 +00:00
if (si->flags & SWP_BLKDEV)
n_ret = swap_alloc_cluster(si, swp_entries);
} else
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
n_goal, swp_entries);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
if (n_ret || size == SWAPFILE_CLUSTER)
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
goto check_out;
cond_resched();
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
spin_lock(&swap_avail_lock);
nextsi:
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
/*
* if we got here, it's likely that si was almost full before,
* and since scan_swap_map_slots() can drop the si->lock,
* multiple callers probably all tried to get a page from the
* same si and it filled up before we could get one; or, the si
* filled up between us dropping swap_avail_lock and taking
* si->lock. Since we dropped the swap_avail_lock, the
* swap_avail_head list may have been modified; so if next is
* still in the swap_avail_head list then try it, otherwise
* start over if we have not gotten any slots.
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
*/
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
if (plist_node_empty(&next->avail_lists[node]))
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
goto start_over;
}
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
spin_unlock(&swap_avail_lock);
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
check_out:
if (n_ret < n_goal)
atomic_long_add((long)(n_goal - n_ret) * size,
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
&nr_swap_pages);
noswap:
mm/swap: allocate swap slots in batches Currently, the swap slots are allocated one page at a time, causing contention to the swap_info lock protecting the swap partition on every page being swapped. This patch adds new functions get_swap_pages and scan_swap_map_slots to request multiple swap slots at once. This will reduces the lock contention on the swap_info lock. Also scan_swap_map_slots can operate more efficiently as swap slots often occurs in clusters close to each other on a swap device and it is quicker to allocate them together. Link: http://lkml.kernel.org/r/9fec2845544371f62c3763d43510045e33d286a6.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:33 +00:00
return n_ret;
}
static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
{
struct swap_info_struct *p;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
unsigned long offset;
if (!entry.val)
goto out;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
p = swp_swap_info(entry);
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
if (!p)
goto bad_nofile;
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
if (data_race(!(p->flags & SWP_USED)))
goto bad_device;
offset = swp_offset(entry);
if (offset >= p->max)
goto bad_offset;
if (data_race(!p->swap_map[swp_offset(entry)]))
goto bad_free;
return p;
bad_free:
pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
goto out;
bad_offset:
pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
goto out;
bad_device:
pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
goto out;
bad_nofile:
pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
out:
return NULL;
}
static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
struct swap_info_struct *q)
{
struct swap_info_struct *p;
p = _swap_info_get(entry);
if (p != q) {
if (q != NULL)
spin_unlock(&q->lock);
if (p != NULL)
spin_lock(&p->lock);
}
return p;
}
static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
unsigned long offset,
unsigned char usage)
{
unsigned char count;
unsigned char has_cache;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
count = p->swap_map[offset];
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
has_cache = count & SWAP_HAS_CACHE;
count &= ~SWAP_HAS_CACHE;
if (usage == SWAP_HAS_CACHE) {
VM_BUG_ON(!has_cache);
has_cache = 0;
} else if (count == SWAP_MAP_SHMEM) {
/*
* Or we could insist on shmem.c using a special
* swap_shmem_free() and free_shmem_swap_and_cache()...
*/
count = 0;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
} else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
if (count == COUNT_CONTINUED) {
if (swap_count_continued(p, offset, count))
count = SWAP_MAP_MAX | COUNT_CONTINUED;
else
count = SWAP_MAP_MAX;
} else
count--;
}
usage = count | has_cache;
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
if (usage)
WRITE_ONCE(p->swap_map[offset], usage);
else
WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
return usage;
}
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
/*
* Check whether swap entry is valid in the swap device. If so,
* return pointer to swap_info_struct, and keep the swap entry valid
* via preventing the swap device from being swapoff, until
* put_swap_device() is called. Otherwise return NULL.
*
* Notice that swapoff or swapoff+swapon can still happen before the
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
* percpu_ref_tryget_live() in get_swap_device() or after the
* percpu_ref_put() in put_swap_device() if there isn't any other way
* to prevent swapoff, such as page lock, page table lock, etc. The
* caller must be prepared for that. For example, the following
* situation is possible.
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
*
* CPU1 CPU2
* do_swap_page()
* ... swapoff+swapon
* __read_swap_cache_async()
* swapcache_prepare()
* __swap_duplicate()
* // check swap_map
* // verify PTE not changed
*
* In __swap_duplicate(), the swap_map need to be checked before
* changing partly because the specified swap entry may be for another
* swap device which has been swapoff. And in do_swap_page(), after
* the page is read from the swap device, the PTE is verified not
* changed with the page table locked to check whether the swap device
* has been swapoff or swapoff+swapon.
*/
struct swap_info_struct *get_swap_device(swp_entry_t entry)
{
struct swap_info_struct *si;
unsigned long offset;
if (!entry.val)
goto out;
si = swp_swap_info(entry);
if (!si)
goto bad_nofile;
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
if (!percpu_ref_tryget_live(&si->users))
goto out;
/*
* Guarantee the si->users are checked before accessing other
* fields of swap_info_struct.
*
* Paired with the spin_unlock() after setup_swap_info() in
* enable_swap_info().
*/
smp_rmb();
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
offset = swp_offset(entry);
if (offset >= si->max)
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
goto put_out;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
return si;
bad_nofile:
pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
out:
return NULL;
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
put_out:
pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
percpu_ref_put(&si->users);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
return NULL;
}
static unsigned char __swap_entry_free(struct swap_info_struct *p,
swp_entry_t entry)
{
struct swap_cluster_info *ci;
unsigned long offset = swp_offset(entry);
unsigned char usage;
ci = lock_cluster_or_swap_info(p, offset);
usage = __swap_entry_free_locked(p, offset, 1);
unlock_cluster_or_swap_info(p, ci);
if (!usage)
free_swap_slot(entry);
return usage;
}
static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
{
struct swap_cluster_info *ci;
unsigned long offset = swp_offset(entry);
unsigned char count;
ci = lock_cluster(p, offset);
count = p->swap_map[offset];
VM_BUG_ON(count != SWAP_HAS_CACHE);
p->swap_map[offset] = 0;
dec_cluster_info_page(p, p->cluster_info, offset);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster(ci);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
mem_cgroup_uncharge_swap(entry, 1);
swap_range_free(p, offset, 1);
}
/*
* Caller has made sure that the swap device corresponding to entry
* is still around or has not been recycled.
*/
void swap_free(swp_entry_t entry)
{
struct swap_info_struct *p;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
p = _swap_info_get(entry);
if (p)
__swap_entry_free(p, entry);
}
/*
* Called after dropping swapcache to decrease refcnt to swap entries.
*/
void put_swap_folio(struct folio *folio, swp_entry_t entry)
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
{
unsigned long offset = swp_offset(entry);
unsigned long idx = offset / SWAPFILE_CLUSTER;
struct swap_cluster_info *ci;
struct swap_info_struct *si;
unsigned char *map;
mm, THP, swap: support to clear swap cache flag for THP swapped out Patch series "mm, THP, swap: Delay splitting THP after swapped out", v3. This is the second step of THP (Transparent Huge Page) swap optimization. In the first step, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. In the second step, the splitting is delayed further to after the swapping out finished. The plan is to delay splitting THP step by step, finally avoid splitting THP for the THP swapping out and swap out/in the THP as a whole. In the patchset, more operations for the anonymous THP reclaiming, such as TLB flushing, writing the THP to the swap device, removing the THP from the swap cache are batched. So that the performance of anonymous THP swapping out are improved. During the development, the following scenarios/code paths have been checked, - swap out/in - swap off - write protect page fault - madvise_free - process exit - split huge page With the patchset, the swap out throughput improves 42% (from about 5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case with 16 processes. At the same time, the IPI (reflect TLB flushing) reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. Below is the part of the cover letter for the first step patchset of THP swap optimization which applies to all steps. ========================= Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce TLB flushing and lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patch (of 12): Previously, swapcache_free_cluster() is used only in the error path of shrink_page_list() to free the swap cluster just allocated if the THP (Transparent Huge Page) is failed to be split. In this patch, it is enhanced to clear the swap cache flag (SWAP_HAS_CACHE) for the swap cluster that holds the contents of THP swapped out. This will be used in delaying splitting THP after swapping out support. Because there is no THP swapping in as a whole support yet, after clearing the swap cache flag, the swap cluster backing the THP swapped out will be split. So that the swap slots in the swap cluster can be swapped in as normal pages later. Link: http://lkml.kernel.org/r/20170724051840.2309-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:12 +00:00
unsigned int i, free_entries = 0;
unsigned char val;
int size = swap_entry_size(folio_nr_pages(folio));
mm/swapfile.c: replace some #ifdef with IS_ENABLED() In mm/swapfile.c, THP (Transparent Huge Page) swap specific code is enclosed by #ifdef CONFIG_THP_SWAP/#endif to avoid code dilating when THP isn't enabled. But #ifdef/#endif in .c file hurt the code readability, so Dave suggested to use IS_ENABLED(CONFIG_THP_SWAP) instead and let compiler to do the dirty job for us. This has potential to remove some duplicated code too. From output of `size`, text data bss dec hex filename THP=y: 26269 2076 340 28685 700d mm/swapfile.o ifdef/endif: 24115 2028 340 26483 6773 mm/swapfile.o IS_ENABLED: 24179 2028 340 26547 67b3 mm/swapfile.o IS_ENABLED() based solution works quite well, almost as good as that of #ifdef/#endif. And from the diffstat, the removed lines are more than added lines. One #ifdef for split_swap_cluster() is kept. Because it is a public function with a stub implementation for CONFIG_THP_SWAP=n in swap.h. Link: http://lkml.kernel.org/r/20180720071845.17920-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:05 +00:00
mm, THP, swap: support to clear swap cache flag for THP swapped out Patch series "mm, THP, swap: Delay splitting THP after swapped out", v3. This is the second step of THP (Transparent Huge Page) swap optimization. In the first step, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. In the second step, the splitting is delayed further to after the swapping out finished. The plan is to delay splitting THP step by step, finally avoid splitting THP for the THP swapping out and swap out/in the THP as a whole. In the patchset, more operations for the anonymous THP reclaiming, such as TLB flushing, writing the THP to the swap device, removing the THP from the swap cache are batched. So that the performance of anonymous THP swapping out are improved. During the development, the following scenarios/code paths have been checked, - swap out/in - swap off - write protect page fault - madvise_free - process exit - split huge page With the patchset, the swap out throughput improves 42% (from about 5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case with 16 processes. At the same time, the IPI (reflect TLB flushing) reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. Below is the part of the cover letter for the first step patchset of THP swap optimization which applies to all steps. ========================= Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce TLB flushing and lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patch (of 12): Previously, swapcache_free_cluster() is used only in the error path of shrink_page_list() to free the swap cluster just allocated if the THP (Transparent Huge Page) is failed to be split. In this patch, it is enhanced to clear the swap cache flag (SWAP_HAS_CACHE) for the swap cluster that holds the contents of THP swapped out. This will be used in delaying splitting THP after swapping out support. Because there is no THP swapping in as a whole support yet, after clearing the swap cache flag, the swap cluster backing the THP swapped out will be split. So that the swap slots in the swap cluster can be swapped in as normal pages later. Link: http://lkml.kernel.org/r/20170724051840.2309-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:12 +00:00
si = _swap_info_get(entry);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
if (!si)
return;
mm/swapfile.c: put_swap_page: share more between huge/normal code path In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. The `free_entries == 0` case is merged into the more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24123 2004 340 26467 6763 mm/swapfile.o unified: 24485 2004 340 26829 68cd mm/swapfile.o Dig on step deeper with `size -A mm/swapfile.o` for base and unified kernel and compare the result, yields, -.text 17723 0 +.text 17835 0 -.orc_unwind_ip 1380 0 +.orc_unwind_ip 1480 0 -.orc_unwind 2070 0 +.orc_unwind 2220 0 -Total 26686 +Total 27048 The total difference is the same. The text segment difference is much smaller: 112. More difference comes from the ORC unwinder segments: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is used, this costs nothing. Link: http://lkml.kernel.org/r/20180720071845.17920-9-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Acked-by: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:29 +00:00
ci = lock_cluster_or_swap_info(si, offset);
if (size == SWAPFILE_CLUSTER) {
VM_BUG_ON(!cluster_is_huge(ci));
map = si->swap_map + offset;
for (i = 0; i < SWAPFILE_CLUSTER; i++) {
val = map[i];
VM_BUG_ON(!(val & SWAP_HAS_CACHE));
if (val == SWAP_HAS_CACHE)
free_entries++;
}
cluster_clear_huge(ci);
if (free_entries == SWAPFILE_CLUSTER) {
mm/swapfile.c: put_swap_page: share more between huge/normal code path In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. The `free_entries == 0` case is merged into the more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24123 2004 340 26467 6763 mm/swapfile.o unified: 24485 2004 340 26829 68cd mm/swapfile.o Dig on step deeper with `size -A mm/swapfile.o` for base and unified kernel and compare the result, yields, -.text 17723 0 +.text 17835 0 -.orc_unwind_ip 1380 0 +.orc_unwind_ip 1480 0 -.orc_unwind 2070 0 +.orc_unwind 2220 0 -Total 26686 +Total 27048 The total difference is the same. The text segment difference is much smaller: 112. More difference comes from the ORC unwinder segments: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is used, this costs nothing. Link: http://lkml.kernel.org/r/20180720071845.17920-9-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Acked-by: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:29 +00:00
unlock_cluster_or_swap_info(si, ci);
spin_lock(&si->lock);
mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
swap_free_cluster(si, idx);
spin_unlock(&si->lock);
return;
}
}
mm/swapfile.c: put_swap_page: share more between huge/normal code path In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. The `free_entries == 0` case is merged into the more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24123 2004 340 26467 6763 mm/swapfile.o unified: 24485 2004 340 26829 68cd mm/swapfile.o Dig on step deeper with `size -A mm/swapfile.o` for base and unified kernel and compare the result, yields, -.text 17723 0 +.text 17835 0 -.orc_unwind_ip 1380 0 +.orc_unwind_ip 1480 0 -.orc_unwind 2070 0 +.orc_unwind 2220 0 -Total 26686 +Total 27048 The total difference is the same. The text segment difference is much smaller: 112. More difference comes from the ORC unwinder segments: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is used, this costs nothing. Link: http://lkml.kernel.org/r/20180720071845.17920-9-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Acked-by: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:29 +00:00
for (i = 0; i < size; i++, entry.val++) {
if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
unlock_cluster_or_swap_info(si, ci);
free_swap_slot(entry);
if (i == size - 1)
return;
lock_cluster_or_swap_info(si, offset);
mm, THP, swap: support to clear swap cache flag for THP swapped out Patch series "mm, THP, swap: Delay splitting THP after swapped out", v3. This is the second step of THP (Transparent Huge Page) swap optimization. In the first step, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. In the second step, the splitting is delayed further to after the swapping out finished. The plan is to delay splitting THP step by step, finally avoid splitting THP for the THP swapping out and swap out/in the THP as a whole. In the patchset, more operations for the anonymous THP reclaiming, such as TLB flushing, writing the THP to the swap device, removing the THP from the swap cache are batched. So that the performance of anonymous THP swapping out are improved. During the development, the following scenarios/code paths have been checked, - swap out/in - swap off - write protect page fault - madvise_free - process exit - split huge page With the patchset, the swap out throughput improves 42% (from about 5.81GB/s to about 8.25GB/s) in the vm-scalability swap-w-seq test case with 16 processes. At the same time, the IPI (reflect TLB flushing) reduced about 78.9%. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. Below is the part of the cover letter for the first step patchset of THP swap optimization which applies to all steps. ========================= Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce TLB flushing and lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patch (of 12): Previously, swapcache_free_cluster() is used only in the error path of shrink_page_list() to free the swap cluster just allocated if the THP (Transparent Huge Page) is failed to be split. In this patch, it is enhanced to clear the swap cache flag (SWAP_HAS_CACHE) for the swap cluster that holds the contents of THP swapped out. This will be used in delaying splitting THP after swapping out support. Because there is no THP swapping in as a whole support yet, after clearing the swap cache flag, the swap cluster backing the THP swapped out will be split. So that the swap slots in the swap cluster can be swapped in as normal pages later. Link: http://lkml.kernel.org/r/20170724051840.2309-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:12 +00:00
}
}
mm/swapfile.c: put_swap_page: share more between huge/normal code path In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. The `free_entries == 0` case is merged into the more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24123 2004 340 26467 6763 mm/swapfile.o unified: 24485 2004 340 26829 68cd mm/swapfile.o Dig on step deeper with `size -A mm/swapfile.o` for base and unified kernel and compare the result, yields, -.text 17723 0 +.text 17835 0 -.orc_unwind_ip 1380 0 +.orc_unwind_ip 1480 0 -.orc_unwind 2070 0 +.orc_unwind 2220 0 -Total 26686 +Total 27048 The total difference is the same. The text segment difference is much smaller: 112. More difference comes from the ORC unwinder segments: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is used, this costs nothing. Link: http://lkml.kernel.org/r/20180720071845.17920-9-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Acked-by: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:29 +00:00
unlock_cluster_or_swap_info(si, ci);
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
}
mm/swapfile.c: replace some #ifdef with IS_ENABLED() In mm/swapfile.c, THP (Transparent Huge Page) swap specific code is enclosed by #ifdef CONFIG_THP_SWAP/#endif to avoid code dilating when THP isn't enabled. But #ifdef/#endif in .c file hurt the code readability, so Dave suggested to use IS_ENABLED(CONFIG_THP_SWAP) instead and let compiler to do the dirty job for us. This has potential to remove some duplicated code too. From output of `size`, text data bss dec hex filename THP=y: 26269 2076 340 28685 700d mm/swapfile.o ifdef/endif: 24115 2028 340 26483 6773 mm/swapfile.o IS_ENABLED: 24179 2028 340 26547 67b3 mm/swapfile.o IS_ENABLED() based solution works quite well, almost as good as that of #ifdef/#endif. And from the diffstat, the removed lines are more than added lines. One #ifdef for split_swap_cluster() is kept. Because it is a public function with a stub implementation for CONFIG_THP_SWAP=n in swap.h. Link: http://lkml.kernel.org/r/20180720071845.17920-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:05 +00:00
#ifdef CONFIG_THP_SWAP
int split_swap_cluster(swp_entry_t entry)
{
struct swap_info_struct *si;
struct swap_cluster_info *ci;
unsigned long offset = swp_offset(entry);
si = _swap_info_get(entry);
if (!si)
return -EBUSY;
ci = lock_cluster(si, offset);
cluster_clear_huge(ci);
unlock_cluster(ci);
return 0;
}
mm/swapfile.c: replace some #ifdef with IS_ENABLED() In mm/swapfile.c, THP (Transparent Huge Page) swap specific code is enclosed by #ifdef CONFIG_THP_SWAP/#endif to avoid code dilating when THP isn't enabled. But #ifdef/#endif in .c file hurt the code readability, so Dave suggested to use IS_ENABLED(CONFIG_THP_SWAP) instead and let compiler to do the dirty job for us. This has potential to remove some duplicated code too. From output of `size`, text data bss dec hex filename THP=y: 26269 2076 340 28685 700d mm/swapfile.o ifdef/endif: 24115 2028 340 26483 6773 mm/swapfile.o IS_ENABLED: 24179 2028 340 26547 67b3 mm/swapfile.o IS_ENABLED() based solution works quite well, almost as good as that of #ifdef/#endif. And from the diffstat, the removed lines are more than added lines. One #ifdef for split_swap_cluster() is kept. Because it is a public function with a stub implementation for CONFIG_THP_SWAP=n in swap.h. Link: http://lkml.kernel.org/r/20180720071845.17920-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Suggested-and-acked-by: Dave Hansen <dave.hansen@linux.intel.com> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Matthew Wilcox <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-22 04:52:05 +00:00
#endif
mm, THP, swap: delay splitting THP during swap out Patch series "THP swap: Delay splitting THP during swapping out", v11. This patchset is to optimize the performance of Transparent Huge Page (THP) swap. Recently, the performance of the storage devices improved so fast that we cannot saturate the disk bandwidth with single logical CPU when do page swap out even on a high-end server machine. Because the performance of the storage device improved faster than that of single logical CPU. And it seems that the trend will not change in the near future. On the other hand, the THP becomes more and more popular because of increased memory size. So it becomes necessary to optimize THP swap performance. The advantages of the THP swap support include: - Batch the swap operations for the THP to reduce lock acquiring/releasing, including allocating/freeing the swap space, adding/deleting to/from the swap cache, and writing/reading the swap space, etc. This will help improve the performance of the THP swap. - The THP swap space read/write will be 2M sequential IO. It is particularly helpful for the swap read, which are usually 4k random IO. This will improve the performance of the THP swap too. - It will help the memory fragmentation, especially when the THP is heavily used by the applications. The 2M continuous pages will be free up after THP swapping out. - It will improve the THP utilization on the system with the swap turned on. Because the speed for khugepaged to collapse the normal pages into the THP is quite slow. After the THP is split during the swapping out, it will take quite long time for the normal pages to collapse back into the THP after being swapped in. The high THP utilization helps the efficiency of the page based memory management too. There are some concerns regarding THP swap in, mainly because possible enlarged read/write IO size (for swap in/out) may put more overhead on the storage device. To deal with that, the THP swap in should be turned on only when necessary. For example, it can be selected via "always/never/madvise" logic, to be turned on globally, turned off globally, or turned on only for VMA with MADV_HUGEPAGE, etc. This patchset is the first step for the THP swap support. The plan is to delay splitting THP step by step, finally avoid splitting THP during the THP swapping out and swap out/in the THP as a whole. As the first step, in this patchset, the splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP and adding the THP into the swap cache. This will reduce lock acquiring/releasing for the locks used for the swap cache management. With the patchset, the swap out throughput improves 15.5% (from about 3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case with 8 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case creates 8 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used up. This patch (of 5): In this patch, splitting huge page is delayed from almost the first step of swapping out to after allocating the swap space for the THP (Transparent Huge Page) and adding the THP into the swap cache. This will batch the corresponding operation, thus improve THP swap out throughput. This is the first step for the THP swap optimization. The plan is to delay splitting the THP step by step and avoid splitting the THP finally. In this patch, one swap cluster is used to hold the contents of each THP swapped out. So, the size of the swap cluster is changed to that of the THP (Transparent Huge Page) on x86_64 architecture (512). For other architectures which want such THP swap optimization, ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for the architecture. In effect, this will enlarge swap cluster size by 2 times on x86_64. Which may make it harder to find a free cluster when the swap space becomes fragmented. So that, this may reduce the continuous swap space allocation and sequential write in theory. The performance test in 0day shows no regressions caused by this. In the future of THP swap optimization, some information of the swapped out THP (such as compound map count) will be recorded in the swap_cluster_info data structure. The mem cgroup swap accounting functions are enhanced to support charge or uncharge a swap cluster backing a THP as a whole. The swap cluster allocate/free functions are added to allocate/free a swap cluster for a THP. A fair simple algorithm is used for swap cluster allocation, that is, only the first swap device in priority list will be tried to allocate the swap cluster. The function will fail if the trying is not successful, and the caller will fallback to allocate a single swap slot instead. This works good enough for normal cases. If the difference of the number of the free swap clusters among multiple swap devices is significant, it is possible that some THPs are split earlier than necessary. For example, this could be caused by big size difference among multiple swap devices. The swap cache functions is enhanced to support add/delete THP to/from the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be enhanced in the future with multi-order radix tree. But because we will split the THP soon during swapping out, that optimization doesn't make much sense for this first step. The THP splitting functions are enhanced to support to split THP in swap cache during swapping out. The page lock will be held during allocating the swap cluster, adding the THP into the swap cache and splitting the THP. So in the code path other than swapping out, if the THP need to be split, the PageSwapCache(THP) will be always false. The swap cluster is only available for SSD, so the THP swap optimization in this patchset has no effect for HDD. [ying.huang@intel.com: fix two issues in THP optimize patch] Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com [hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size] Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option] Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h] Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:37:18 +00:00
mm/swapfile.c: sort swap entries before free To reduce the lock contention of swap_info_struct->lock when freeing swap entry. The freed swap entries will be collected in a per-CPU buffer firstly, and be really freed later in batch. During the batch freeing, if the consecutive swap entries in the per-CPU buffer belongs to same swap device, the swap_info_struct->lock needs to be acquired/released only once, so that the lock contention could be reduced greatly. But if there are multiple swap devices, it is possible that the lock may be unnecessarily released/acquired because the swap entries belong to the same swap device are non-consecutive in the per-CPU buffer. To solve the issue, the per-CPU buffer is sorted according to the swap device before freeing the swap entries. With the patch, the memory (some swapped out) free time reduced 11.6% (from 2.65s to 2.35s) in the vm-scalability swap-w-rand test case with 16 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test swapping, the test case creates 16 processes, which allocate and write to the anonymous pages until the RAM and part of the swap device is used up, finally the memory (some swapped out) is freed before exit. [akpm@linux-foundation.org: tweak comment] Link: http://lkml.kernel.org/r/20170525005916.25249-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:40:31 +00:00
static int swp_entry_cmp(const void *ent1, const void *ent2)
{
const swp_entry_t *e1 = ent1, *e2 = ent2;
return (int)swp_type(*e1) - (int)swp_type(*e2);
}
void swapcache_free_entries(swp_entry_t *entries, int n)
{
struct swap_info_struct *p, *prev;
int i;
if (n <= 0)
return;
prev = NULL;
p = NULL;
mm/swapfile.c: sort swap entries before free To reduce the lock contention of swap_info_struct->lock when freeing swap entry. The freed swap entries will be collected in a per-CPU buffer firstly, and be really freed later in batch. During the batch freeing, if the consecutive swap entries in the per-CPU buffer belongs to same swap device, the swap_info_struct->lock needs to be acquired/released only once, so that the lock contention could be reduced greatly. But if there are multiple swap devices, it is possible that the lock may be unnecessarily released/acquired because the swap entries belong to the same swap device are non-consecutive in the per-CPU buffer. To solve the issue, the per-CPU buffer is sorted according to the swap device before freeing the swap entries. With the patch, the memory (some swapped out) free time reduced 11.6% (from 2.65s to 2.35s) in the vm-scalability swap-w-rand test case with 16 processes. The test is done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test swapping, the test case creates 16 processes, which allocate and write to the anonymous pages until the RAM and part of the swap device is used up, finally the memory (some swapped out) is freed before exit. [akpm@linux-foundation.org: tweak comment] Link: http://lkml.kernel.org/r/20170525005916.25249-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 22:40:31 +00:00
/*
* Sort swap entries by swap device, so each lock is only taken once.
* nr_swapfiles isn't absolutely correct, but the overhead of sort() is
* so low that it isn't necessary to optimize further.
*/
if (nr_swapfiles > 1)
sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
for (i = 0; i < n; ++i) {
p = swap_info_get_cont(entries[i], prev);
if (p)
swap_entry_free(p, entries[i]);
prev = p;
}
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
if (p)
spin_unlock(&p->lock);
}
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
int __swap_count(swp_entry_t entry)
{
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
struct swap_info_struct *si;
pgoff_t offset = swp_offset(entry);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
int count = 0;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
si = get_swap_device(entry);
if (si) {
count = swap_count(si->swap_map[offset]);
put_swap_device(si);
}
return count;
}
/*
* How many references to @entry are currently swapped out?
* This does not give an exact answer when swap count is continued,
* but does include the high COUNT_CONTINUED flag to allow for that.
*/
mm, swap: Fix a race in free_swap_and_cache() Before using cluster lock in free_swap_and_cache(), the swap_info_struct->lock will be held during freeing the swap entry and acquiring page lock, so the page swap count will not change when testing page information later. But after using cluster lock, the cluster lock (or swap_info_struct->lock) will be held only during freeing the swap entry. So before acquiring the page lock, the page swap count may be changed in another thread. If the page swap count is not 0, we should not delete the page from the swap cache. This is fixed via checking page swap count again after acquiring the page lock. I found the race when I review the code, so I didn't trigger the race via a test program. If the race occurs for an anonymous page shared by multiple processes via fork, multiple pages will be allocated and swapped in from the swap device for the previously shared one page. That is, the user-visible runtime effect is more memory will be used and the access latency for the page will be higher, that is, the performance regression. Link: http://lkml.kernel.org/r/20170301143905.12846-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Tim Chen <tim.c.chen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:52:49 +00:00
static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
{
pgoff_t offset = swp_offset(entry);
struct swap_cluster_info *ci;
int count;
mm, swap: Fix a race in free_swap_and_cache() Before using cluster lock in free_swap_and_cache(), the swap_info_struct->lock will be held during freeing the swap entry and acquiring page lock, so the page swap count will not change when testing page information later. But after using cluster lock, the cluster lock (or swap_info_struct->lock) will be held only during freeing the swap entry. So before acquiring the page lock, the page swap count may be changed in another thread. If the page swap count is not 0, we should not delete the page from the swap cache. This is fixed via checking page swap count again after acquiring the page lock. I found the race when I review the code, so I didn't trigger the race via a test program. If the race occurs for an anonymous page shared by multiple processes via fork, multiple pages will be allocated and swapped in from the swap device for the previously shared one page. That is, the user-visible runtime effect is more memory will be used and the access latency for the page will be higher, that is, the performance regression. Link: http://lkml.kernel.org/r/20170301143905.12846-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Tim Chen <tim.c.chen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:52:49 +00:00
ci = lock_cluster_or_swap_info(si, offset);
count = swap_count(si->swap_map[offset]);
unlock_cluster_or_swap_info(si, ci);
return count;
}
/*
* How many references to @entry are currently swapped out?
* This does not give an exact answer when swap count is continued,
* but does include the high COUNT_CONTINUED flag to allow for that.
*/
int __swp_swapcount(swp_entry_t entry)
{
int count = 0;
struct swap_info_struct *si;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
si = get_swap_device(entry);
if (si) {
mm, swap: Fix a race in free_swap_and_cache() Before using cluster lock in free_swap_and_cache(), the swap_info_struct->lock will be held during freeing the swap entry and acquiring page lock, so the page swap count will not change when testing page information later. But after using cluster lock, the cluster lock (or swap_info_struct->lock) will be held only during freeing the swap entry. So before acquiring the page lock, the page swap count may be changed in another thread. If the page swap count is not 0, we should not delete the page from the swap cache. This is fixed via checking page swap count again after acquiring the page lock. I found the race when I review the code, so I didn't trigger the race via a test program. If the race occurs for an anonymous page shared by multiple processes via fork, multiple pages will be allocated and swapped in from the swap device for the previously shared one page. That is, the user-visible runtime effect is more memory will be used and the access latency for the page will be higher, that is, the performance regression. Link: http://lkml.kernel.org/r/20170301143905.12846-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Tim Chen <tim.c.chen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 21:52:49 +00:00
count = swap_swapcount(si, entry);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
put_swap_device(si);
}
return count;
}
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
/*
* How many references to @entry are currently swapped out?
* This considers COUNT_CONTINUED so it returns exact answer.
