linux/fs/dax.c

1473 lines
40 KiB
C
Raw Normal View History

/*
* fs/dax.c - Direct Access filesystem code
* Copyright (c) 2013-2014 Intel Corporation
* Author: Matthew Wilcox <matthew.r.wilcox@intel.com>
* Author: Ross Zwisler <ross.zwisler@linux.intel.com>
*
* This program is free software; you can redistribute it and/or modify it
* under the terms and conditions of the GNU General Public License,
* version 2, as published by the Free Software Foundation.
*
* This program is distributed in the hope it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
* more details.
*/
#include <linux/atomic.h>
#include <linux/blkdev.h>
#include <linux/buffer_head.h>
#include <linux/dax.h>
#include <linux/fs.h>
#include <linux/genhd.h>
#include <linux/highmem.h>
#include <linux/memcontrol.h>
#include <linux/mm.h>
#include <linux/mutex.h>
#include <linux/pagevec.h>
#include <linux/sched.h>
#include <linux/sched/signal.h>
#include <linux/uio.h>
#include <linux/vmstat.h>
#include <linux/pfn_t.h>
#include <linux/sizes.h>
#include <linux/mmu_notifier.h>
#include <linux/iomap.h>
#include "internal.h"
dax: add tracepoint infrastructure, PMD tracing Tracepoints are the standard way to capture debugging and tracing information in many parts of the kernel, including the XFS and ext4 filesystems. Create a tracepoint header for FS DAX and add the first DAX tracepoints to the PMD fault handler. This allows the tracing for DAX to be done in the same way as the filesystem tracing so that developers can look at them together and get a coherent idea of what the system is doing. I added both an entry and exit tracepoint because future patches will add tracepoints to child functions of dax_iomap_pmd_fault() like dax_pmd_load_hole() and dax_pmd_insert_mapping(). We want those messages to be wrapped by the parent function tracepoints so the code flow is more easily understood. Having entry and exit tracepoints for faults also allows us to easily see what filesystems functions were called during the fault. These filesystem functions get executed via iomap_begin() and iomap_end() calls, for example, and will have their own tracepoints. For PMD faults we primarily want to understand the type of mapping, the fault flags, the faulting address and whether it fell back to 4k faults. If it fell back to 4k faults the tracepoints should let us understand why. I named the new tracepoint header file "fs_dax.h" to allow for device DAX to have its own separate tracing header in the same directory at some point. Here is an example output for these events from a successful PMD fault: big-1441 [005] .... 32.582758: xfs_filemap_pmd_fault: dev 259:0 ino 0x1003 big-1441 [005] .... 32.582776: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 big-1441 [005] .... 32.583292: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 NOPAGE Link: http://lkml.kernel.org/r/1484085142-2297-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dave Chinner <david@fromorbit.com> Reviewed-by: Jan Kara <jack@suse.cz> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:39:50 +00:00
#define CREATE_TRACE_POINTS
#include <trace/events/fs_dax.h>
/* We choose 4096 entries - same as per-zone page wait tables */
#define DAX_WAIT_TABLE_BITS 12
#define DAX_WAIT_TABLE_ENTRIES (1 << DAX_WAIT_TABLE_BITS)
static wait_queue_head_t wait_table[DAX_WAIT_TABLE_ENTRIES];
static int __init init_dax_wait_table(void)
{
int i;
for (i = 0; i < DAX_WAIT_TABLE_ENTRIES; i++)
init_waitqueue_head(wait_table + i);
return 0;
}
fs_initcall(init_dax_wait_table);
/*
* We use lowest available bit in exceptional entry for locking, one bit for
* the entry size (PMD) and two more to tell us if the entry is a zero page or
* an empty entry that is just used for locking. In total four special bits.
*
* If the PMD bit isn't set the entry has size PAGE_SIZE, and if the ZERO_PAGE
* and EMPTY bits aren't set the entry is a normal DAX entry with a filesystem
* block allocation.
*/
#define RADIX_DAX_SHIFT (RADIX_TREE_EXCEPTIONAL_SHIFT + 4)
#define RADIX_DAX_ENTRY_LOCK (1 << RADIX_TREE_EXCEPTIONAL_SHIFT)
#define RADIX_DAX_PMD (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 1))
#define RADIX_DAX_ZERO_PAGE (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 2))
#define RADIX_DAX_EMPTY (1 << (RADIX_TREE_EXCEPTIONAL_SHIFT + 3))
static unsigned long dax_radix_sector(void *entry)
{
return (unsigned long)entry >> RADIX_DAX_SHIFT;
}
static void *dax_radix_locked_entry(sector_t sector, unsigned long flags)
{
return (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY | flags |
((unsigned long)sector << RADIX_DAX_SHIFT) |
RADIX_DAX_ENTRY_LOCK);
}
static unsigned int dax_radix_order(void *entry)
{
if ((unsigned long)entry & RADIX_DAX_PMD)
return PMD_SHIFT - PAGE_SHIFT;
return 0;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
static int dax_is_pmd_entry(void *entry)
{
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return (unsigned long)entry & RADIX_DAX_PMD;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
static int dax_is_pte_entry(void *entry)
{
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return !((unsigned long)entry & RADIX_DAX_PMD);
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
static int dax_is_zero_entry(void *entry)
{
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
return (unsigned long)entry & RADIX_DAX_ZERO_PAGE;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
static int dax_is_empty_entry(void *entry)
{
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return (unsigned long)entry & RADIX_DAX_EMPTY;
}
/*
* DAX radix tree locking
*/
struct exceptional_entry_key {
struct address_space *mapping;
pgoff_t entry_start;
};
struct wait_exceptional_entry_queue {
wait_queue_entry_t wait;
struct exceptional_entry_key key;
};
static wait_queue_head_t *dax_entry_waitqueue(struct address_space *mapping,
pgoff_t index, void *entry, struct exceptional_entry_key *key)
{
unsigned long hash;
/*
* If 'entry' is a PMD, align the 'index' that we use for the wait
* queue to the start of that PMD. This ensures that all offsets in
* the range covered by the PMD map to the same bit lock.
*/
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (dax_is_pmd_entry(entry))
index &= ~((1UL << (PMD_SHIFT - PAGE_SHIFT)) - 1);
key->mapping = mapping;
key->entry_start = index;
hash = hash_long((unsigned long)mapping ^ index, DAX_WAIT_TABLE_BITS);
return wait_table + hash;
}
static int wake_exceptional_entry_func(wait_queue_entry_t *wait, unsigned int mode,
int sync, void *keyp)
{
struct exceptional_entry_key *key = keyp;
struct wait_exceptional_entry_queue *ewait =
container_of(wait, struct wait_exceptional_entry_queue, wait);
if (key->mapping != ewait->key.mapping ||
key->entry_start != ewait->key.entry_start)
return 0;
return autoremove_wake_function(wait, mode, sync, NULL);
}
/*
* We do not necessarily hold the mapping->tree_lock when we call this
* function so it is possible that 'entry' is no longer a valid item in the
* radix tree. This is okay because all we really need to do is to find the
* correct waitqueue where tasks might be waiting for that old 'entry' and
* wake them.
*/
static void dax_wake_mapping_entry_waiter(struct address_space *mapping,
pgoff_t index, void *entry, bool wake_all)
{
struct exceptional_entry_key key;
wait_queue_head_t *wq;
wq = dax_entry_waitqueue(mapping, index, entry, &key);
/*
* Checking for locked entry and prepare_to_wait_exclusive() happens
* under mapping->tree_lock, ditto for entry handling in our callers.
* So at this point all tasks that could have seen our entry locked
* must be in the waitqueue and the following check will see them.
*/
if (waitqueue_active(wq))
__wake_up(wq, TASK_NORMAL, wake_all ? 0 : 1, &key);
}
/*
* Check whether the given slot is locked. The function must be called with
* mapping->tree_lock held
*/
static inline int slot_locked(struct address_space *mapping, void **slot)
{
unsigned long entry = (unsigned long)
radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
return entry & RADIX_DAX_ENTRY_LOCK;
}
/*
* Mark the given slot is locked. The function must be called with
* mapping->tree_lock held
*/
static inline void *lock_slot(struct address_space *mapping, void **slot)
{
unsigned long entry = (unsigned long)
radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
entry |= RADIX_DAX_ENTRY_LOCK;
radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
return (void *)entry;
}
/*
* Mark the given slot is unlocked. The function must be called with
* mapping->tree_lock held
*/
static inline void *unlock_slot(struct address_space *mapping, void **slot)
{
unsigned long entry = (unsigned long)
radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
entry &= ~(unsigned long)RADIX_DAX_ENTRY_LOCK;
radix_tree_replace_slot(&mapping->page_tree, slot, (void *)entry);
return (void *)entry;
}
/*
* Lookup entry in radix tree, wait for it to become unlocked if it is
* exceptional entry and return it. The caller must call
* put_unlocked_mapping_entry() when he decided not to lock the entry or
* put_locked_mapping_entry() when he locked the entry and now wants to
* unlock it.
*
* The function must be called with mapping->tree_lock held.
*/
static void *get_unlocked_mapping_entry(struct address_space *mapping,
pgoff_t index, void ***slotp)
{
void *entry, **slot;
struct wait_exceptional_entry_queue ewait;
wait_queue_head_t *wq;
init_wait(&ewait.wait);
ewait.wait.func = wake_exceptional_entry_func;
for (;;) {
entry = __radix_tree_lookup(&mapping->page_tree, index, NULL,
&slot);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (!entry ||
WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)) ||
!slot_locked(mapping, slot)) {
if (slotp)
*slotp = slot;
return entry;
}
wq = dax_entry_waitqueue(mapping, index, entry, &ewait.key);
prepare_to_wait_exclusive(wq, &ewait.wait,
TASK_UNINTERRUPTIBLE);
spin_unlock_irq(&mapping->tree_lock);
schedule();
finish_wait(wq, &ewait.wait);
spin_lock_irq(&mapping->tree_lock);
}
}
static void dax_unlock_mapping_entry(struct address_space *mapping,
pgoff_t index)
{
void *entry, **slot;
spin_lock_irq(&mapping->tree_lock);
entry = __radix_tree_lookup(&mapping->page_tree, index, NULL, &slot);
if (WARN_ON_ONCE(!entry || !radix_tree_exceptional_entry(entry) ||
!slot_locked(mapping, slot))) {
spin_unlock_irq(&mapping->tree_lock);
return;
}
unlock_slot(mapping, slot);
spin_unlock_irq(&mapping->tree_lock);
dax_wake_mapping_entry_waiter(mapping, index, entry, false);
}
static void put_locked_mapping_entry(struct address_space *mapping,
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
pgoff_t index)
{
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
dax_unlock_mapping_entry(mapping, index);
}
/*
* Called when we are done with radix tree entry we looked up via
* get_unlocked_mapping_entry() and which we didn't lock in the end.