*/
int swp_swapcount(swp_entry_t entry)
{
int count, tmp_count, n;
struct swap_info_struct *p;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci;
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
struct page *page;
pgoff_t offset;
unsigned char *map;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
p = _swap_info_get(entry);
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
if (!p)
return 0;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
offset = swp_offset(entry);
ci = lock_cluster_or_swap_info(p, offset);
count = swap_count(p->swap_map[offset]);
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
if (!(count & COUNT_CONTINUED))
goto out;
count &= ~COUNT_CONTINUED;
n = SWAP_MAP_MAX + 1;
page = vmalloc_to_page(p->swap_map + offset);
offset &= ~PAGE_MASK;
VM_BUG_ON(page_private(page) != SWP_CONTINUED);
do {
page = list_next_entry(page, lru);
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
map = kmap_atomic(page);
tmp_count = map[offset];
kunmap_atomic(map);
count += (tmp_count & ~COUNT_CONTINUED) * n;
n *= (SWAP_CONT_MAX + 1);
} while (tmp_count & COUNT_CONTINUED);
out:
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster_or_swap_info(p, ci);
mm: /proc/pid/smaps:: show proportional swap share of the mapping We want to know per-process workingset size for smart memory management on userland and we use swap(ex, zram) heavily to maximize memory efficiency so workingset includes swap as well as RSS. On such system, if there are lots of shared anonymous pages, it's really hard to figure out exactly how many each process consumes memory(ie, rss + wap) if the system has lots of shared anonymous memory(e.g, android). This patch introduces SwapPss field on /proc/<pid>/smaps so we can get more exact workingset size per process. Bongkyu tested it. Result is below. 1. 50M used swap SwapTotal: 461976 kB SwapFree: 411192 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 48236 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 141184 2. 240M used swap SwapTotal: 461976 kB SwapFree: 216808 kB $ adb shell cat /proc/*/smaps | grep "SwapPss:" | awk '{sum += $2} END {print sum}'; 230315 $ adb shell cat /proc/*/smaps | grep "Swap:" | awk '{sum += $2} END {print sum}'; 1387744 [akpm@linux-foundation.org: simplify kunmap_atomic() call] Signed-off-by: Minchan Kim <minchan@kernel.org> Reported-by: Bongkyu Kim <bongkyu.kim@lge.com> Tested-by: Bongkyu Kim <bongkyu.kim@lge.com> Cc: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Jerome Marchand <jmarchan@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-08 22:00:24 +00:00
return count;
}
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
swp_entry_t entry)
{
struct swap_cluster_info *ci;
unsigned char *map = si->swap_map;
unsigned long roffset = swp_offset(entry);
unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
int i;
bool ret = false;
ci = lock_cluster_or_swap_info(si, offset);
if (!ci || !cluster_is_huge(ci)) {
if (swap_count(map[roffset]))
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
ret = true;
goto unlock_out;
}
for (i = 0; i < SWAPFILE_CLUSTER; i++) {
if (swap_count(map[offset + i])) {
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
ret = true;
break;
}
}
unlock_out:
unlock_cluster_or_swap_info(si, ci);
return ret;
}
static bool folio_swapped(struct folio *folio)
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
{
swp_entry_t entry = folio_swap_entry(folio);
struct swap_info_struct *si = _swap_info_get(entry);
if (!si)
return false;
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
return swap_swapcount(si, entry) != 0;
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
return swap_page_trans_huge_swapped(si, entry);
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
}
/**
* folio_free_swap() - Free the swap space used for this folio.
* @folio: The folio to remove.
*
* If swap is getting full, or if there are no more mappings of this folio,
* then call folio_free_swap to free its swap space.
*
* Return: true if we were able to release the swap space.
*/
bool folio_free_swap(struct folio *folio)
{
VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
if (!folio_test_swapcache(folio))
return false;
if (folio_test_writeback(folio))
return false;
if (folio_swapped(folio))
return false;
/*
* Once hibernation has begun to create its image of memory,
* there's a danger that one of the calls to folio_free_swap()
* - most probably a call from __try_to_reclaim_swap() while
* hibernation is allocating its own swap pages for the image,
* but conceivably even a call from memory reclaim - will free
* the swap from a folio which has already been recorded in the
* image as a clean swapcache folio, and then reuse its swap for
* another page of the image. On waking from hibernation, the
* original folio might be freed under memory pressure, then
* later read back in from swap, now with the wrong data.
*
* Hibernation suspends storage while it is writing the image
mm: avoid livelock on !__GFP_FS allocations Colin Cross reported; Under the following conditions, __alloc_pages_slowpath can loop forever: gfp_mask & __GFP_WAIT is true gfp_mask & __GFP_FS is false reclaim and compaction make no progress order <= PAGE_ALLOC_COSTLY_ORDER These conditions happen very often during suspend and resume, when pm_restrict_gfp_mask() effectively converts all GFP_KERNEL allocations into __GFP_WAIT. The oom killer is not run because gfp_mask & __GFP_FS is false, but should_alloc_retry will always return true when order is less than PAGE_ALLOC_COSTLY_ORDER. In his fix, he avoided retrying the allocation if reclaim made no progress and __GFP_FS was not set. The problem is that this would result in GFP_NOIO allocations failing that previously succeeded which would be very unfortunate. The big difference between GFP_NOIO and suspend converting GFP_KERNEL to behave like GFP_NOIO is that normally flushers will be cleaning pages and kswapd reclaims pages allowing GFP_NOIO to succeed after a short delay. The same does not necessarily apply during suspend as the storage device may be suspended. This patch special cases the suspend case to fail the page allocation if reclaim cannot make progress and adds some documentation on how gfp_allowed_mask is currently used. Failing allocations like this may cause suspend to abort but that is better than a livelock. [mgorman@suse.de: Rework fix to be suspend specific] [rientjes@google.com: Move suspended device check to should_alloc_retry] Reported-by: Colin Cross <ccross@android.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: David Rientjes <rientjes@google.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 23:07:15 +00:00
* to disk so check that here.
*/
mm: avoid livelock on !__GFP_FS allocations Colin Cross reported; Under the following conditions, __alloc_pages_slowpath can loop forever: gfp_mask & __GFP_WAIT is true gfp_mask & __GFP_FS is false reclaim and compaction make no progress order <= PAGE_ALLOC_COSTLY_ORDER These conditions happen very often during suspend and resume, when pm_restrict_gfp_mask() effectively converts all GFP_KERNEL allocations into __GFP_WAIT. The oom killer is not run because gfp_mask & __GFP_FS is false, but should_alloc_retry will always return true when order is less than PAGE_ALLOC_COSTLY_ORDER. In his fix, he avoided retrying the allocation if reclaim made no progress and __GFP_FS was not set. The problem is that this would result in GFP_NOIO allocations failing that previously succeeded which would be very unfortunate. The big difference between GFP_NOIO and suspend converting GFP_KERNEL to behave like GFP_NOIO is that normally flushers will be cleaning pages and kswapd reclaims pages allowing GFP_NOIO to succeed after a short delay. The same does not necessarily apply during suspend as the storage device may be suspended. This patch special cases the suspend case to fail the page allocation if reclaim cannot make progress and adds some documentation on how gfp_allowed_mask is currently used. Failing allocations like this may cause suspend to abort but that is better than a livelock. [mgorman@suse.de: Rework fix to be suspend specific] [rientjes@google.com: Move suspended device check to should_alloc_retry] Reported-by: Colin Cross <ccross@android.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: David Rientjes <rientjes@google.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Pekka Enberg <penberg@cs.helsinki.fi> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-01-10 23:07:15 +00:00
if (pm_suspended_storage())
return false;
delete_from_swap_cache(folio);
folio_set_dirty(folio);
return true;
}
/*
* Free the swap entry like above, but also try to
* free the page cache entry if it is the last user.
*/
int free_swap_and_cache(swp_entry_t entry)
{
struct swap_info_struct *p;
unsigned char count;
if (non_swap_entry(entry))
return 1;
[PATCH] Swapless page migration: add R/W migration entries Implement read/write migration ptes We take the upper two swapfiles for the two types of migration ptes and define a series of macros in swapops.h. The VM is modified to handle the migration entries. migration entries can only be encountered when the page they are pointing to is locked. This limits the number of places one has to fix. We also check in copy_pte_range and in mprotect_pte_range() for migration ptes. We check for migration ptes in do_swap_cache and call a function that will then wait on the page lock. This allows us to effectively stop all accesses to apge. Migration entries are created by try_to_unmap if called for migration and removed by local functions in migrate.c From: Hugh Dickins <hugh@veritas.com> Several times while testing swapless page migration (I've no NUMA, just hacking it up to migrate recklessly while running load), I've hit the BUG_ON(!PageLocked(p)) in migration_entry_to_page. This comes from an orphaned migration entry, unrelated to the current correctly locked migration, but hit by remove_anon_migration_ptes as it checks an address in each vma of the anon_vma list. Such an orphan may be left behind if an earlier migration raced with fork: copy_one_pte can duplicate a migration entry from parent to child, after remove_anon_migration_ptes has checked the child vma, but before it has removed it from the parent vma. (If the process were later to fault on this orphaned entry, it would hit the same BUG from migration_entry_wait.) This could be fixed by locking anon_vma in copy_one_pte, but we'd rather not. There's no such problem with file pages, because vma_prio_tree_add adds child vma after parent vma, and the page table locking at each end is enough to serialize. Follow that example with anon_vma: add new vmas to the tail instead of the head. (There's no corresponding problem when inserting migration entries, because a missed pte will leave the page count and mapcount high, which is allowed for. And there's no corresponding problem when migrating via swap, because a leftover swap entry will be correctly faulted. But the swapless method has no refcounting of its entries.) From: Ingo Molnar <mingo@elte.hu> pte_unmap_unlock() takes the pte pointer as an argument. From: Hugh Dickins <hugh@veritas.com> Several times while testing swapless page migration, gcc has tried to exec a pointer instead of a string: smells like COW mappings are not being properly write-protected on fork. The protection in copy_one_pte looks very convincing, until at last you realize that the second arg to make_migration_entry is a boolean "write", and SWP_MIGRATION_READ is 30. Anyway, it's better done like in change_pte_range, using is_write_migration_entry and make_migration_entry_read. From: Hugh Dickins <hugh@veritas.com> Remove unnecessary obfuscation from sys_swapon's range check on swap type, which blew up causing memory corruption once swapless migration made MAX_SWAPFILES no longer 2 ^ MAX_SWAPFILES_SHIFT. Signed-off-by: Hugh Dickins <hugh@veritas.com> Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Christoph Lameter <clameter@engr.sgi.com> Signed-off-by: Ingo Molnar <mingo@elte.hu> From: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-23 09:03:35 +00:00
p = _swap_info_get(entry);
if (p) {
count = __swap_entry_free(p, entry);
mm, THP, swap: support to reclaim swap space for THP swapped out The normal swap slot reclaiming can be done when the swap count reaches SWAP_HAS_CACHE. But for the swap slot which is backing a THP, all swap slots backing one THP must be reclaimed together, because the swap slot may be used again when the THP is swapped out again later. So the swap slots backing one THP can be reclaimed together when the swap count for all swap slots for the THP reached SWAP_HAS_CACHE. In the patch, the functions to check whether the swap count for all swap slots backing one THP reached SWAP_HAS_CACHE are implemented and used when checking whether a swap slot can be reclaimed. To make it easier to determine whether a swap slot is backing a THP, a new swap cluster flag named CLUSTER_FLAG_HUGE is added to mark a swap cluster which is backing a THP (Transparent Huge Page). Because THP swap in as a whole isn't supported now. After deleting the THP from the swap cache (for example, swapping out finished), the CLUSTER_FLAG_HUGE flag will be cleared. So that, the normal pages inside THP can be swapped in individually. [ying.huang@intel.com: fix swap_page_trans_huge_swapped on HDD] Link: http://lkml.kernel.org/r/874ltsm0bi.fsf@yhuang-dev.intel.com Link: http://lkml.kernel.org/r/20170724051840.2309-3-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Michal Hocko <mhocko@kernel.org> Cc: Ross Zwisler <ross.zwisler@intel.com> [for brd.c, zram_drv.c, pmem.c] Cc: Vishal L Verma <vishal.l.verma@intel.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:22:16 +00:00
if (count == SWAP_HAS_CACHE &&
mm/swapfile.c: use __try_to_reclaim_swap() in free_swap_and_cache() The code path to reclaim the swap entry in free_swap_and_cache() is almost same as that of __try_to_reclaim_swap(). The largest difference is just coding style. So the support to the additional requirement of free_swap_and_cache() is added into __try_to_reclaim_swap(). free_swap_and_cache() is changed to call __try_to_reclaim_swap(), and delete the duplicated code. This will improve code readability and reduce the potential bugs. There are 2 functionality differences between __try_to_reclaim_swap() and swap entry reclaim code of free_swap_and_cache(). - free_swap_and_cache() only reclaims the swap entry if the page is unmapped or swap is getting full. The support has been added into __try_to_reclaim_swap(). - try_to_free_swap() (called by __try_to_reclaim_swap()) checks pm_suspended_storage(), while free_swap_and_cache() not. I think this is OK. Because the page and the swap entry can be reclaimed later eventually. Link: http://lkml.kernel.org/r/20180827075535.17406-2-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-26 22:03:46 +00:00
!swap_page_trans_huge_swapped(p, entry))
__try_to_reclaim_swap(p, swp_offset(entry),
TTRS_UNMAPPED | TTRS_FULL);
}
return p != NULL;
}
#ifdef CONFIG_HIBERNATION
swp_entry_t get_swap_page_of_type(int type)
{
struct swap_info_struct *si = swap_type_to_swap_info(type);
swp_entry_t entry = {0};
if (!si)
goto fail;
/* This is called for allocating swap entry, not cache */
spin_lock(&si->lock);
if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
atomic_long_dec(&nr_swap_pages);
spin_unlock(&si->lock);
fail:
return entry;
}
/*
2006-12-07 04:34:07 +00:00
* Find the swap type that corresponds to given device (if any).
*
2006-12-07 04:34:07 +00:00
* @offset - number of the PAGE_SIZE-sized block of the device, starting
* from 0, in which the swap header is expected to be located.
*
* This is needed for the suspend to disk (aka swsusp).
*/
int swap_type_of(dev_t device, sector_t offset)
{
int type;
if (!device)
return -1;
2006-12-07 04:34:07 +00:00
spin_lock(&swap_lock);
for (type = 0; type < nr_swapfiles; type++) {
struct swap_info_struct *sis = swap_info[type];
2006-12-07 04:34:07 +00:00
if (!(sis->flags & SWP_WRITEOK))
continue;
if (device == sis->bdev->bd_dev) {
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
struct swap_extent *se = first_se(sis);
2006-12-07 04:34:07 +00:00
if (se->start_block == offset) {
spin_unlock(&swap_lock);
return type;
2006-12-07 04:34:07 +00:00
}
}
}
spin_unlock(&swap_lock);
return -ENODEV;
}
2006-12-07 04:34:07 +00:00
int find_first_swap(dev_t *device)
{
int type;
2006-12-07 04:34:07 +00:00
spin_lock(&swap_lock);
for (type = 0; type < nr_swapfiles; type++) {
struct swap_info_struct *sis = swap_info[type];
if (!(sis->flags & SWP_WRITEOK))
continue;
*device = sis->bdev->bd_dev;
spin_unlock(&swap_lock);
return type;
}
spin_unlock(&swap_lock);
return -ENODEV;
}
/*
* Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
* corresponding to given index in swap_info (swap type).
*/
sector_t swapdev_block(int type, pgoff_t offset)
{
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
struct swap_info_struct *si = swap_type_to_swap_info(type);
struct swap_extent *se;
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
if (!si || !(si->flags & SWP_WRITEOK))
return 0;
se = offset_to_swap_extent(si, offset);
return se->start_block + (offset - se->start_page);
}
/*
* Return either the total number of swap pages of given type, or the number
* of free pages of that type (depending on @free)
*
* This is needed for software suspend
*/
unsigned int count_swap_pages(int type, int free)
{
unsigned int n = 0;
spin_lock(&swap_lock);
if ((unsigned int)type < nr_swapfiles) {
struct swap_info_struct *sis = swap_info[type];
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&sis->lock);
if (sis->flags & SWP_WRITEOK) {
n = sis->pages;
if (free)
n -= sis->inuse_pages;
}
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&sis->lock);
}
spin_unlock(&swap_lock);
return n;
}
#endif /* CONFIG_HIBERNATION */
static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
{
return pte_same(pte_swp_clear_flags(pte), swp_pte);
}
/*
* No need to decide whether this PTE shares the swap entry with others,
* just let do_wp_page work it out if a write is requested later - to
* force COW, vm_page_prot omits write permission from any private vma.
*/
static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
unsigned long addr, swp_entry_t entry, struct folio *folio)
{
struct page *page = folio_file_page(folio, swp_offset(entry));
struct page *swapcache;
spinlock_t *ptl;
pte_t *pte, new_pte;
bool hwposioned = false;
int ret = 1;
swapcache = page;
page = ksm_might_need_to_copy(page, vma, addr);
if (unlikely(!page))
return -ENOMEM;
else if (unlikely(PTR_ERR(page) == -EHWPOISON))
hwposioned = true;
pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
ret = 0;
goto out;
}
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
if (unlikely(hwposioned || !PageUptodate(page))) {
swp_entry_t swp_entry;
mm/swapfile: unuse_pte can map random data if swap read fails Patch series "A few fixup patches for mm", v4. This series contains a few patches to avoid mapping random data if swap read fails and fix lost swap bits in unuse_pte. Also we free hwpoison and swapin error entry in madvise_free_pte_range and so on. More details can be found in the respective changelogs. This patch (of 5): There is a bug in unuse_pte(): when swap page happens to be unreadable, page filled with random data is mapped into user address space. In case of error, a special swap entry indicating swap read fails is set to the page table. So the swapcache page can be freed and the user won't end up with a permanently mounted swap because a sector is bad. And if the page is accessed later, the user process will be killed so that corrupted data is never consumed. On the other hand, if the page is never accessed, the user won't even notice it. Link: https://lkml.kernel.org/r/20220519125030.21486-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20220519125030.21486-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Acked-by: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Howells <dhowells@redhat.com> Cc: NeilBrown <neilb@suse.de> Cc: Alistair Popple <apopple@nvidia.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Peter Xu <peterx@redhat.com> Cc: Ralph Campbell <rcampbell@nvidia.com> Cc: Naoya Horiguchi <naoya.horiguchi@nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 12:50:26 +00:00
dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
if (hwposioned) {
swp_entry = make_hwpoison_entry(swapcache);
page = swapcache;
} else {
swp_entry = make_swapin_error_entry();
}
new_pte = swp_entry_to_pte(swp_entry);
mm/swapfile: unuse_pte can map random data if swap read fails Patch series "A few fixup patches for mm", v4. This series contains a few patches to avoid mapping random data if swap read fails and fix lost swap bits in unuse_pte. Also we free hwpoison and swapin error entry in madvise_free_pte_range and so on. More details can be found in the respective changelogs. This patch (of 5): There is a bug in unuse_pte(): when swap page happens to be unreadable, page filled with random data is mapped into user address space. In case of error, a special swap entry indicating swap read fails is set to the page table. So the swapcache page can be freed and the user won't end up with a permanently mounted swap because a sector is bad. And if the page is accessed later, the user process will be killed so that corrupted data is never consumed. On the other hand, if the page is never accessed, the user won't even notice it. Link: https://lkml.kernel.org/r/20220519125030.21486-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20220519125030.21486-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Acked-by: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Howells <dhowells@redhat.com> Cc: NeilBrown <neilb@suse.de> Cc: Alistair Popple <apopple@nvidia.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Peter Xu <peterx@redhat.com> Cc: Ralph Campbell <rcampbell@nvidia.com> Cc: Naoya Horiguchi <naoya.horiguchi@nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 12:50:26 +00:00
ret = 0;
goto setpte;
mm/swapfile: unuse_pte can map random data if swap read fails Patch series "A few fixup patches for mm", v4. This series contains a few patches to avoid mapping random data if swap read fails and fix lost swap bits in unuse_pte. Also we free hwpoison and swapin error entry in madvise_free_pte_range and so on. More details can be found in the respective changelogs. This patch (of 5): There is a bug in unuse_pte(): when swap page happens to be unreadable, page filled with random data is mapped into user address space. In case of error, a special swap entry indicating swap read fails is set to the page table. So the swapcache page can be freed and the user won't end up with a permanently mounted swap because a sector is bad. And if the page is accessed later, the user process will be killed so that corrupted data is never consumed. On the other hand, if the page is never accessed, the user won't even notice it. Link: https://lkml.kernel.org/r/20220519125030.21486-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20220519125030.21486-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Acked-by: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: David Howells <dhowells@redhat.com> Cc: NeilBrown <neilb@suse.de> Cc: Alistair Popple <apopple@nvidia.com> Cc: Suren Baghdasaryan <surenb@google.com> Cc: Peter Xu <peterx@redhat.com> Cc: Ralph Campbell <rcampbell@nvidia.com> Cc: Naoya Horiguchi <naoya.horiguchi@nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-19 12:50:26 +00:00
}
mm/page-flags: reuse PG_mappedtodisk as PG_anon_exclusive for PageAnon() pages The basic question we would like to have a reliable and efficient answer to is: is this anonymous page exclusive to a single process or might it be shared? We need that information for ordinary/single pages, hugetlb pages, and possibly each subpage of a THP. Introduce a way to mark an anonymous page as exclusive, with the ultimate goal of teaching our COW logic to not do "wrong COWs", whereby GUP pins lose consistency with the pages mapped into the page table, resulting in reported memory corruptions. Most pageflags already have semantics for anonymous pages, however, PG_mappedtodisk should never apply to pages in the swapcache, so let's reuse that flag. As PG_has_hwpoisoned also uses that flag on the second tail page of a compound page, convert it to PG_error instead, which is marked as PF_NO_TAIL, so never used for tail pages. Use custom page flag modification functions such that we can do additional sanity checks. The semantics we'll put into some kernel doc in the future are: " PG_anon_exclusive is *usually* only expressive in combination with a page table entry. Depending on the page table entry type it might store the following information: Is what's mapped via this page table entry exclusive to the single process and can be mapped writable without further checks? If not, it might be shared and we might have to COW. For now, we only expect PTE-mapped THPs to make use of PG_anon_exclusive in subpages. For other anonymous compound folios (i.e., hugetlb), only the head page is logically mapped and holds this information. For example, an exclusive, PMD-mapped THP only has PG_anon_exclusive set on the head page. When replacing the PMD by a page table full of PTEs, PG_anon_exclusive, if set on the head page, will be set on all tail pages accordingly. Note that converting from a PTE-mapping to a PMD mapping using the same compound page is currently not possible and consequently doesn't require care. If GUP wants to take a reliable pin (FOLL_PIN) on an anonymous page, it should only pin if the relevant PG_anon_exclusive is set. In that case, the pin will be fully reliable and stay consistent with the pages mapped into the page table, as the bit cannot get cleared (e.g., by fork(), KSM) while the page is pinned. For anonymous pages that are mapped R/W, PG_anon_exclusive can be assumed to always be set because such pages cannot possibly be shared. The page table lock protecting the page table entry is the primary synchronization mechanism for PG_anon_exclusive; GUP-fast that does not take the PT lock needs special care when trying to clear the flag. Page table entry types and PG_anon_exclusive: * Present: PG_anon_exclusive applies. * Swap: the information is lost. PG_anon_exclusive was cleared. * Migration: the entry holds this information instead. PG_anon_exclusive was cleared. * Device private: PG_anon_exclusive applies. * Device exclusive: PG_anon_exclusive applies. * HW Poison: PG_anon_exclusive is stale and not changed. If the page may be pinned (FOLL_PIN), clearing PG_anon_exclusive is not allowed and the flag will stick around until the page is freed and folio->mapping is cleared. " We won't be clearing PG_anon_exclusive on destructive unmapping (i.e., zapping) of page table entries, page freeing code will handle that when also invalidate page->mapping to not indicate PageAnon() anymore. Letting information about exclusivity stick around will be an important property when adding sanity checks to unpinning code. Note that we properly clear the flag in free_pages_prepare() via PAGE_FLAGS_CHECK_AT_PREP for each individual subpage of a compound page, so there is no need to manually clear the flag. Link: https://lkml.kernel.org/r/20220428083441.37290-12-david@redhat.com Signed-off-by: David Hildenbrand <david@redhat.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: David Rientjes <rientjes@google.com> Cc: Don Dutile <ddutile@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jan Kara <jack@suse.cz> Cc: Jann Horn <jannh@google.com> Cc: Jason Gunthorpe <jgg@nvidia.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Khalid Aziz <khalid.aziz@oracle.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Liang Zhang <zhangliang5@huawei.com> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport <rppt@linux.ibm.com> Cc: Nadav Amit <namit@vmware.com> Cc: Oded Gabbay <oded.gabbay@gmail.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com> Cc: Peter Xu <peterx@redhat.com> Cc: Rik van Riel <riel@surriel.com> Cc: Roman Gushchin <guro@fb.com> Cc: Shakeel Butt <shakeelb@google.com> Cc: Yang Shi <shy828301@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 01:20:44 +00:00
/* See do_swap_page() */
BUG_ON(!PageAnon(page) && PageMappedToDisk(page));
BUG_ON(PageAnon(page) && PageAnonExclusive(page));
dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
get_page(page);
mm: memcontrol: rewrite charge API These patches rework memcg charge lifetime to integrate more naturally with the lifetime of user pages. This drastically simplifies the code and reduces charging and uncharging overhead. The most expensive part of charging and uncharging is the page_cgroup bit spinlock, which is removed entirely after this series. Here are the top-10 profile entries of a stress test that reads a 128G sparse file on a freshly booted box, without even a dedicated cgroup (i.e. executing in the root memcg). Before: 15.36% cat [kernel.kallsyms] [k] copy_user_generic_string 13.31% cat [kernel.kallsyms] [k] memset 11.48% cat [kernel.kallsyms] [k] do_mpage_readpage 4.23% cat [kernel.kallsyms] [k] get_page_from_freelist 2.38% cat [kernel.kallsyms] [k] put_page 2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge 2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common 1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn After: 15.67% cat [kernel.kallsyms] [k] copy_user_generic_string 13.48% cat [kernel.kallsyms] [k] memset 11.42% cat [kernel.kallsyms] [k] do_mpage_readpage 3.98% cat [kernel.kallsyms] [k] get_page_from_freelist 2.46% cat [kernel.kallsyms] [k] put_page 2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn 1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk 1.30% cat [kernel.kallsyms] [k] kfree As you can see, the memcg footprint has shrunk quite a bit. text data bss dec hex filename 37970 9892 400 48262 bc86 mm/memcontrol.o.old 35239 9892 400 45531 b1db mm/memcontrol.o This patch (of 4): The memcg charge API charges pages before they are rmapped - i.e. have an actual "type" - and so every callsite needs its own set of charge and uncharge functions to know what type is being operated on. Worse, uncharge has to happen from a context that is still type-specific, rather than at the end of the page's lifetime with exclusive access, and so requires a lot of synchronization. Rewrite the charge API to provide a generic set of try_charge(), commit_charge() and cancel_charge() transaction operations, much like what's currently done for swap-in: mem_cgroup_try_charge() attempts to reserve a charge, reclaiming pages from the memcg if necessary. mem_cgroup_commit_charge() commits the page to the charge once it has a valid page->mapping and PageAnon() reliably tells the type. mem_cgroup_cancel_charge() aborts the transaction. This reduces the charge API and enables subsequent patches to drastically simplify uncharging. As pages need to be committed after rmap is established but before they are added to the LRU, page_add_new_anon_rmap() must stop doing LRU additions again. Revive lru_cache_add_active_or_unevictable(). [hughd@google.com: fix shmem_unuse] [hughd@google.com: Add comments on the private use of -EAGAIN] Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Michal Hocko <mhocko@suse.cz> Cc: Tejun Heo <tj@kernel.org> Cc: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 21:19:20 +00:00
if (page == swapcache) {
mm/swap: remember PG_anon_exclusive via a swp pte bit Patch series "mm: COW fixes part 3: reliable GUP R/W FOLL_GET of anonymous pages", v2. This series fixes memory corruptions when a GUP R/W reference (FOLL_WRITE | FOLL_GET) was taken on an anonymous page and COW logic fails to detect exclusivity of the page to then replacing the anonymous page by a copy in the page table: The GUP reference lost synchronicity with the pages mapped into the page tables. This series focuses on x86, arm64, s390x and ppc64/book3s -- other architectures are fairly easy to support by implementing __HAVE_ARCH_PTE_SWP_EXCLUSIVE. This primarily fixes the O_DIRECT memory corruptions that can happen on concurrent swapout, whereby we lose DMA reads to a page (modifying the user page by writing to it). O_DIRECT currently uses FOLL_GET for short-term (!FOLL_LONGTERM) DMA from/to a user page. In the long run, we want to convert it to properly use FOLL_PIN, and John is working on it, but that might take a while and might not be easy to backport. In the meantime, let's restore what used to work before we started modifying our COW logic: make R/W FOLL_GET references reliable as long as there is no fork() after GUP involved. This is just the natural follow-up of part 2, that will also further reduce "wrong COW" on the swapin path, for example, when we cannot remove a page from the swapcache due to concurrent writeback, or if we have two threads faulting on the same swapped-out page. Fixing O_DIRECT is just a nice side-product This issue, including other related COW issues, has been summarized in [3] under 2): " 2. Intra Process Memory Corruptions due to Wrong COW (FOLL_GET) It was discovered that we can create a memory corruption by reading a file via O_DIRECT to a part (e.g., first 512 bytes) of a page, concurrently writing to an unrelated part (e.g., last byte) of the same page, and concurrently write-protecting the page via clear_refs SOFTDIRTY tracking [6]. For the reproducer, the issue is that O_DIRECT grabs a reference of the target page (via FOLL_GET) and clear_refs write-protects the relevant page table entry. On successive write access to the page from the process itself, we wrongly COW the page when resolving the write fault, resulting in a loss of synchronicity and consequently a memory corruption. While some people might think that using clear_refs in this combination is a corner cases, it turns out to be a more generic problem unfortunately. For example, it was just recently discovered that we can similarly create a memory corruption without clear_refs, simply by concurrently swapping out the buffer pages [7]. Note that we nowadays even use the swap infrastructure in Linux without an actual swap disk/partition: the prime example is zram which is enabled as default under Fedora [10]. The root issue is that a write-fault on a page that has additional references results in a COW and thereby a loss of synchronicity and consequently a memory corruption if two parties believe they are referencing the same page. " We don't particularly care about R/O FOLL_GET references: they were never reliable and O_DIRECT doesn't expect to observe modifications from a page after DMA was started. Note that: * this only fixes the issue on x86, arm64, s390x and ppc64/book3s ("enterprise architectures"). Other architectures have to implement __HAVE_ARCH_PTE_SWP_EXCLUSIVE to achieve the same. * this does *not * consider any kind of fork() after taking the reference: fork() after GUP never worked reliably with FOLL_GET. * Not losing PG_anon_exclusive during swapout was the last remaining piece. KSM already makes sure that there are no other references on a page before considering it for sharing. Page migration maintains PG_anon_exclusive and simply fails when there are additional references (freezing the refcount fails). Only swapout code dropped the PG_anon_exclusive flag because it requires more work to remember + restore it. With this series in place, most COW issues of [3] are fixed on said architectures. Other architectures can implement __HAVE_ARCH_PTE_SWP_EXCLUSIVE fairly easily. [1] https://lkml.kernel.org/r/20220329160440.193848-1-david@redhat.com [2] https://lkml.kernel.org/r/20211217113049.23850-1-david@redhat.com [3] https://lore.kernel.org/r/3ae33b08-d9ef-f846-56fb-645e3b9b4c66@redhat.com This patch (of 8): Currently, we clear PG_anon_exclusive in try_to_unmap() and forget about it. We do this, to keep fork() logic on swap entries easy and efficient: for example, if we wouldn't clear it when unmapping, we'd have to lookup the page in the swapcache for each and every swap entry during fork() and clear PG_anon_exclusive if set. Instead, we want to store that information directly in the swap pte, protected by the page table lock, similarly to how we handle SWP_MIGRATION_READ_EXCLUSIVE for migration entries. However, for actual swap entries, we don't want to mess with the swap type (e.g., still one bit) because it overcomplicates swap code. In try_to_unmap(), we already reject to unmap in case the page might be pinned, because we must not lose PG_anon_exclusive on pinned pages ever. Checking if there are other unexpected references reliably *before* completely unmapping a page is unfortunately not really possible: THP heavily overcomplicate the situation. Once fully unmapped it's easier -- we, for example, make sure that there are no unexpected references *after* unmapping a page before starting writeback on that page. So, we currently might end up unmapping a page and clearing PG_anon_exclusive if that page has additional references, for example, due to a FOLL_GET. do_swap_page() has to re-determine if a page is exclusive, which will easily fail if there are other references on a page, most prominently GUP references via FOLL_GET. This can currently result in memory corruptions when taking a FOLL_GET | FOLL_WRITE reference on a page even when fork() is never involved: try_to_unmap() will succeed, and when refaulting the page, it cannot be marked exclusive and will get replaced by a copy in the page tables on the next write access, resulting in writes via the GUP reference to the page being lost. In an ideal world, everybody that uses GUP and wants to modify page content, such as O_DIRECT, would properly use FOLL_PIN. However, that conversion will take a while. It's easier to fix what used to work in the past (FOLL_GET | FOLL_WRITE) remembering PG_anon_exclusive. In addition, by remembering PG_anon_exclusive we can further reduce unnecessary COW in some cases, so it's the natural thing to do. So let's transfer the PG_anon_exclusive information to the swap pte and store it via an architecture-dependant pte bit; use that information when restoring the swap pte in do_swap_page() and unuse_pte(). During fork(), we simply have to clear the pte bit and are done. Of course, there is one corner case to handle: swap backends that don't support concurrent page modifications while the page is under writeback. Special case these, and drop the exclusive marker. Add a comment why that is just fine (also, reuse_swap_page() would have done the same in the past). In the future, we'll hopefully have all architectures support __HAVE_ARCH_PTE_SWP_EXCLUSIVE, such that we can get rid of the empty stubs and the define completely. Then, we can also convert SWP_MIGRATION_READ_EXCLUSIVE. For architectures it's fairly easy to support: either simply use a yet unused pte bit that can be used for swap entries, steal one from the arch type bits if they exceed 5, or steal one from the offset bits. Note: R/O FOLL_GET references were never really reliable, especially when taking one on a shared page and then writing to the page (e.g., GUP after fork()). FOLL_GET, including R/W references, were never really reliable once fork was involved (e.g., GUP before fork(), GUP during fork()). KSM steps back in case it stumbles over unexpected references and is, therefore, fine. [david@redhat.com: fix SWP_STABLE_WRITES test] Link: https://lkml.kernel.org/r/ac725bcb-313a-4fff-250a-68ba9a8f85fb@redhat.comLink: https://lkml.kernel.org/r/20220329164329.208407-1-david@redhat.com Link: https://lkml.kernel.org/r/20220329164329.208407-2-david@redhat.com Signed-off-by: David Hildenbrand <david@redhat.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Hugh Dickins <hughd@google.com> Cc: Shakeel Butt <shakeelb@google.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Jason Gunthorpe <jgg@nvidia.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport <rppt@linux.ibm.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Jann Horn <jannh@google.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Nadav Amit <namit@vmware.com> Cc: Rik van Riel <riel@surriel.com> Cc: Roman Gushchin <guro@fb.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Peter Xu <peterx@redhat.com> Cc: Don Dutile <ddutile@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Liang Zhang <zhangliang5@huawei.com> Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com> Cc: Oded Gabbay <oded.gabbay@gmail.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will@kernel.org> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Heiko Carstens <hca@linux.ibm.com> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 01:20:45 +00:00
rmap_t rmap_flags = RMAP_NONE;
/*
* See do_swap_page(): PageWriteback() would be problematic.