*/
static void put_unlocked_mapping_entry(struct address_space *mapping,
pgoff_t index, void *entry)
{
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (!entry)
return;
/* We have to wake up next waiter for the radix tree entry lock */
dax_wake_mapping_entry_waiter(mapping, index, entry, false);
}
/*
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
* Find radix tree entry at given index. If it points to an exceptional entry,
* return it with the radix tree entry locked. If the radix tree doesn't
* contain given index, create an empty exceptional entry for the index and
* return with it locked.
*
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
* When requesting an entry with size RADIX_DAX_PMD, grab_mapping_entry() will
* either return that locked entry or will return an error. This error will
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
* happen if there are any 4k entries within the 2MiB range that we are
* requesting.
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
*
* We always favor 4k entries over 2MiB entries. There isn't a flow where we
* evict 4k entries in order to 'upgrade' them to a 2MiB entry. A 2MiB
* insertion will fail if it finds any 4k entries already in the tree, and a
* 4k insertion will cause an existing 2MiB entry to be unmapped and
* downgraded to 4k entries. This happens for both 2MiB huge zero pages as
* well as 2MiB empty entries.
*
* The exception to this downgrade path is for 2MiB DAX PMD entries that have
* real storage backing them. We will leave these real 2MiB DAX entries in
* the tree, and PTE writes will simply dirty the entire 2MiB DAX entry.
*
* Note: Unlike filemap_fault() we don't honor FAULT_FLAG_RETRY flags. For
* persistent memory the benefit is doubtful. We can add that later if we can
* show it helps.
*/
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
static void *grab_mapping_entry(struct address_space *mapping, pgoff_t index,
unsigned long size_flag)
{
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
bool pmd_downgrade = false; /* splitting 2MiB entry into 4k entries? */
void *entry, **slot;
restart:
spin_lock_irq(&mapping->tree_lock);
entry = get_unlocked_mapping_entry(mapping, index, &slot);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (WARN_ON_ONCE(entry && !radix_tree_exceptional_entry(entry))) {
entry = ERR_PTR(-EIO);
goto out_unlock;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (entry) {
if (size_flag & RADIX_DAX_PMD) {
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (dax_is_pte_entry(entry)) {
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
put_unlocked_mapping_entry(mapping, index,
entry);
entry = ERR_PTR(-EEXIST);
goto out_unlock;
}
} else { /* trying to grab a PTE entry */
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (dax_is_pmd_entry(entry) &&
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
(dax_is_zero_entry(entry) ||
dax_is_empty_entry(entry))) {
pmd_downgrade = true;
}
}
}
/* No entry for given index? Make sure radix tree is big enough. */
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (!entry || pmd_downgrade) {
int err;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (pmd_downgrade) {
/*
* Make sure 'entry' remains valid while we drop
* mapping->tree_lock.
*/
entry = lock_slot(mapping, slot);
}
spin_unlock_irq(&mapping->tree_lock);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
* Besides huge zero pages the only other thing that gets
* downgraded are empty entries which don't need to be
* unmapped.
*/
if (pmd_downgrade && dax_is_zero_entry(entry))
unmap_mapping_range(mapping,
(index << PAGE_SHIFT) & PMD_MASK, PMD_SIZE, 0);
err = radix_tree_preload(
mapping_gfp_mask(mapping) & ~__GFP_HIGHMEM);
if (err) {
if (pmd_downgrade)
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
put_locked_mapping_entry(mapping, index);
return ERR_PTR(err);
}
spin_lock_irq(&mapping->tree_lock);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
dax: fix radix tree insertion race While running generic/340 in my test setup I hit the following race. It can happen with kernels that support FS DAX PMDs, so v4.10 thru v4.11-rc5. Thread 1 Thread 2 -------- -------- dax_iomap_pmd_fault() grab_mapping_entry() spin_lock_irq() get_unlocked_mapping_entry() 'entry' is NULL, can't call lock_slot() spin_unlock_irq() radix_tree_preload() dax_iomap_pmd_fault() grab_mapping_entry() spin_lock_irq() get_unlocked_mapping_entry() ... lock_slot() spin_unlock_irq() dax_pmd_insert_mapping() <inserts a PMD mapping> spin_lock_irq() __radix_tree_insert() fails with -EEXIST <fall back to 4k fault, and die horribly when inserting a 4k entry where a PMD exists> The issue is that we have to drop mapping->tree_lock while calling radix_tree_preload(), but since we didn't have a radix tree entry to lock (unlike in the pmd_downgrade case) we have no protection against Thread 2 coming along and inserting a PMD at the same index. For 4k entries we handled this with a special-case response to -EEXIST coming from the __radix_tree_insert(), but this doesn't save us for PMDs because the -EEXIST case can also mean that we collided with a 4k entry in the radix tree at a different index, but one that is covered by our PMD range. So, correctly handle both the 4k and 2M collision cases by explicitly re-checking the radix tree for an entry at our index once we reacquire mapping->tree_lock. This patch has made it through a clean xfstests run with the current v4.11-rc5 based linux/master, and it also ran generic/340 500 times in a loop. It used to fail within the first 10 iterations. Link: http://lkml.kernel.org/r/20170406212944.2866-1-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: <stable@vger.kernel.org> [4.10+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-04-07 23:04:57 +00:00
if (!entry) {
/*
* We needed to drop the page_tree lock while calling
* radix_tree_preload() and we didn't have an entry to
* lock. See if another thread inserted an entry at
* our index during this time.
*/
entry = __radix_tree_lookup(&mapping->page_tree, index,
NULL, &slot);
if (entry) {
radix_tree_preload_end();
spin_unlock_irq(&mapping->tree_lock);
goto restart;
}
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (pmd_downgrade) {
radix_tree_delete(&mapping->page_tree, index);
mapping->nrexceptional--;
dax_wake_mapping_entry_waiter(mapping, index, entry,
true);
}
entry = dax_radix_locked_entry(0, size_flag | RADIX_DAX_EMPTY);
err = __radix_tree_insert(&mapping->page_tree, index,
dax_radix_order(entry), entry);
radix_tree_preload_end();
if (err) {
spin_unlock_irq(&mapping->tree_lock);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
dax: fix radix tree insertion race While running generic/340 in my test setup I hit the following race. It can happen with kernels that support FS DAX PMDs, so v4.10 thru v4.11-rc5. Thread 1 Thread 2 -------- -------- dax_iomap_pmd_fault() grab_mapping_entry() spin_lock_irq() get_unlocked_mapping_entry() 'entry' is NULL, can't call lock_slot() spin_unlock_irq() radix_tree_preload() dax_iomap_pmd_fault() grab_mapping_entry() spin_lock_irq() get_unlocked_mapping_entry() ... lock_slot() spin_unlock_irq() dax_pmd_insert_mapping() <inserts a PMD mapping> spin_lock_irq() __radix_tree_insert() fails with -EEXIST <fall back to 4k fault, and die horribly when inserting a 4k entry where a PMD exists> The issue is that we have to drop mapping->tree_lock while calling radix_tree_preload(), but since we didn't have a radix tree entry to lock (unlike in the pmd_downgrade case) we have no protection against Thread 2 coming along and inserting a PMD at the same index. For 4k entries we handled this with a special-case response to -EEXIST coming from the __radix_tree_insert(), but this doesn't save us for PMDs because the -EEXIST case can also mean that we collided with a 4k entry in the radix tree at a different index, but one that is covered by our PMD range. So, correctly handle both the 4k and 2M collision cases by explicitly re-checking the radix tree for an entry at our index once we reacquire mapping->tree_lock. This patch has made it through a clean xfstests run with the current v4.11-rc5 based linux/master, and it also ran generic/340 500 times in a loop. It used to fail within the first 10 iterations. Link: http://lkml.kernel.org/r/20170406212944.2866-1-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: <stable@vger.kernel.org> [4.10+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-04-07 23:04:57 +00:00
* Our insertion of a DAX entry failed, most likely
* because we were inserting a PMD entry and it
* collided with a PTE sized entry at a different
* index in the PMD range. We haven't inserted
* anything into the radix tree and have no waiters to
* wake.
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
*/
return ERR_PTR(err);
}
/* Good, we have inserted empty locked entry into the tree. */
mapping->nrexceptional++;
spin_unlock_irq(&mapping->tree_lock);
return entry;
}
entry = lock_slot(mapping, slot);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
out_unlock:
spin_unlock_irq(&mapping->tree_lock);
return entry;
}
static int __dax_invalidate_mapping_entry(struct address_space *mapping,
pgoff_t index, bool trunc)
{
int ret = 0;
void *entry;
struct radix_tree_root *page_tree = &mapping->page_tree;
spin_lock_irq(&mapping->tree_lock);
entry = get_unlocked_mapping_entry(mapping, index, NULL);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (!entry || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry)))
goto out;
if (!trunc &&
(radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_DIRTY) ||
radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE)))
goto out;
radix_tree_delete(page_tree, index);
mapping->nrexceptional--;
ret = 1;
out:
put_unlocked_mapping_entry(mapping, index, entry);
spin_unlock_irq(&mapping->tree_lock);
return ret;
}
/*
* Delete exceptional DAX entry at @index from @mapping. Wait for radix tree
* entry to get unlocked before deleting it.
*/
int dax_delete_mapping_entry(struct address_space *mapping, pgoff_t index)
{
int ret = __dax_invalidate_mapping_entry(mapping, index, true);
/*
* This gets called from truncate / punch_hole path. As such, the caller
* must hold locks protecting against concurrent modifications of the
* radix tree (usually fs-private i_mmap_sem for writing). Since the
* caller has seen exceptional entry for this index, we better find it
* at that index as well...
*/
WARN_ON_ONCE(!ret);
return ret;
}
/*
* Invalidate exceptional DAX entry if it is clean.
*/
int dax_invalidate_mapping_entry_sync(struct address_space *mapping,
pgoff_t index)
{
return __dax_invalidate_mapping_entry(mapping, index, false);
}
static int copy_user_dax(struct block_device *bdev, struct dax_device *dax_dev,
sector_t sector, size_t size, struct page *to,
unsigned long vaddr)
{
void *vto, *kaddr;
pgoff_t pgoff;
pfn_t pfn;
long rc;
int id;
rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
if (rc)
return rc;
id = dax_read_lock();
rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
if (rc < 0) {
dax_read_unlock(id);
return rc;
}
vto = kmap_atomic(to);
copy_user_page(vto, (void __force *)kaddr, vaddr, to);
kunmap_atomic(vto);
dax_read_unlock(id);
return 0;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
* By this point grab_mapping_entry() has ensured that we have a locked entry
* of the appropriate size so we don't have to worry about downgrading PMDs to
* PTEs. If we happen to be trying to insert a PTE and there is a PMD
* already in the tree, we will skip the insertion and just dirty the PMD as
* appropriate.
*/
static void *dax_insert_mapping_entry(struct address_space *mapping,
struct vm_fault *vmf,
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
void *entry, sector_t sector,
unsigned long flags)
{
struct radix_tree_root *page_tree = &mapping->page_tree;
void *new_entry;
pgoff_t index = vmf->pgoff;
if (vmf->flags & FAULT_FLAG_WRITE)
__mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (dax_is_zero_entry(entry) && !(flags & RADIX_DAX_ZERO_PAGE)) {
/* we are replacing a zero page with block mapping */
if (dax_is_pmd_entry(entry))
unmap_mapping_range(mapping,
(vmf->pgoff << PAGE_SHIFT) & PMD_MASK,
PMD_SIZE, 0);
else /* pte entry */
unmap_mapping_range(mapping, vmf->pgoff << PAGE_SHIFT,
PAGE_SIZE, 0);
}
spin_lock_irq(&mapping->tree_lock);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
new_entry = dax_radix_locked_entry(sector, flags);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (dax_is_zero_entry(entry) || dax_is_empty_entry(entry)) {
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
* Only swap our new entry into the radix tree if the current
* entry is a zero page or an empty entry. If a normal PTE or
* PMD entry is already in the tree, we leave it alone. This
* means that if we are trying to insert a PTE and the
* existing entry is a PMD, we will just leave the PMD in the
* tree and dirty it if necessary.
*/
struct radix_tree_node *node;
void **slot;
void *ret;
ret = __radix_tree_lookup(page_tree, index, &node, &slot);
WARN_ON_ONCE(ret != entry);
__radix_tree_replace(page_tree, node, slot,
new_entry, NULL, NULL);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
entry = new_entry;
}
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (vmf->flags & FAULT_FLAG_WRITE)
radix_tree_tag_set(page_tree, index, PAGECACHE_TAG_DIRTY);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
spin_unlock_irq(&mapping->tree_lock);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
return entry;
}
static inline unsigned long
pgoff_address(pgoff_t pgoff, struct vm_area_struct *vma)
{
unsigned long address;
address = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
VM_BUG_ON_VMA(address < vma->vm_start || address >= vma->vm_end, vma);
return address;
}
/* Walk all mappings of a given index of a file and writeprotect them */
static void dax_mapping_entry_mkclean(struct address_space *mapping,
pgoff_t index, unsigned long pfn)
{
struct vm_area_struct *vma;
dax: wrprotect pmd_t in dax_mapping_entry_mkclean Currently dax_mapping_entry_mkclean() fails to clean and write protect the pmd_t of a DAX PMD entry during an *sync operation. This can result in data loss in the following sequence: 1) mmap write to DAX PMD, dirtying PMD radix tree entry and making the pmd_t dirty and writeable 2) fsync, flushing out PMD data and cleaning the radix tree entry. We currently fail to mark the pmd_t as clean and write protected. 3) more mmap writes to the PMD. These don't cause any page faults since the pmd_t is dirty and writeable. The radix tree entry remains clean. 4) fsync, which fails to flush the dirty PMD data because the radix tree entry was clean. 5) crash - dirty data that should have been fsync'd as part of 4) could still have been in the processor cache, and is lost. Fix this by marking the pmd_t clean and write protected in dax_mapping_entry_mkclean(), which is called as part of the fsync operation 2). This will cause the writes in step 3) above to generate page faults where we'll re-dirty the PMD radix tree entry, resulting in flushes in the fsync that happens in step 4). Fixes: 4b4bb46d00b3 ("dax: clear dirty entry tags on cache flush") Link: http://lkml.kernel.org/r/1482272586-21177-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:57:24 +00:00
pte_t pte, *ptep = NULL;
pmd_t *pmdp = NULL;
spinlock_t *ptl;
i_mmap_lock_read(mapping);
vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
unsigned long address, start, end;
cond_resched();
if (!(vma->vm_flags & VM_SHARED))
continue;
address = pgoff_address(index, vma);
/*
* Note because we provide start/end to follow_pte_pmd it will
* call mmu_notifier_invalidate_range_start() on our behalf
* before taking any lock.
*/
if (follow_pte_pmd(vma->vm_mm, address, &start, &end, &ptep, &pmdp, &ptl))
continue;
dax: wrprotect pmd_t in dax_mapping_entry_mkclean Currently dax_mapping_entry_mkclean() fails to clean and write protect the pmd_t of a DAX PMD entry during an *sync operation. This can result in data loss in the following sequence: 1) mmap write to DAX PMD, dirtying PMD radix tree entry and making the pmd_t dirty and writeable 2) fsync, flushing out PMD data and cleaning the radix tree entry. We currently fail to mark the pmd_t as clean and write protected. 3) more mmap writes to the PMD. These don't cause any page faults since the pmd_t is dirty and writeable. The radix tree entry remains clean. 4) fsync, which fails to flush the dirty PMD data because the radix tree entry was clean. 5) crash - dirty data that should have been fsync'd as part of 4) could still have been in the processor cache, and is lost. Fix this by marking the pmd_t clean and write protected in dax_mapping_entry_mkclean(), which is called as part of the fsync operation 2). This will cause the writes in step 3) above to generate page faults where we'll re-dirty the PMD radix tree entry, resulting in flushes in the fsync that happens in step 4). Fixes: 4b4bb46d00b3 ("dax: clear dirty entry tags on cache flush") Link: http://lkml.kernel.org/r/1482272586-21177-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:57:24 +00:00
if (pmdp) {
#ifdef CONFIG_FS_DAX_PMD
pmd_t pmd;
if (pfn != pmd_pfn(*pmdp))
goto unlock_pmd;
if (!pmd_dirty(*pmdp) && !pmd_write(*pmdp))
goto unlock_pmd;
flush_cache_page(vma, address, pfn);
pmd = pmdp_huge_clear_flush(vma, address, pmdp);
pmd = pmd_wrprotect(pmd);
pmd = pmd_mkclean(pmd);
set_pmd_at(vma->vm_mm, address, pmdp, pmd);
mmu_notifier_invalidate_range(vma->vm_mm, start, end);
dax: wrprotect pmd_t in dax_mapping_entry_mkclean Currently dax_mapping_entry_mkclean() fails to clean and write protect the pmd_t of a DAX PMD entry during an *sync operation. This can result in data loss in the following sequence: 1) mmap write to DAX PMD, dirtying PMD radix tree entry and making the pmd_t dirty and writeable 2) fsync, flushing out PMD data and cleaning the radix tree entry. We currently fail to mark the pmd_t as clean and write protected. 3) more mmap writes to the PMD. These don't cause any page faults since the pmd_t is dirty and writeable. The radix tree entry remains clean. 4) fsync, which fails to flush the dirty PMD data because the radix tree entry was clean. 5) crash - dirty data that should have been fsync'd as part of 4) could still have been in the processor cache, and is lost. Fix this by marking the pmd_t clean and write protected in dax_mapping_entry_mkclean(), which is called as part of the fsync operation 2). This will cause the writes in step 3) above to generate page faults where we'll re-dirty the PMD radix tree entry, resulting in flushes in the fsync that happens in step 4). Fixes: 4b4bb46d00b3 ("dax: clear dirty entry tags on cache flush") Link: http://lkml.kernel.org/r/1482272586-21177-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:57:24 +00:00
unlock_pmd:
spin_unlock(ptl);
#endif
} else {
if (pfn != pte_pfn(*ptep))
goto unlock_pte;
if (!pte_dirty(*ptep) && !pte_write(*ptep))
goto unlock_pte;
flush_cache_page(vma, address, pfn);
pte = ptep_clear_flush(vma, address, ptep);
pte = pte_wrprotect(pte);
pte = pte_mkclean(pte);
set_pte_at(vma->vm_mm, address, ptep, pte);
mmu_notifier_invalidate_range(vma->vm_mm, start, end);
dax: wrprotect pmd_t in dax_mapping_entry_mkclean Currently dax_mapping_entry_mkclean() fails to clean and write protect the pmd_t of a DAX PMD entry during an *sync operation. This can result in data loss in the following sequence: 1) mmap write to DAX PMD, dirtying PMD radix tree entry and making the pmd_t dirty and writeable 2) fsync, flushing out PMD data and cleaning the radix tree entry. We currently fail to mark the pmd_t as clean and write protected. 3) more mmap writes to the PMD. These don't cause any page faults since the pmd_t is dirty and writeable. The radix tree entry remains clean. 4) fsync, which fails to flush the dirty PMD data because the radix tree entry was clean. 5) crash - dirty data that should have been fsync'd as part of 4) could still have been in the processor cache, and is lost. Fix this by marking the pmd_t clean and write protected in dax_mapping_entry_mkclean(), which is called as part of the fsync operation 2). This will cause the writes in step 3) above to generate page faults where we'll re-dirty the PMD radix tree entry, resulting in flushes in the fsync that happens in step 4). Fixes: 4b4bb46d00b3 ("dax: clear dirty entry tags on cache flush") Link: http://lkml.kernel.org/r/1482272586-21177-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Jan Kara <jack@suse.cz> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Dave Hansen <dave.hansen@intel.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-01-11 00:57:24 +00:00
unlock_pte:
pte_unmap_unlock(ptep, ptl);
}
mmu_notifier_invalidate_range_end(vma->vm_mm, start, end);
}
i_mmap_unlock_read(mapping);
}
static int dax_writeback_one(struct block_device *bdev,
struct dax_device *dax_dev, struct address_space *mapping,
pgoff_t index, void *entry)
{
struct radix_tree_root *page_tree = &mapping->page_tree;
void *entry2, **slot, *kaddr;
long ret = 0, id;
sector_t sector;
pgoff_t pgoff;
size_t size;
pfn_t pfn;
/*
* A page got tagged dirty in DAX mapping? Something is seriously
* wrong.