* However, we do a wait_on_page_writeback() just before this
* call and have the page locked.
*/
VM_BUG_ON_PAGE(PageWriteback(page), page);
if (pte_swp_exclusive(*pte))
rmap_flags |= RMAP_EXCLUSIVE;
page_add_anon_rmap(page, vma, addr, rmap_flags);
mm: memcontrol: rewrite charge API These patches rework memcg charge lifetime to integrate more naturally with the lifetime of user pages. This drastically simplifies the code and reduces charging and uncharging overhead. The most expensive part of charging and uncharging is the page_cgroup bit spinlock, which is removed entirely after this series. Here are the top-10 profile entries of a stress test that reads a 128G sparse file on a freshly booted box, without even a dedicated cgroup (i.e. executing in the root memcg). Before: 15.36% cat [kernel.kallsyms] [k] copy_user_generic_string 13.31% cat [kernel.kallsyms] [k] memset 11.48% cat [kernel.kallsyms] [k] do_mpage_readpage 4.23% cat [kernel.kallsyms] [k] get_page_from_freelist 2.38% cat [kernel.kallsyms] [k] put_page 2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge 2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common 1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn After: 15.67% cat [kernel.kallsyms] [k] copy_user_generic_string 13.48% cat [kernel.kallsyms] [k] memset 11.42% cat [kernel.kallsyms] [k] do_mpage_readpage 3.98% cat [kernel.kallsyms] [k] get_page_from_freelist 2.46% cat [kernel.kallsyms] [k] put_page 2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn 1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk 1.30% cat [kernel.kallsyms] [k] kfree As you can see, the memcg footprint has shrunk quite a bit. text data bss dec hex filename 37970 9892 400 48262 bc86 mm/memcontrol.o.old 35239 9892 400 45531 b1db mm/memcontrol.o This patch (of 4): The memcg charge API charges pages before they are rmapped - i.e. have an actual "type" - and so every callsite needs its own set of charge and uncharge functions to know what type is being operated on. Worse, uncharge has to happen from a context that is still type-specific, rather than at the end of the page's lifetime with exclusive access, and so requires a lot of synchronization. Rewrite the charge API to provide a generic set of try_charge(), commit_charge() and cancel_charge() transaction operations, much like what's currently done for swap-in: mem_cgroup_try_charge() attempts to reserve a charge, reclaiming pages from the memcg if necessary. mem_cgroup_commit_charge() commits the page to the charge once it has a valid page->mapping and PageAnon() reliably tells the type. mem_cgroup_cancel_charge() aborts the transaction. This reduces the charge API and enables subsequent patches to drastically simplify uncharging. As pages need to be committed after rmap is established but before they are added to the LRU, page_add_new_anon_rmap() must stop doing LRU additions again. Revive lru_cache_add_active_or_unevictable(). [hughd@google.com: fix shmem_unuse] [hughd@google.com: Add comments on the private use of -EAGAIN] Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Michal Hocko <mhocko@suse.cz> Cc: Tejun Heo <tj@kernel.org> Cc: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 21:19:20 +00:00
} else { /* ksm created a completely new copy */
mm/rmap: drop "compound" parameter from page_add_new_anon_rmap() New anonymous pages are always mapped natively: only THP/khugepaged code maps a new compound anonymous page and passes "true". Otherwise, we're just dealing with simple, non-compound pages. Let's give the interface clearer semantics and document these. Remove the PageTransCompound() sanity check from page_add_new_anon_rmap(). Link: https://lkml.kernel.org/r/20220428083441.37290-9-david@redhat.com Signed-off-by: David Hildenbrand <david@redhat.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christoph Hellwig <hch@lst.de> Cc: David Rientjes <rientjes@google.com> Cc: Don Dutile <ddutile@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jan Kara <jack@suse.cz> Cc: Jann Horn <jannh@google.com> Cc: Jason Gunthorpe <jgg@nvidia.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Khalid Aziz <khalid.aziz@oracle.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Liang Zhang <zhangliang5@huawei.com> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: Mike Rapoport <rppt@linux.ibm.com> Cc: Nadav Amit <namit@vmware.com> Cc: Oded Gabbay <oded.gabbay@gmail.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Pedro Demarchi Gomes <pedrodemargomes@gmail.com> Cc: Peter Xu <peterx@redhat.com> Cc: Rik van Riel <riel@surriel.com> Cc: Roman Gushchin <guro@fb.com> Cc: Shakeel Butt <shakeelb@google.com> Cc: Yang Shi <shy828301@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 01:20:43 +00:00
page_add_new_anon_rmap(page, vma, addr);
mm/vmscan: protect the workingset on anonymous LRU In current implementation, newly created or swap-in anonymous page is started on active list. Growing active list results in rebalancing active/inactive list so old pages on active list are demoted to inactive list. Hence, the page on active list isn't protected at all. Following is an example of this situation. Assume that 50 hot pages on active list. Numbers denote the number of pages on active/inactive list (active | inactive). 1. 50 hot pages on active list 50(h) | 0 2. workload: 50 newly created (used-once) pages 50(uo) | 50(h) 3. workload: another 50 newly created (used-once) pages 50(uo) | 50(uo), swap-out 50(h) This patch tries to fix this issue. Like as file LRU, newly created or swap-in anonymous pages will be inserted to the inactive list. They are promoted to active list if enough reference happens. This simple modification changes the above example as following. 1. 50 hot pages on active list 50(h) | 0 2. workload: 50 newly created (used-once) pages 50(h) | 50(uo) 3. workload: another 50 newly created (used-once) pages 50(h) | 50(uo), swap-out 50(uo) As you can see, hot pages on active list would be protected. Note that, this implementation has a drawback that the page cannot be promoted and will be swapped-out if re-access interval is greater than the size of inactive list but less than the size of total(active+inactive). To solve this potential issue, following patch will apply workingset detection similar to the one that's already applied to file LRU. Signed-off-by: Joonsoo Kim <iamjoonsoo.kim@lge.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Hugh Dickins <hughd@google.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Link: http://lkml.kernel.org/r/1595490560-15117-3-git-send-email-iamjoonsoo.kim@lge.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-12 01:30:40 +00:00
lru_cache_add_inactive_or_unevictable(page, vma);
mm: memcontrol: rewrite charge API These patches rework memcg charge lifetime to integrate more naturally with the lifetime of user pages. This drastically simplifies the code and reduces charging and uncharging overhead. The most expensive part of charging and uncharging is the page_cgroup bit spinlock, which is removed entirely after this series. Here are the top-10 profile entries of a stress test that reads a 128G sparse file on a freshly booted box, without even a dedicated cgroup (i.e. executing in the root memcg). Before: 15.36% cat [kernel.kallsyms] [k] copy_user_generic_string 13.31% cat [kernel.kallsyms] [k] memset 11.48% cat [kernel.kallsyms] [k] do_mpage_readpage 4.23% cat [kernel.kallsyms] [k] get_page_from_freelist 2.38% cat [kernel.kallsyms] [k] put_page 2.32% cat [kernel.kallsyms] [k] __mem_cgroup_commit_charge 2.18% kswapd0 [kernel.kallsyms] [k] __mem_cgroup_uncharge_common 1.92% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.86% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.62% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn After: 15.67% cat [kernel.kallsyms] [k] copy_user_generic_string 13.48% cat [kernel.kallsyms] [k] memset 11.42% cat [kernel.kallsyms] [k] do_mpage_readpage 3.98% cat [kernel.kallsyms] [k] get_page_from_freelist 2.46% cat [kernel.kallsyms] [k] put_page 2.13% kswapd0 [kernel.kallsyms] [k] shrink_page_list 1.88% cat [kernel.kallsyms] [k] __radix_tree_lookup 1.67% cat [kernel.kallsyms] [k] __pagevec_lru_add_fn 1.39% kswapd0 [kernel.kallsyms] [k] free_pcppages_bulk 1.30% cat [kernel.kallsyms] [k] kfree As you can see, the memcg footprint has shrunk quite a bit. text data bss dec hex filename 37970 9892 400 48262 bc86 mm/memcontrol.o.old 35239 9892 400 45531 b1db mm/memcontrol.o This patch (of 4): The memcg charge API charges pages before they are rmapped - i.e. have an actual "type" - and so every callsite needs its own set of charge and uncharge functions to know what type is being operated on. Worse, uncharge has to happen from a context that is still type-specific, rather than at the end of the page's lifetime with exclusive access, and so requires a lot of synchronization. Rewrite the charge API to provide a generic set of try_charge(), commit_charge() and cancel_charge() transaction operations, much like what's currently done for swap-in: mem_cgroup_try_charge() attempts to reserve a charge, reclaiming pages from the memcg if necessary. mem_cgroup_commit_charge() commits the page to the charge once it has a valid page->mapping and PageAnon() reliably tells the type. mem_cgroup_cancel_charge() aborts the transaction. This reduces the charge API and enables subsequent patches to drastically simplify uncharging. As pages need to be committed after rmap is established but before they are added to the LRU, page_add_new_anon_rmap() must stop doing LRU additions again. Revive lru_cache_add_active_or_unevictable(). [hughd@google.com: fix shmem_unuse] [hughd@google.com: Add comments on the private use of -EAGAIN] Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Michal Hocko <mhocko@suse.cz> Cc: Tejun Heo <tj@kernel.org> Cc: Vladimir Davydov <vdavydov@parallels.com> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-08-08 21:19:20 +00:00
}
new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
if (pte_swp_soft_dirty(*pte))
new_pte = pte_mksoft_dirty(new_pte);
if (pte_swp_uffd_wp(*pte))
new_pte = pte_mkuffd_wp(new_pte);
setpte:
set_pte_at(vma->vm_mm, addr, pte, new_pte);
swap_free(entry);
out:
pte_unmap_unlock(pte, ptl);
if (page != swapcache) {
unlock_page(page);
put_page(page);
}
return ret;
}
static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
unsigned long addr, unsigned long end,
unsigned int type)
{
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
swp_entry_t entry;
pte_t *pte;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
struct swap_info_struct *si;
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
int ret = 0;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
si = swap_info[type];
pte = pte_offset_map(pmd, addr);
do {
struct folio *folio;
unsigned long offset;
unsigned char swp_count;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (!is_swap_pte(*pte))
continue;
entry = pte_to_swp_entry(*pte);
if (swp_type(entry) != type)
continue;
offset = swp_offset(entry);
pte_unmap(pte);
folio = swap_cache_get_folio(entry, vma, addr);
if (!folio) {
struct page *page;
struct vm_fault vmf = {
.vma = vma,
.address = addr,
userfaultfd: provide unmasked address on page-fault Userfaultfd is supposed to provide the full address (i.e., unmasked) of the faulting access back to userspace. However, that is not the case for quite some time. Even running "userfaultfd_demo" from the userfaultfd man page provides the wrong output (and contradicts the man page). Notice that "UFFD_EVENT_PAGEFAULT event" shows the masked address (7fc5e30b3000) and not the first read address (0x7fc5e30b300f). Address returned by mmap() = 0x7fc5e30b3000 fault_handler_thread(): poll() returns: nready = 1; POLLIN = 1; POLLERR = 0 UFFD_EVENT_PAGEFAULT event: flags = 0; address = 7fc5e30b3000 (uffdio_copy.copy returned 4096) Read address 0x7fc5e30b300f in main(): A Read address 0x7fc5e30b340f in main(): A Read address 0x7fc5e30b380f in main(): A Read address 0x7fc5e30b3c0f in main(): A The exact address is useful for various reasons and specifically for prefetching decisions. If it is known that the memory is populated by certain objects whose size is not page-aligned, then based on the faulting address, the uffd-monitor can decide whether to prefetch and prefault the adjacent page. This bug has been for quite some time in the kernel: since commit 1a29d85eb0f1 ("mm: use vmf->address instead of of vmf->virtual_address") vmf->virtual_address"), which dates back to 2016. A concern has been raised that existing userspace application might rely on the old/wrong behavior in which the address is masked. Therefore, it was suggested to provide the masked address unless the user explicitly asks for the exact address. Add a new userfaultfd feature UFFD_FEATURE_EXACT_ADDRESS to direct userfaultfd to provide the exact address. Add a new "real_address" field to vmf to hold the unmasked address. Provide the address to userspace accordingly. Initialize real_address in various code-paths to be consistent with address, even when it is not used, to be on the safe side. [namit@vmware.com: initialize real_address on all code paths, per Jan] Link: https://lkml.kernel.org/r/20220226022655.350562-1-namit@vmware.com [akpm@linux-foundation.org: fix typo in comment, per Jan] Link: https://lkml.kernel.org/r/20220218041003.3508-1-namit@vmware.com Signed-off-by: Nadav Amit <namit@vmware.com> Acked-by: Peter Xu <peterx@redhat.com> Reviewed-by: David Hildenbrand <david@redhat.com> Acked-by: Mike Rapoport <rppt@linux.ibm.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Andrea Arcangeli <aarcange@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-03-22 21:45:32 +00:00
.real_address = addr,
.pmd = pmd,
};
mm: swap: properly update readahead statistics in unuse_pte_range() In unuse_pte_range() we blindly swap-in pages without checking if the swap entry is already present in the swap cache. By doing this, the hit/miss ratio used by the swap readahead heuristic is not properly updated and this leads to non-optimal performance during swapoff. Tracing the distribution of the readahead size returned by the swap readahead heuristic during swapoff shows that a small readahead size is used most of the time as if we had only misses (this happens both with cluster and vma readahead), for example: r::swapin_nr_pages(unsigned long offset):unsigned long:$retval COUNT EVENT 36948 $retval = 8 44151 $retval = 4 49290 $retval = 1 527771 $retval = 2 Checking if the swap entry is present in the swap cache, instead, allows to properly update the readahead statistics and the heuristic behaves in a better way during swapoff, selecting a bigger readahead size: r::swapin_nr_pages(unsigned long offset):unsigned long:$retval COUNT EVENT 1618 $retval = 1 4960 $retval = 2 41315 $retval = 4 103521 $retval = 8 In terms of swapoff performance the result is the following: Testing environment =================== - Host: CPU: 1.8GHz Intel Core i7-8565U (quad-core, 8MB cache) HDD: PC401 NVMe SK hynix 512GB MEM: 16GB - Guest (kvm): 8GB of RAM virtio block driver 16GB swap file on ext4 (/swapfile) Test case ========= - allocate 85% of memory - `systemctl hibernate` to force all the pages to be swapped-out to the swap file - resume the system - measure the time that swapoff takes to complete: # /usr/bin/time swapoff /swapfile Result (swapoff time) ====== 5.6 vanilla 5.6 w/ this patch ----------- ----------------- cluster-readahead 22.09s 12.19s vma-readahead 18.20s 15.33s Conclusion ========== The specific use case this patch is addressing is to improve swapoff performance in cloud environments when a VM has been hibernated, resumed and all the memory needs to be forced back to RAM by disabling swap. This change allows to better exploits the advantages of the readahead heuristic during swapoff and this improvement allows to to speed up the resume process of such VMs. [andrea.righi@canonical.com: update changelog] Link: http://lkml.kernel.org/r/20200418084705.GA147642@xps-13 Signed-off-by: Andrea Righi <andrea.righi@canonical.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Anchal Agarwal <anchalag@amazon.com> Cc: Hugh Dickins <hughd@google.com> Cc: Vineeth Remanan Pillai <vpillai@digitalocean.com> Cc: Kelley Nielsen <kelleynnn@gmail.com> Link: http://lkml.kernel.org/r/20200416180132.GB3352@xps-13 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:48:43 +00:00
page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
&vmf);
if (page)
folio = page_folio(page);
mm: swap: properly update readahead statistics in unuse_pte_range() In unuse_pte_range() we blindly swap-in pages without checking if the swap entry is already present in the swap cache. By doing this, the hit/miss ratio used by the swap readahead heuristic is not properly updated and this leads to non-optimal performance during swapoff. Tracing the distribution of the readahead size returned by the swap readahead heuristic during swapoff shows that a small readahead size is used most of the time as if we had only misses (this happens both with cluster and vma readahead), for example: r::swapin_nr_pages(unsigned long offset):unsigned long:$retval COUNT EVENT 36948 $retval = 8 44151 $retval = 4 49290 $retval = 1 527771 $retval = 2 Checking if the swap entry is present in the swap cache, instead, allows to properly update the readahead statistics and the heuristic behaves in a better way during swapoff, selecting a bigger readahead size: r::swapin_nr_pages(unsigned long offset):unsigned long:$retval COUNT EVENT 1618 $retval = 1 4960 $retval = 2 41315 $retval = 4 103521 $retval = 8 In terms of swapoff performance the result is the following: Testing environment =================== - Host: CPU: 1.8GHz Intel Core i7-8565U (quad-core, 8MB cache) HDD: PC401 NVMe SK hynix 512GB MEM: 16GB - Guest (kvm): 8GB of RAM virtio block driver 16GB swap file on ext4 (/swapfile) Test case ========= - allocate 85% of memory - `systemctl hibernate` to force all the pages to be swapped-out to the swap file - resume the system - measure the time that swapoff takes to complete: # /usr/bin/time swapoff /swapfile Result (swapoff time) ====== 5.6 vanilla 5.6 w/ this patch ----------- ----------------- cluster-readahead 22.09s 12.19s vma-readahead 18.20s 15.33s Conclusion ========== The specific use case this patch is addressing is to improve swapoff performance in cloud environments when a VM has been hibernated, resumed and all the memory needs to be forced back to RAM by disabling swap. This change allows to better exploits the advantages of the readahead heuristic during swapoff and this improvement allows to to speed up the resume process of such VMs. [andrea.righi@canonical.com: update changelog] Link: http://lkml.kernel.org/r/20200418084705.GA147642@xps-13 Signed-off-by: Andrea Righi <andrea.righi@canonical.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Anchal Agarwal <anchalag@amazon.com> Cc: Hugh Dickins <hughd@google.com> Cc: Vineeth Remanan Pillai <vpillai@digitalocean.com> Cc: Kelley Nielsen <kelleynnn@gmail.com> Link: http://lkml.kernel.org/r/20200416180132.GB3352@xps-13 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:48:43 +00:00
}
if (!folio) {
swp_count = READ_ONCE(si->swap_map[offset]);
if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
goto try_next;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
return -ENOMEM;
}
folio_lock(folio);
folio_wait_writeback(folio);
ret = unuse_pte(vma, pmd, addr, entry, folio);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (ret < 0) {
folio_unlock(folio);
folio_put(folio);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
goto out;
}
folio_free_swap(folio);
folio_unlock(folio);
folio_put(folio);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
try_next:
pte = pte_offset_map(pmd, addr);
} while (pte++, addr += PAGE_SIZE, addr != end);
pte_unmap(pte - 1);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
ret = 0;
out:
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
return ret;
}
static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
unsigned long addr, unsigned long end,
unsigned int type)
{
pmd_t *pmd;
unsigned long next;
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
int ret;
pmd = pmd_offset(pud, addr);
do {
cond_resched();
next = pmd_addr_end(addr, end);
mm: thp: fix pmd_bad() triggering in code paths holding mmap_sem read mode In some cases it may happen that pmd_none_or_clear_bad() is called with the mmap_sem hold in read mode. In those cases the huge page faults can allocate hugepmds under pmd_none_or_clear_bad() and that can trigger a false positive from pmd_bad() that will not like to see a pmd materializing as trans huge. It's not khugepaged causing the problem, khugepaged holds the mmap_sem in write mode (and all those sites must hold the mmap_sem in read mode to prevent pagetables to go away from under them, during code review it seems vm86 mode on 32bit kernels requires that too unless it's restricted to 1 thread per process or UP builds). The race is only with the huge pagefaults that can convert a pmd_none() into a pmd_trans_huge(). Effectively all these pmd_none_or_clear_bad() sites running with mmap_sem in read mode are somewhat speculative with the page faults, and the result is always undefined when they run simultaneously. This is probably why it wasn't common to run into this. For example if the madvise(MADV_DONTNEED) runs zap_page_range() shortly before the page fault, the hugepage will not be zapped, if the page fault runs first it will be zapped. Altering pmd_bad() not to error out if it finds hugepmds won't be enough to fix this, because zap_pmd_range would then proceed to call zap_pte_range (which would be incorrect if the pmd become a pmd_trans_huge()). The simplest way to fix this is to read the pmd in the local stack (regardless of what we read, no need of actual CPU barriers, only compiler barrier needed), and be sure it is not changing under the code that computes its value. Even if the real pmd is changing under the value we hold on the stack, we don't care. If we actually end up in zap_pte_range it means the pmd was not none already and it was not huge, and it can't become huge from under us (khugepaged locking explained above). All we need is to enforce that there is no way anymore that in a code path like below, pmd_trans_huge can be false, but pmd_none_or_clear_bad can run into a hugepmd. The overhead of a barrier() is just a compiler tweak and should not be measurable (I only added it for THP builds). I don't exclude different compiler versions may have prevented the race too by caching the value of *pmd on the stack (that hasn't been verified, but it wouldn't be impossible considering pmd_none_or_clear_bad, pmd_bad, pmd_trans_huge, pmd_none are all inlines and there's no external function called in between pmd_trans_huge and pmd_none_or_clear_bad). if (pmd_trans_huge(*pmd)) { if (next-addr != HPAGE_PMD_SIZE) { VM_BUG_ON(!rwsem_is_locked(&tlb->mm->mmap_sem)); split_huge_page_pmd(vma->vm_mm, pmd); } else if (zap_huge_pmd(tlb, vma, pmd, addr)) continue; /* fall through */ } if (pmd_none_or_clear_bad(pmd)) Because this race condition could be exercised without special privileges this was reported in CVE-2012-1179. The race was identified and fully explained by Ulrich who debugged it. I'm quoting his accurate explanation below, for reference. ====== start quote ======= mapcount 0 page_mapcount 1 kernel BUG at mm/huge_memory.c:1384! At some point prior to the panic, a "bad pmd ..." message similar to the following is logged on the console: mm/memory.c:145: bad pmd ffff8800376e1f98(80000000314000e7). The "bad pmd ..." message is logged by pmd_clear_bad() before it clears the page's PMD table entry. 143 void pmd_clear_bad(pmd_t *pmd) 144 { -> 145 pmd_ERROR(*pmd); 146 pmd_clear(pmd); 147 } After the PMD table entry has been cleared, there is an inconsistency between the actual number of PMD table entries that are mapping the page and the page's map count (_mapcount field in struct page). When the page is subsequently reclaimed, __split_huge_page() detects this inconsistency. 1381 if (mapcount != page_mapcount(page)) 1382 printk(KERN_ERR "mapcount %d page_mapcount %d\n", 1383 mapcount, page_mapcount(page)); -> 1384 BUG_ON(mapcount != page_mapcount(page)); The root cause of the problem is a race of two threads in a multithreaded process. Thread B incurs a page fault on a virtual address that has never been accessed (PMD entry is zero) while Thread A is executing an madvise() system call on a virtual address within the same 2 MB (huge page) range. virtual address space .---------------------. | | | | .-|---------------------| | | | | | |<-- B(fault) | | | 2 MB | |/////////////////////|-. huge < |/////////////////////| > A(range) page | |/////////////////////|-' | | | | | | '-|---------------------| | | | | '---------------------' - Thread A is executing an madvise(..., MADV_DONTNEED) system call on the virtual address range "A(range)" shown in the picture. sys_madvise // Acquire the semaphore in shared mode. down_read(&current->mm->mmap_sem) ... madvise_vma switch (behavior) case MADV_DONTNEED: madvise_dontneed zap_page_range unmap_vmas unmap_page_range zap_pud_range zap_pmd_range // // Assume that this huge page has never been accessed. // I.e. content of the PMD entry is zero (not mapped). // if (pmd_trans_huge(*pmd)) { // We don't get here due to the above assumption. } // // Assume that Thread B incurred a page fault and .---------> // sneaks in here as shown below. | // | if (pmd_none_or_clear_bad(pmd)) | { | if (unlikely(pmd_bad(*pmd))) | pmd_clear_bad | { | pmd_ERROR | // Log "bad pmd ..." message here. | pmd_clear | // Clear the page's PMD entry. | // Thread B incremented the map count | // in page_add_new_anon_rmap(), but | // now the page is no longer mapped | // by a PMD entry (-> inconsistency). | } | } | v - Thread B is handling a page fault on virtual address "B(fault)" shown in the picture. ... do_page_fault __do_page_fault // Acquire the semaphore in shared mode. down_read_trylock(&mm->mmap_sem) ... handle_mm_fault if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) // We get here due to the above assumption (PMD entry is zero). do_huge_pmd_anonymous_page alloc_hugepage_vma // Allocate a new transparent huge page here. ... __do_huge_pmd_anonymous_page ... spin_lock(&mm->page_table_lock) ... page_add_new_anon_rmap // Here we increment the page's map count (starts at -1). atomic_set(&page->_mapcount, 0) set_pmd_at // Here we set the page's PMD entry which will be cleared // when Thread A calls pmd_clear_bad(). ... spin_unlock(&mm->page_table_lock) The mmap_sem does not prevent the race because both threads are acquiring it in shared mode (down_read). Thread B holds the page_table_lock while the page's map count and PMD table entry are updated. However, Thread A does not synchronize on that lock. ====== end quote ======= [akpm@linux-foundation.org: checkpatch fixes] Reported-by: Ulrich Obergfell <uobergfe@redhat.com> Signed-off-by: Andrea Arcangeli <aarcange@redhat.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Hugh Dickins <hughd@google.com> Cc: Dave Jones <davej@redhat.com> Acked-by: Larry Woodman <lwoodman@redhat.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: <stable@vger.kernel.org> [2.6.38+] Cc: Mark Salter <msalter@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-21 23:33:42 +00:00
if (pmd_none_or_trans_huge_or_clear_bad(pmd))
continue;
ret = unuse_pte_range(vma, pmd, addr, next, type);
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
if (ret)
return ret;
} while (pmd++, addr = next, addr != end);
return 0;
}
static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
unsigned long addr, unsigned long end,
unsigned int type)
{
pud_t *pud;
unsigned long next;
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
int ret;
pud = pud_offset(p4d, addr);
do {
next = pud_addr_end(addr, end);
if (pud_none_or_clear_bad(pud))
continue;
ret = unuse_pmd_range(vma, pud, addr, next, type);
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
if (ret)
return ret;
} while (pud++, addr = next, addr != end);
return 0;
}
static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
unsigned long addr, unsigned long end,
unsigned int type)
{
p4d_t *p4d;
unsigned long next;
int ret;
p4d = p4d_offset(pgd, addr);
do {
next = p4d_addr_end(addr, end);
if (p4d_none_or_clear_bad(p4d))
continue;
ret = unuse_pud_range(vma, p4d, addr, next, type);
if (ret)
return ret;
} while (p4d++, addr = next, addr != end);
return 0;
}
static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
{
pgd_t *pgd;
unsigned long addr, end, next;
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
int ret;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
addr = vma->vm_start;
end = vma->vm_end;
pgd = pgd_offset(vma->vm_mm, addr);
do {
next = pgd_addr_end(addr, end);
if (pgd_none_or_clear_bad(pgd))
continue;
ret = unuse_p4d_range(vma, pgd, addr, next, type);
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
if (ret)
return ret;
} while (pgd++, addr = next, addr != end);
return 0;
}
static int unuse_mm(struct mm_struct *mm, unsigned int type)
{
struct vm_area_struct *vma;
Memory controller: memory accounting Add the accounting hooks. The accounting is carried out for RSS and Page Cache (unmapped) pages. There is now a common limit and accounting for both. The RSS accounting is accounted at page_add_*_rmap() and page_remove_rmap() time. Page cache is accounted at add_to_page_cache(), __delete_from_page_cache(). Swap cache is also accounted for. Each page's page_cgroup is protected with the last bit of the page_cgroup pointer, this makes handling of race conditions involving simultaneous mappings of a page easier. A reference count is kept in the page_cgroup to deal with cases where a page might be unmapped from the RSS of all tasks, but still lives in the page cache. Credits go to Vaidyanathan Srinivasan for helping with reference counting work of the page cgroup. Almost all of the page cache accounting code has help from Vaidyanathan Srinivasan. [hugh@veritas.com: fix swapoff breakage] [akpm@linux-foundation.org: fix locking] Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Paul Menage <menage@google.com> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: Kirill Korotaev <dev@sw.ru> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: David Rientjes <rientjes@google.com> Cc: <Valdis.Kletnieks@vt.edu> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-07 08:13:53 +00:00
int ret = 0;
VMA_ITERATOR(vmi, mm, 0);
mmap locking API: use coccinelle to convert mmap_sem rwsem call sites This change converts the existing mmap_sem rwsem calls to use the new mmap locking API instead. The change is generated using coccinelle with the following rule: // spatch --sp-file mmap_lock_api.cocci --in-place --include-headers --dir . @@ expression mm; @@ ( -init_rwsem +mmap_init_lock | -down_write +mmap_write_lock | -down_write_killable +mmap_write_lock_killable | -down_write_trylock +mmap_write_trylock | -up_write +mmap_write_unlock | -downgrade_write +mmap_write_downgrade | -down_read +mmap_read_lock | -down_read_killable +mmap_read_lock_killable | -down_read_trylock +mmap_read_trylock | -up_read +mmap_read_unlock ) -(&mm->mmap_sem) +(mm) Signed-off-by: Michel Lespinasse <walken@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com> Reviewed-by: Vlastimil Babka <vbabka@suse.cz> Cc: Davidlohr Bueso <dbueso@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Jerome Glisse <jglisse@redhat.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Liam Howlett <Liam.Howlett@oracle.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ying Han <yinghan@google.com> Link: http://lkml.kernel.org/r/20200520052908.204642-5-walken@google.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-09 04:33:25 +00:00
mmap_read_lock(mm);
for_each_vma(vmi, vma) {
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (vma->anon_vma) {
ret = unuse_vma(vma, type);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (ret)
break;
}
cond_resched();
}
mmap locking API: use coccinelle to convert mmap_sem rwsem call sites This change converts the existing mmap_sem rwsem calls to use the new mmap locking API instead. The change is generated using coccinelle with the following rule: // spatch --sp-file mmap_lock_api.cocci --in-place --include-headers --dir . @@ expression mm; @@ ( -init_rwsem +mmap_init_lock | -down_write +mmap_write_lock | -down_write_killable +mmap_write_lock_killable | -down_write_trylock +mmap_write_trylock | -up_write +mmap_write_unlock | -downgrade_write +mmap_write_downgrade | -down_read +mmap_read_lock | -down_read_killable +mmap_read_lock_killable | -down_read_trylock +mmap_read_trylock | -up_read +mmap_read_unlock ) -(&mm->mmap_sem) +(mm) Signed-off-by: Michel Lespinasse <walken@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com> Reviewed-by: Vlastimil Babka <vbabka@suse.cz> Cc: Davidlohr Bueso <dbueso@suse.de> Cc: David Rientjes <rientjes@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Jerome Glisse <jglisse@redhat.com> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Liam Howlett <Liam.Howlett@oracle.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Ying Han <yinghan@google.com> Link: http://lkml.kernel.org/r/20200520052908.204642-5-walken@google.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-09 04:33:25 +00:00
mmap_read_unlock(mm);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
return ret;
}
/*
* Scan swap_map from current position to next entry still in use.
* Return 0 if there are no inuse entries after prev till end of
* the map.
*/
static unsigned int find_next_to_unuse(struct swap_info_struct *si,
unsigned int prev)
{
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
unsigned int i;
unsigned char count;
/*
* No need for swap_lock here: we're just looking
* for whether an entry is in use, not modifying it; false
* hits are okay, and sys_swapoff() has already prevented new
* allocations from this area (while holding swap_lock).
*/
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
for (i = prev + 1; i < si->max; i++) {
count = READ_ONCE(si->swap_map[i]);
if (count && swap_count(count) != SWAP_MAP_BAD)
break;
if ((i % LATENCY_LIMIT) == 0)
cond_resched();
}
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (i == si->max)
i = 0;
return i;
}
static int try_to_unuse(unsigned int type)
{
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
struct mm_struct *prev_mm;
struct mm_struct *mm;
struct list_head *p;
int retval = 0;
struct swap_info_struct *si = swap_info[type];
struct folio *folio;
swp_entry_t entry;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
unsigned int i;
if (!READ_ONCE(si->inuse_pages))
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
return 0;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
retry:
retval = shmem_unuse(type);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (retval)
return retval;
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
prev_mm = &init_mm;
mmget(prev_mm);
spin_lock(&mmlist_lock);
p = &init_mm.mmlist;
while (READ_ONCE(si->inuse_pages) &&
!signal_pending(current) &&
(p = p->next) != &init_mm.mmlist) {
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
mm = list_entry(p, struct mm_struct, mmlist);
if (!mmget_not_zero(mm))
continue;
spin_unlock(&mmlist_lock);
mmput(prev_mm);
prev_mm = mm;
retval = unuse_mm(mm, type);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
if (retval) {
mmput(prev_mm);
return retval;
}
/*
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
* Make sure that we aren't completely killing
* interactive performance.
*/
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
cond_resched();
spin_lock(&mmlist_lock);
}
spin_unlock(&mmlist_lock);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
mmput(prev_mm);
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
i = 0;
while (READ_ONCE(si->inuse_pages) &&
!signal_pending(current) &&
(i = find_next_to_unuse(si, i)) != 0) {
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
entry = swp_entry(type, i);
folio = filemap_get_folio(swap_address_space(entry), i);
if (!folio)
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
continue;
mm: try_to_unuse check removing right swap There's a possible race in try_to_unuse() which Nick Piggin led me to two years ago. Where it does lock_page() after read_swap_cache_async(), what if another task removed that page from swapcache just before we locked it? It would sail though the (*swap_map > 1) tests doing nothing (because it could not have been removed from swapcache before its swap references were gone), until it reaches the delete_from_swap_cache(page) near the bottom. Now imagine that this page has been allocated to swap on a different swap area while we dropped page lock (perhaps at the top, perhaps in unuse_mm): we could wrongly remove from swap cache before the page has been written to swap, so a subsequent do_swap_page() would read in stale data from swap. I think this case could not happen before: remove_exclusive_swap_page() refused while page count was raised. But now with reuse_swap_page() and try_to_free_swap() removing from swap cache without minding page count, I think it could happen - the previous patch argued that it was safe because try_to_unuse() already ignored page count, but overlooked that it might be breaking the assumptions in try_to_unuse() itself. Signed-off-by: Hugh Dickins <hugh@veritas.com> Cc: Lee Schermerhorn <lee.schermerhorn@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Robin Holt <holt@sgi.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-06 22:39:37 +00:00
/*
* It is conceivable that a racing task removed this folio from
* swap cache just before we acquired the page lock. The folio
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
* might even be back in swap cache on another swap area. But
* that is okay, folio_free_swap() only removes stale folios.
*/
folio_lock(folio);
folio_wait_writeback(folio);
folio_free_swap(folio);
folio_unlock(folio);
folio_put(folio);
}
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
/*
* Lets check again to see if there are still swap entries in the map.
* If yes, we would need to do retry the unuse logic again.
* Under global memory pressure, swap entries can be reinserted back
* into process space after the mmlist loop above passes over them.
*
* Limit the number of retries? No: when mmget_not_zero()
* above fails, that mm is likely to be freeing swap from
* exit_mmap(), which proceeds at its own independent pace;
* and even shmem_writepage() could have been preempted after
* folio_alloc_swap(), temporarily hiding that swap. It's easy
* and robust (though cpu-intensive) just to keep retrying.
mm: rid swapoff of quadratic complexity This patch was initially posted by Kelley Nielsen. Reposting the patch with all review comments addressed and with minor modifications and optimizations. Also, folding in the fixes offered by Hugh Dickins and Huang Ying. Tests were rerun and commit message updated with new results. try_to_unuse() is of quadratic complexity, with a lot of wasted effort. It unuses swap entries one by one, potentially iterating over all the page tables for all the processes in the system for each one. This new proposed implementation of try_to_unuse simplifies its complexity to linear. It iterates over the system's mms once, unusing all the affected entries as it walks each set of page tables. It also makes similar changes to shmem_unuse. Improvement swapoff was called on a swap partition containing about 6G of data, in a VM(8cpu, 16G RAM), and calls to unuse_pte_range() were counted. Present implementation....about 1200M calls(8min, avg 80% cpu util). Prototype.................about 9.0K calls(3min, avg 5% cpu util). Details In shmem_unuse(), iterate over the shmem_swaplist and, for each shmem_inode_info that contains a swap entry, pass it to shmem_unuse_inode(), along with the swap type. In shmem_unuse_inode(), iterate over its associated xarray, and store the index and value of each swap entry in an array for passing to shmem_swapin_page() outside of the RCU critical section. In try_to_unuse(), instead of iterating over the entries in the type and unusing them one by one, perhaps walking all the page tables for all the processes for each one, iterate over the mmlist, making one pass. Pass each mm to unuse_mm() to begin its page table walk, and during the walk, unuse all the ptes that have backing store in the swap type received by try_to_unuse(). After the walk, check the type for orphaned swap entries with find_next_to_unuse(), and remove them from the swap cache. If find_next_to_unuse() starts over at the beginning of the type, repeat the check of the shmem_swaplist and the walk a maximum of three times. Change unuse_mm() and the intervening walk functions down to unuse_pte_range() to take the type as a parameter, and to iterate over their entire range, calling the next function down on every iteration. In unuse_pte_range(), make a swap entry from each pte in the range using the passed in type. If it has backing store in the type, call swapin_readahead() to retrieve the page and pass it to unuse_pte(). Pass the count of pages_to_unuse down the page table walks in try_to_unuse(), and return from the walk when the desired number of pages has been swapped back in. Link: http://lkml.kernel.org/r/20190114153129.4852-2-vpillai@digitalocean.com Signed-off-by: Vineeth Remanan Pillai <vpillai@digitalocean.com> Signed-off-by: Kelley Nielsen <kelleynnn@gmail.com> Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@surriel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:47:03 +00:00
*/
if (READ_ONCE(si->inuse_pages)) {
if (!signal_pending(current))
goto retry;
return -EINTR;
}
return 0;
}
/*
* After a successful try_to_unuse, if no swap is now in use, we know
* we can empty the mmlist. swap_lock must be held on entry and exit.
* Note that mmlist_lock nests inside swap_lock, and an mm must be
* added to the mmlist just after page_duplicate - before would be racy.
*/
static void drain_mmlist(void)
{
struct list_head *p, *next;
unsigned int type;
for (type = 0; type < nr_swapfiles; type++)
if (swap_info[type]->inuse_pages)
return;
spin_lock(&mmlist_lock);
list_for_each_safe(p, next, &init_mm.mmlist)
list_del_init(p);
spin_unlock(&mmlist_lock);
}
/*
* Free all of a swapdev's extent information
*/
static void destroy_swap_extents(struct swap_info_struct *sis)
{
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
struct rb_node *rb = sis->swap_extent_root.rb_node;
struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
rb_erase(rb, &sis->swap_extent_root);
kfree(se);
}
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
if (sis->flags & SWP_ACTIVATED) {
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
struct file *swap_file = sis->swap_file;
struct address_space *mapping = swap_file->f_mapping;
sis->flags &= ~SWP_ACTIVATED;
if (mapping->a_ops->swap_deactivate)
mapping->a_ops->swap_deactivate(swap_file);
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
}
}
/*
* Add a block range (and the corresponding page range) into this swapdev's
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
* extent tree.
*
* This function rather assumes that it is called in ascending page order.
*/
int
add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
unsigned long nr_pages, sector_t start_block)
{
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
struct swap_extent *se;
struct swap_extent *new_se;
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
/*
* place the new node at the right most since the
* function is called in ascending page order.
*/
while (*link) {
parent = *link;
link = &parent->rb_right;
}
if (parent) {
se = rb_entry(parent, struct swap_extent, rb_node);
BUG_ON(se->start_page + se->nr_pages != start_page);
if (se->start_block + se->nr_pages == start_block) {
/* Merge it */
se->nr_pages += nr_pages;
return 0;
}
}
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
/* No merge, insert a new extent. */
new_se = kmalloc(sizeof(*se), GFP_KERNEL);
if (new_se == NULL)
return -ENOMEM;
new_se->start_page = start_page;
new_se->nr_pages = nr_pages;
new_se->start_block = start_block;
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
rb_link_node(&new_se->rb_node, parent, link);
rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
return 1;
}
EXPORT_SYMBOL_GPL(add_swap_extent);
/*
* A `swap extent' is a simple thing which maps a contiguous range of pages
* onto a contiguous range of disk blocks. A rbtree of swap extents is
* built at swapon time and is then used at swap_writepage/swap_readpage
* time for locating where on disk a page belongs.
*
* If the swapfile is an S_ISBLK block device, a single extent is installed.
* This is done so that the main operating code can treat S_ISBLK and S_ISREG
* swap files identically.
*
* Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
* extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
* swapfiles are handled *identically* after swapon time.
*
* For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
* and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
* blocks are found which do not fall within the PAGE_SIZE alignment
* requirements, they are simply tossed out - we will never use those blocks
* for swapping.
*
* For all swap devices we set S_SWAPFILE across the life of the swapon. This
* prevents users from writing to the swap device, which will corrupt memory.
*
* The amount of disk space which a single swap extent represents varies.
* Typically it is in the 1-4 megabyte range. So we can have hundreds of
* extents in the rbtree. - akpm.
*/
static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
{
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
struct file *swap_file = sis->swap_file;
struct address_space *mapping = swap_file->f_mapping;
struct inode *inode = mapping->host;
int ret;
if (S_ISBLK(inode->i_mode)) {
ret = add_swap_extent(sis, 0, sis->max, 0);
*span = sis->pages;
return ret;
}
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
if (mapping->a_ops->swap_activate) {
ret = mapping->a_ops->swap_activate(sis, swap_file, span);
if (ret < 0)
return ret;
sis->flags |= SWP_ACTIVATED;
mm: introduce ->swap_rw and use it for reads from SWP_FS_OPS swap-space swap currently uses ->readpage to read swap pages. This can only request one page at a time from the filesystem, which is not most efficient. swap uses ->direct_IO for writes which while this is adequate is an inappropriate over-loading. ->direct_IO may need to had handle allocate space for holes or other details that are not relevant for swap. So this patch introduces a new address_space operation: ->swap_rw. In this patch it is used for reads, and a subsequent patch will switch writes to use it. No filesystem yet supports ->swap_rw, but that is not a problem because no filesystem actually works with filesystem-based swap. Only two filesystems set SWP_FS_OPS: - cifs sets the flag, but ->direct_IO always fails so swap cannot work. - nfs sets the flag, but ->direct_IO calls generic_write_checks() which has failed on swap files for several releases. To ensure that a NULL ->swap_rw isn't called, ->activate_swap() for both NFS and cifs are changed to fail if ->swap_rw is not set. This can be removed if/when the function is added. Future patches will restore swap-over-NFS functionality. To submit an async read with ->swap_rw() we need to allocate a structure to hold the kiocb and other details. swap_readpage() cannot handle transient failure, so we create a mempool to provide the structures. Link: https://lkml.kernel.org/r/164859778125.29473.13430559328221330589.stgit@noble.brown Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: David Howells <dhowells@redhat.com> Tested-by: Geert Uytterhoeven <geert+renesas@glider.be> Cc: Hugh Dickins <hughd@google.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Trond Myklebust <trond.myklebust@hammerspace.com> Cc: Miaohe Lin <linmiaohe@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-05-10 01:20:48 +00:00
if ((sis->flags & SWP_FS_OPS) &&
sio_pool_init() != 0) {
destroy_swap_extents(sis);
return -ENOMEM;
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
}
return ret;
mm: add support for a filesystem to activate swap files and use direct_IO for writing swap pages Currently swapfiles are managed entirely by the core VM by using ->bmap to allocate space and write to the blocks directly. This effectively ensures that the underlying blocks are allocated and avoids the need for the swap subsystem to locate what physical blocks store offsets within a file. If the swap subsystem is to use the filesystem information to locate the blocks, it is critical that information such as block groups, block bitmaps and the block descriptor table that map the swap file were resident in memory. This patch adds address_space_operations that the VM can call when activating or deactivating swap backed by a file. int swap_activate(struct file *); int swap_deactivate(struct file *); The ->swap_activate() method is used to communicate to the file that the VM relies on it, and the address_space should take adequate measures such as reserving space in the underlying device, reserving memory for mempools and pinning information such as the block descriptor table in memory. The ->swap_deactivate() method is called on sys_swapoff() if ->swap_activate() returned success. After a successful swapfile ->swap_activate, the swapfile is marked SWP_FILE and swapper_space.a_ops will proxy to sis->swap_file->f_mappings->a_ops using ->direct_io to write swapcache pages and ->readpage to read. It is perfectly possible that direct_IO be used to read the swap pages but it is an unnecessary complication. Similarly, it is possible that ->writepage be used instead of direct_io to write the pages but filesystem developers have stated that calling writepage from the VM is undesirable for a variety of reasons and using direct_IO opens up the possibility of writing back batches of swap pages in the future. [a.p.zijlstra@chello.nl: Original patch] Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: David S. Miller <davem@davemloft.net> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Paris <eparis@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Neil Brown <neilb@suse.de> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Trond Myklebust <Trond.Myklebust@netapp.com> Cc: Xiaotian Feng <dfeng@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 23:44:55 +00:00
}
return generic_swapfile_activate(sis, swap_file, span);
}
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
static int swap_node(struct swap_info_struct *p)
{
struct block_device *bdev;
if (p->bdev)
bdev = p->bdev;
else
bdev = p->swap_file->f_inode->i_sb->s_bdev;
return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
}
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
static void setup_swap_info(struct swap_info_struct *p, int prio,
unsigned char *swap_map,
struct swap_cluster_info *cluster_info)
{
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
int i;
if (prio >= 0)
p->prio = prio;
else
p->prio = --least_priority;
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
/*
* the plist prio is negated because plist ordering is
* low-to-high, while swap ordering is high-to-low
*/
p->list.prio = -p->prio;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
for_each_node(i) {
if (p->prio >= 0)
p->avail_lists[i].prio = -p->prio;
else {
if (swap_node(p) == i)
p->avail_lists[i].prio = 1;
else
p->avail_lists[i].prio = -p->prio;
}
}
p->swap_map = swap_map;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
p->cluster_info = cluster_info;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
}
static void _enable_swap_info(struct swap_info_struct *p)
{
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
p->flags |= SWP_WRITEOK;
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
atomic_long_add(p->pages, &nr_swap_pages);
total_swap_pages += p->pages;
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
assert_spin_locked(&swap_lock);
/*
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
* both lists are plists, and thus priority ordered.
* swap_active_head needs to be priority ordered for swapoff(),
* which on removal of any swap_info_struct with an auto-assigned
* (i.e. negative) priority increments the auto-assigned priority
* of any lower-priority swap_info_structs.
* swap_avail_head needs to be priority ordered for folio_alloc_swap(),
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
* which allocates swap pages from the highest available priority
* swap_info_struct.
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
*/
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
plist_add(&p->list, &swap_active_head);
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
add_to_avail_list(p);
}
static void enable_swap_info(struct swap_info_struct *p, int prio,
unsigned char *swap_map,
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
struct swap_cluster_info *cluster_info,
unsigned long *frontswap_map)
{
if (IS_ENABLED(CONFIG_FRONTSWAP))
frontswap_init(p->type, frontswap_map);
spin_lock(&swap_lock);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&p->lock);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
setup_swap_info(p, prio, swap_map, cluster_info);
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
/*
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
* Finished initializing swap device, now it's safe to reference it.
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
*/
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
percpu_ref_resurrect(&p->users);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
spin_lock(&swap_lock);
spin_lock(&p->lock);
_enable_swap_info(p);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
}
static void reinsert_swap_info(struct swap_info_struct *p)
{
spin_lock(&swap_lock);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&p->lock);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
_enable_swap_info(p);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
}
mm/swap: add cache for swap slots allocation We add per cpu caches for swap slots that can be allocated and freed quickly without the need to touch the swap info lock. Two separate caches are maintained for swap slots allocated and swap slots returned. This is to allow the swap slots to be returned to the global pool in a batch so they will have a chance to be coaelesced with other slots in a cluster. We do not reuse the slots that are returned right away, as it may increase fragmentation of the slots. The swap allocation cache is protected by a mutex as we may sleep when searching for empty slots in cache. The swap free cache is protected by a spin lock as we cannot sleep in the free path. We refill the swap slots cache when we run out of slots, and we disable the swap slots cache and drain the slots if the global number of slots fall below a low watermark threshold. We re-enable the cache agian when the slots available are above a high watermark. [ying.huang@intel.com: use raw_cpu_ptr over this_cpu_ptr for swap slots access] [tim.c.chen@linux.intel.com: add comments on locks in swap_slots.h] Link: http://lkml.kernel.org/r/20170118180327.GA24225@linux.intel.com Link: http://lkml.kernel.org/r/35de301a4eaa8daa2977de6e987f2c154385eb66.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:39 +00:00
bool has_usable_swap(void)
{
bool ret = true;
spin_lock(&swap_lock);
if (plist_head_empty(&swap_active_head))
ret = false;
spin_unlock(&swap_lock);
return ret;
}
SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
{
struct swap_info_struct *p = NULL;
unsigned char *swap_map;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
struct swap_cluster_info *cluster_info;
unsigned long *frontswap_map;
struct file *swap_file, *victim;
struct address_space *mapping;
struct inode *inode;
struct filename *pathname;
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
int err, found = 0;
swap: fix set_blocksize race during swapon/swapoff Fix race between swapoff and swapon. Swapoff used old_block_size from swap_info outside of swapon_mutex so it could be overwritten by concurrent swapon. The race has visible effect only if more than one swap block device exists with different block sizes (e.g. /dev/sda1 with block size 4096 and /dev/sdb1 with 512). In such case it leads to setting the blocksize of swapped off device with wrong blocksize. The bug can be triggered with multiple concurrent swapoff and swapon: 0. Swap for some device is on. 1. swapoff: First the swapoff is called on this device and "struct swap_info_struct *p" is assigned. This is done under swap_lock however this lock is released for the call try_to_unuse(). 2. swapon: After the assignment above (and before acquiring swapon_mutex & swap_lock by swapoff) the swapon is called on the same device. The p->old_block_size is assigned to the value of block_size the device. This block size should be the same as previous but sometimes it is not. The swapon ends successfully. 3. swapoff: Swapoff resumes, grabs the locks and mutex and continues to disable this swap device. Now it sets the block size to value taken from swap_info which was overwritten by swapon in 2. Signed-off-by: Krzysztof Kozlowski <k.kozlowski@samsung.com> Reported-by: Weijie Yang <weijie.yang.kh@gmail.com> Cc: Bob Liu <bob.liu@oracle.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Shaohua Li <shli@fusionio.com> Cc: Minchan Kim <minchan@kernel.org> Acked-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-10-16 20:47:06 +00:00
unsigned int old_block_size;
if (!capable(CAP_SYS_ADMIN))
return -EPERM;
BUG_ON(!current->mm);
pathname = getname(specialfile);
if (IS_ERR(pathname))
return PTR_ERR(pathname);
victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
err = PTR_ERR(victim);
if (IS_ERR(victim))
goto out;
mapping = victim->f_mapping;
spin_lock(&swap_lock);
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
plist_for_each_entry(p, &swap_active_head, list) {
if (p->flags & SWP_WRITEOK) {
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
if (p->swap_file->f_mapping == mapping) {
found = 1;
break;
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
}
}
}
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
if (!found) {
err = -EINVAL;
spin_unlock(&swap_lock);
goto out_dput;
}
if (!security_vm_enough_memory_mm(current->mm, p->pages))
vm_unacct_memory(p->pages);
else {
err = -ENOMEM;
spin_unlock(&swap_lock);
goto out_dput;
}
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
del_from_avail_list(p);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&p->lock);
if (p->prio < 0) {
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
struct swap_info_struct *si = p;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
int nid;
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
plist_for_each_entry_continue(si, &swap_active_head, list) {
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
si->prio++;
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
si->list.prio--;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
for_each_node(nid) {
if (si->avail_lists[nid].prio != 1)
si->avail_lists[nid].prio--;
}
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
}
least_priority++;
}
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
plist_del(&p->list, &swap_active_head);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
atomic_long_sub(p->pages, &nr_swap_pages);
total_swap_pages -= p->pages;
p->flags &= ~SWP_WRITEOK;
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
disable_swap_slots_cache_lock();
mm, oom: fix race when specifying a thread as the oom origin test_set_oom_score_adj() and compare_swap_oom_score_adj() are used to specify that current should be killed first if an oom condition occurs in between the two calls. The usage is short oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX); ... compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj); to store the thread's oom_score_adj, temporarily change it to the maximum score possible, and then restore the old value if it is still the same. This happens to still be racy, however, if the user writes OOM_SCORE_ADJ_MAX to /proc/pid/oom_score_adj in between the two calls. The compare_swap_oom_score_adj() will then incorrectly reset the old value prior to the write of OOM_SCORE_ADJ_MAX. To fix this, introduce a new oom_flags_t member in struct signal_struct that will be used for per-thread oom killer flags. KSM and swapoff can now use a bit in this member to specify that threads should be killed first in oom conditions without playing around with oom_score_adj. This also allows the correct oom_score_adj to always be shown when reading /proc/pid/oom_score. Signed-off-by: David Rientjes <rientjes@google.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Michal Hocko <mhocko@suse.cz> Cc: Anton Vorontsov <anton.vorontsov@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-12 00:02:56 +00:00
set_current_oom_origin();
err = try_to_unuse(p->type);
mm, oom: fix race when specifying a thread as the oom origin test_set_oom_score_adj() and compare_swap_oom_score_adj() are used to specify that current should be killed first if an oom condition occurs in between the two calls. The usage is short oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX); ... compare_swap_oom_score_adj(OOM_SCORE_ADJ_MAX, oom_score_adj); to store the thread's oom_score_adj, temporarily change it to the maximum score possible, and then restore the old value if it is still the same. This happens to still be racy, however, if the user writes OOM_SCORE_ADJ_MAX to /proc/pid/oom_score_adj in between the two calls. The compare_swap_oom_score_adj() will then incorrectly reset the old value prior to the write of OOM_SCORE_ADJ_MAX. To fix this, introduce a new oom_flags_t member in struct signal_struct that will be used for per-thread oom killer flags. KSM and swapoff can now use a bit in this member to specify that threads should be killed first in oom conditions without playing around with oom_score_adj. This also allows the correct oom_score_adj to always be shown when reading /proc/pid/oom_score. Signed-off-by: David Rientjes <rientjes@google.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Reviewed-by: Michal Hocko <mhocko@suse.cz> Cc: Anton Vorontsov <anton.vorontsov@linaro.org> Cc: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-12-12 00:02:56 +00:00
clear_current_oom_origin();
if (err) {
/* re-insert swap space back into swap_list */
reinsert_swap_info(p);
reenable_swap_slots_cache_unlock();
goto out_dput;
}
reenable_swap_slots_cache_unlock();
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
/*
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
* Wait for swap operations protected by get/put_swap_device()
* to complete.
*
* We need synchronize_rcu() here to protect the accessing to
* the swap cache data structure.