*/
if (WARN_ON(!radix_tree_exceptional_entry(entry)))
return -EIO;
spin_lock_irq(&mapping->tree_lock);
entry2 = get_unlocked_mapping_entry(mapping, index, &slot);
/* Entry got punched out / reallocated? */
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (!entry2 || WARN_ON_ONCE(!radix_tree_exceptional_entry(entry2)))
goto put_unlocked;
/*
* Entry got reallocated elsewhere? No need to writeback. We have to
* compare sectors as we must not bail out due to difference in lockbit
* or entry type.
*/
if (dax_radix_sector(entry2) != dax_radix_sector(entry))
goto put_unlocked;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (WARN_ON_ONCE(dax_is_empty_entry(entry) ||
dax_is_zero_entry(entry))) {
ret = -EIO;
goto put_unlocked;
}
/* Another fsync thread may have already written back this entry */
if (!radix_tree_tag_get(page_tree, index, PAGECACHE_TAG_TOWRITE))
goto put_unlocked;
/* Lock the entry to serialize with page faults */
entry = lock_slot(mapping, slot);
/*
* We can clear the tag now but we have to be careful so that concurrent
* dax_writeback_one() calls for the same index cannot finish before we
* actually flush the caches. This is achieved as the calls will look
* at the entry only under tree_lock and once they do that they will
* see the entry locked and wait for it to unlock.
*/
radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_TOWRITE);
spin_unlock_irq(&mapping->tree_lock);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
* Even if dax_writeback_mapping_range() was given a wbc->range_start
* in the middle of a PMD, the 'index' we are given will be aligned to
* the start index of the PMD, as will the sector we pull from
* 'entry'. This allows us to flush for PMD_SIZE and not have to
* worry about partial PMD writebacks.
*/
sector = dax_radix_sector(entry);
size = PAGE_SIZE << dax_radix_order(entry);
id = dax_read_lock();
ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
if (ret)
goto dax_unlock;
/*
* dax_direct_access() may sleep, so cannot hold tree_lock over
* its invocation.
*/
ret = dax_direct_access(dax_dev, pgoff, size / PAGE_SIZE, &kaddr, &pfn);
if (ret < 0)
goto dax_unlock;
if (WARN_ON_ONCE(ret < size / PAGE_SIZE)) {
ret = -EIO;
goto dax_unlock;
}
dax_mapping_entry_mkclean(mapping, index, pfn_t_to_pfn(pfn));
dax_flush(dax_dev, pgoff, kaddr, size);
/*
* After we have flushed the cache, we can clear the dirty tag. There
* cannot be new dirty data in the pfn after the flush has completed as
* the pfn mappings are writeprotected and fault waits for mapping
* entry lock.
*/
spin_lock_irq(&mapping->tree_lock);
radix_tree_tag_clear(page_tree, index, PAGECACHE_TAG_DIRTY);
spin_unlock_irq(&mapping->tree_lock);
trace_dax_writeback_one(mapping->host, index, size >> PAGE_SHIFT);
dax_unlock:
dax_read_unlock(id);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
put_locked_mapping_entry(mapping, index);
return ret;
put_unlocked:
put_unlocked_mapping_entry(mapping, index, entry2);
spin_unlock_irq(&mapping->tree_lock);
return ret;
}
/*
* Flush the mapping to the persistent domain within the byte range of [start,
* end]. This is required by data integrity operations to ensure file data is
* on persistent storage prior to completion of the operation.
*/
int dax_writeback_mapping_range(struct address_space *mapping,
struct block_device *bdev, struct writeback_control *wbc)
{
struct inode *inode = mapping->host;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
pgoff_t start_index, end_index;
pgoff_t indices[PAGEVEC_SIZE];
struct dax_device *dax_dev;
struct pagevec pvec;
bool done = false;
int i, ret = 0;
if (WARN_ON_ONCE(inode->i_blkbits != PAGE_SHIFT))
return -EIO;
if (!mapping->nrexceptional || wbc->sync_mode != WB_SYNC_ALL)
return 0;
dax_dev = dax_get_by_host(bdev->bd_disk->disk_name);
if (!dax_dev)
return -EIO;
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
start_index = wbc->range_start >> PAGE_SHIFT;
end_index = wbc->range_end >> PAGE_SHIFT;
trace_dax_writeback_range(inode, start_index, end_index);
tag_pages_for_writeback(mapping, start_index, end_index);
pagevec_init(&pvec, 0);
while (!done) {
pvec.nr = find_get_entries_tag(mapping, start_index,
PAGECACHE_TAG_TOWRITE, PAGEVEC_SIZE,
pvec.pages, indices);
if (pvec.nr == 0)
break;
for (i = 0; i < pvec.nr; i++) {
if (indices[i] > end_index) {
done = true;
break;
}
ret = dax_writeback_one(bdev, dax_dev, mapping,
indices[i], pvec.pages[i]);
if (ret < 0) {
mapping_set_error(mapping, ret);
goto out;
}
}
start_index = indices[pvec.nr - 1] + 1;
}
out:
put_dax(dax_dev);
trace_dax_writeback_range_done(inode, start_index, end_index);
return (ret < 0 ? ret : 0);
}
EXPORT_SYMBOL_GPL(dax_writeback_mapping_range);
static int dax_insert_mapping(struct address_space *mapping,
struct block_device *bdev, struct dax_device *dax_dev,
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
sector_t sector, size_t size, void *entry,
struct vm_area_struct *vma, struct vm_fault *vmf)
{
unsigned long vaddr = vmf->address;
void *ret, *kaddr;
pgoff_t pgoff;
int id, rc;
pfn_t pfn;
rc = bdev_dax_pgoff(bdev, sector, size, &pgoff);
if (rc)
return rc;
id = dax_read_lock();
rc = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
if (rc < 0) {
dax_read_unlock(id);
return rc;
}
dax_read_unlock(id);
ret = dax_insert_mapping_entry(mapping, vmf, entry, sector, 0);
if (IS_ERR(ret))
return PTR_ERR(ret);
trace_dax_insert_mapping(mapping->host, vmf, ret);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
if (vmf->flags & FAULT_FLAG_WRITE)
return vm_insert_mixed_mkwrite(vma, vaddr, pfn);
else
return vm_insert_mixed(vma, vaddr, pfn);
}
/*
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
* The user has performed a load from a hole in the file. Allocating a new
* page in the file would cause excessive storage usage for workloads with
* sparse files. Instead we insert a read-only mapping of the 4k zero page.
* If this page is ever written to we will re-fault and change the mapping to
* point to real DAX storage instead.