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
*/
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
percpu_ref_kill(&p->users);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
synchronize_rcu();
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
wait_for_completion(&p->comp);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
flush_work(&p->discard_work);
destroy_swap_extents(p);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
if (p->flags & SWP_CONTINUED)
free_swap_count_continuations(p);
if (!p->bdev || !bdev_nonrot(p->bdev))
atomic_dec(&nr_rotate_swap);
mutex_lock(&swapon_mutex);
spin_lock(&swap_lock);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&p->lock);
drain_mmlist();
/* wait for anyone still in scan_swap_map_slots */
p->highest_bit = 0; /* cuts scans short */
while (p->flags >= SWP_SCANNING) {
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
schedule_timeout_uninterruptible(1);
spin_lock(&swap_lock);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock(&p->lock);
}
swap_file = p->swap_file;
swap: fix set_blocksize race during swapon/swapoff Fix race between swapoff and swapon. Swapoff used old_block_size from swap_info outside of swapon_mutex so it could be overwritten by concurrent swapon. The race has visible effect only if more than one swap block device exists with different block sizes (e.g. /dev/sda1 with block size 4096 and /dev/sdb1 with 512). In such case it leads to setting the blocksize of swapped off device with wrong blocksize. The bug can be triggered with multiple concurrent swapoff and swapon: 0. Swap for some device is on. 1. swapoff: First the swapoff is called on this device and "struct swap_info_struct *p" is assigned. This is done under swap_lock however this lock is released for the call try_to_unuse(). 2. swapon: After the assignment above (and before acquiring swapon_mutex & swap_lock by swapoff) the swapon is called on the same device. The p->old_block_size is assigned to the value of block_size the device. This block size should be the same as previous but sometimes it is not. The swapon ends successfully. 3. swapoff: Swapoff resumes, grabs the locks and mutex and continues to disable this swap device. Now it sets the block size to value taken from swap_info which was overwritten by swapon in 2. Signed-off-by: Krzysztof Kozlowski <k.kozlowski@samsung.com> Reported-by: Weijie Yang <weijie.yang.kh@gmail.com> Cc: Bob Liu <bob.liu@oracle.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Shaohua Li <shli@fusionio.com> Cc: Minchan Kim <minchan@kernel.org> Acked-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-10-16 20:47:06 +00:00
old_block_size = p->old_block_size;
p->swap_file = NULL;
p->max = 0;
swap_map = p->swap_map;
p->swap_map = NULL;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
cluster_info = p->cluster_info;
p->cluster_info = NULL;
frontswap_map = frontswap_map_get(p);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&p->lock);
spin_unlock(&swap_lock);
arch_swap_invalidate_area(p->type);
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
frontswap_invalidate_area(p->type);
frontswap_map_set(p, NULL);
mutex_unlock(&swapon_mutex);
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
free_percpu(p->percpu_cluster);
p->percpu_cluster = NULL;
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
free_percpu(p->cluster_next_cpu);
p->cluster_next_cpu = NULL;
vfree(swap_map);
mm, swap: use kvzalloc to allocate some swap data structures Now vzalloc() is used in swap code to allocate various data structures, such as swap cache, swap slots cache, cluster info, etc. Because the size may be too large on some system, so that normal kzalloc() may fail. But using kzalloc() has some advantages, for example, less memory fragmentation, less TLB pressure, etc. So change the data structure allocation in swap code to use kvzalloc() which will try kzalloc() firstly, and fallback to vzalloc() if kzalloc() failed. In general, although kmalloc() will reduce the number of high-order pages in short term, vmalloc() will cause more pain for memory fragmentation in the long term. And the swap data structure allocation that is changed in this patch is expected to be long term allocation. From Dave Hansen: "for example, we have a two-page data structure. vmalloc() takes two effectively random order-0 pages, probably from two different 2M pages and pins them. That "kills" two 2M pages. kmalloc(), allocating two *contiguous* pages, will not cross a 2M boundary. That means it will only "kill" the possibility of a single 2M page. More 2M pages == less fragmentation. The allocation in this patch occurs during swap on time, which is usually done during system boot, so usually we have high opportunity to allocate the contiguous pages successfully. The allocation for swap_map[] in struct swap_info_struct is not changed, because that is usually quite large and vmalloc_to_page() is used for it. That makes it a little harder to change. Link: http://lkml.kernel.org/r/20170407064911.25447-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 22:57:40 +00:00
kvfree(cluster_info);
kvfree(frontswap_map);
/* Destroy swap account information */
swap: change swap_info singly-linked list to list_head The logic controlling the singly-linked list of swap_info_struct entries for all active, i.e. swapon'ed, swap targets is rather complex, because: - it stores the entries in priority order - there is a pointer to the highest priority entry - there is a pointer to the highest priority not-full entry - there is a highest_priority_index variable set outside the swap_lock - swap entries of equal priority should be used equally this complexity leads to bugs such as: https://lkml.org/lkml/2014/2/13/181 where different priority swap targets are incorrectly used equally. That bug probably could be solved with the existing singly-linked lists, but I think it would only add more complexity to the already difficult to understand get_swap_page() swap_list iteration logic. The first patch changes from a singly-linked list to a doubly-linked list using list_heads; the highest_priority_index and related code are removed and get_swap_page() starts each iteration at the highest priority swap_info entry, even if it's full. While this does introduce unnecessary list iteration (i.e. Schlemiel the painter's algorithm) in the case where one or more of the highest priority entries are full, the iteration and manipulation code is much simpler and behaves correctly re: the above bug; and the fourth patch removes the unnecessary iteration. The second patch adds some minor plist helper functions; nothing new really, just functions to match existing regular list functions. These are used by the next two patches. The third patch adds plist_requeue(), which is used by get_swap_page() in the next patch - it performs the requeueing of same-priority entries (which moves the entry to the end of its priority in the plist), so that all equal-priority swap_info_structs get used equally. The fourth patch converts the main list into a plist, and adds a new plist that contains only swap_info entries that are both active and not full. As Mel suggested using plists allows removing all the ordering code from swap - plists handle ordering automatically. The list naming is also clarified now that there are two lists, with the original list changed from swap_list_head to swap_active_head and the new list named swap_avail_head. A new spinlock is also added for the new list, so swap_info entries can be added or removed from the new list immediately as they become full or not full. This patch (of 4): Replace the singly-linked list tracking active, i.e. swapon'ed, swap_info_struct entries with a doubly-linked list using struct list_heads. Simplify the logic iterating and manipulating the list of entries, especially get_swap_page(), by using standard list_head functions, and removing the highest priority iteration logic. The change fixes the bug: https://lkml.org/lkml/2014/2/13/181 in which different priority swap entries after the highest priority entry are incorrectly used equally in pairs. The swap behavior is now as advertised, i.e. different priority swap entries are used in order, and equal priority swap targets are used concurrently. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:53 +00:00
swap_cgroup_swapoff(p->type);
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
exit_swap_address_space(p->type);
memcg: swap cgroup for remembering usage For accounting swap, we need a record per swap entry, at least. This patch adds following function. - swap_cgroup_swapon() .... called from swapon - swap_cgroup_swapoff() ... called at the end of swapoff. - swap_cgroup_record() .... record information of swap entry. - swap_cgroup_lookup() .... lookup information of swap entry. This patch just implements "how to record information". No actual method for limit the usage of swap. These routine uses flat table to record and lookup. "wise" lookup system like radix-tree requires requires memory allocation at new records but swap-out is usually called under memory shortage (or memcg hits limit.) So, I used static allocation. (maybe dynamic allocation is not very hard but it adds additional memory allocation in memory shortage path.) Note1: In this, we use pointer to record information and this means 8bytes per swap entry. I think we can reduce this when we create "id of cgroup" in the range of 0-65535 or 0-255. Reported-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Reviewed-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Tested-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Reported-by: Hugh Dickins <hugh@veritas.com> Reported-by: Balbir Singh <balbir@linux.vnet.ibm.com> Reported-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Pavel Emelianov <xemul@openvz.org> Cc: Li Zefan <lizf@cn.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-01-08 02:07:58 +00:00
inode = mapping->host;
if (S_ISBLK(inode->i_mode)) {
struct block_device *bdev = I_BDEV(inode);
swap: fix set_blocksize race during swapon/swapoff Fix race between swapoff and swapon. Swapoff used old_block_size from swap_info outside of swapon_mutex so it could be overwritten by concurrent swapon. The race has visible effect only if more than one swap block device exists with different block sizes (e.g. /dev/sda1 with block size 4096 and /dev/sdb1 with 512). In such case it leads to setting the blocksize of swapped off device with wrong blocksize. The bug can be triggered with multiple concurrent swapoff and swapon: 0. Swap for some device is on. 1. swapoff: First the swapoff is called on this device and "struct swap_info_struct *p" is assigned. This is done under swap_lock however this lock is released for the call try_to_unuse(). 2. swapon: After the assignment above (and before acquiring swapon_mutex & swap_lock by swapoff) the swapon is called on the same device. The p->old_block_size is assigned to the value of block_size the device. This block size should be the same as previous but sometimes it is not. The swapon ends successfully. 3. swapoff: Swapoff resumes, grabs the locks and mutex and continues to disable this swap device. Now it sets the block size to value taken from swap_info which was overwritten by swapon in 2. Signed-off-by: Krzysztof Kozlowski <k.kozlowski@samsung.com> Reported-by: Weijie Yang <weijie.yang.kh@gmail.com> Cc: Bob Liu <bob.liu@oracle.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Shaohua Li <shli@fusionio.com> Cc: Minchan Kim <minchan@kernel.org> Acked-by: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-10-16 20:47:06 +00:00
set_blocksize(bdev, old_block_size);
block: make blkdev_get/put() handle exclusive access Over time, block layer has accumulated a set of APIs dealing with bdev open, close, claim and release. * blkdev_get/put() are the primary open and close functions. * bd_claim/release() deal with exclusive open. * open/close_bdev_exclusive() are combination of open and claim and the other way around, respectively. * bd_link/unlink_disk_holder() to create and remove holder/slave symlinks. * open_by_devnum() wraps bdget() + blkdev_get(). The interface is a bit confusing and the decoupling of open and claim makes it impossible to properly guarantee exclusive access as in-kernel open + claim sequence can disturb the existing exclusive open even before the block layer knows the current open if for another exclusive access. Reorganize the interface such that, * blkdev_get() is extended to include exclusive access management. @holder argument is added and, if is @FMODE_EXCL specified, it will gain exclusive access atomically w.r.t. other exclusive accesses. * blkdev_put() is similarly extended. It now takes @mode argument and if @FMODE_EXCL is set, it releases an exclusive access. Also, when the last exclusive claim is released, the holder/slave symlinks are removed automatically. * bd_claim/release() and close_bdev_exclusive() are no longer necessary and either made static or removed. * bd_link_disk_holder() remains the same but bd_unlink_disk_holder() is no longer necessary and removed. * open_bdev_exclusive() becomes a simple wrapper around lookup_bdev() and blkdev_get(). It also has an unexpected extra bdev_read_only() test which probably should be moved into blkdev_get(). * open_by_devnum() is modified to take @holder argument and pass it to blkdev_get(). Most of bdev open/close operations are unified into blkdev_get/put() and most exclusive accesses are tested atomically at the open time (as it should). This cleans up code and removes some, both valid and invalid, but unnecessary all the same, corner cases. open_bdev_exclusive() and open_by_devnum() can use further cleanup - rename to blkdev_get_by_path() and blkdev_get_by_devt() and drop special features. Well, let's leave them for another day. Most conversions are straight-forward. drbd conversion is a bit more involved as there was some reordering, but the logic should stay the same. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Neil Brown <neilb@suse.de> Acked-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp> Acked-by: Mike Snitzer <snitzer@redhat.com> Acked-by: Philipp Reisner <philipp.reisner@linbit.com> Cc: Peter Osterlund <petero2@telia.com> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Jan Kara <jack@suse.cz> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <joel.becker@oracle.com> Cc: Alex Elder <aelder@sgi.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: dm-devel@redhat.com Cc: drbd-dev@lists.linbit.com Cc: Leo Chen <leochen@broadcom.com> Cc: Scott Branden <sbranden@broadcom.com> Cc: Chris Mason <chris.mason@oracle.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Dave Kleikamp <shaggy@linux.vnet.ibm.com> Cc: Joern Engel <joern@logfs.org> Cc: reiserfs-devel@vger.kernel.org Cc: Alexander Viro <viro@zeniv.linux.org.uk>
2010-11-13 10:55:17 +00:00
blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
}
inode_lock(inode);
inode->i_flags &= ~S_SWAPFILE;
inode_unlock(inode);
filp_close(swap_file, NULL);
/*
* Clear the SWP_USED flag after all resources are freed so that swapon
* can reuse this swap_info in alloc_swap_info() safely. It is ok to
* not hold p->lock after we cleared its SWP_WRITEOK.
*/
spin_lock(&swap_lock);
p->flags = 0;
spin_unlock(&swap_lock);
err = 0;
atomic_inc(&proc_poll_event);
wake_up_interruptible(&proc_poll_wait);
out_dput:
filp_close(victim, NULL);
out:
putname(pathname);
return err;
}
#ifdef CONFIG_PROC_FS
static __poll_t swaps_poll(struct file *file, poll_table *wait)
{
struct seq_file *seq = file->private_data;
poll_wait(file, &proc_poll_wait, wait);
if (seq->poll_event != atomic_read(&proc_poll_event)) {
seq->poll_event = atomic_read(&proc_poll_event);
return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
}
return EPOLLIN | EPOLLRDNORM;
}
/* iterator */
static void *swap_start(struct seq_file *swap, loff_t *pos)
{
struct swap_info_struct *si;
int type;
loff_t l = *pos;
mutex_lock(&swapon_mutex);
if (!l)
return SEQ_START_TOKEN;
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
if (!(si->flags & SWP_USED) || !si->swap_map)
continue;
if (!--l)
return si;
}
return NULL;
}
static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
{
struct swap_info_struct *si = v;
int type;
if (v == SEQ_START_TOKEN)
type = 0;
else
type = si->type + 1;
mm/swapfile.c: swap_next should increase position index If seq_file .next fuction does not change position index, read after some lseek can generate unexpected output. In Aug 2018 NeilBrown noticed commit 1f4aace60b0e ("fs/seq_file.c: simplify seq_file iteration code and interface") "Some ->next functions do not increment *pos when they return NULL... Note that such ->next functions are buggy and should be fixed. A simple demonstration is dd if=/proc/swaps bs=1000 skip=1 Choose any block size larger than the size of /proc/swaps. This will always show the whole last line of /proc/swaps" Described problem is still actual. If you make lseek into middle of last output line following read will output end of last line and whole last line once again. $ dd if=/proc/swaps bs=1 # usual output Filename Type Size Used Priority /dev/dm-0 partition 4194812 97536 -2 104+0 records in 104+0 records out 104 bytes copied $ dd if=/proc/swaps bs=40 skip=1 # last line was generated twice dd: /proc/swaps: cannot skip to specified offset v/dm-0 partition 4194812 97536 -2 /dev/dm-0 partition 4194812 97536 -2 3+1 records in 3+1 records out 131 bytes copied https://bugzilla.kernel.org/show_bug.cgi?id=206283 Link: http://lkml.kernel.org/r/bd8cfd7b-ac95-9b91-f9e7-e8438bd5047d@virtuozzo.com Signed-off-by: Vasily Averin <vvs@virtuozzo.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Cc: Jann Horn <jannh@google.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Kees Cook <keescook@chromium.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-01-31 06:13:39 +00:00
++(*pos);
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
for (; (si = swap_type_to_swap_info(type)); type++) {
if (!(si->flags & SWP_USED) || !si->swap_map)
continue;
return si;
}
return NULL;
}
static void swap_stop(struct seq_file *swap, void *v)
{
mutex_unlock(&swapon_mutex);
}
static int swap_show(struct seq_file *swap, void *v)
{
struct swap_info_struct *si = v;
struct file *file;
int len;
unsigned long bytes, inuse;
if (si == SEQ_START_TOKEN) {
mm/mempool: minor coding style tweaks Various coding style tweaks to various files under mm/ [daizhiyuan@phytium.com.cn: mm/swapfile: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614223624-16055-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/sparse: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614227288-19363-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/vmscan: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614227649-19853-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/compaction: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614228218-20770-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/oom_kill: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614228360-21168-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/shmem: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614228504-21491-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/page_alloc: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614228613-21754-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/filemap: minor coding style tweaks] Link: https://lkml.kernel.org/r/1614228936-22337-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/mlock: minor coding style tweaks] Link: https://lkml.kernel.org/r/1613956588-2453-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/frontswap: minor coding style tweaks] Link: https://lkml.kernel.org/r/1613962668-15045-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/vmalloc: minor coding style tweaks] Link: https://lkml.kernel.org/r/1613963379-15988-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/memory_hotplug: minor coding style tweaks] Link: https://lkml.kernel.org/r/1613971784-24878-1-git-send-email-daizhiyuan@phytium.com.cn [daizhiyuan@phytium.com.cn: mm/mempolicy: minor coding style tweaks] Link: https://lkml.kernel.org/r/1613972228-25501-1-git-send-email-daizhiyuan@phytium.com.cn Link: https://lkml.kernel.org/r/1614222374-13805-1-git-send-email-daizhiyuan@phytium.com.cn Signed-off-by: Zhiyuan Dai <daizhiyuan@phytium.com.cn> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-05 01:40:12 +00:00
seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
return 0;
}
2020-06-02 04:49:26 +00:00
bytes = si->pages << (PAGE_SHIFT - 10);
inuse = READ_ONCE(si->inuse_pages) << (PAGE_SHIFT - 10);
2020-06-02 04:49:26 +00:00
file = si->swap_file;
len = seq_file_path(swap, file, " \t\n\\");
seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
len < 40 ? 40 - len : 1, " ",
S_ISBLK(file_inode(file)->i_mode) ?
"partition" : "file\t",
2020-06-02 04:49:26 +00:00
bytes, bytes < 10000000 ? "\t" : "",
inuse, inuse < 10000000 ? "\t" : "",
si->prio);
return 0;
}
static const struct seq_operations swaps_op = {
.start = swap_start,
.next = swap_next,
.stop = swap_stop,
.show = swap_show
};
static int swaps_open(struct inode *inode, struct file *file)
{
struct seq_file *seq;
int ret;
ret = seq_open(file, &swaps_op);
if (ret)
return ret;
seq = file->private_data;
seq->poll_event = atomic_read(&proc_poll_event);
return 0;
}
static const struct proc_ops swaps_proc_ops = {
proc: faster open/read/close with "permanent" files Now that "struct proc_ops" exist we can start putting there stuff which could not fly with VFS "struct file_operations"... Most of fs/proc/inode.c file is dedicated to make open/read/.../close reliable in the event of disappearing /proc entries which usually happens if module is getting removed. Files like /proc/cpuinfo which never disappear simply do not need such protection. Save 2 atomic ops, 1 allocation, 1 free per open/read/close sequence for such "permanent" files. Enable "permanent" flag for /proc/cpuinfo /proc/kmsg /proc/modules /proc/slabinfo /proc/stat /proc/sysvipc/* /proc/swaps More will come once I figure out foolproof way to prevent out module authors from marking their stuff "permanent" for performance reasons when it is not. This should help with scalability: benchmark is "read /proc/cpuinfo R times by N threads scattered over the system". N R t, s (before) t, s (after) ----------------------------------------------------- 64 4096 1.582458 1.530502 -3.2% 256 4096 6.371926 6.125168 -3.9% 1024 4096 25.64888 24.47528 -4.6% Benchmark source: #include <chrono> #include <iostream> #include <thread> #include <vector> #include <sys/types.h> #include <sys/stat.h> #include <fcntl.h> #include <unistd.h> const int NR_CPUS = sysconf(_SC_NPROCESSORS_ONLN); int N; const char *filename; int R; int xxx = 0; int glue(int n) { cpu_set_t m; CPU_ZERO(&m); CPU_SET(n, &m); return sched_setaffinity(0, sizeof(cpu_set_t), &m); } void f(int n) { glue(n % NR_CPUS); while (*(volatile int *)&xxx == 0) { } for (int i = 0; i < R; i++) { int fd = open(filename, O_RDONLY); char buf[4096]; ssize_t rv = read(fd, buf, sizeof(buf)); asm volatile ("" :: "g" (rv)); close(fd); } } int main(int argc, char *argv[]) { if (argc < 4) { std::cerr << "usage: " << argv[0] << ' ' << "N /proc/filename R "; return 1; } N = atoi(argv[1]); filename = argv[2]; R = atoi(argv[3]); for (int i = 0; i < NR_CPUS; i++) { if (glue(i) == 0) break; } std::vector<std::thread> T; T.reserve(N); for (int i = 0; i < N; i++) { T.emplace_back(f, i); } auto t0 = std::chrono::system_clock::now(); { *(volatile int *)&xxx = 1; for (auto& t: T) { t.join(); } } auto t1 = std::chrono::system_clock::now(); std::chrono::duration<double> dt = t1 - t0; std::cout << dt.count() << ' '; return 0; } P.S.: Explicit randomization marker is added because adding non-function pointer will silently disable structure layout randomization. [akpm@linux-foundation.org: coding style fixes] Reported-by: kbuild test robot <lkp@intel.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Joe Perches <joe@perches.com> Link: http://lkml.kernel.org/r/20200222201539.GA22576@avx2 Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-04-07 03:09:01 +00:00
.proc_flags = PROC_ENTRY_PERMANENT,
.proc_open = swaps_open,
.proc_read = seq_read,
.proc_lseek = seq_lseek,
.proc_release = seq_release,
.proc_poll = swaps_poll,
};
static int __init procswaps_init(void)
{
proc_create("swaps", 0, NULL, &swaps_proc_ops);
return 0;
}
__initcall(procswaps_init);
#endif /* CONFIG_PROC_FS */
#ifdef MAX_SWAPFILES_CHECK
static int __init max_swapfiles_check(void)
{
MAX_SWAPFILES_CHECK();
return 0;
}
late_initcall(max_swapfiles_check);
#endif
static struct swap_info_struct *alloc_swap_info(void)
{
struct swap_info_struct *p;
struct swap_info_struct *defer = NULL;
unsigned int type;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
int i;
p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
if (!p)
return ERR_PTR(-ENOMEM);
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
if (percpu_ref_init(&p->users, swap_users_ref_free,
PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
kvfree(p);
return ERR_PTR(-ENOMEM);
}
spin_lock(&swap_lock);
for (type = 0; type < nr_swapfiles; type++) {
if (!(swap_info[type]->flags & SWP_USED))
break;
}
[PATCH] Swapless page migration: add R/W migration entries Implement read/write migration ptes We take the upper two swapfiles for the two types of migration ptes and define a series of macros in swapops.h. The VM is modified to handle the migration entries. migration entries can only be encountered when the page they are pointing to is locked. This limits the number of places one has to fix. We also check in copy_pte_range and in mprotect_pte_range() for migration ptes. We check for migration ptes in do_swap_cache and call a function that will then wait on the page lock. This allows us to effectively stop all accesses to apge. Migration entries are created by try_to_unmap if called for migration and removed by local functions in migrate.c From: Hugh Dickins <hugh@veritas.com> Several times while testing swapless page migration (I've no NUMA, just hacking it up to migrate recklessly while running load), I've hit the BUG_ON(!PageLocked(p)) in migration_entry_to_page. This comes from an orphaned migration entry, unrelated to the current correctly locked migration, but hit by remove_anon_migration_ptes as it checks an address in each vma of the anon_vma list. Such an orphan may be left behind if an earlier migration raced with fork: copy_one_pte can duplicate a migration entry from parent to child, after remove_anon_migration_ptes has checked the child vma, but before it has removed it from the parent vma. (If the process were later to fault on this orphaned entry, it would hit the same BUG from migration_entry_wait.) This could be fixed by locking anon_vma in copy_one_pte, but we'd rather not. There's no such problem with file pages, because vma_prio_tree_add adds child vma after parent vma, and the page table locking at each end is enough to serialize. Follow that example with anon_vma: add new vmas to the tail instead of the head. (There's no corresponding problem when inserting migration entries, because a missed pte will leave the page count and mapcount high, which is allowed for. And there's no corresponding problem when migrating via swap, because a leftover swap entry will be correctly faulted. But the swapless method has no refcounting of its entries.) From: Ingo Molnar <mingo@elte.hu> pte_unmap_unlock() takes the pte pointer as an argument. From: Hugh Dickins <hugh@veritas.com> Several times while testing swapless page migration, gcc has tried to exec a pointer instead of a string: smells like COW mappings are not being properly write-protected on fork. The protection in copy_one_pte looks very convincing, until at last you realize that the second arg to make_migration_entry is a boolean "write", and SWP_MIGRATION_READ is 30. Anyway, it's better done like in change_pte_range, using is_write_migration_entry and make_migration_entry_read. From: Hugh Dickins <hugh@veritas.com> Remove unnecessary obfuscation from sys_swapon's range check on swap type, which blew up causing memory corruption once swapless migration made MAX_SWAPFILES no longer 2 ^ MAX_SWAPFILES_SHIFT. Signed-off-by: Hugh Dickins <hugh@veritas.com> Acked-by: Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by: Hugh Dickins <hugh@veritas.com> Signed-off-by: Christoph Lameter <clameter@engr.sgi.com> Signed-off-by: Ingo Molnar <mingo@elte.hu> From: Hugh Dickins <hugh@veritas.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-23 09:03:35 +00:00
if (type >= MAX_SWAPFILES) {
spin_unlock(&swap_lock);
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
percpu_ref_exit(&p->users);
kvfree(p);
return ERR_PTR(-EPERM);
}
if (type >= nr_swapfiles) {
p->type = type;
/*
mm, swap: remove unnecessary smp_rmb() in swap_type_to_swap_info() Before commit c10d38cc8d3e ("mm, swap: bounds check swap_info array accesses to avoid NULL derefs"), the typical code to reference the swap_info[] is as follows, type = swp_type(swp_entry); if (type >= nr_swapfiles) /* handle invalid swp_entry */; p = swap_info[type]; /* access fields of *p. OOPS! p may be NULL! */ Because the ordering isn't guaranteed, it's possible that swap_info[type] is read before "nr_swapfiles". And that may result in NULL pointer dereference. So after commit c10d38cc8d3e, the code becomes, struct swap_info_struct *swap_type_to_swap_info(int type) { if (type >= READ_ONCE(nr_swapfiles)) return NULL; smp_rmb(); return READ_ONCE(swap_info[type]); } /* users */ type = swp_type(swp_entry); p = swap_type_to_swap_info(type); if (!p) /* handle invalid swp_entry */; /* dereference p */ Where the value of swap_info[type] (that is, "p") is checked to be non-zero before being dereferenced. So, the NULL deferencing becomes impossible even if "nr_swapfiles" is read after swap_info[type]. Therefore, the "smp_rmb()" becomes unnecessary. And, we don't even need to read "nr_swapfiles" here. Because the non-zero checking for "p" is sufficient. We just need to make sure we will not access out of the boundary of the array. With the change, nr_swapfiles will only be accessed with swap_lock held, except in swapcache_free_entries(). Where the absolute correctness of the value isn't needed, as described in the comments. We still need to guarantee swap_info[type] is read before being dereferenced. That can be satisfied via the data dependency ordering enforced by READ_ONCE(swap_info[type]). This needs to be paired with proper write barriers. So smp_store_release() is used in alloc_swap_info() to guarantee the fields of *swap_info[type] is initialized before swap_info[type] itself being written. Note that the fields of *swap_info[type] is initialized to be 0 via kvzalloc() firstly. The assignment and deferencing of swap_info[type] is like rcu_assign_pointer() and rcu_dereference(). Link: https://lkml.kernel.org/r/20210520073301.1676294-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Dan Carpenter <dan.carpenter@oracle.com> Cc: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Cc: Miaohe Lin <linmiaohe@huawei.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:37:09 +00:00
* Publish the swap_info_struct after initializing it.
* Note that kvzalloc() above zeroes all its fields.
*/
mm, swap: remove unnecessary smp_rmb() in swap_type_to_swap_info() Before commit c10d38cc8d3e ("mm, swap: bounds check swap_info array accesses to avoid NULL derefs"), the typical code to reference the swap_info[] is as follows, type = swp_type(swp_entry); if (type >= nr_swapfiles) /* handle invalid swp_entry */; p = swap_info[type]; /* access fields of *p. OOPS! p may be NULL! */ Because the ordering isn't guaranteed, it's possible that swap_info[type] is read before "nr_swapfiles". And that may result in NULL pointer dereference. So after commit c10d38cc8d3e, the code becomes, struct swap_info_struct *swap_type_to_swap_info(int type) { if (type >= READ_ONCE(nr_swapfiles)) return NULL; smp_rmb(); return READ_ONCE(swap_info[type]); } /* users */ type = swp_type(swp_entry); p = swap_type_to_swap_info(type); if (!p) /* handle invalid swp_entry */; /* dereference p */ Where the value of swap_info[type] (that is, "p") is checked to be non-zero before being dereferenced. So, the NULL deferencing becomes impossible even if "nr_swapfiles" is read after swap_info[type]. Therefore, the "smp_rmb()" becomes unnecessary. And, we don't even need to read "nr_swapfiles" here. Because the non-zero checking for "p" is sufficient. We just need to make sure we will not access out of the boundary of the array. With the change, nr_swapfiles will only be accessed with swap_lock held, except in swapcache_free_entries(). Where the absolute correctness of the value isn't needed, as described in the comments. We still need to guarantee swap_info[type] is read before being dereferenced. That can be satisfied via the data dependency ordering enforced by READ_ONCE(swap_info[type]). This needs to be paired with proper write barriers. So smp_store_release() is used in alloc_swap_info() to guarantee the fields of *swap_info[type] is initialized before swap_info[type] itself being written. Note that the fields of *swap_info[type] is initialized to be 0 via kvzalloc() firstly. The assignment and deferencing of swap_info[type] is like rcu_assign_pointer() and rcu_dereference(). Link: https://lkml.kernel.org/r/20210520073301.1676294-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Dan Carpenter <dan.carpenter@oracle.com> Cc: Andrea Parri <andrea.parri@amarulasolutions.com> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@kernel.org> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Cc: Miaohe Lin <linmiaohe@huawei.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:37:09 +00:00
smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
nr_swapfiles++;
} else {
defer = p;
p = swap_info[type];
/*
* Do not memset this entry: a racing procfs swap_next()
* would be relying on p->type to remain valid.
*/
}
mm, swap: use rbtree for swap_extent swap_extent is used to map swap page offset to backing device's block offset. For a continuous block range, one swap_extent is used and all these swap_extents are managed in a linked list. These swap_extents are used by map_swap_entry() during swap's read and write path. To find out the backing device's block offset for a page offset, the swap_extent list will be traversed linearly, with curr_swap_extent being used as a cache to speed up the search. This works well as long as swap_extents are not huge or when the number of processes that access swap device are few, but when the swap device has many extents and there are a number of processes accessing the swap device concurrently, it can be a problem. On one of our servers, the disk's remaining size is tight: $df -h Filesystem Size Used Avail Use% Mounted on ... ... /dev/nvme0n1p1 1.8T 1.3T 504G 72% /home/t4 When creating a 80G swapfile there, there are as many as 84656 swap extents. The end result is, kernel spends abou 30% time in map_swap_entry() and swap throughput is only 70MB/s. As a comparison, when I used smaller sized swapfile, like 4G whose swap_extent dropped to 2000, swap throughput is back to 400-500MB/s and map_swap_entry() is about 3%. One downside of using rbtree for swap_extent is, 'struct rbtree' takes 24 bytes while 'struct list_head' takes 16 bytes, that's 8 bytes more for each swap_extent. For a swapfile that has 80k swap_extents, that means 625KiB more memory consumed. Test: Since it's not possible to reboot that server, I can not test this patch diretly there. Instead, I tested it on another server with NVMe disk. I created a 20G swapfile on an NVMe backed XFS fs. By default, the filesystem is quite clean and the created swapfile has only 2 extents. Testing vanilla and this patch shows no obvious performance difference when swapfile is not fragmented. To see the patch's effects, I used some tweaks to manually fragment the swapfile by breaking the extent at 1M boundary. This made the swapfile have 20K extents. nr_task=4 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 165191 90.77% 171798 90.21% patched 858993 +420% 2.16% 715827 +317% 0.77% nr_task=8 kernel swapout(KB/s) map_swap_entry(perf) swapin(KB/s) map_swap_entry(perf) vanilla 306783 92.19% 318145 87.76% patched 954437 +211% 2.35% 1073741 +237% 1.57% swapout: the throughput of swap out, in KB/s, higher is better 1st map_swap_entry: cpu cycles percent sampled by perf swapin: the throughput of swap in, in KB/s, higher is better. 2nd map_swap_entry: cpu cycles percent sampled by perf nr_task=1 doesn't show any difference, this is due to the curr_swap_extent can be effectively used to cache the correct swap extent for single task workload. [akpm@linux-foundation.org: s/BUG_ON(1)/BUG()/] Link: http://lkml.kernel.org/r/20190523142404.GA181@aaronlu Signed-off-by: Aaron Lu <ziqian.lzq@antfin.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:41 +00:00
p->swap_extent_root = RB_ROOT;
swap: change swap_list_head to plist, add swap_avail_head Originally get_swap_page() started iterating through the singly-linked list of swap_info_structs using swap_list.next or highest_priority_index, which both were intended to point to the highest priority active swap target that was not full. The first patch in this series changed the singly-linked list to a doubly-linked list, and removed the logic to start at the highest priority non-full entry; it starts scanning at the highest priority entry each time, even if the entry is full. Replace the manually ordered swap_list_head with a plist, swap_active_head. Add a new plist, swap_avail_head. The original swap_active_head plist contains all active swap_info_structs, as before, while the new swap_avail_head plist contains only swap_info_structs that are active and available, i.e. not full. Add a new spinlock, swap_avail_lock, to protect the swap_avail_head list. Mel Gorman suggested using plists since they internally handle ordering the list entries based on priority, which is exactly what swap was doing manually. All the ordering code is now removed, and swap_info_struct entries and simply added to their corresponding plist and automatically ordered correctly. Using a new plist for available swap_info_structs simplifies and optimizes get_swap_page(), which no longer has to iterate over full swap_info_structs. Using a new spinlock for swap_avail_head plist allows each swap_info_struct to add or remove themselves from the plist when they become full or not-full; previously they could not do so because the swap_info_struct->lock is held when they change from full<->not-full, and the swap_lock protecting the main swap_active_head must be ordered before any swap_info_struct->lock. Signed-off-by: Dan Streetman <ddstreet@ieee.org> Acked-by: Mel Gorman <mgorman@suse.de> Cc: Shaohua Li <shli@fusionio.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Dan Streetman <ddstreet@ieee.org> Cc: Michal Hocko <mhocko@suse.cz> Cc: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com> Cc: Weijie Yang <weijieut@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Bob Liu <bob.liu@oracle.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-06-04 23:09:59 +00:00
plist_node_init(&p->list, 0);
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
for_each_node(i)
plist_node_init(&p->avail_lists[i], 0);
p->flags = SWP_USED;
spin_unlock(&swap_lock);
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
if (defer) {
percpu_ref_exit(&defer->users);
kvfree(defer);
}
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_lock_init(&p->lock);
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
spin_lock_init(&p->cont_lock);
mm/swapfile: use percpu_ref to serialize against concurrent swapoff Patch series "close various race windows for swap", v6. When I was investigating the swap code, I found some possible race windows. This series aims to fix all these races. But using current get/put_swap_device() to guard against concurrent swapoff for swap_readpage() looks terrible because swap_readpage() may take really long time. And to reduce the performance overhead on the hot-path as much as possible, it appears we can use the percpu_ref to close this race window(as suggested by Huang, Ying). The patch 1 adds percpu_ref support for swap and most of the remaining patches try to use this to close various race windows. More details can be found in the respective changelogs. This patch (of 4): Using current get/put_swap_device() to guard against concurrent swapoff for some swap ops, e.g. swap_readpage(), looks terrible because they might take really long time. This patch adds the percpu_ref support to serialize against concurrent swapoff(as suggested by Huang, Ying). Also we remove the SWP_VALID flag because it's used together with RCU solution. Link: https://lkml.kernel.org/r/20210426123316.806267-1-linmiaohe@huawei.com Link: https://lkml.kernel.org/r/20210426123316.806267-2-linmiaohe@huawei.com Signed-off-by: Miaohe Lin <linmiaohe@huawei.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alex Shi <alexs@kernel.org> Cc: David Hildenbrand <david@redhat.com> Cc: Dennis Zhou <dennis@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Yu Zhao <yuzhao@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-06-29 02:36:46 +00:00
init_completion(&p->comp);
return p;
}
static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
{
int error;
if (S_ISBLK(inode->i_mode)) {
p->bdev = blkdev_get_by_dev(inode->i_rdev,
FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
if (IS_ERR(p->bdev)) {
error = PTR_ERR(p->bdev);
p->bdev = NULL;
return error;
}
p->old_block_size = block_size(p->bdev);
error = set_blocksize(p->bdev, PAGE_SIZE);
if (error < 0)
return error;
/*
* Zoned block devices contain zones that have a sequential
* write only restriction. Hence zoned block devices are not
* suitable for swapping. Disallow them here.