*/
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
static int dax_load_hole(struct address_space *mapping, void *entry,
struct vm_fault *vmf)
{
struct inode *inode = mapping->host;
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
unsigned long vaddr = vmf->address;
int ret = VM_FAULT_NOPAGE;
struct page *zero_page;
void *entry2;
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
zero_page = ZERO_PAGE(0);
if (unlikely(!zero_page)) {
ret = VM_FAULT_OOM;
goto out;
}
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
entry2 = dax_insert_mapping_entry(mapping, vmf, entry, 0,
RADIX_DAX_ZERO_PAGE);
if (IS_ERR(entry2)) {
ret = VM_FAULT_SIGBUS;
goto out;
}
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
vm_insert_mixed(vmf->vma, vaddr, page_to_pfn_t(zero_page));
out:
trace_dax_load_hole(inode, vmf, ret);
return ret;
}
static bool dax_range_is_aligned(struct block_device *bdev,
unsigned int offset, unsigned int length)
{
unsigned short sector_size = bdev_logical_block_size(bdev);
if (!IS_ALIGNED(offset, sector_size))
return false;
if (!IS_ALIGNED(length, sector_size))
return false;
return true;
}
int __dax_zero_page_range(struct block_device *bdev,
struct dax_device *dax_dev, sector_t sector,
unsigned int offset, unsigned int size)
{
if (dax_range_is_aligned(bdev, offset, size)) {
sector_t start_sector = sector + (offset >> 9);
return blkdev_issue_zeroout(bdev, start_sector,
libnvdimm for 4.12 * Region media error reporting: A libnvdimm region device is the parent to one or more namespaces. To date, media errors have been reported via the "badblocks" attribute attached to pmem block devices for namespaces in "raw" or "memory" mode. Given that namespaces can be in "device-dax" or "btt-sector" mode this new interface reports media errors generically, i.e. independent of namespace modes or state. This subsequently allows userspace tooling to craft "ACPI 6.1 Section 9.20.7.6 Function Index 4 - Clear Uncorrectable Error" requests and submit them via the ioctl path for NVDIMM root bus devices. * Introduce 'struct dax_device' and 'struct dax_operations': Prompted by a request from Linus and feedback from Christoph this allows for dax capable drivers to publish their own custom dax operations. This fixes the broken assumption that all dax operations are related to a persistent memory device, and makes it easier for other architectures and platforms to add customized persistent memory support. * 'libnvdimm' core updates: A new "deep_flush" sysfs attribute is available for storage appliance applications to manually trigger memory controllers to drain write-pending buffers that would otherwise be flushed automatically by the platform ADR (asynchronous-DRAM-refresh) mechanism at a power loss event. Support for "locked" DIMMs is included to prevent namespaces from surfacing when the namespace label data area is locked. Finally, fixes for various reported deadlocks and crashes, also tagged for -stable. * ACPI / nfit driver updates: General updates of the nfit driver to add DSM command overrides, ACPI 6.1 health state flags support, DSM payload debug available by default, and various fixes. Acknowledgements that came after the branch was pushed: commmit 565851c972b5 "device-dax: fix sysfs attribute deadlock" Tested-by: Yi Zhang <yizhan@redhat.com> commit 23f498448362 "libnvdimm: rework region badblocks clearing" Tested-by: Toshi Kani <toshi.kani@hpe.com> -----BEGIN PGP SIGNATURE----- iQIcBAABAgAGBQJZDONJAAoJEB7SkWpmfYgC3SsP/2KrLvTUcz646ViuPOgZ2cC4 W6wAx6cvDSt+H52kLnFEsYoFt7WAj20ggPirb/Bc5jkGlvwE0lT9Xtmso9GpVkYT J9ZJ9pP/4YaAD3II1gmTwaUjYi0FxoOdx3Eb92yuWkO/8ylz4b2Nu3cBpYwyziGQ nIfEVwDXRLE86u6x0bWuf6TlVuvsbdiAI55CDqDMVQC6xIOLbSez7b8QIHlpiKEb Mw+xqdQva0esoreZEOXEhWNO+qtfILx8/ceBEGTNMp4e/JjZ2FbrSNplM+9bH5k7 ywqP8lW+mBEw0fmBBkYoVG/xyesiiBb55JLnbi8Ew+7IUxw8a3iV7wftRi62lHcK zAjsHe4L+MansgtZsCL8wluvIPaktAdtB4xr7l9VNLKRYRUG73jEWU0gcUNryHIL BkQJ52pUS1PkClyAsWbBBHl1I/CvzVPd21VW0YELmLR4OywKy1c+eKw2bcYgjrb4 59HZSv6S6EoKaQC+2qvVNpePil7cdfg5V2ubH/ki9HoYVyoxDptEWHnvf0NNatIH Y7mNcOPvhOksJmnKSyHbDjtRur7WoHIlC9D7UjEFkSBWsKPjxJHoidN4SnCMRtjQ WKQU0seoaKj04b68Bs/Qm9NozVgnsPFIUDZeLMikLFX2Jt7YSPu+Jmi2s4re6WLh TmJQ3Ly9t3o3/weHSzmn =Ox0s -----END PGP SIGNATURE----- Merge tag 'libnvdimm-for-4.12' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm Pull libnvdimm updates from Dan Williams: "The bulk of this has been in multiple -next releases. There were a few late breaking fixes and small features that got added in the last couple days, but the whole set has received a build success notification from the kbuild robot. Change summary: - Region media error reporting: A libnvdimm region device is the parent to one or more namespaces. To date, media errors have been reported via the "badblocks" attribute attached to pmem block devices for namespaces in "raw" or "memory" mode. Given that namespaces can be in "device-dax" or "btt-sector" mode this new interface reports media errors generically, i.e. independent of namespace modes or state. This subsequently allows userspace tooling to craft "ACPI 6.1 Section 9.20.7.6 Function Index 4 - Clear Uncorrectable Error" requests and submit them via the ioctl path for NVDIMM root bus devices. - Introduce 'struct dax_device' and 'struct dax_operations': Prompted by a request from Linus and feedback from Christoph this allows for dax capable drivers to publish their own custom dax operations. This fixes the broken assumption that all dax operations are related to a persistent memory device, and makes it easier for other architectures and platforms to add customized persistent memory support. - 'libnvdimm' core updates: A new "deep_flush" sysfs attribute is available for storage appliance applications to manually trigger memory controllers to drain write-pending buffers that would otherwise be flushed automatically by the platform ADR (asynchronous-DRAM-refresh) mechanism at a power loss event. Support for "locked" DIMMs is included to prevent namespaces from surfacing when the namespace label data area is locked. Finally, fixes for various reported deadlocks and crashes, also tagged for -stable. - ACPI / nfit driver updates: General updates of the nfit driver to add DSM command overrides, ACPI 6.1 health state flags support, DSM payload debug available by default, and various fixes. Acknowledgements that came after the branch was pushed: - commmit 565851c972b5 "device-dax: fix sysfs attribute deadlock": Tested-by: Yi Zhang <yizhan@redhat.com> - commit 23f498448362 "libnvdimm: rework region badblocks clearing" Tested-by: Toshi Kani <toshi.kani@hpe.com>" * tag 'libnvdimm-for-4.12' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm: (52 commits) libnvdimm, pfn: fix 'npfns' vs section alignment libnvdimm: handle locked label storage areas libnvdimm: convert NDD_ flags to use bitops, introduce NDD_LOCKED brd: fix uninitialized use of brd->dax_dev block, dax: use correct format string in bdev_dax_supported device-dax: fix sysfs attribute deadlock libnvdimm: restore "libnvdimm: band aid btt vs clear poison locking" libnvdimm: fix nvdimm_bus_lock() vs device_lock() ordering libnvdimm: rework region badblocks clearing acpi, nfit: kill ACPI_NFIT_DEBUG libnvdimm: fix clear length of nvdimm_forget_poison() libnvdimm, pmem: fix a NULL pointer BUG in nd_pmem_notify libnvdimm, region: sysfs trigger for nvdimm_flush() libnvdimm: fix phys_addr for nvdimm_clear_poison x86, dax, pmem: remove indirection around memcpy_from_pmem() block: remove block_device_operations ->direct_access() block, dax: convert bdev_dax_supported() to dax_direct_access() filesystem-dax: convert to dax_direct_access() Revert "block: use DAX for partition table reads" ext2, ext4, xfs: retrieve dax_device for iomap operations ...
2017-05-06 01:49:20 +00:00
size >> 9, GFP_NOFS, 0);
} else {
pgoff_t pgoff;
long rc, id;
void *kaddr;
pfn_t pfn;
rc = bdev_dax_pgoff(bdev, sector, PAGE_SIZE, &pgoff);
if (rc)
return rc;
id = dax_read_lock();
rc = dax_direct_access(dax_dev, pgoff, 1, &kaddr,
&pfn);
if (rc < 0) {
dax_read_unlock(id);
return rc;
}
memset(kaddr + offset, 0, size);
dax_flush(dax_dev, pgoff, kaddr + offset, size);
dax_read_unlock(id);
}
return 0;
}
EXPORT_SYMBOL_GPL(__dax_zero_page_range);
static sector_t dax_iomap_sector(struct iomap *iomap, loff_t pos)
{
return iomap->blkno + (((pos & PAGE_MASK) - iomap->offset) >> 9);
}
static loff_t
dax_iomap_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
struct iomap *iomap)
{
struct block_device *bdev = iomap->bdev;
struct dax_device *dax_dev = iomap->dax_dev;
struct iov_iter *iter = data;
loff_t end = pos + length, done = 0;
ssize_t ret = 0;
int id;
if (iov_iter_rw(iter) == READ) {
end = min(end, i_size_read(inode));
if (pos >= end)
return 0;
if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
return iov_iter_zero(min(length, end - pos), iter);
}
if (WARN_ON_ONCE(iomap->type != IOMAP_MAPPED))
return -EIO;
/*
* Write can allocate block for an area which has a hole page mapped
* into page tables. We have to tear down these mappings so that data
* written by write(2) is visible in mmap.
*/
if (iomap->flags & IOMAP_F_NEW) {
invalidate_inode_pages2_range(inode->i_mapping,
pos >> PAGE_SHIFT,
(end - 1) >> PAGE_SHIFT);
}
id = dax_read_lock();
while (pos < end) {
unsigned offset = pos & (PAGE_SIZE - 1);
const size_t size = ALIGN(length + offset, PAGE_SIZE);
const sector_t sector = dax_iomap_sector(iomap, pos);
ssize_t map_len;
pgoff_t pgoff;
void *kaddr;
pfn_t pfn;
fs: break out of iomap_file_buffered_write on fatal signals Tetsuo has noticed that an OOM stress test which performs large write requests can cause the full memory reserves depletion. He has tracked this down to the following path __alloc_pages_nodemask+0x436/0x4d0 alloc_pages_current+0x97/0x1b0 __page_cache_alloc+0x15d/0x1a0 mm/filemap.c:728 pagecache_get_page+0x5a/0x2b0 mm/filemap.c:1331 grab_cache_page_write_begin+0x23/0x40 mm/filemap.c:2773 iomap_write_begin+0x50/0xd0 fs/iomap.c:118 iomap_write_actor+0xb5/0x1a0 fs/iomap.c:190 ? iomap_write_end+0x80/0x80 fs/iomap.c:150 iomap_apply+0xb3/0x130 fs/iomap.c:79 iomap_file_buffered_write+0x68/0xa0 fs/iomap.c:243 ? iomap_write_end+0x80/0x80 xfs_file_buffered_aio_write+0x132/0x390 [xfs] ? remove_wait_queue+0x59/0x60 xfs_file_write_iter+0x90/0x130 [xfs] __vfs_write+0xe5/0x140 vfs_write+0xc7/0x1f0 ? syscall_trace_enter+0x1d0/0x380 SyS_write+0x58/0xc0 do_syscall_64+0x6c/0x200 entry_SYSCALL64_slow_path+0x25/0x25 the oom victim has access to all memory reserves to make a forward progress to exit easier. But iomap_file_buffered_write and other callers of iomap_apply loop to complete the full request. We need to check for fatal signals and back off with a short write instead. As the iomap_apply delegates all the work down to the actor we have to hook into those. All callers that work with the page cache are calling iomap_write_begin so we will check for signals there. dax_iomap_actor has to handle the situation explicitly because it copies data to the userspace directly. Other callers like iomap_page_mkwrite work on a single page or iomap_fiemap_actor do not allocate memory based on the given len. Fixes: 68a9f5e7007c ("xfs: implement iomap based buffered write path") Link: http://lkml.kernel.org/r/20170201092706.9966-2-mhocko@kernel.org Signed-off-by: Michal Hocko <mhocko@suse.com> Reported-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Reviewed-by: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: <stable@vger.kernel.org> [4.8+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-03 21:13:26 +00:00
if (fatal_signal_pending(current)) {
ret = -EINTR;
break;
}
ret = bdev_dax_pgoff(bdev, sector, size, &pgoff);
if (ret)
break;
map_len = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size),
&kaddr, &pfn);
if (map_len < 0) {
ret = map_len;
break;
}
map_len = PFN_PHYS(map_len);
kaddr += offset;
map_len -= offset;
if (map_len > end - pos)
map_len = end - pos;
/*
* The userspace address for the memory copy has already been
* validated via access_ok() in either vfs_read() or
* vfs_write(), depending on which operation we are doing.