*/
if (bdev_is_zoned(p->bdev))
return -EINVAL;
p->flags |= SWP_BLKDEV;
} else if (S_ISREG(inode->i_mode)) {
p->bdev = inode->i_sb->s_bdev;
}
return 0;
}
/*
* Find out how many pages are allowed for a single swap device. There
* are two limiting factors:
* 1) the number of bits for the swap offset in the swp_entry_t type, and
* 2) the number of bits in the swap pte, as defined by the different
* architectures.
*
* In order to find the largest possible bit mask, a swap entry with
* swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
* decoded to a swp_entry_t again, and finally the swap offset is
* extracted.
*
* This will mask all the bits from the initial ~0UL mask that can't
* be encoded in either the swp_entry_t or the architecture definition
* of a swap pte.
*/
unsigned long generic_max_swapfile_size(void)
{
return swp_offset(pte_to_swp_entry(
swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
}
/* Can be overridden by an architecture for additional checks. */
mm/swap: cache maximum swapfile size when init swap We used to have swapfile_maximum_size() fetching a maximum value of swapfile size per-arch. As the caller of max_swapfile_size() grows, this patch introduce a variable "swapfile_maximum_size" and cache the value of old max_swapfile_size(), so that we don't need to calculate the value every time. Caching the value in swapfile_init() is safe because when reaching the phase we should have initialized all the relevant information. Here the major arch to take care of is x86, which defines the max swapfile size based on L1TF mitigation. Here both X86_BUG_L1TF or l1tf_mitigation should have been setup properly when reaching swapfile_init(). As a reference, the code path looks like this for x86: - start_kernel - setup_arch - early_cpu_init - early_identify_cpu --> setup X86_BUG_L1TF - parse_early_param - l1tf_cmdline --> set l1tf_mitigation - check_bugs - l1tf_select_mitigation --> set l1tf_mitigation - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) The swapfile size only depends on swp pte format on non-x86 archs, so caching it is safe too. Since at it, rename max_swapfile_size() to arch_max_swapfile_size() because arch can define its own function, so it's more straightforward to have "arch_" as its prefix. At the meantime, export swapfile_maximum_size to replace the old usages of max_swapfile_size(). [peterx@redhat.com: declare arch_max_swapfile_size) in swapfile.h] Link: https://lkml.kernel.org/r/YxTh1GuC6ro5fKL5@xz-m1.local Link: https://lkml.kernel.org/r/20220811161331.37055-7-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:30 +00:00
__weak unsigned long arch_max_swapfile_size(void)
{
return generic_max_swapfile_size();
}
static unsigned long read_swap_header(struct swap_info_struct *p,
union swap_header *swap_header,
struct inode *inode)
{
int i;
unsigned long maxpages;
unsigned long swapfilepages;
unsigned long last_page;
if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
pr_err("Unable to find swap-space signature\n");
return 0;
}
/* swap partition endianness hack... */
if (swab32(swap_header->info.version) == 1) {
swab32s(&swap_header->info.version);
swab32s(&swap_header->info.last_page);
swab32s(&swap_header->info.nr_badpages);
if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
return 0;
for (i = 0; i < swap_header->info.nr_badpages; i++)
swab32s(&swap_header->info.badpages[i]);
}
/* Check the swap header's sub-version */
if (swap_header->info.version != 1) {
pr_warn("Unable to handle swap header version %d\n",
swap_header->info.version);
return 0;
}
p->lowest_bit = 1;
p->cluster_next = 1;
p->cluster_nr = 0;
mm/swap: cache maximum swapfile size when init swap We used to have swapfile_maximum_size() fetching a maximum value of swapfile size per-arch. As the caller of max_swapfile_size() grows, this patch introduce a variable "swapfile_maximum_size" and cache the value of old max_swapfile_size(), so that we don't need to calculate the value every time. Caching the value in swapfile_init() is safe because when reaching the phase we should have initialized all the relevant information. Here the major arch to take care of is x86, which defines the max swapfile size based on L1TF mitigation. Here both X86_BUG_L1TF or l1tf_mitigation should have been setup properly when reaching swapfile_init(). As a reference, the code path looks like this for x86: - start_kernel - setup_arch - early_cpu_init - early_identify_cpu --> setup X86_BUG_L1TF - parse_early_param - l1tf_cmdline --> set l1tf_mitigation - check_bugs - l1tf_select_mitigation --> set l1tf_mitigation - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) The swapfile size only depends on swp pte format on non-x86 archs, so caching it is safe too. Since at it, rename max_swapfile_size() to arch_max_swapfile_size() because arch can define its own function, so it's more straightforward to have "arch_" as its prefix. At the meantime, export swapfile_maximum_size to replace the old usages of max_swapfile_size(). [peterx@redhat.com: declare arch_max_swapfile_size) in swapfile.h] Link: https://lkml.kernel.org/r/YxTh1GuC6ro5fKL5@xz-m1.local Link: https://lkml.kernel.org/r/20220811161331.37055-7-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:30 +00:00
maxpages = swapfile_maximum_size;
last_page = swap_header->info.last_page;
if (!last_page) {
pr_warn("Empty swap-file\n");
return 0;
}
if (last_page > maxpages) {
pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
maxpages << (PAGE_SHIFT - 10),
last_page << (PAGE_SHIFT - 10));
}
if (maxpages > last_page) {
maxpages = last_page + 1;
/* p->max is an unsigned int: don't overflow it */
if ((unsigned int)maxpages == 0)
maxpages = UINT_MAX;
}
p->highest_bit = maxpages - 1;
if (!maxpages)
return 0;
swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
if (swapfilepages && maxpages > swapfilepages) {
pr_warn("Swap area shorter than signature indicates\n");
return 0;
}
if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
return 0;
if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
return 0;
return maxpages;
}
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
#define SWAP_CLUSTER_INFO_COLS \
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
#define SWAP_CLUSTER_SPACE_COLS \
DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
#define SWAP_CLUSTER_COLS \
max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
static int setup_swap_map_and_extents(struct swap_info_struct *p,
union swap_header *swap_header,
unsigned char *swap_map,
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
struct swap_cluster_info *cluster_info,
unsigned long maxpages,
sector_t *span)
{
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unsigned int j, k;
unsigned int nr_good_pages;
int nr_extents;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
unsigned long i, idx;
nr_good_pages = maxpages - 1; /* omit header page */
cluster_list_init(&p->free_clusters);
cluster_list_init(&p->discard_clusters);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
for (i = 0; i < swap_header->info.nr_badpages; i++) {
unsigned int page_nr = swap_header->info.badpages[i];
if (page_nr == 0 || page_nr > swap_header->info.last_page)
return -EINVAL;
if (page_nr < maxpages) {
swap_map[page_nr] = SWAP_MAP_BAD;
nr_good_pages--;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* Haven't marked the cluster free yet, no list
* operation involved
*/
inc_cluster_info_page(p, cluster_info, page_nr);
}
}
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/* Haven't marked the cluster free yet, no list operation involved */
for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
inc_cluster_info_page(p, cluster_info, i);
if (nr_good_pages) {
swap_map[0] = SWAP_MAP_BAD;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* Not mark the cluster free yet, no list
* operation involved
*/
inc_cluster_info_page(p, cluster_info, 0);
p->max = maxpages;
p->pages = nr_good_pages;
nr_extents = setup_swap_extents(p, span);
if (nr_extents < 0)
return nr_extents;
nr_good_pages = p->pages;
}
if (!nr_good_pages) {
pr_warn("Empty swap-file\n");
return -EINVAL;
}
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
if (!cluster_info)
return nr_extents;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
/*
* Reduce false cache line sharing between cluster_info and
* sharing same address space.
*/
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
j = (k + col) % SWAP_CLUSTER_COLS;
for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
idx = i * SWAP_CLUSTER_COLS + j;
if (idx >= nr_clusters)
continue;
if (cluster_count(&cluster_info[idx]))
continue;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
cluster_list_add_tail(&p->free_clusters, cluster_info,
idx);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
}
}
return nr_extents;
}
SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
{
struct swap_info_struct *p;
struct filename *name;
struct file *swap_file = NULL;
struct address_space *mapping;
struct dentry *dentry;
int prio;
int error;
union swap_header *swap_header;
int nr_extents;
sector_t span;
unsigned long maxpages;
unsigned char *swap_map = NULL;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
struct swap_cluster_info *cluster_info = NULL;
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
unsigned long *frontswap_map = NULL;
struct page *page = NULL;
struct inode *inode = NULL;
bool inced_nr_rotate_swap = false;
if (swap_flags & ~SWAP_FLAGS_VALID)
return -EINVAL;
if (!capable(CAP_SYS_ADMIN))
return -EPERM;
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
if (!swap_avail_heads)
return -ENOMEM;
p = alloc_swap_info();
if (IS_ERR(p))
return PTR_ERR(p);
swap: make swap discard async swap can do cluster discard for SSD, which is good, but there are some problems here: 1. swap do the discard just before page reclaim gets a swap entry and writes the disk sectors. This is useless for high end SSD, because an overwrite to a sector implies a discard to original sector too. A discard + overwrite == overwrite. 2. the purpose of doing discard is to improve SSD firmware garbage collection. Idealy we should send discard as early as possible, so firmware can do something smart. Sending discard just after swap entry is freed is considered early compared to sending discard before write. Of course, if workload is already bound to gc speed, sending discard earlier or later doesn't make 3. block discard is a sync API, which will delay scan_swap_map() significantly. 4. Write and discard command can be executed parallel in PCIe SSD. Making swap discard async can make execution more efficiently. This patch makes swap discard async and moves discard to where swap entry is freed. Discard and write have no dependence now, so above issues can be avoided. Idealy we should do discard for any freed sectors, but some SSD discard is very slow. This patch still does discard for a whole cluster. My test does a several round of 'mmap, write, unmap', which will trigger a lot of swap discard. In a fusionio card, with this patch, the test runtime is reduced to 18% of the time without it, so around 5.5x faster. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:30 +00:00
INIT_WORK(&p->discard_work, swap_discard_work);
name = getname(specialfile);
if (IS_ERR(name)) {
error = PTR_ERR(name);
name = NULL;
goto bad_swap;
}
swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
if (IS_ERR(swap_file)) {
error = PTR_ERR(swap_file);
swap_file = NULL;
goto bad_swap;
}
p->swap_file = swap_file;
mapping = swap_file->f_mapping;
dentry = swap_file->f_path.dentry;
inode = mapping->host;
error = claim_swapfile(p, inode);
if (unlikely(error))
goto bad_swap;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
inode_lock(inode);
if (d_unlinked(dentry) || cant_mount(dentry)) {
error = -ENOENT;
goto bad_swap_unlock_inode;
}
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
if (IS_SWAPFILE(inode)) {
error = -EBUSY;
goto bad_swap_unlock_inode;
}
/*
* Read the swap header.
*/
if (!mapping->a_ops->read_folio) {
error = -EINVAL;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
}
page = read_mapping_page(mapping, 0, swap_file);
if (IS_ERR(page)) {
error = PTR_ERR(page);
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
}
swap_header = kmap(page);
maxpages = read_swap_header(p, swap_header, inode);
if (unlikely(!maxpages)) {
error = -EINVAL;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
}
/* OK, set up the swap map and apply the bad block list */
swap_map = vzalloc(maxpages);
if (!swap_map) {
error = -ENOMEM;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
}
mm: support anonymous stable page During developemnt for zram-swap asynchronous writeback, I found strange corruption of compressed page, resulting in: Modules linked in: zram(E) CPU: 3 PID: 1520 Comm: zramd-1 Tainted: G E 4.8.0-mm1-00320-ge0d4894c9c38-dirty #3274 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014 task: ffff88007620b840 task.stack: ffff880078090000 RIP: set_freeobj.part.43+0x1c/0x1f RSP: 0018:ffff880078093ca8 EFLAGS: 00010246 RAX: 0000000000000018 RBX: ffff880076798d88 RCX: ffffffff81c408c8 RDX: 0000000000000018 RSI: 0000000000000000 RDI: 0000000000000246 RBP: ffff880078093cb0 R08: 0000000000000000 R09: 0000000000000000 R10: ffff88005bc43030 R11: 0000000000001df3 R12: ffff880076798d88 R13: 000000000005bc43 R14: ffff88007819d1b8 R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff88007e380000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fc934048f20 CR3: 0000000077b01000 CR4: 00000000000406e0 Call Trace: obj_malloc+0x22b/0x260 zs_malloc+0x1e4/0x580 zram_bvec_rw+0x4cd/0x830 [zram] page_requests_rw+0x9c/0x130 [zram] zram_thread+0xe6/0x173 [zram] kthread+0xca/0xe0 ret_from_fork+0x25/0x30 With investigation, it reveals currently stable page doesn't support anonymous page. IOW, reuse_swap_page can reuse the page without waiting writeback completion so it can overwrite page zram is compressing. Unfortunately, zram has used per-cpu stream feature from v4.7. It aims for increasing cache hit ratio of scratch buffer for compressing. Downside of that approach is that zram should ask memory space for compressed page in per-cpu context which requires stricted gfp flag which could be failed. If so, it retries to allocate memory space out of per-cpu context so it could get memory this time and compress the data again, copies it to the memory space. In this scenario, zram assumes the data should never be changed but it is not true unless stable page supports. So, If the data is changed under us, zram can make buffer overrun because second compression size could be bigger than one we got in previous trial and blindly, copy bigger size object to smaller buffer which is buffer overrun. The overrun breaks zsmalloc free object chaining so system goes crash like above. I think below is same problem. https://bugzilla.suse.com/show_bug.cgi?id=997574 Unfortunately, reuse_swap_page should be atomic so that we cannot wait on writeback in there so the approach in this patch is simply return false if we found it needs stable page. Although it increases memory footprint temporarily, it happens rarely and it should be reclaimed easily althoug it happened. Also, It would be better than waiting of IO completion, which is critial path for application latency. Fixes: da9556a2367c ("zram: user per-cpu compression streams") Link: http://lkml.kernel.org/r/20161120233015.GA14113@bbox Link: http://lkml.kernel.org/r/1482366980-3782-2-git-send-email-minchan@kernel.org Signed-off-by: Minchan Kim <minchan@kernel.org> Acked-by: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: Takashi Iwai <tiwai@suse.de> Cc: Hyeoncheol Lee <cheol.lee@lge.com> Cc: <yjay.kim@lge.com> Cc: Sangseok Lee <sangseok.lee@lge.com> Cc: <stable@vger.kernel.org> [4.7+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:58:15 +00:00
if (p->bdev && bdev_stable_writes(p->bdev))
mm: support anonymous stable page During developemnt for zram-swap asynchronous writeback, I found strange corruption of compressed page, resulting in: Modules linked in: zram(E) CPU: 3 PID: 1520 Comm: zramd-1 Tainted: G E 4.8.0-mm1-00320-ge0d4894c9c38-dirty #3274 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Ubuntu-1.8.2-1ubuntu1 04/01/2014 task: ffff88007620b840 task.stack: ffff880078090000 RIP: set_freeobj.part.43+0x1c/0x1f RSP: 0018:ffff880078093ca8 EFLAGS: 00010246 RAX: 0000000000000018 RBX: ffff880076798d88 RCX: ffffffff81c408c8 RDX: 0000000000000018 RSI: 0000000000000000 RDI: 0000000000000246 RBP: ffff880078093cb0 R08: 0000000000000000 R09: 0000000000000000 R10: ffff88005bc43030 R11: 0000000000001df3 R12: ffff880076798d88 R13: 000000000005bc43 R14: ffff88007819d1b8 R15: 0000000000000001 FS: 0000000000000000(0000) GS:ffff88007e380000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fc934048f20 CR3: 0000000077b01000 CR4: 00000000000406e0 Call Trace: obj_malloc+0x22b/0x260 zs_malloc+0x1e4/0x580 zram_bvec_rw+0x4cd/0x830 [zram] page_requests_rw+0x9c/0x130 [zram] zram_thread+0xe6/0x173 [zram] kthread+0xca/0xe0 ret_from_fork+0x25/0x30 With investigation, it reveals currently stable page doesn't support anonymous page. IOW, reuse_swap_page can reuse the page without waiting writeback completion so it can overwrite page zram is compressing. Unfortunately, zram has used per-cpu stream feature from v4.7. It aims for increasing cache hit ratio of scratch buffer for compressing. Downside of that approach is that zram should ask memory space for compressed page in per-cpu context which requires stricted gfp flag which could be failed. If so, it retries to allocate memory space out of per-cpu context so it could get memory this time and compress the data again, copies it to the memory space. In this scenario, zram assumes the data should never be changed but it is not true unless stable page supports. So, If the data is changed under us, zram can make buffer overrun because second compression size could be bigger than one we got in previous trial and blindly, copy bigger size object to smaller buffer which is buffer overrun. The overrun breaks zsmalloc free object chaining so system goes crash like above. I think below is same problem. https://bugzilla.suse.com/show_bug.cgi?id=997574 Unfortunately, reuse_swap_page should be atomic so that we cannot wait on writeback in there so the approach in this patch is simply return false if we found it needs stable page. Although it increases memory footprint temporarily, it happens rarely and it should be reclaimed easily althoug it happened. Also, It would be better than waiting of IO completion, which is critial path for application latency. Fixes: da9556a2367c ("zram: user per-cpu compression streams") Link: http://lkml.kernel.org/r/20161120233015.GA14113@bbox Link: http://lkml.kernel.org/r/1482366980-3782-2-git-send-email-minchan@kernel.org Signed-off-by: Minchan Kim <minchan@kernel.org> Acked-by: Hugh Dickins <hughd@google.com> Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: Takashi Iwai <tiwai@suse.de> Cc: Hyeoncheol Lee <cheol.lee@lge.com> Cc: <yjay.kim@lge.com> Cc: Sangseok Lee <sangseok.lee@lge.com> Cc: <stable@vger.kernel.org> [4.7+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:58:15 +00:00
p->flags |= SWP_STABLE_WRITES;
if (p->bdev && bdev_synchronous(p->bdev))
p->flags |= SWP_SYNCHRONOUS_IO;
if (p->bdev && bdev_nonrot(p->bdev)) {
int cpu;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unsigned long ci, nr_cluster;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
p->flags |= SWP_SOLIDSTATE;
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
p->cluster_next_cpu = alloc_percpu(unsigned int);
if (!p->cluster_next_cpu) {
error = -ENOMEM;
goto bad_swap_unlock_inode;
}
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* select a random position to start with to help wear leveling
* SSD
*/
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
for_each_possible_cpu(cpu) {
per_cpu(*p->cluster_next_cpu, cpu) =
get_random_u32_inclusive(1, p->highest_bit);
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
}
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
treewide: kvzalloc() -> kvcalloc() The kvzalloc() function has a 2-factor argument form, kvcalloc(). This patch replaces cases of: kvzalloc(a * b, gfp) with: kvcalloc(a * b, gfp) as well as handling cases of: kvzalloc(a * b * c, gfp) with: kvzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kvcalloc(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kvzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kvzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kvzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kvzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kvzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kvzalloc( - sizeof(char) * COUNT + COUNT , ...) | kvzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kvzalloc + kvcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kvzalloc + kvcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kvzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kvzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kvzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kvzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kvzalloc(C1 * C2 * C3, ...) | kvzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kvzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kvzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kvzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kvzalloc(sizeof(THING) * C2, ...) | kvzalloc(sizeof(TYPE) * C2, ...) | kvzalloc(C1 * C2 * C3, ...) | kvzalloc(C1 * C2, ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - (E1) * E2 + E1, E2 , ...) | - kvzalloc + kvcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kvzalloc + kvcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:04:48 +00:00
cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
mm, swap: use kvzalloc to allocate some swap data structures Now vzalloc() is used in swap code to allocate various data structures, such as swap cache, swap slots cache, cluster info, etc. Because the size may be too large on some system, so that normal kzalloc() may fail. But using kzalloc() has some advantages, for example, less memory fragmentation, less TLB pressure, etc. So change the data structure allocation in swap code to use kvzalloc() which will try kzalloc() firstly, and fallback to vzalloc() if kzalloc() failed. In general, although kmalloc() will reduce the number of high-order pages in short term, vmalloc() will cause more pain for memory fragmentation in the long term. And the swap data structure allocation that is changed in this patch is expected to be long term allocation. From Dave Hansen: "for example, we have a two-page data structure. vmalloc() takes two effectively random order-0 pages, probably from two different 2M pages and pins them. That "kills" two 2M pages. kmalloc(), allocating two *contiguous* pages, will not cross a 2M boundary. That means it will only "kill" the possibility of a single 2M page. More 2M pages == less fragmentation. The allocation in this patch occurs during swap on time, which is usually done during system boot, so usually we have high opportunity to allocate the contiguous pages successfully. The allocation for swap_map[] in struct swap_info_struct is not changed, because that is usually quite large and vmalloc_to_page() is used for it. That makes it a little harder to change. Link: http://lkml.kernel.org/r/20170407064911.25447-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 22:57:40 +00:00
GFP_KERNEL);
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
if (!cluster_info) {
error = -ENOMEM;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
}
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
for (ci = 0; ci < nr_cluster; ci++)
spin_lock_init(&((cluster_info + ci)->lock));
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
p->percpu_cluster = alloc_percpu(struct percpu_cluster);
if (!p->percpu_cluster) {
error = -ENOMEM;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
}
for_each_possible_cpu(cpu) {
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
struct percpu_cluster *cluster;
cluster = per_cpu_ptr(p->percpu_cluster, cpu);
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
cluster_set_null(&cluster->index);
}
} else {
atomic_inc(&nr_rotate_swap);
inced_nr_rotate_swap = true;
}
error = swap_cgroup_swapon(p->type, maxpages);
if (error)
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
cluster_info, maxpages, &span);
if (unlikely(nr_extents < 0)) {
error = nr_extents;
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
}
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
/* frontswap enabled? set up bit-per-page map for frontswap */
mm, frontswap: convert frontswap_enabled to static key I have noticed that frontswap.h first declares "frontswap_enabled" as extern bool variable, and then overrides it with "#define frontswap_enabled (1)" for CONFIG_FRONTSWAP=Y or (0) when disabled. The bool variable isn't actually instantiated anywhere. This all looks like an unfinished attempt to make frontswap_enabled reflect whether a backend is instantiated. But in the current state, all frontswap hooks call unconditionally into frontswap.c just to check if frontswap_ops is non-NULL. This should at least be checked inline, but we can further eliminate the overhead when CONFIG_FRONTSWAP is enabled and no backend registered, using a static key that is initially disabled, and gets enabled only upon first backend registration. Thus, checks for "frontswap_enabled" are replaced with "frontswap_enabled()" wrapping the static key check. There are two exceptions: - xen's selfballoon_process() was testing frontswap_enabled in code guarded by #ifdef CONFIG_FRONTSWAP, which was effectively always true when reachable. The patch just removes this check. Using frontswap_enabled() does not sound correct here, as this can be true even without xen's own backend being registered. - in SYSCALL_DEFINE2(swapon), change the check to IS_ENABLED(CONFIG_FRONTSWAP) as it seems the bitmap allocation cannot currently be postponed until a backend is registered. This means that frontswap will still have some memory overhead by being configured, but without a backend. After the patch, we can expect that some functions in frontswap.c are called only when frontswap_ops is non-NULL. Change the checks there to VM_BUG_ONs. While at it, convert other BUG_ONs to VM_BUG_ONs as frontswap has been stable for some time. [akpm@linux-foundation.org: coding-style fixes] Link: http://lkml.kernel.org/r/1463152235-9717-1-git-send-email-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: David Vrabel <david.vrabel@citrix.com> Cc: Juergen Gross <jgross@suse.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-07-26 22:24:42 +00:00
if (IS_ENABLED(CONFIG_FRONTSWAP))
treewide: kvzalloc() -> kvcalloc() The kvzalloc() function has a 2-factor argument form, kvcalloc(). This patch replaces cases of: kvzalloc(a * b, gfp) with: kvcalloc(a * b, gfp) as well as handling cases of: kvzalloc(a * b * c, gfp) with: kvzalloc(array3_size(a, b, c), gfp) as it's slightly less ugly than: kvcalloc(array_size(a, b), c, gfp) This does, however, attempt to ignore constant size factors like: kvzalloc(4 * 1024, gfp) though any constants defined via macros get caught up in the conversion. Any factors with a sizeof() of "unsigned char", "char", and "u8" were dropped, since they're redundant. The Coccinelle script used for this was: // Fix redundant parens around sizeof(). @@ type TYPE; expression THING, E; @@ ( kvzalloc( - (sizeof(TYPE)) * E + sizeof(TYPE) * E , ...) | kvzalloc( - (sizeof(THING)) * E + sizeof(THING) * E , ...) ) // Drop single-byte sizes and redundant parens. @@ expression COUNT; typedef u8; typedef __u8; @@ ( kvzalloc( - sizeof(u8) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(__u8) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(char) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(unsigned char) * (COUNT) + COUNT , ...) | kvzalloc( - sizeof(u8) * COUNT + COUNT , ...) | kvzalloc( - sizeof(__u8) * COUNT + COUNT , ...) | kvzalloc( - sizeof(char) * COUNT + COUNT , ...) | kvzalloc( - sizeof(unsigned char) * COUNT + COUNT , ...) ) // 2-factor product with sizeof(type/expression) and identifier or constant. @@ type TYPE; expression THING; identifier COUNT_ID; constant COUNT_CONST; @@ ( - kvzalloc + kvcalloc ( - sizeof(TYPE) * (COUNT_ID) + COUNT_ID, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * COUNT_ID + COUNT_ID, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * (COUNT_CONST) + COUNT_CONST, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * COUNT_CONST + COUNT_CONST, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (COUNT_ID) + COUNT_ID, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * COUNT_ID + COUNT_ID, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (COUNT_CONST) + COUNT_CONST, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * COUNT_CONST + COUNT_CONST, sizeof(THING) , ...) ) // 2-factor product, only identifiers. @@ identifier SIZE, COUNT; @@ - kvzalloc + kvcalloc ( - SIZE * COUNT + COUNT, SIZE , ...) // 3-factor product with 1 sizeof(type) or sizeof(expression), with // redundant parens removed. @@ expression THING; identifier STRIDE, COUNT; type TYPE; @@ ( kvzalloc( - sizeof(TYPE) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(TYPE) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(TYPE)) , ...) | kvzalloc( - sizeof(THING) * (COUNT) * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * (COUNT) * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * COUNT * (STRIDE) + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) | kvzalloc( - sizeof(THING) * COUNT * STRIDE + array3_size(COUNT, STRIDE, sizeof(THING)) , ...) ) // 3-factor product with 2 sizeof(variable), with redundant parens removed. @@ expression THING1, THING2; identifier COUNT; type TYPE1, TYPE2; @@ ( kvzalloc( - sizeof(TYPE1) * sizeof(TYPE2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(TYPE2)) , ...) | kvzalloc( - sizeof(THING1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(THING1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(THING1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * COUNT + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) | kvzalloc( - sizeof(TYPE1) * sizeof(THING2) * (COUNT) + array3_size(COUNT, sizeof(TYPE1), sizeof(THING2)) , ...) ) // 3-factor product, only identifiers, with redundant parens removed. @@ identifier STRIDE, SIZE, COUNT; @@ ( kvzalloc( - (COUNT) * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * (STRIDE) * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * STRIDE * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - (COUNT) * (STRIDE) * (SIZE) + array3_size(COUNT, STRIDE, SIZE) , ...) | kvzalloc( - COUNT * STRIDE * SIZE + array3_size(COUNT, STRIDE, SIZE) , ...) ) // Any remaining multi-factor products, first at least 3-factor products, // when they're not all constants... @@ expression E1, E2, E3; constant C1, C2, C3; @@ ( kvzalloc(C1 * C2 * C3, ...) | kvzalloc( - (E1) * E2 * E3 + array3_size(E1, E2, E3) , ...) | kvzalloc( - (E1) * (E2) * E3 + array3_size(E1, E2, E3) , ...) | kvzalloc( - (E1) * (E2) * (E3) + array3_size(E1, E2, E3) , ...) | kvzalloc( - E1 * E2 * E3 + array3_size(E1, E2, E3) , ...) ) // And then all remaining 2 factors products when they're not all constants, // keeping sizeof() as the second factor argument. @@ expression THING, E1, E2; type TYPE; constant C1, C2, C3; @@ ( kvzalloc(sizeof(THING) * C2, ...) | kvzalloc(sizeof(TYPE) * C2, ...) | kvzalloc(C1 * C2 * C3, ...) | kvzalloc(C1 * C2, ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * (E2) + E2, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(TYPE) * E2 + E2, sizeof(TYPE) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * (E2) + E2, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - sizeof(THING) * E2 + E2, sizeof(THING) , ...) | - kvzalloc + kvcalloc ( - (E1) * E2 + E1, E2 , ...) | - kvzalloc + kvcalloc ( - (E1) * (E2) + E1, E2 , ...) | - kvzalloc + kvcalloc ( - E1 * E2 + E1, E2 , ...) ) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-06-12 21:04:48 +00:00
frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
sizeof(long),
mm, swap: use kvzalloc to allocate some swap data structures Now vzalloc() is used in swap code to allocate various data structures, such as swap cache, swap slots cache, cluster info, etc. Because the size may be too large on some system, so that normal kzalloc() may fail. But using kzalloc() has some advantages, for example, less memory fragmentation, less TLB pressure, etc. So change the data structure allocation in swap code to use kvzalloc() which will try kzalloc() firstly, and fallback to vzalloc() if kzalloc() failed. In general, although kmalloc() will reduce the number of high-order pages in short term, vmalloc() will cause more pain for memory fragmentation in the long term. And the swap data structure allocation that is changed in this patch is expected to be long term allocation. From Dave Hansen: "for example, we have a two-page data structure. vmalloc() takes two effectively random order-0 pages, probably from two different 2M pages and pins them. That "kills" two 2M pages. kmalloc(), allocating two *contiguous* pages, will not cross a 2M boundary. That means it will only "kill" the possibility of a single 2M page. More 2M pages == less fragmentation. The allocation in this patch occurs during swap on time, which is usually done during system boot, so usually we have high opportunity to allocate the contiguous pages successfully. The allocation for swap_map[] in struct swap_info_struct is not changed, because that is usually quite large and vmalloc_to_page() is used for it. That makes it a little harder to change. Link: http://lkml.kernel.org/r/20170407064911.25447-1-ying.huang@intel.com Signed-off-by: Huang Ying <ying.huang@intel.com> Acked-by: Tim Chen <tim.c.chen@intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Acked-by: Rik van Riel <riel@redhat.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 22:57:40 +00:00
GFP_KERNEL);
if ((swap_flags & SWAP_FLAG_DISCARD) &&
p->bdev && bdev_max_discard_sectors(p->bdev)) {
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* When discard is enabled for swap with no particular
* policy flagged, we set all swap discard flags here in
* order to sustain backward compatibility with older
* swapon(8) releases.
*/
p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
SWP_PAGE_DISCARD);
swap: discard while swapping only if SWAP_FLAG_DISCARD_PAGES Considering the use cases where the swap device supports discard: a) and can do it quickly; b) but it's slow to do in small granularities (or concurrent with other I/O); c) but the implementation is so horrendous that you don't even want to send one down; And assuming that the sysadmin considers it useful to send the discards down at all, we would (probably) want the following solutions: i. do the fine-grained discards for freed swap pages, if device is capable of doing so optimally; ii. do single-time (batched) swap area discards, either at swapon or via something like fstrim (not implemented yet); iii. allow doing both single-time and fine-grained discards; or iv. turn it off completely (default behavior) As implemented today, one can only enable/disable discards for swap, but one cannot select, for instance, solution (ii) on a swap device like (b) even though the single-time discard is regarded to be interesting, or necessary to the workload because it would imply (1), and the device is not capable of performing it optimally. This patch addresses the scenario depicted above by introducing a way to ensure the (probably) wanted solutions (i, ii, iii and iv) can be flexibly flagged through swapon(8) to allow a sysadmin to select the best suitable swap discard policy accordingly to system constraints. This patch introduces SWAP_FLAG_DISCARD_PAGES and SWAP_FLAG_DISCARD_ONCE new flags to allow more flexibe swap discard policies being flagged through swapon(8). The default behavior is to keep both single-time, or batched, area discards (SWAP_FLAG_DISCARD_ONCE) and fine-grained discards for page-clusters (SWAP_FLAG_DISCARD_PAGES) enabled, in order to keep consistentcy with older kernel behavior, as well as maintain compatibility with older swapon(8). However, through the new introduced flags the best suitable discard policy can be selected accordingly to any given swap device constraint. [akpm@linux-foundation.org: tweak comments] Signed-off-by: Rafael Aquini <aquini@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Karel Zak <kzak@redhat.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:02:46 +00:00
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
/*
* By flagging sys_swapon, a sysadmin can tell us to
* either do single-time area discards only, or to just
* perform discards for released swap page-clusters.
* Now it's time to adjust the p->flags accordingly.
*/
if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
p->flags &= ~SWP_PAGE_DISCARD;
else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
p->flags &= ~SWP_AREA_DISCARD;
/* issue a swapon-time discard if it's still required */
if (p->flags & SWP_AREA_DISCARD) {
int err = discard_swap(p);
if (unlikely(err))
pr_err("swapon: discard_swap(%p): %d\n",
p, err);
swap: discard while swapping only if SWAP_FLAG_DISCARD_PAGES Considering the use cases where the swap device supports discard: a) and can do it quickly; b) but it's slow to do in small granularities (or concurrent with other I/O); c) but the implementation is so horrendous that you don't even want to send one down; And assuming that the sysadmin considers it useful to send the discards down at all, we would (probably) want the following solutions: i. do the fine-grained discards for freed swap pages, if device is capable of doing so optimally; ii. do single-time (batched) swap area discards, either at swapon or via something like fstrim (not implemented yet); iii. allow doing both single-time and fine-grained discards; or iv. turn it off completely (default behavior) As implemented today, one can only enable/disable discards for swap, but one cannot select, for instance, solution (ii) on a swap device like (b) even though the single-time discard is regarded to be interesting, or necessary to the workload because it would imply (1), and the device is not capable of performing it optimally. This patch addresses the scenario depicted above by introducing a way to ensure the (probably) wanted solutions (i, ii, iii and iv) can be flexibly flagged through swapon(8) to allow a sysadmin to select the best suitable swap discard policy accordingly to system constraints. This patch introduces SWAP_FLAG_DISCARD_PAGES and SWAP_FLAG_DISCARD_ONCE new flags to allow more flexibe swap discard policies being flagged through swapon(8). The default behavior is to keep both single-time, or batched, area discards (SWAP_FLAG_DISCARD_ONCE) and fine-grained discards for page-clusters (SWAP_FLAG_DISCARD_PAGES) enabled, in order to keep consistentcy with older kernel behavior, as well as maintain compatibility with older swapon(8). However, through the new introduced flags the best suitable discard policy can be selected accordingly to any given swap device constraint. [akpm@linux-foundation.org: tweak comments] Signed-off-by: Rafael Aquini <aquini@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Karel Zak <kzak@redhat.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:02:46 +00:00
}
}
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
error = init_swap_address_space(p->type, maxpages);
if (error)
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
goto bad_swap_unlock_inode;
mm/swap: split swap cache into 64MB trunks The patch is to improve the scalability of the swap out/in via using fine grained locks for the swap cache. In current kernel, one address space will be used for each swap device. And in the common configuration, the number of the swap device is very small (one is typical). This causes the heavy lock contention on the radix tree of the address space if multiple tasks swap out/in concurrently. But in fact, there is no dependency between pages in the swap cache. So that, we can split the one shared address space for each swap device into several address spaces to reduce the lock contention. In the patch, the shared address space is split into 64MB trunks. 64MB is chosen to balance the memory space usage and effect of lock contention reduction. The size of struct address_space on x86_64 architecture is 408B, so with the patch, 6528B more memory will be used for every 1GB swap space on x86_64 architecture. One address space is still shared for the swap entries in the same 64M trunks. To avoid lock contention for the first round of swap space allocation, the order of the swap clusters in the initial free clusters list is changed. The swap space distance between the consecutive swap clusters in the free cluster list is at least 64M. After the first round of allocation, the swap clusters are expected to be freed randomly, so the lock contention should be reduced effectively. Link: http://lkml.kernel.org/r/735bab895e64c930581ffb0a05b661e01da82bc5.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:26 +00:00
/*
* Flush any pending IO and dirty mappings before we start using this
* swap device.
*/
inode->i_flags |= S_SWAPFILE;
error = inode_drain_writes(inode);
if (error) {
inode->i_flags &= ~S_SWAPFILE;
goto free_swap_address_space;
}
mutex_lock(&swapon_mutex);
prio = -1;
if (swap_flags & SWAP_FLAG_PREFER)
prio =
(swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
swap: change block allocation algorithm for SSD I'm using a fast SSD to do swap. scan_swap_map() sometimes uses up to 20~30% CPU time (when cluster is hard to find, the CPU time can be up to 80%), which becomes a bottleneck. scan_swap_map() scans a byte array to search a 256 page cluster, which is very slow. Here I introduced a simple algorithm to search cluster. Since we only care about 256 pages cluster, we can just use a counter to track if a cluster is free. Every 256 pages use one int to store the counter. If the counter of a cluster is 0, the cluster is free. All free clusters will be added to a list, so searching cluster is very efficient. With this, scap_swap_map() overhead disappears. This might help low end SD card swap too. Because if the cluster is aligned, SD firmware can do flash erase more efficiently. We only enable the algorithm for SSD. Hard disk swap isn't fast enough and has downside with the algorithm which might introduce regression (see below). The patch slightly changes which cluster is choosen. It always adds free cluster to list tail. This can help wear leveling for low end SSD too. And if no cluster found, the scan_swap_map() will do search from the end of last cluster. So if no cluster found, the scan_swap_map() will do search from the end of last free cluster, which is random. For SSD, this isn't a problem at all. Another downside is the cluster must be aligned to 256 pages, which will reduce the chance to find a cluster. I would expect this isn't a big problem for SSD because of the non-seek penality. (And this is the reason I only enable the algorithm for SSD). Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:28 +00:00
enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
(p->flags & SWP_SOLIDSTATE) ? "SS" : "",
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
(p->flags & SWP_DISCARDABLE) ? "D" : "",
swap: discard while swapping only if SWAP_FLAG_DISCARD_PAGES Considering the use cases where the swap device supports discard: a) and can do it quickly; b) but it's slow to do in small granularities (or concurrent with other I/O); c) but the implementation is so horrendous that you don't even want to send one down; And assuming that the sysadmin considers it useful to send the discards down at all, we would (probably) want the following solutions: i. do the fine-grained discards for freed swap pages, if device is capable of doing so optimally; ii. do single-time (batched) swap area discards, either at swapon or via something like fstrim (not implemented yet); iii. allow doing both single-time and fine-grained discards; or iv. turn it off completely (default behavior) As implemented today, one can only enable/disable discards for swap, but one cannot select, for instance, solution (ii) on a swap device like (b) even though the single-time discard is regarded to be interesting, or necessary to the workload because it would imply (1), and the device is not capable of performing it optimally. This patch addresses the scenario depicted above by introducing a way to ensure the (probably) wanted solutions (i, ii, iii and iv) can be flexibly flagged through swapon(8) to allow a sysadmin to select the best suitable swap discard policy accordingly to system constraints. This patch introduces SWAP_FLAG_DISCARD_PAGES and SWAP_FLAG_DISCARD_ONCE new flags to allow more flexibe swap discard policies being flagged through swapon(8). The default behavior is to keep both single-time, or batched, area discards (SWAP_FLAG_DISCARD_ONCE) and fine-grained discards for page-clusters (SWAP_FLAG_DISCARD_PAGES) enabled, in order to keep consistentcy with older kernel behavior, as well as maintain compatibility with older swapon(8). However, through the new introduced flags the best suitable discard policy can be selected accordingly to any given swap device constraint. [akpm@linux-foundation.org: tweak comments] Signed-off-by: Rafael Aquini <aquini@redhat.com> Acked-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> Cc: Hugh Dickins <hughd@google.com> Cc: Shaohua Li <shli@kernel.org> Cc: Karel Zak <kzak@redhat.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Mel Gorman <mel@csn.ul.ie> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-03 22:02:46 +00:00
(p->flags & SWP_AREA_DISCARD) ? "s" : "",
(p->flags & SWP_PAGE_DISCARD) ? "c" : "",
mm: frontswap: core swap subsystem hooks and headers This patch, 2of4, contains the changes to the core swap subsystem. This includes: (1) makes available core swap data structures (swap_lock, swap_list and swap_info) that are needed by frontswap.c but we don't need to expose them to the dozens of files that include swap.h so we create a new swapfile.h just to extern-ify these and modify their declarations to non-static (2) adds frontswap-related elements to swap_info_struct. Frontswap_map points to vzalloc'ed one-bit-per-swap-page metadata that indicates whether the swap page is in frontswap or in the device and frontswap_pages counts how many pages are in frontswap. (3) adds hooks in the swap subsystem and extends try_to_unuse so that frontswap_shrink can do a "partial swapoff". Note that a failed frontswap_map allocation is safe... failure is noted by lack of "FS" in the subsequent printk. --- [v14: rebase to 3.4-rc2] [v10: no change] [v9: akpm@linux-foundation.org: mark some statics __read_mostly] [v9: akpm@linux-foundation.org: add clarifying comments] [v9: akpm@linux-foundation.org: no need to loop repeating try_to_unuse] [v9: error27@gmail.com: remove superfluous check for NULL] [v8: rebase to 3.0-rc4] [v8: kamezawa.hiroyu@jp.fujitsu.com: change counter to atomic_t to avoid races] [v8: kamezawa.hiroyu@jp.fujitsu.com: comment to clarify informational counters] [v7: rebase to 3.0-rc3] [v7: JBeulich@novell.com: add new swap struct elements only if config'd] [v6: rebase to 3.0-rc1] [v6: lliubbo@gmail.com: fix null pointer deref if vzalloc fails] [v6: konrad.wilk@oracl.com: various checks and code clarifications/comments] [v5: no change from v4] [v4: rebase to 2.6.39] Signed-off-by: Dan Magenheimer <dan.magenheimer@oracle.com> Reviewed-by: Kamezawa Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Acked-by: Jan Beulich <JBeulich@novell.com> Acked-by: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Jeremy Fitzhardinge <jeremy@goop.org> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Nitin Gupta <ngupta@vflare.org> Cc: Matthew Wilcox <matthew@wil.cx> Cc: Chris Mason <chris.mason@oracle.com> Cc: Rik Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> [v11: Rebased, fixed mm/swapfile.c context change] Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
2012-04-09 23:08:06 +00:00
(frontswap_map) ? "FS" : "");
mutex_unlock(&swapon_mutex);
atomic_inc(&proc_poll_event);
wake_up_interruptible(&proc_poll_wait);
error = 0;
goto out;
free_swap_address_space:
exit_swap_address_space(p->type);
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
bad_swap_unlock_inode:
inode_unlock(inode);
bad_swap:
swap: make cluster allocation per-cpu swap cluster allocation is to get better request merge to improve performance. But the cluster is shared globally, if multiple tasks are doing swap, this will cause interleave disk access. While multiple tasks swap is quite common, for example, each numa node has a kswapd thread doing swap and multiple threads/processes doing direct page reclaim. ioscheduler can't help too much here, because tasks don't send swapout IO down to block layer in the meantime. Block layer does merge some IOs, but a lot not, depending on how many tasks are doing swapout concurrently. In practice, I've seen a lot of small size IO in swapout workloads. We makes the cluster allocation per-cpu here. The interleave disk access issue goes away. All tasks swapout to their own cluster, so swapout will become sequential, which can be easily merged to big size IO. If one CPU can't get its per-cpu cluster (for example, there is no free cluster anymore in the swap), it will fallback to scan swap_map. The CPU can still continue swap. We don't need recycle free swap entries of other CPUs. In my test (swap to a 2-disk raid0 partition), this improves around 10% swapout throughput, and request size is increased significantly. How does this impact swap readahead is uncertain though. On one side, page reclaim always isolates and swaps several adjancent pages, this will make page reclaim write the pages sequentially and benefit readahead. On the other side, several CPU write pages interleave means the pages don't live _sequentially_ but relatively _near_. In the per-cpu allocation case, if adjancent pages are written by different cpus, they will live relatively _far_. So how this impacts swap readahead depends on how many pages page reclaim isolates and swaps one time. If the number is big, this patch will benefit swap readahead. Of course, this is about sequential access pattern. The patch has no impact for random access pattern, because the new cluster allocation algorithm is just for SSD. Alternative solution is organizing swap layout to be per-mm instead of this per-cpu approach. In the per-mm layout, we allocate a disk range for each mm, so pages of one mm live in swap disk adjacently. per-mm layout has potential issues of lock contention if multiple reclaimers are swap pages from one mm. For a sequential workload, per-mm layout is better to implement swap readahead, because pages from the mm are adjacent in disk. But per-cpu layout isn't very bad in this workload, as page reclaim always isolates and swaps several pages one time, such pages will still live in disk sequentially and readahead can utilize this. For a random workload, per-mm layout isn't beneficial of request merge, because it's quite possible pages from different mm are swapout in the meantime and IO can't be merged in per-mm layout. while with per-cpu layout we can merge requests from any mm. Considering random workload is more popular in workloads with swap (and per-cpu approach isn't too bad for sequential workload too), I'm choosing per-cpu layout. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:32 +00:00
free_percpu(p->percpu_cluster);
p->percpu_cluster = NULL;
swap: reduce lock contention on swap cache from swap slots allocation In some swap scalability test, it is found that there are heavy lock contention on swap cache even if we have split one swap cache radix tree per swap device to one swap cache radix tree every 64 MB trunk in commit 4b3ef9daa4fc ("mm/swap: split swap cache into 64MB trunks"). The reason is as follow. After the swap device becomes fragmented so that there's no free swap cluster, the swap device will be scanned linearly to find the free swap slots. swap_info_struct->cluster_next is the next scanning base that is shared by all CPUs. So nearby free swap slots will be allocated for different CPUs. The probability for multiple CPUs to operate on the same 64 MB trunk is high. This causes the lock contention on the swap cache. To solve the issue, in this patch, for SSD swap device, a percpu version next scanning base (cluster_next_cpu) is added. Every CPU will use its own per-cpu next scanning base. And after finishing scanning a 64MB trunk, the per-cpu scanning base will be changed to the beginning of another randomly selected 64MB trunk. In this way, the probability for multiple CPUs to operate on the same 64 MB trunk is reduced greatly. Thus the lock contention is reduced too. For HDD, because sequential access is more important for IO performance, the original shared next scanning base is used. To test the patch, we have run 16-process pmbench memory benchmark on a 2-socket server machine with 48 cores. One ram disk is configured as the swap device per socket. The pmbench working-set size is much larger than the available memory so that swapping is triggered. The memory read/write ratio is 80/20 and the accessing pattern is random. In the original implementation, the lock contention on the swap cache is heavy. The perf profiling data of the lock contention code path is as following, _raw_spin_lock_irq.add_to_swap_cache.add_to_swap.shrink_page_list: 7.91 _raw_spin_lock_irqsave.__remove_mapping.shrink_page_list: 7.11 _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 2.51 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 1.66 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 1.29 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.03 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 0.93 After applying this patch, it becomes, _raw_spin_lock.swapcache_free_entries.free_swap_slot.__swap_entry_free: 3.58 _raw_spin_lock_irq.shrink_inactive_list.shrink_lruvec.shrink_node: 2.3 _raw_spin_lock_irqsave.swap_cgroup_record.mem_cgroup_uncharge_swap: 2.26 _raw_spin_lock_irq.shrink_active_list.shrink_lruvec.shrink_node: 1.8 _raw_spin_lock.free_pcppages_bulk.drain_pages_zone.drain_pages: 1.19 The lock contention on the swap cache is almost eliminated. And the pmbench score increases 18.5%. The swapin throughput increases 18.7% from 2.96 GB/s to 3.51 GB/s. While the swapout throughput increases 18.5% from 2.99 GB/s to 3.54 GB/s. We need really fast disk to show the benefit. I have tried this on 2 Intel P3600 NVMe disks. The performance improvement is only about 1%. The improvement should be better on the faster disks, such as Intel Optane disk. [ying.huang@intel.com: fix cluster_next_cpu allocation and freeing, per Daniel] Link: http://lkml.kernel.org/r/20200525002648.336325-1-ying.huang@intel.com [ying.huang@intel.com: v4] Link: http://lkml.kernel.org/r/20200529010840.928819-1-ying.huang@intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/20200520031502.175659-1-ying.huang@intel.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-02 04:49:22 +00:00
free_percpu(p->cluster_next_cpu);
p->cluster_next_cpu = NULL;
if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
set_blocksize(p->bdev, p->old_block_size);
blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
}
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
inode = NULL;
destroy_swap_extents(p);
swap_cgroup_swapoff(p->type);
spin_lock(&swap_lock);
p->swap_file = NULL;
p->flags = 0;
spin_unlock(&swap_lock);
vfree(swap_map);
kvfree(cluster_info);
kvfree(frontswap_map);
if (inced_nr_rotate_swap)
atomic_dec(&nr_rotate_swap);
mm/swapfile.c: move inode_lock out of claim_swapfile claim_swapfile() currently keeps the inode locked when it is successful, or the file is already swapfile (with -EBUSY). And, on the other error cases, it does not lock the inode. This inconsistency of the lock state and return value is quite confusing and actually causing a bad unlock balance as below in the "bad_swap" section of __do_sys_swapon(). This commit fixes this issue by moving the inode_lock() and IS_SWAPFILE check out of claim_swapfile(). The inode is unlocked in "bad_swap_unlock_inode" section, so that the inode is ensured to be unlocked at "bad_swap". Thus, error handling codes after the locking now jumps to "bad_swap_unlock_inode" instead of "bad_swap". ===================================== WARNING: bad unlock balance detected! 5.5.0-rc7+ #176 Not tainted ------------------------------------- swapon/4294 is trying to release lock (&sb->s_type->i_mutex_key) at: __do_sys_swapon+0x94b/0x3550 but there are no more locks to release! other info that might help us debug this: no locks held by swapon/4294. stack backtrace: CPU: 5 PID: 4294 Comm: swapon Not tainted 5.5.0-rc7-BTRFS-ZNS+ #176 Hardware name: ASUS All Series/H87-PRO, BIOS 2102 07/29/2014 Call Trace: dump_stack+0xa1/0xea print_unlock_imbalance_bug.cold+0x114/0x123 lock_release+0x562/0xed0 up_write+0x2d/0x490 __do_sys_swapon+0x94b/0x3550 __x64_sys_swapon+0x54/0x80 do_syscall_64+0xa4/0x4b0 entry_SYSCALL_64_after_hwframe+0x49/0xbe RIP: 0033:0x7f15da0a0dc7 Fixes: 1638045c3677 ("mm: set S_SWAPFILE on blockdev swap devices") Signed-off-by: Naohiro Aota <naohiro.aota@wdc.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Tested-by: Qais Youef <qais.yousef@arm.com> Reviewed-by: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: <stable@vger.kernel.org> Link: http://lkml.kernel.org/r/20200206090132.154869-1-naohiro.aota@wdc.com Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-03-29 02:17:15 +00:00
if (swap_file)
filp_close(swap_file, NULL);
out:
if (page && !IS_ERR(page)) {
kunmap(page);
mm, fs: get rid of PAGE_CACHE_* and page_cache_{get,release} macros PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time ago with promise that one day it will be possible to implement page cache with bigger chunks than PAGE_SIZE. This promise never materialized. And unlikely will. We have many places where PAGE_CACHE_SIZE assumed to be equal to PAGE_SIZE. And it's constant source of confusion on whether PAGE_CACHE_* or PAGE_* constant should be used in a particular case, especially on the border between fs and mm. Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much breakage to be doable. Let's stop pretending that pages in page cache are special. They are not. The changes are pretty straight-forward: - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>; - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN}; - page_cache_get() -> get_page(); - page_cache_release() -> put_page(); This patch contains automated changes generated with coccinelle using script below. For some reason, coccinelle doesn't patch header files. I've called spatch for them manually. The only adjustment after coccinelle is revert of changes to PAGE_CAHCE_ALIGN definition: we are going to drop it later. There are few places in the code where coccinelle didn't reach. I'll fix them manually in a separate patch. Comments and documentation also will be addressed with the separate patch. virtual patch @@ expression E; @@ - E << (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ expression E; @@ - E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) + E @@ @@ - PAGE_CACHE_SHIFT + PAGE_SHIFT @@ @@ - PAGE_CACHE_SIZE + PAGE_SIZE @@ @@ - PAGE_CACHE_MASK + PAGE_MASK @@ expression E; @@ - PAGE_CACHE_ALIGN(E) + PAGE_ALIGN(E) @@ expression E; @@ - page_cache_get(E) + get_page(E) @@ expression E; @@ - page_cache_release(E) + put_page(E) Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Michal Hocko <mhocko@suse.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-04-01 12:29:47 +00:00
put_page(page);
}
if (name)
putname(name);
if (inode)
inode_unlock(inode);
if (!error)
enable_swap_slots_cache();
return error;
}
void si_swapinfo(struct sysinfo *val)
{
unsigned int type;
unsigned long nr_to_be_unused = 0;
spin_lock(&swap_lock);
for (type = 0; type < nr_swapfiles; type++) {
struct swap_info_struct *si = swap_info[type];
if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
nr_to_be_unused += READ_ONCE(si->inuse_pages);
}
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
val->totalswap = total_swap_pages + nr_to_be_unused;
spin_unlock(&swap_lock);
}
/*
* Verify that a swap entry is valid and increment its swap map count.
*
* Returns error code in following case.
* - success -> 0
* - swp_entry is invalid -> EINVAL
* - swp_entry is migration entry -> EINVAL
* - swap-cache reference is requested but there is already one. -> EEXIST
* - swap-cache reference is requested but the entry is not used. -> ENOENT
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
* - swap-mapped reference requested but needs continued swap count. -> ENOMEM
*/
static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
{
struct swap_info_struct *p;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci;
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
unsigned long offset;
unsigned char count;
unsigned char has_cache;
int err;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
p = get_swap_device(entry);
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
if (!p)
return -EINVAL;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
offset = swp_offset(entry);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
ci = lock_cluster_or_swap_info(p, offset);
count = p->swap_map[offset];
swap: fix races exposed by swap discard The previous patch can expose races, according to Hugh: swapoff was sometimes failing with "Cannot allocate memory", coming from try_to_unuse()'s -ENOMEM: it needs to allow for swap_duplicate() failing on a free entry temporarily SWAP_MAP_BAD while being discarded. We should use ACCESS_ONCE() there, and whenever accessing swap_map locklessly; but rather than peppering it throughout try_to_unuse(), just declare *swap_map with volatile. try_to_unuse() is accustomed to *swap_map going down racily, but not necessarily to it jumping up from 0 to SWAP_MAP_BAD: we'll be safer to prevent that transition once SWP_WRITEOK is switched off, when it's a waste of time to issue discards anyway (swapon can do a whole discard). Another issue is: In swapin_readahead(), read_swap_cache_async() can read a bad swap entry, because we don't check if readahead swap entry is bad. This doesn't break anything but such swapin page is wasteful and can only be freed at page reclaim. We should avoid read such swap entry. And in discard, we mark swap entry SWAP_MAP_BAD and then switch it to normal when discard is finished. If readahead reads such swap entry, we have the same issue, so we much check if swap entry is bad too. Thanks Hugh to inspire swapin_readahead could use bad swap entry. [include Hugh's patch 'swap: fix swapoff ENOMEMs from discard'] Signed-off-by: Shaohua Li <shli@fusionio.com> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Kyungmin Park <kmpark@infradead.org> Cc: Rafael Aquini <aquini@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11 21:20:31 +00:00
/*
* swapin_readahead() doesn't check if a swap entry is valid, so the
* swap entry could be SWAP_MAP_BAD. Check here with lock held.
*/
if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
err = -ENOENT;
goto unlock_out;
}
has_cache = count & SWAP_HAS_CACHE;
count &= ~SWAP_HAS_CACHE;
err = 0;
if (usage == SWAP_HAS_CACHE) {
/* set SWAP_HAS_CACHE if there is no cache and entry is used */
if (!has_cache && count)
has_cache = SWAP_HAS_CACHE;
else if (has_cache) /* someone else added cache */
err = -EEXIST;
else /* no users remaining */
err = -ENOENT;
} else if (count || has_cache) {
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
count += usage;
else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
err = -EINVAL;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
else if (swap_count_continued(p, offset, count))
count = COUNT_CONTINUED;
else
err = -ENOMEM;
} else
err = -ENOENT; /* unused swap entry */
mm/swapfile: fix and annotate various data races swap_info_struct si.highest_bit, si.swap_map[offset] and si.flags could be accessed concurrently separately as noticed by KCSAN, === si.highest_bit === write to 0xffff8d5abccdc4d4 of 4 bytes by task 5353 on cpu 24: swap_range_alloc+0x81/0x130 swap_range_alloc at mm/swapfile.c:681 scan_swap_map_slots+0x371/0xb90 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff8d5abccdc4d4 of 4 bytes by task 6672 on cpu 70: scan_swap_map_slots+0x4a6/0xb90 scan_swap_map_slots at mm/swapfile.c:892 get_swap_pages+0x39d/0x5c0 get_swap_page+0xf2/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 Reported by Kernel Concurrency Sanitizer on: CPU: 70 PID: 6672 Comm: oom01 Tainted: G W L 5.5.0-next-20200205+ #3 Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019 === si.swap_map[offset] === write to 0xffffbc370c29a64c of 1 bytes by task 6856 on cpu 86: __swap_entry_free_locked+0x8c/0x100 __swap_entry_free_locked at mm/swapfile.c:1209 (discriminator 4) __swap_entry_free.constprop.20+0x69/0xb0 free_swap_and_cache+0x53/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 read to 0xffffbc370c29a64c of 1 bytes by task 6855 on cpu 20: _swap_info_get+0x81/0xa0 _swap_info_get at mm/swapfile.c:1140 free_swap_and_cache+0x40/0xa0 unmap_page_range+0x7f8/0x1d70 unmap_single_vma+0xcd/0x170 unmap_vmas+0x18b/0x220 exit_mmap+0xee/0x220 mmput+0x10e/0x270 do_exit+0x59b/0xf40 do_group_exit+0x8b/0x180 === si.flags === write to 0xffff956c8fc6c400 of 8 bytes by task 6087 on cpu 23: scan_swap_map_slots+0x6fe/0xb50 scan_swap_map_slots at mm/swapfile.c:887 get_swap_pages+0x39d/0x5c0 get_swap_page+0x377/0x524 add_to_swap+0xe4/0x1c0 shrink_page_list+0x1795/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 read to 0xffff956c8fc6c400 of 8 bytes by task 6207 on cpu 63: _swap_info_get+0x41/0xa0 __swap_info_get at mm/swapfile.c:1114 put_swap_page+0x84/0x490 __remove_mapping+0x384/0x5f0 shrink_page_list+0xff1/0x2870 shrink_inactive_list+0x316/0x880 shrink_lruvec+0x8dc/0x1380 shrink_node+0x317/0xd80 do_try_to_free_pages+0x1f7/0xa10 try_to_free_pages+0x26c/0x5e0 __alloc_pages_slowpath+0x458/0x1290 The writes are under si->lock but the reads are not. For si.highest_bit and si.swap_map[offset], data race could trigger logic bugs, so fix them by having WRITE_ONCE() for the writes and READ_ONCE() for the reads except those isolated reads where they compare against zero which a data race would cause no harm. Thus, annotate them as intentional data races using the data_race() macro. For si.flags, the readers are only interested in a single bit where a data race there would cause no issue there. [cai@lca.pw: add a missing annotation for si->flags in memory.c] Link: http://lkml.kernel.org/r/1581612647-5958-1-git-send-email-cai@lca.pw Signed-off-by: Qian Cai <cai@lca.pw> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Marco Elver <elver@google.com> Cc: Hugh Dickins <hughd@google.com> Link: http://lkml.kernel.org/r/1581095163-12198-1-git-send-email-cai@lca.pw Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-15 00:31:31 +00:00
WRITE_ONCE(p->swap_map[offset], count | has_cache);
unlock_out:
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster_or_swap_info(p, ci);
put_swap_device(p);
return err;
}
/*
* Help swapoff by noting that swap entry belongs to shmem/tmpfs
* (in which case its reference count is never incremented).
*/
void swap_shmem_alloc(swp_entry_t entry)
{
__swap_duplicate(entry, SWAP_MAP_SHMEM);
}
/*
* Increase reference count of swap entry by 1.
* Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
* but could not be atomically allocated. Returns 0, just as if it succeeded,
* if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
* might occur if a page table entry has got corrupted.
*/
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
int swap_duplicate(swp_entry_t entry)
{
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
int err = 0;
while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
err = add_swap_count_continuation(entry, GFP_ATOMIC);
return err;
}
/*
* @entry: swap entry for which we allocate swap cache.
*
* Called when allocating swap cache for existing swap entry,
* This can return error codes. Returns 0 at success.
* -EEXIST means there is a swap cache.
* Note: return code is different from swap_duplicate().