*/
if (iov_iter_rw(iter) == WRITE)
map_len = dax_copy_from_iter(dax_dev, pgoff, kaddr,
map_len, iter);
else
map_len = copy_to_iter(kaddr, map_len, iter);
if (map_len <= 0) {
ret = map_len ? map_len : -EFAULT;
break;
}
pos += map_len;
length -= map_len;
done += map_len;
}
dax_read_unlock(id);
return done ? done : ret;
}
/**
* dax_iomap_rw - Perform I/O to a DAX file
* @iocb: The control block for this I/O
* @iter: The addresses to do I/O from or to
* @ops: iomap ops passed from the file system
*
* This function performs read and write operations to directly mapped
* persistent memory. The callers needs to take care of read/write exclusion
* and evicting any page cache pages in the region under I/O.
*/
ssize_t
dax_iomap_rw(struct kiocb *iocb, struct iov_iter *iter,
const struct iomap_ops *ops)
{
struct address_space *mapping = iocb->ki_filp->f_mapping;
struct inode *inode = mapping->host;
loff_t pos = iocb->ki_pos, ret = 0, done = 0;
unsigned flags = 0;
if (iov_iter_rw(iter) == WRITE) {
lockdep_assert_held_exclusive(&inode->i_rwsem);
flags |= IOMAP_WRITE;
} else {
lockdep_assert_held(&inode->i_rwsem);
}
while (iov_iter_count(iter)) {
ret = iomap_apply(inode, pos, iov_iter_count(iter), flags, ops,
iter, dax_iomap_actor);
if (ret <= 0)
break;
pos += ret;
done += ret;
}
iocb->ki_pos += done;
return done ? done : ret;
}
EXPORT_SYMBOL_GPL(dax_iomap_rw);
static int dax_fault_return(int error)
{
if (error == 0)
return VM_FAULT_NOPAGE;
if (error == -ENOMEM)
return VM_FAULT_OOM;
return VM_FAULT_SIGBUS;
}
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
static int dax_iomap_pte_fault(struct vm_fault *vmf,
const struct iomap_ops *ops)
{
struct address_space *mapping = vmf->vma->vm_file->f_mapping;
struct inode *inode = mapping->host;
unsigned long vaddr = vmf->address;
loff_t pos = (loff_t)vmf->pgoff << PAGE_SHIFT;
sector_t sector;
struct iomap iomap = { 0 };
unsigned flags = IOMAP_FAULT;
int error, major = 0;
int vmf_ret = 0;
void *entry;
dax: add tracepoints to dax_iomap_pte_fault() Patch series "second round of tracepoints for DAX". This second round of DAX tracepoint patches adds tracing to the PTE fault path (dax_iomap_pte_fault(), dax_pfn_mkwrite(), dax_load_hole(), dax_insert_mapping()) and to the writeback path (dax_writeback_mapping_range(), dax_writeback_one()). The purpose of this tracing is to give us a high level view of what DAX is doing, whether faults are being serviced by PMDs or PTEs, and by real storage or by zero pages covering holes. I do have some patches nearly ready which also add tracing to grab_mapping_entry() and dax_insert_mapping_entry(). These are more targeted at logging how we are interacting with the radix tree, how we use empty entries for locking, whether we "downgrade" huge zero pages to 4k PTE sized allocations, etc. In the end it seemed to me that this might be too detailed to have as constantly present tracepoints, but if anyone sees value in having tracepoints like this in the DAX code permanently (Jan?), please let me know and I'll add those last two patches. All these tracepoints were done to be consistent with the style of the XFS tracepoints and with the existing DAX PMD tracepoints. This patch (of 6): Add tracepoints to dax_iomap_pte_fault(), following the same logging conventions as the rest of DAX. Here is an example fault that initially tries to be serviced by the PMD fault handler but which falls back to PTEs because the VMA isn't large enough to hold a PMD: small-1086 [005] .... 71.140014: xfs_filemap_huge_fault: dev 259:0 ino 0x1003 small-1086 [005] .... 71.140027: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 small-1086 [005] .... 71.140028: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 FALLBACK small-1086 [005] .... 71.140035: dax_pte_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 small-1086 [005] .... 71.140396: dax_pte_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 MAJOR|NOPAGE Link: http://lkml.kernel.org/r/20170221195116.13278-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 23:00:00 +00:00
trace_dax_pte_fault(inode, vmf, vmf_ret);
/*
* Check whether offset isn't beyond end of file now. Caller is supposed
* to hold locks serializing us with truncate / punch hole so this is
* a reliable test.
*/
dax: add tracepoints to dax_iomap_pte_fault() Patch series "second round of tracepoints for DAX". This second round of DAX tracepoint patches adds tracing to the PTE fault path (dax_iomap_pte_fault(), dax_pfn_mkwrite(), dax_load_hole(), dax_insert_mapping()) and to the writeback path (dax_writeback_mapping_range(), dax_writeback_one()). The purpose of this tracing is to give us a high level view of what DAX is doing, whether faults are being serviced by PMDs or PTEs, and by real storage or by zero pages covering holes. I do have some patches nearly ready which also add tracing to grab_mapping_entry() and dax_insert_mapping_entry(). These are more targeted at logging how we are interacting with the radix tree, how we use empty entries for locking, whether we "downgrade" huge zero pages to 4k PTE sized allocations, etc. In the end it seemed to me that this might be too detailed to have as constantly present tracepoints, but if anyone sees value in having tracepoints like this in the DAX code permanently (Jan?), please let me know and I'll add those last two patches. All these tracepoints were done to be consistent with the style of the XFS tracepoints and with the existing DAX PMD tracepoints. This patch (of 6): Add tracepoints to dax_iomap_pte_fault(), following the same logging conventions as the rest of DAX. Here is an example fault that initially tries to be serviced by the PMD fault handler but which falls back to PTEs because the VMA isn't large enough to hold a PMD: small-1086 [005] .... 71.140014: xfs_filemap_huge_fault: dev 259:0 ino 0x1003 small-1086 [005] .... 71.140027: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 small-1086 [005] .... 71.140028: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 FALLBACK small-1086 [005] .... 71.140035: dax_pte_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 small-1086 [005] .... 71.140396: dax_pte_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 MAJOR|NOPAGE Link: http://lkml.kernel.org/r/20170221195116.13278-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 23:00:00 +00:00
if (pos >= i_size_read(inode)) {
vmf_ret = VM_FAULT_SIGBUS;
goto out;
}
if ((vmf->flags & FAULT_FLAG_WRITE) && !vmf->cow_page)
flags |= IOMAP_WRITE;
entry = grab_mapping_entry(mapping, vmf->pgoff, 0);
if (IS_ERR(entry)) {
vmf_ret = dax_fault_return(PTR_ERR(entry));
goto out;
}
dax: fix race between colliding PMD & PTE entries We currently have two related PMD vs PTE races in the DAX code. These can both be easily triggered by having two threads reading and writing simultaneously to the same private mapping, with the key being that private mapping reads can be handled with PMDs but private mapping writes are always handled with PTEs so that we can COW. Here is the first race: CPU 0 CPU 1 (private mapping write) __handle_mm_fault() create_huge_pmd() - FALLBACK handle_pte_fault() passes check for pmd_devmap() (private mapping read) __handle_mm_fault() create_huge_pmd() dax_iomap_pmd_fault() inserts PMD dax_iomap_pte_fault() does a PTE fault, but we already have a DAX PMD installed in our page tables at this spot. Here's the second race: CPU 0 CPU 1 (private mapping read) __handle_mm_fault() passes check for pmd_none() create_huge_pmd() dax_iomap_pmd_fault() inserts PMD (private mapping write) __handle_mm_fault() create_huge_pmd() - FALLBACK (private mapping read) __handle_mm_fault() passes check for pmd_none() create_huge_pmd() handle_pte_fault() dax_iomap_pte_fault() inserts PTE dax_iomap_pmd_fault() inserts PMD, but we already have a PTE at this spot. The core of the issue is that while there is isolation between faults to the same range in the DAX fault handlers via our DAX entry locking, there is no isolation between faults in the code in mm/memory.c. This means for instance that this code in __handle_mm_fault() can run: if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) { ret = create_huge_pmd(&vmf); But by the time we actually get to run the fault handler called by create_huge_pmd(), the PMD is no longer pmd_none() because a racing PTE fault has installed a normal PMD here as a parent. This is the cause of the 2nd race. The first race is similar - there is the following check in handle_pte_fault(): } else { /* See comment in pte_alloc_one_map() */ if (pmd_devmap(*vmf->pmd) || pmd_trans_unstable(vmf->pmd)) return 0; So if a pmd_devmap() PMD (a DAX PMD) has been installed at vmf->pmd, we will bail and retry the fault. This is correct, but there is nothing preventing the PMD from being installed after this check but before we actually get to the DAX PTE fault handlers. In my testing these races result in the following types of errors: BUG: Bad rss-counter state mm:ffff8800a817d280 idx:1 val:1 BUG: non-zero nr_ptes on freeing mm: 15 Fix this issue by having the DAX fault handlers verify that it is safe to continue their fault after they have taken an entry lock to block other racing faults. [ross.zwisler@linux.intel.com: improve fix for colliding PMD & PTE entries] Link: http://lkml.kernel.org/r/20170526195932.32178-1-ross.zwisler@linux.intel.com Link: http://lkml.kernel.org/r/20170522215749.23516-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reported-by: Pawel Lebioda <pawel.lebioda@intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Pawel Lebioda <pawel.lebioda@intel.com> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Xiong Zhou <xzhou@redhat.com> Cc: Eryu Guan <eguan@redhat.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-02 21:46:37 +00:00
/*
* It is possible, particularly with mixed reads & writes to private
* mappings, that we have raced with a PMD fault that overlaps with
* the PTE we need to set up. If so just return and the fault will be
* retried.
*/
if (pmd_trans_huge(*vmf->pmd) || pmd_devmap(*vmf->pmd)) {
vmf_ret = VM_FAULT_NOPAGE;
goto unlock_entry;
}
/*
* Note that we don't bother to use iomap_apply here: DAX required
* the file system block size to be equal the page size, which means
* that we never have to deal with more than a single extent here.
*/
error = ops->iomap_begin(inode, pos, PAGE_SIZE, flags, &iomap);
dax: add tracepoints to dax_iomap_pte_fault() Patch series "second round of tracepoints for DAX". This second round of DAX tracepoint patches adds tracing to the PTE fault path (dax_iomap_pte_fault(), dax_pfn_mkwrite(), dax_load_hole(), dax_insert_mapping()) and to the writeback path (dax_writeback_mapping_range(), dax_writeback_one()). The purpose of this tracing is to give us a high level view of what DAX is doing, whether faults are being serviced by PMDs or PTEs, and by real storage or by zero pages covering holes. I do have some patches nearly ready which also add tracing to grab_mapping_entry() and dax_insert_mapping_entry(). These are more targeted at logging how we are interacting with the radix tree, how we use empty entries for locking, whether we "downgrade" huge zero pages to 4k PTE sized allocations, etc. In the end it seemed to me that this might be too detailed to have as constantly present tracepoints, but if anyone sees value in having tracepoints like this in the DAX code permanently (Jan?), please let me know and I'll add those last two patches. All these tracepoints were done to be consistent with the style of the XFS tracepoints and with the existing DAX PMD tracepoints. This patch (of 6): Add tracepoints to dax_iomap_pte_fault(), following the same logging conventions as the rest of DAX. Here is an example fault that initially tries to be serviced by the PMD fault handler but which falls back to PTEs because the VMA isn't large enough to hold a PMD: small-1086 [005] .... 71.140014: xfs_filemap_huge_fault: dev 259:0 ino 0x1003 small-1086 [005] .... 71.140027: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 small-1086 [005] .... 71.140028: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 FALLBACK small-1086 [005] .... 71.140035: dax_pte_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 small-1086 [005] .... 71.140396: dax_pte_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 MAJOR|NOPAGE Link: http://lkml.kernel.org/r/20170221195116.13278-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 23:00:00 +00:00
if (error) {
vmf_ret = dax_fault_return(error);
goto unlock_entry;
dax: add tracepoints to dax_iomap_pte_fault() Patch series "second round of tracepoints for DAX". This second round of DAX tracepoint patches adds tracing to the PTE fault path (dax_iomap_pte_fault(), dax_pfn_mkwrite(), dax_load_hole(), dax_insert_mapping()) and to the writeback path (dax_writeback_mapping_range(), dax_writeback_one()). The purpose of this tracing is to give us a high level view of what DAX is doing, whether faults are being serviced by PMDs or PTEs, and by real storage or by zero pages covering holes. I do have some patches nearly ready which also add tracing to grab_mapping_entry() and dax_insert_mapping_entry(). These are more targeted at logging how we are interacting with the radix tree, how we use empty entries for locking, whether we "downgrade" huge zero pages to 4k PTE sized allocations, etc. In the end it seemed to me that this might be too detailed to have as constantly present tracepoints, but if anyone sees value in having tracepoints like this in the DAX code permanently (Jan?), please let me know and I'll add those last two patches. All these tracepoints were done to be consistent with the style of the XFS tracepoints and with the existing DAX PMD tracepoints. This patch (of 6): Add tracepoints to dax_iomap_pte_fault(), following the same logging conventions as the rest of DAX. Here is an example fault that initially tries to be serviced by the PMD fault handler but which falls back to PTEs because the VMA isn't large enough to hold a PMD: small-1086 [005] .... 71.140014: xfs_filemap_huge_fault: dev 259:0 ino 0x1003 small-1086 [005] .... 71.140027: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 small-1086 [005] .... 71.140028: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 FALLBACK small-1086 [005] .... 71.140035: dax_pte_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 small-1086 [005] .... 71.140396: dax_pte_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 MAJOR|NOPAGE Link: http://lkml.kernel.org/r/20170221195116.13278-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 23:00:00 +00:00
}
if (WARN_ON_ONCE(iomap.offset + iomap.length < pos + PAGE_SIZE)) {
error = -EIO; /* fs corruption? */
goto error_finish_iomap;
}
sector = dax_iomap_sector(&iomap, pos);
if (vmf->cow_page) {
switch (iomap.type) {
case IOMAP_HOLE:
case IOMAP_UNWRITTEN:
clear_user_highpage(vmf->cow_page, vaddr);
break;
case IOMAP_MAPPED:
error = copy_user_dax(iomap.bdev, iomap.dax_dev,
sector, PAGE_SIZE, vmf->cow_page, vaddr);
break;
default:
WARN_ON_ONCE(1);
error = -EIO;
break;
}
if (error)
goto error_finish_iomap;
__SetPageUptodate(vmf->cow_page);
vmf_ret = finish_fault(vmf);
if (!vmf_ret)
vmf_ret = VM_FAULT_DONE_COW;
goto finish_iomap;
}
switch (iomap.type) {
case IOMAP_MAPPED:
if (iomap.flags & IOMAP_F_NEW) {
count_vm_event(PGMAJFAULT);
count_memcg_event_mm(vmf->vma->vm_mm, PGMAJFAULT);
major = VM_FAULT_MAJOR;
}
error = dax_insert_mapping(mapping, iomap.bdev, iomap.dax_dev,
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
sector, PAGE_SIZE, entry, vmf->vma, vmf);
/* -EBUSY is fine, somebody else faulted on the same PTE */
if (error == -EBUSY)
error = 0;
break;
case IOMAP_UNWRITTEN:
case IOMAP_HOLE:
if (!(vmf->flags & FAULT_FLAG_WRITE)) {
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
vmf_ret = dax_load_hole(mapping, entry, vmf);
goto finish_iomap;
}
/*FALLTHRU*/
default:
WARN_ON_ONCE(1);
error = -EIO;
break;
}
error_finish_iomap:
vmf_ret = dax_fault_return(error) | major;
finish_iomap:
if (ops->iomap_end) {
int copied = PAGE_SIZE;
if (vmf_ret & VM_FAULT_ERROR)
copied = 0;
/*
* The fault is done by now and there's no way back (other
* thread may be already happily using PTE we have installed).
* Just ignore error from ->iomap_end since we cannot do much
* with it.
*/
ops->iomap_end(inode, pos, PAGE_SIZE, copied, flags, &iomap);
}
unlock_entry:
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
put_locked_mapping_entry(mapping, vmf->pgoff);
out:
dax: add tracepoints to dax_iomap_pte_fault() Patch series "second round of tracepoints for DAX". This second round of DAX tracepoint patches adds tracing to the PTE fault path (dax_iomap_pte_fault(), dax_pfn_mkwrite(), dax_load_hole(), dax_insert_mapping()) and to the writeback path (dax_writeback_mapping_range(), dax_writeback_one()). The purpose of this tracing is to give us a high level view of what DAX is doing, whether faults are being serviced by PMDs or PTEs, and by real storage or by zero pages covering holes. I do have some patches nearly ready which also add tracing to grab_mapping_entry() and dax_insert_mapping_entry(). These are more targeted at logging how we are interacting with the radix tree, how we use empty entries for locking, whether we "downgrade" huge zero pages to 4k PTE sized allocations, etc. In the end it seemed to me that this might be too detailed to have as constantly present tracepoints, but if anyone sees value in having tracepoints like this in the DAX code permanently (Jan?), please let me know and I'll add those last two patches. All these tracepoints were done to be consistent with the style of the XFS tracepoints and with the existing DAX PMD tracepoints. This patch (of 6): Add tracepoints to dax_iomap_pte_fault(), following the same logging conventions as the rest of DAX. Here is an example fault that initially tries to be serviced by the PMD fault handler but which falls back to PTEs because the VMA isn't large enough to hold a PMD: small-1086 [005] .... 71.140014: xfs_filemap_huge_fault: dev 259:0 ino 0x1003 small-1086 [005] .... 71.140027: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 small-1086 [005] .... 71.140028: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 vm_start 0x10200000 vm_end 0x10500000 pgoff 0x220 max_pgoff 0x1400 FALLBACK small-1086 [005] .... 71.140035: dax_pte_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 small-1086 [005] .... 71.140396: dax_pte_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10420000 pgoff 0x220 MAJOR|NOPAGE Link: http://lkml.kernel.org/r/20170221195116.13278-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-08 23:00:00 +00:00
trace_dax_pte_fault_done(inode, vmf, vmf_ret);
return vmf_ret;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
#ifdef CONFIG_FS_DAX_PMD
/*
* The 'colour' (ie low bits) within a PMD of a page offset. This comes up
* more often than one might expect in the below functions.