*/
int swapcache_prepare(swp_entry_t entry)
{
return __swap_duplicate(entry, SWAP_HAS_CACHE);
}
struct swap_info_struct *swp_swap_info(swp_entry_t entry)
{
mm, swap: bounds check swap_info array accesses to avoid NULL derefs Dan Carpenter reports a potential NULL dereference in get_swap_page_of_type: Smatch complains that the NULL checks on "si" aren't consistent. This seems like a real bug because we have not ensured that the type is valid and so "si" can be NULL. Add the missing check for NULL, taking care to use a read barrier to ensure CPU1 observes CPU0's updates in the correct order: CPU0 CPU1 alloc_swap_info() if (type >= nr_swapfiles) swap_info[type] = p /* handle invalid entry */ smp_wmb() smp_rmb() ++nr_swapfiles p = swap_info[type] Without smp_rmb, CPU1 might observe CPU0's write to nr_swapfiles before CPU0's write to swap_info[type] and read NULL from swap_info[type]. Ying Huang noticed other places in swapfile.c don't order these reads properly. Introduce swap_type_to_swap_info to encourage correct usage. Use READ_ONCE and WRITE_ONCE to follow the Linux Kernel Memory Model (see tools/memory-model/Documentation/explanation.txt). This ordering need not be enforced in places where swap_lock is held (e.g. si_swapinfo) because swap_lock serializes updates to nr_swapfiles and the swap_info array. Link: http://lkml.kernel.org/r/20190131024410.29859-1-daniel.m.jordan@oracle.com Fixes: ec8acf20afb8 ("swap: add per-partition lock for swapfile") Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com> Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Suggested-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Andi Kleen <ak@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Paul McKenney <paulmck@linux.vnet.ibm.com> Cc: Shaohua Li <shli@kernel.org> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Tejun Heo <tj@kernel.org> Cc: Will Deacon <will.deacon@arm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-03-05 23:48:19 +00:00
return swap_type_to_swap_info(swp_type(entry));
}
struct swap_info_struct *page_swap_info(struct page *page)
{
swp_entry_t entry = { .val = page_private(page) };
return swp_swap_info(entry);
}
/*
mm/util: Add folio_mapping() and folio_file_mapping() These are the folio equivalent of page_mapping() and page_file_mapping(). Add an out-of-line page_mapping() wrapper around folio_mapping() in order to prevent the page_folio() call from bloating every caller of page_mapping(). Adjust page_file_mapping() and page_mapping_file() to use folios internally. Rename __page_file_mapping() to swapcache_mapping() and change it to take a folio. This ends up saving 122 bytes of text overall. folio_mapping() is 45 bytes shorter than page_mapping() was, but the new page_mapping() wrapper is 30 bytes. The major reduction is a few bytes less in dozens of nfs functions (which call page_file_mapping()). Most of these appear to be a slight change in gcc's register allocation decisions, which allow: 48 8b 56 08 mov 0x8(%rsi),%rdx 48 8d 42 ff lea -0x1(%rdx),%rax 83 e2 01 and $0x1,%edx 48 0f 44 c6 cmove %rsi,%rax to become: 48 8b 46 08 mov 0x8(%rsi),%rax 48 8d 78 ff lea -0x1(%rax),%rdi a8 01 test $0x1,%al 48 0f 44 fe cmove %rsi,%rdi for a reduction of a single byte. Once the NFS client is converted to use folios, this entire sequence will disappear. Also add folio_mapping() documentation. Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Jeff Layton <jlayton@kernel.org> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: William Kucharski <william.kucharski@oracle.com> Reviewed-by: David Howells <dhowells@redhat.com>
2020-12-10 15:55:05 +00:00
* out-of-line methods to avoid include hell.
*/
mm/util: Add folio_mapping() and folio_file_mapping() These are the folio equivalent of page_mapping() and page_file_mapping(). Add an out-of-line page_mapping() wrapper around folio_mapping() in order to prevent the page_folio() call from bloating every caller of page_mapping(). Adjust page_file_mapping() and page_mapping_file() to use folios internally. Rename __page_file_mapping() to swapcache_mapping() and change it to take a folio. This ends up saving 122 bytes of text overall. folio_mapping() is 45 bytes shorter than page_mapping() was, but the new page_mapping() wrapper is 30 bytes. The major reduction is a few bytes less in dozens of nfs functions (which call page_file_mapping()). Most of these appear to be a slight change in gcc's register allocation decisions, which allow: 48 8b 56 08 mov 0x8(%rsi),%rdx 48 8d 42 ff lea -0x1(%rdx),%rax 83 e2 01 and $0x1,%edx 48 0f 44 c6 cmove %rsi,%rax to become: 48 8b 46 08 mov 0x8(%rsi),%rax 48 8d 78 ff lea -0x1(%rax),%rdi a8 01 test $0x1,%al 48 0f 44 fe cmove %rsi,%rdi for a reduction of a single byte. Once the NFS client is converted to use folios, this entire sequence will disappear. Also add folio_mapping() documentation. Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Jeff Layton <jlayton@kernel.org> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: William Kucharski <william.kucharski@oracle.com> Reviewed-by: David Howells <dhowells@redhat.com>
2020-12-10 15:55:05 +00:00
struct address_space *swapcache_mapping(struct folio *folio)
{
mm/util: Add folio_mapping() and folio_file_mapping() These are the folio equivalent of page_mapping() and page_file_mapping(). Add an out-of-line page_mapping() wrapper around folio_mapping() in order to prevent the page_folio() call from bloating every caller of page_mapping(). Adjust page_file_mapping() and page_mapping_file() to use folios internally. Rename __page_file_mapping() to swapcache_mapping() and change it to take a folio. This ends up saving 122 bytes of text overall. folio_mapping() is 45 bytes shorter than page_mapping() was, but the new page_mapping() wrapper is 30 bytes. The major reduction is a few bytes less in dozens of nfs functions (which call page_file_mapping()). Most of these appear to be a slight change in gcc's register allocation decisions, which allow: 48 8b 56 08 mov 0x8(%rsi),%rdx 48 8d 42 ff lea -0x1(%rdx),%rax 83 e2 01 and $0x1,%edx 48 0f 44 c6 cmove %rsi,%rax to become: 48 8b 46 08 mov 0x8(%rsi),%rax 48 8d 78 ff lea -0x1(%rax),%rdi a8 01 test $0x1,%al 48 0f 44 fe cmove %rsi,%rdi for a reduction of a single byte. Once the NFS client is converted to use folios, this entire sequence will disappear. Also add folio_mapping() documentation. Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Jeff Layton <jlayton@kernel.org> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: William Kucharski <william.kucharski@oracle.com> Reviewed-by: David Howells <dhowells@redhat.com>
2020-12-10 15:55:05 +00:00
return page_swap_info(&folio->page)->swap_file->f_mapping;
}
mm/util: Add folio_mapping() and folio_file_mapping() These are the folio equivalent of page_mapping() and page_file_mapping(). Add an out-of-line page_mapping() wrapper around folio_mapping() in order to prevent the page_folio() call from bloating every caller of page_mapping(). Adjust page_file_mapping() and page_mapping_file() to use folios internally. Rename __page_file_mapping() to swapcache_mapping() and change it to take a folio. This ends up saving 122 bytes of text overall. folio_mapping() is 45 bytes shorter than page_mapping() was, but the new page_mapping() wrapper is 30 bytes. The major reduction is a few bytes less in dozens of nfs functions (which call page_file_mapping()). Most of these appear to be a slight change in gcc's register allocation decisions, which allow: 48 8b 56 08 mov 0x8(%rsi),%rdx 48 8d 42 ff lea -0x1(%rdx),%rax 83 e2 01 and $0x1,%edx 48 0f 44 c6 cmove %rsi,%rax to become: 48 8b 46 08 mov 0x8(%rsi),%rax 48 8d 78 ff lea -0x1(%rax),%rdi a8 01 test $0x1,%al 48 0f 44 fe cmove %rsi,%rdi for a reduction of a single byte. Once the NFS client is converted to use folios, this entire sequence will disappear. Also add folio_mapping() documentation. Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Jeff Layton <jlayton@kernel.org> Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: William Kucharski <william.kucharski@oracle.com> Reviewed-by: David Howells <dhowells@redhat.com>
2020-12-10 15:55:05 +00:00
EXPORT_SYMBOL_GPL(swapcache_mapping);
pgoff_t __page_file_index(struct page *page)
{
swp_entry_t swap = { .val = page_private(page) };
return swp_offset(swap);
}
EXPORT_SYMBOL_GPL(__page_file_index);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
/*
* add_swap_count_continuation - called when a swap count is duplicated
* beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
* page of the original vmalloc'ed swap_map, to hold the continuation count
* (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
* again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
*
* These continuation pages are seldom referenced: the common paths all work
* on the original swap_map, only referring to a continuation page when the
* low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
*
* add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
* page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
* can be called after dropping locks.
*/
int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
{
struct swap_info_struct *si;
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
struct swap_cluster_info *ci;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
struct page *head;
struct page *page;
struct page *list_page;
pgoff_t offset;
unsigned char count;
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
int ret = 0;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
/*
* When debugging, it's easier to use __GFP_ZERO here; but it's better
* for latency not to zero a page while GFP_ATOMIC and holding locks.
*/
page = alloc_page(gfp_mask | __GFP_HIGHMEM);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
si = get_swap_device(entry);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
if (!si) {
/*
* An acceptable race has occurred since the failing
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
* __swap_duplicate(): the swap device may be swapoff
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
*/
goto outer;
}
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
spin_lock(&si->lock);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
offset = swp_offset(entry);
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
ci = lock_cluster(si, offset);
count = swap_count(si->swap_map[offset]);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
/*
* The higher the swap count, the more likely it is that tasks
* will race to add swap count continuation: we need to avoid
* over-provisioning.
*/
goto out;
}
if (!page) {
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
ret = -ENOMEM;
goto out;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
/*
* We are fortunate that although vmalloc_to_page uses pte_offset_map,
* no architecture is using highmem pages for kernel page tables: so it
* will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
*/
head = vmalloc_to_page(si->swap_map + offset);
offset &= ~PAGE_MASK;
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
spin_lock(&si->cont_lock);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
/*
* Page allocation does not initialize the page's lru field,
* but it does always reset its private field.
*/
if (!page_private(head)) {
BUG_ON(count & COUNT_CONTINUED);
INIT_LIST_HEAD(&head->lru);
set_page_private(head, SWP_CONTINUED);
si->flags |= SWP_CONTINUED;
}
list_for_each_entry(list_page, &head->lru, lru) {
unsigned char *map;
/*
* If the previous map said no continuation, but we've found
* a continuation page, free our allocation and use this one.
*/
if (!(count & COUNT_CONTINUED))
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
goto out_unlock_cont;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
map = kmap_atomic(list_page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
count = *map;
kunmap_atomic(map);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
/*
* If this continuation count now has some space in it,
* free our allocation and use this one.
*/
if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
goto out_unlock_cont;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
list_add_tail(&page->lru, &head->lru);
page = NULL; /* now it's attached, don't free it */
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
out_unlock_cont:
spin_unlock(&si->cont_lock);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
out:
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
unlock_cluster(ci);
swap: add per-partition lock for swapfile swap_lock is heavily contended when I test swap to 3 fast SSD (even slightly slower than swap to 2 such SSD). The main contention comes from swap_info_get(). This patch tries to fix the gap with adding a new per-partition lock. Global data like nr_swapfiles, total_swap_pages, least_priority and swap_list are still protected by swap_lock. nr_swap_pages is an atomic now, it can be changed without swap_lock. In theory, it's possible get_swap_page() finds no swap pages but actually there are free swap pages. But sounds not a big problem. Accessing partition specific data (like scan_swap_map and so on) is only protected by swap_info_struct.lock. Changing swap_info_struct.flags need hold swap_lock and swap_info_struct.lock, because scan_scan_map() will check it. read the flags is ok with either the locks hold. If both swap_lock and swap_info_struct.lock must be hold, we always hold the former first to avoid deadlock. swap_entry_free() can change swap_list. To delete that code, we add a new highest_priority_index. Whenever get_swap_page() is called, we check it. If it's valid, we use it. It's a pity get_swap_page() still holds swap_lock(). But in practice, swap_lock() isn't heavily contended in my test with this patch (or I can say there are other much more heavier bottlenecks like TLB flush). And BTW, looks get_swap_page() doesn't really need the lock. We never free swap_info[] and we check SWAP_WRITEOK flag. The only risk without the lock is we could swapout to some low priority swap, but we can quickly recover after several rounds of swap, so sounds not a big deal to me. But I'd prefer to fix this if it's a real problem. "swap: make each swap partition have one address_space" improved the swapout speed from 1.7G/s to 2G/s. This patch further improves the speed to 2.3G/s, so around 15% improvement. It's a multi-process test, so TLB flush isn't the biggest bottleneck before the patches. [arnd@arndb.de: fix it for nommu] [hughd@google.com: add missing unlock] [minchan@kernel.org: get rid of lockdep whinge on sys_swapon] Signed-off-by: Shaohua Li <shli@fusionio.com> Cc: Hugh Dickins <hughd@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Minchan Kim <minchan.kim@gmail.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Seth Jennings <sjenning@linux.vnet.ibm.com> Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Cc: Dan Magenheimer <dan.magenheimer@oracle.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Hugh Dickins <hughd@google.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-02-23 00:34:38 +00:00
spin_unlock(&si->lock);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
put_swap_device(si);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
outer:
if (page)
__free_page(page);
mm, swap: fix race between swapoff and some swap operations When swapin is performed, after getting the swap entry information from the page table, system will swap in the swap entry, without any lock held to prevent the swap device from being swapoff. This may cause the race like below, CPU 1 CPU 2 ----- ----- do_swap_page swapin_readahead __read_swap_cache_async swapoff swapcache_prepare p->swap_map = NULL __swap_duplicate p->swap_map[?] /* !!! NULL pointer access */ Because swapoff is usually done when system shutdown only, the race may not hit many people in practice. But it is still a race need to be fixed. To fix the race, get_swap_device() is added to check whether the specified swap entry is valid in its swap device. If so, it will keep the swap entry valid via preventing the swap device from being swapoff, until put_swap_device() is called. Because swapoff() is very rare code path, to make the normal path runs as fast as possible, rcu_read_lock/unlock() and synchronize_rcu() instead of reference count is used to implement get/put_swap_device(). >From get_swap_device() to put_swap_device(), RCU reader side is locked, so synchronize_rcu() in swapoff() will wait until put_swap_device() is called. In addition to swap_map, cluster_info, etc. data structure in the struct swap_info_struct, the swap cache radix tree will be freed after swapoff, so this patch fixes the race between swap cache looking up and swapoff too. Races between some other swap cache usages and swapoff are fixed too via calling synchronize_rcu() between clearing PageSwapCache() and freeing swap cache data structure. Another possible method to fix this is to use preempt_off() + stop_machine() to prevent the swap device from being swapoff when its data structure is being accessed. The overhead in hot-path of both methods is similar. The advantages of RCU based method are, 1. stop_machine() may disturb the normal execution code path on other CPUs. 2. File cache uses RCU to protect its radix tree. If the similar mechanism is used for swap cache too, it is easier to share code between them. 3. RCU is used to protect swap cache in total_swapcache_pages() and exit_swap_address_space() already. The two mechanisms can be merged to simplify the logic. Link: http://lkml.kernel.org/r/20190522015423.14418-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Reviewed-by: Andrea Parri <andrea.parri@amarulasolutions.com> Not-nacked-by: Hugh Dickins <hughd@google.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Daniel Jordan <daniel.m.jordan@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Tim Chen <tim.c.chen@linux.intel.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Jérôme Glisse <jglisse@redhat.com> Cc: Yang Shi <yang.shi@linux.alibaba.com> Cc: David Rientjes <rientjes@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jan Kara <jack@suse.cz> Cc: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-07-12 03:55:33 +00:00
return ret;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
/*
* swap_count_continued - when the original swap_map count is incremented
* from SWAP_MAP_MAX, check if there is already a continuation page to carry
* into, carry if so, or else fail until a new continuation page is allocated;
* when the original swap_map count is decremented from 0 with continuation,
* borrow from the continuation and report whether it still holds more.
mm/swap: add cluster lock This patch is to reduce the lock contention of swap_info_struct->lock via using a more fine grained lock in swap_cluster_info for some swap operations. swap_info_struct->lock is heavily contended if multiple processes reclaim pages simultaneously. Because there is only one lock for each swap device. While in common configuration, there is only one or several swap devices in the system. The lock protects almost all swap related operations. In fact, many swap operations only access one element of swap_info_struct->swap_map array. And there is no dependency between different elements of swap_info_struct->swap_map. So a fine grained lock can be used to allow parallel access to the different elements of swap_info_struct->swap_map. In this patch, a spinlock is added to swap_cluster_info to protect the elements of swap_info_struct->swap_map in the swap cluster and the fields of swap_cluster_info. This reduced locking contention for swap_info_struct->swap_map access greatly. Because of the added spinlock, the size of swap_cluster_info increases from 4 bytes to 8 bytes on the 64 bit and 32 bit system. This will use additional 4k RAM for every 1G swap space. Because the size of swap_cluster_info is much smaller than the size of the cache line (8 vs 64 on x86_64 architecture), there may be false cache line sharing between spinlocks in swap_cluster_info. To avoid the false sharing in the first round of the swap cluster allocation, the order of the swap clusters in the free clusters list is changed. So that, the swap_cluster_info sharing the same cache line will be placed as far as possible. After the first round of allocation, the order of the clusters in free clusters list is expected to be random. So the false sharing should be not serious. Compared with a previous implementation using bit_spin_lock, the sequential swap out throughput improved about 3.2%. Test was done on a Xeon E5 v3 system. The swap device used is a RAM simulated PMEM (persistent memory) device. To test the sequential swapping out, the test case created 32 processes, which sequentially allocate and write to the anonymous pages until the RAM and part of the swap device is used. [ying.huang@intel.com: v5] Link: http://lkml.kernel.org/r/878tqeuuic.fsf_-_@yhuang-dev.intel.com [minchan@kernel.org: initialize spinlock for swap_cluster_info] Link: http://lkml.kernel.org/r/1486434945-29753-1-git-send-email-minchan@kernel.org [hughd@google.com: annotate nested locking for cluster lock] Link: http://lkml.kernel.org/r/alpine.LSU.2.11.1702161050540.21773@eggly.anvils Link: http://lkml.kernel.org/r/dbb860bbd825b1aaba18988015e8963f263c3f0d.1484082593.git.tim.c.chen@linux.intel.com Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com> Signed-off-by: Minchan Kim <minchan@kernel.org> Signed-off-by: Hugh Dickins <hughd@google.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Christian Borntraeger <borntraeger@de.ibm.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Hillf Danton <hillf.zj@alibaba-inc.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Jonathan Corbet <corbet@lwn.net> escreveu: Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Shaohua Li <shli@kernel.org> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:45:22 +00:00
* Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
* lock.
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
*/
static bool swap_count_continued(struct swap_info_struct *si,
pgoff_t offset, unsigned char count)
{
struct page *head;
struct page *page;
unsigned char *map;
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
bool ret;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
head = vmalloc_to_page(si->swap_map + offset);
if (page_private(head) != SWP_CONTINUED) {
BUG_ON(count & COUNT_CONTINUED);
return false; /* need to add count continuation */
}
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
spin_lock(&si->cont_lock);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
offset &= ~PAGE_MASK;
page = list_next_entry(head, lru);
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
goto init_map; /* jump over SWAP_CONT_MAX checks */
if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
/*
* Think of how you add 1 to 999
*/
while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
kunmap_atomic(map);
page = list_next_entry(page, lru);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
BUG_ON(page == head);
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
if (*map == SWAP_CONT_MAX) {
kunmap_atomic(map);
page = list_next_entry(page, lru);
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
if (page == head) {
ret = false; /* add count continuation */
goto out;
}
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
init_map: *map = 0; /* we didn't zero the page */
}
*map += 1;
kunmap_atomic(map);
while ((page = list_prev_entry(page, lru)) != head) {
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
*map = COUNT_CONTINUED;
kunmap_atomic(map);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
ret = true; /* incremented */
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
} else { /* decrementing */
/*
* Think of how you subtract 1 from 1000
*/
BUG_ON(count != COUNT_CONTINUED);
while (*map == COUNT_CONTINUED) {
kunmap_atomic(map);
page = list_next_entry(page, lru);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
BUG_ON(page == head);
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
BUG_ON(*map == 0);
*map -= 1;
if (*map == 0)
count = 0;
kunmap_atomic(map);
while ((page = list_prev_entry(page, lru)) != head) {
map = kmap_atomic(page) + offset;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
*map = SWAP_CONT_MAX | count;
count = COUNT_CONTINUED;
kunmap_atomic(map);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
ret = count == COUNT_CONTINUED;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
mm, swap: fix race between swap count continuation operations One page may store a set of entries of the sis->swap_map (swap_info_struct->swap_map) in multiple swap clusters. If some of the entries has sis->swap_map[offset] > SWAP_MAP_MAX, multiple pages will be used to store the set of entries of the sis->swap_map. And the pages are linked with page->lru. This is called swap count continuation. To access the pages which store the set of entries of the sis->swap_map simultaneously, previously, sis->lock is used. But to improve the scalability of __swap_duplicate(), swap cluster lock may be used in swap_count_continued() now. This may race with add_swap_count_continuation() which operates on a nearby swap cluster, in which the sis->swap_map entries are stored in the same page. The race can cause wrong swap count in practice, thus cause unfreeable swap entries or software lockup, etc. To fix the race, a new spin lock called cont_lock is added to struct swap_info_struct to protect the swap count continuation page list. This is a lock at the swap device level, so the scalability isn't very well. But it is still much better than the original sis->lock, because it is only acquired/released when swap count continuation is used. Which is considered rare in practice. If it turns out that the scalability becomes an issue for some workloads, we can split the lock into some more fine grained locks. Link: http://lkml.kernel.org/r/20171017081320.28133-1-ying.huang@intel.com Fixes: 235b62176712 ("mm/swap: add cluster lock") Signed-off-by: "Huang, Ying" <ying.huang@intel.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Shaohua Li <shli@kernel.org> Cc: Tim Chen <tim.c.chen@intel.com> Cc: Michal Hocko <mhocko@suse.com> Cc: Aaron Lu <aaron.lu@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <ak@linux.intel.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Cc: <stable@vger.kernel.org> [4.11+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-11-02 22:59:50 +00:00
out:
spin_unlock(&si->cont_lock);
return ret;
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
}
/*
* free_swap_count_continuations - swapoff free all the continuation pages
* appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
*/
static void free_swap_count_continuations(struct swap_info_struct *si)
{
pgoff_t offset;
for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
struct page *head;
head = vmalloc_to_page(si->swap_map + offset);
if (page_private(head)) {
struct page *page, *next;
list_for_each_entry_safe(page, next, &head->lru, lru) {
list_del(&page->lru);
swap_info: swap count continuations Swap is duplicated (reference count incremented by one) whenever the same swap page is inserted into another mm (when forking finds a swap entry in place of a pte, or when reclaim unmaps a pte to insert the swap entry). swap_info_struct's vmalloc'ed swap_map is the array of these reference counts: but what happens when the unsigned short (or unsigned char since the preceding patch) is full? (and its high bit is kept for a cache flag) We then lose track of it, never freeing, leaving it in use until swapoff: at which point we _hope_ that a single pass will have found all instances, assume there are no more, and will lose user data if we're wrong. Swapping of KSM pages has not yet been enabled; but it is implemented, and makes it very easy for a user to overflow the maximum swap count: possible with ordinary process pages, but unlikely, even when pid_max has been raised from PID_MAX_DEFAULT. This patch implements swap count continuations: when the count overflows, a continuation page is allocated and linked to the original vmalloc'ed map page, and this used to hold the continuation counts for that entry and its neighbours. These continuation pages are seldom referenced: the common paths all work on the original swap_map, only referring to a continuation page when the low "digit" of a count is incremented or decremented through SWAP_MAP_MAX. Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 01:58:46 +00:00
__free_page(page);
}
}
}
}
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
mm, memcg: inline swap-related functions to improve disabled memcg config Inline mem_cgroup_try_charge_swap, mem_cgroup_uncharge_swap and cgroup_throttle_swaprate functions to perform mem_cgroup_disabled static key check inline before calling the main body of the function. This minimizes the memcg overhead in the pagefault and exit_mmap paths when memcgs are disabled using cgroup_disable=memory command-line option. This change results in ~1% overhead reduction when running PFT test [1] comparing {CONFIG_MEMCG=n} against {CONFIG_MEMCG=y, cgroup_disable=memory} configuration on an 8-core ARM64 Android device. [1] https://lkml.org/lkml/2006/8/29/294 also used in mmtests suite Link: https://lkml.kernel.org/r/20210713010934.299876-3-surenb@google.com Signed-off-by: Suren Baghdasaryan <surenb@google.com> Reviewed-by: Shakeel Butt <shakeelb@google.com> Reviewed-by: Muchun Song <songmuchun@bytedance.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Michal Hocko <mhocko@suse.com> Cc: Tejun Heo <tj@kernel.org> Cc: Roman Gushchin <guro@fb.com> Cc: Yang Shi <shy828301@gmail.com> Cc: Alex Shi <alexs@kernel.org> Cc: Wei Yang <richard.weiyang@gmail.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: David Hildenbrand <david@redhat.com> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Cc: Alistair Popple <apopple@nvidia.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Miaohe Lin <linmiaohe@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-02 21:54:54 +00:00
void __cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
{
struct swap_info_struct *si, *next;
int nid = page_to_nid(page);
if (!(gfp_mask & __GFP_IO))
return;
if (!blk_cgroup_congested())
return;
/*
* We've already scheduled a throttle, avoid taking the global swap
* lock.
*/
if (current->throttle_disk)
return;
spin_lock(&swap_avail_lock);
plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
avail_lists[nid]) {
if (si->bdev) {
blkcg_schedule_throttle(si->bdev->bd_disk, true);
break;
}
}
spin_unlock(&swap_avail_lock);
}
#endif
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
static int __init swapfile_init(void)
{
int nid;
swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
GFP_KERNEL);
if (!swap_avail_heads) {
pr_emerg("Not enough memory for swap heads, swap is disabled\n");
return -ENOMEM;
}
for_each_node(nid)
plist_head_init(&swap_avail_heads[nid]);
mm/swap: cache maximum swapfile size when init swap We used to have swapfile_maximum_size() fetching a maximum value of swapfile size per-arch. As the caller of max_swapfile_size() grows, this patch introduce a variable "swapfile_maximum_size" and cache the value of old max_swapfile_size(), so that we don't need to calculate the value every time. Caching the value in swapfile_init() is safe because when reaching the phase we should have initialized all the relevant information. Here the major arch to take care of is x86, which defines the max swapfile size based on L1TF mitigation. Here both X86_BUG_L1TF or l1tf_mitigation should have been setup properly when reaching swapfile_init(). As a reference, the code path looks like this for x86: - start_kernel - setup_arch - early_cpu_init - early_identify_cpu --> setup X86_BUG_L1TF - parse_early_param - l1tf_cmdline --> set l1tf_mitigation - check_bugs - l1tf_select_mitigation --> set l1tf_mitigation - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) The swapfile size only depends on swp pte format on non-x86 archs, so caching it is safe too. Since at it, rename max_swapfile_size() to arch_max_swapfile_size() because arch can define its own function, so it's more straightforward to have "arch_" as its prefix. At the meantime, export swapfile_maximum_size to replace the old usages of max_swapfile_size(). [peterx@redhat.com: declare arch_max_swapfile_size) in swapfile.h] Link: https://lkml.kernel.org/r/YxTh1GuC6ro5fKL5@xz-m1.local Link: https://lkml.kernel.org/r/20220811161331.37055-7-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:30 +00:00
swapfile_maximum_size = arch_max_swapfile_size();
mm/swap: cache swap migration A/D bits support Introduce a variable swap_migration_ad_supported to cache whether the arch supports swap migration A/D bits. Here one thing to mention is that SWP_MIG_TOTAL_BITS will internally reference the other macro MAX_PHYSMEM_BITS, which is a function call on x86 (constant on all the rest of archs). It's safe to reference it in swapfile_init() because when reaching here we're already during initcalls level 4 so we must have initialized 5-level pgtable for x86_64 (right after early_identify_cpu() finishes). - start_kernel - setup_arch - early_cpu_init - get_cpu_cap --> fetch from CPUID (including X86_FEATURE_LA57) - early_identify_cpu --> clear X86_FEATURE_LA57 (if early lvl5 not enabled (USE_EARLY_PGTABLE_L5)) - arch_call_rest_init - rest_init - kernel_init - kernel_init_freeable - do_basic_setup - do_initcalls --> calls swapfile_init() (initcall level 4) This should slightly speed up the migration swap entry handlings. Link: https://lkml.kernel.org/r/20220811161331.37055-8-peterx@redhat.com Signed-off-by: Peter Xu <peterx@redhat.com> Cc: Alistair Popple <apopple@nvidia.com> Cc: Andi Kleen <andi.kleen@intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: David Hildenbrand <david@redhat.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Hugh Dickins <hughd@google.com> Cc: "Kirill A . Shutemov" <kirill@shutemov.name> Cc: Minchan Kim <minchan@kernel.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-08-11 16:13:31 +00:00
#ifdef CONFIG_MIGRATION
if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
swap_migration_ad_supported = true;
#endif /* CONFIG_MIGRATION */
swap: choose swap device according to numa node If the system has more than one swap device and swap device has the node information, we can make use of this information to decide which swap device to use in get_swap_pages() to get better performance. The current code uses a priority based list, swap_avail_list, to decide which swap device to use and if multiple swap devices share the same priority, they are used round robin. This patch changes the previous single global swap_avail_list into a per-numa-node list, i.e. for each numa node, it sees its own priority based list of available swap devices. Swap device's priority can be promoted on its matching node's swap_avail_list. The current swap device's priority is set as: user can set a >=0 value, or the system will pick one starting from -1 then downwards. The priority value in the swap_avail_list is the negated value of the swap device's due to plist being sorted from low to high. The new policy doesn't change the semantics for priority >=0 cases, the previous starting from -1 then downwards now becomes starting from -2 then downwards and -1 is reserved as the promoted value. Take 4-node EX machine as an example, suppose 4 swap devices are available, each sit on a different node: swapA on node 0 swapB on node 1 swapC on node 2 swapD on node 3 After they are all swapped on in the sequence of ABCD. Current behaviour: their priorities will be: swapA: -1 swapB: -2 swapC: -3 swapD: -4 And their position in the global swap_avail_list will be: swapA -> swapB -> swapC -> swapD prio:1 prio:2 prio:3 prio:4 New behaviour: their priorities will be(note that -1 is skipped): swapA: -2 swapB: -3 swapC: -4 swapD: -5 And their positions in the 4 swap_avail_lists[nid] will be: swap_avail_lists[0]: /* node 0's available swap device list */ swapA -> swapB -> swapC -> swapD prio:1 prio:3 prio:4 prio:5 swap_avali_lists[1]: /* node 1's available swap device list */ swapB -> swapA -> swapC -> swapD prio:1 prio:2 prio:4 prio:5 swap_avail_lists[2]: /* node 2's available swap device list */ swapC -> swapA -> swapB -> swapD prio:1 prio:2 prio:3 prio:5 swap_avail_lists[3]: /* node 3's available swap device list */ swapD -> swapA -> swapB -> swapC prio:1 prio:2 prio:3 prio:4 To see the effect of the patch, a test that starts N process, each mmap a region of anonymous memory and then continually write to it at random position to trigger both swap in and out is used. On a 2 node Skylake EP machine with 64GiB memory, two 170GB SSD drives are used as swap devices with each attached to a different node, the result is: runtime=30m/processes=32/total test size=128G/each process mmap region=4G kernel throughput vanilla 13306 auto-binding 15169 +14% runtime=30m/processes=64/total test size=128G/each process mmap region=2G kernel throughput vanilla 11885 auto-binding 14879 +25% [aaron.lu@intel.com: v2] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com [akpm@linux-foundation.org: use kmalloc_array()] Link: http://lkml.kernel.org/r/20170814053130.GD2369@aaronlu.sh.intel.com Link: http://lkml.kernel.org/r/20170816024439.GA10925@aaronlu.sh.intel.com Signed-off-by: Aaron Lu <aaron.lu@intel.com> Cc: "Chen, Tim C" <tim.c.chen@intel.com> Cc: Huang Ying <ying.huang@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Hugh Dickins <hughd@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-06 23:24:57 +00:00
return 0;
}
subsys_initcall(swapfile_init);