*/
#define PG_PMD_COLOUR ((PMD_SIZE >> PAGE_SHIFT) - 1)
static int dax_pmd_insert_mapping(struct vm_fault *vmf, struct iomap *iomap,
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
loff_t pos, void *entry)
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
{
struct address_space *mapping = vmf->vma->vm_file->f_mapping;
const sector_t sector = dax_iomap_sector(iomap, pos);
struct dax_device *dax_dev = iomap->dax_dev;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
struct block_device *bdev = iomap->bdev;
struct inode *inode = mapping->host;
const size_t size = PMD_SIZE;
void *ret = NULL, *kaddr;
long length = 0;
pgoff_t pgoff;
pfn_t pfn;
int id;
if (bdev_dax_pgoff(bdev, sector, size, &pgoff) != 0)
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
id = dax_read_lock();
length = dax_direct_access(dax_dev, pgoff, PHYS_PFN(size), &kaddr, &pfn);
if (length < 0)
goto unlock_fallback;
length = PFN_PHYS(length);
if (length < size)
goto unlock_fallback;
if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR)
goto unlock_fallback;
if (!pfn_t_devmap(pfn))
goto unlock_fallback;
dax_read_unlock(id);
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
ret = dax_insert_mapping_entry(mapping, vmf, entry, sector,
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
RADIX_DAX_PMD);
if (IS_ERR(ret))
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
trace_dax_pmd_insert_mapping(inode, vmf, length, pfn, ret);
return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd,
pfn, vmf->flags & FAULT_FLAG_WRITE);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
unlock_fallback:
dax_read_unlock(id);
fallback:
trace_dax_pmd_insert_mapping_fallback(inode, vmf, length, pfn, ret);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return VM_FAULT_FALLBACK;
}
static int dax_pmd_load_hole(struct vm_fault *vmf, struct iomap *iomap,
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
void *entry)
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
{
struct address_space *mapping = vmf->vma->vm_file->f_mapping;
unsigned long pmd_addr = vmf->address & PMD_MASK;
struct inode *inode = mapping->host;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
struct page *zero_page;
void *ret = NULL;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
spinlock_t *ptl;
pmd_t pmd_entry;
zero_page = mm_get_huge_zero_page(vmf->vma->vm_mm);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (unlikely(!zero_page))
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
ret = dax_insert_mapping_entry(mapping, vmf, entry, 0,
RADIX_DAX_PMD | RADIX_DAX_ZERO_PAGE);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (IS_ERR(ret))
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
if (!pmd_none(*(vmf->pmd))) {
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
spin_unlock(ptl);
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
}
pmd_entry = mk_pmd(zero_page, vmf->vma->vm_page_prot);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
pmd_entry = pmd_mkhuge(pmd_entry);
set_pmd_at(vmf->vma->vm_mm, pmd_addr, vmf->pmd, pmd_entry);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
spin_unlock(ptl);
trace_dax_pmd_load_hole(inode, vmf, zero_page, ret);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return VM_FAULT_NOPAGE;
fallback:
trace_dax_pmd_load_hole_fallback(inode, vmf, zero_page, ret);
return VM_FAULT_FALLBACK;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
}
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
static int dax_iomap_pmd_fault(struct vm_fault *vmf,
const struct iomap_ops *ops)
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
{
struct vm_area_struct *vma = vmf->vma;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
struct address_space *mapping = vma->vm_file->f_mapping;
unsigned long pmd_addr = vmf->address & PMD_MASK;
bool write = vmf->flags & FAULT_FLAG_WRITE;
unsigned int iomap_flags = (write ? IOMAP_WRITE : 0) | IOMAP_FAULT;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
struct inode *inode = mapping->host;
int result = VM_FAULT_FALLBACK;
struct iomap iomap = { 0 };
pgoff_t max_pgoff, pgoff;
void *entry;
loff_t pos;
int error;
dax: add tracepoint infrastructure, PMD tracing Tracepoints are the standard way to capture debugging and tracing information in many parts of the kernel, including the XFS and ext4 filesystems. Create a tracepoint header for FS DAX and add the first DAX tracepoints to the PMD fault handler. This allows the tracing for DAX to be done in the same way as the filesystem tracing so that developers can look at them together and get a coherent idea of what the system is doing. I added both an entry and exit tracepoint because future patches will add tracepoints to child functions of dax_iomap_pmd_fault() like dax_pmd_load_hole() and dax_pmd_insert_mapping(). We want those messages to be wrapped by the parent function tracepoints so the code flow is more easily understood. Having entry and exit tracepoints for faults also allows us to easily see what filesystems functions were called during the fault. These filesystem functions get executed via iomap_begin() and iomap_end() calls, for example, and will have their own tracepoints. For PMD faults we primarily want to understand the type of mapping, the fault flags, the faulting address and whether it fell back to 4k faults. If it fell back to 4k faults the tracepoints should let us understand why. I named the new tracepoint header file "fs_dax.h" to allow for device DAX to have its own separate tracing header in the same directory at some point. Here is an example output for these events from a successful PMD fault: big-1441 [005] .... 32.582758: xfs_filemap_pmd_fault: dev 259:0 ino 0x1003 big-1441 [005] .... 32.582776: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 big-1441 [005] .... 32.583292: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 NOPAGE Link: http://lkml.kernel.org/r/1484085142-2297-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dave Chinner <david@fromorbit.com> Reviewed-by: Jan Kara <jack@suse.cz> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:39:50 +00:00
/*
* Check whether offset isn't beyond end of file now. Caller is
* supposed to hold locks serializing us with truncate / punch hole so
* this is a reliable test.
*/
pgoff = linear_page_index(vma, pmd_addr);
max_pgoff = (i_size_read(inode) - 1) >> PAGE_SHIFT;
trace_dax_pmd_fault(inode, vmf, max_pgoff, 0);
dax: add tracepoint infrastructure, PMD tracing Tracepoints are the standard way to capture debugging and tracing information in many parts of the kernel, including the XFS and ext4 filesystems. Create a tracepoint header for FS DAX and add the first DAX tracepoints to the PMD fault handler. This allows the tracing for DAX to be done in the same way as the filesystem tracing so that developers can look at them together and get a coherent idea of what the system is doing. I added both an entry and exit tracepoint because future patches will add tracepoints to child functions of dax_iomap_pmd_fault() like dax_pmd_load_hole() and dax_pmd_insert_mapping(). We want those messages to be wrapped by the parent function tracepoints so the code flow is more easily understood. Having entry and exit tracepoints for faults also allows us to easily see what filesystems functions were called during the fault. These filesystem functions get executed via iomap_begin() and iomap_end() calls, for example, and will have their own tracepoints. For PMD faults we primarily want to understand the type of mapping, the fault flags, the faulting address and whether it fell back to 4k faults. If it fell back to 4k faults the tracepoints should let us understand why. I named the new tracepoint header file "fs_dax.h" to allow for device DAX to have its own separate tracing header in the same directory at some point. Here is an example output for these events from a successful PMD fault: big-1441 [005] .... 32.582758: xfs_filemap_pmd_fault: dev 259:0 ino 0x1003 big-1441 [005] .... 32.582776: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 big-1441 [005] .... 32.583292: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 NOPAGE Link: http://lkml.kernel.org/r/1484085142-2297-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dave Chinner <david@fromorbit.com> Reviewed-by: Jan Kara <jack@suse.cz> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:39:50 +00:00
dax: fix deadlock due to misaligned PMD faults In DAX there are two separate places where the 2MiB range of a PMD is defined. The first is in the page tables, where a PMD mapping inserted for a given address spans from (vmf->address & PMD_MASK) to ((vmf->address & PMD_MASK) + PMD_SIZE - 1). That is, from the 2MiB boundary below the address to the 2MiB boundary above the address. So, for example, a fault at address 3MiB (0x30 0000) falls within the PMD that ranges from 2MiB (0x20 0000) to 4MiB (0x40 0000). The second PMD range is in the mapping->page_tree, where a given file offset is covered by a radix tree entry that spans from one 2MiB aligned file offset to another 2MiB aligned file offset. So, for example, the file offset for 3MiB (pgoff 768) falls within the PMD range for the order 9 radix tree entry that ranges from 2MiB (pgoff 512) to 4MiB (pgoff 1024). This system works so long as the addresses and file offsets for a given mapping both have the same offsets relative to the start of each PMD. Consider the case where the starting address for a given file isn't 2MiB aligned - say our faulting address is 3 MiB (0x30 0000), but that corresponds to the beginning of our file (pgoff 0). Now all the PMDs in the mapping are misaligned so that the 2MiB range defined in the page tables never matches up with the 2MiB range defined in the radix tree. The current code notices this case for DAX faults to storage with the following test in dax_pmd_insert_mapping(): if (pfn_t_to_pfn(pfn) & PG_PMD_COLOUR) goto unlock_fallback; This test makes sure that the pfn we get from the driver is 2MiB aligned, and relies on the assumption that the 2MiB alignment of the pfn we get back from the driver matches the 2MiB alignment of the faulting address. However, faults to holes were not checked and we could hit the problem described above. This was reported in response to the NVML nvml/src/test/pmempool_sync TEST5: $ cd nvml/src/test/pmempool_sync $ make TEST5 You can grab NVML here: https://github.com/pmem/nvml/ The dmesg warning you see when you hit this error is: WARNING: CPU: 13 PID: 2900 at fs/dax.c:641 dax_insert_mapping_entry+0x2df/0x310 Where we notice in dax_insert_mapping_entry() that the radix tree entry we are about to replace doesn't match the locked entry that we had previously inserted into the tree. This happens because the initial insertion was done in grab_mapping_entry() using a pgoff calculated from the faulting address (vmf->address), and the replacement in dax_pmd_load_hole() => dax_insert_mapping_entry() is done using vmf->pgoff. In our failure case those two page offsets (one calculated from vmf->address, one using vmf->pgoff) point to different order 9 radix tree entries. This failure case can result in a deadlock because the radix tree unlock also happens on the pgoff calculated from vmf->address. This means that the locked radix tree entry that we swapped in to the tree in dax_insert_mapping_entry() using vmf->pgoff is never unlocked, so all future faults to that 2MiB range will block forever. Fix this by validating that the faulting address's PMD offset matches the PMD offset from the start of the file. This check is done at the very beginning of the fault and covers faults that would have mapped to storage as well as faults to holes. I left the COLOUR check in dax_pmd_insert_mapping() in place in case we ever hit the insanity condition where the alignment of the pfn we get from the driver doesn't match the alignment of the userspace address. Link: http://lkml.kernel.org/r/20170822222436.18926-1-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reported-by: "Slusarz, Marcin" <marcin.slusarz@intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-08-25 22:55:36 +00:00
/*
* Make sure that the faulting address's PMD offset (color) matches
* the PMD offset from the start of the file. This is necessary so
* that a PMD range in the page table overlaps exactly with a PMD
* range in the radix tree.
*/
if ((vmf->pgoff & PG_PMD_COLOUR) !=
((vmf->address >> PAGE_SHIFT) & PG_PMD_COLOUR))
goto fallback;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/* Fall back to PTEs if we're going to COW */
if (write && !(vma->vm_flags & VM_SHARED))
goto fallback;
/* If the PMD would extend outside the VMA */
if (pmd_addr < vma->vm_start)
goto fallback;
if ((pmd_addr + PMD_SIZE) > vma->vm_end)
goto fallback;
dax: add tracepoint infrastructure, PMD tracing Tracepoints are the standard way to capture debugging and tracing information in many parts of the kernel, including the XFS and ext4 filesystems. Create a tracepoint header for FS DAX and add the first DAX tracepoints to the PMD fault handler. This allows the tracing for DAX to be done in the same way as the filesystem tracing so that developers can look at them together and get a coherent idea of what the system is doing. I added both an entry and exit tracepoint because future patches will add tracepoints to child functions of dax_iomap_pmd_fault() like dax_pmd_load_hole() and dax_pmd_insert_mapping(). We want those messages to be wrapped by the parent function tracepoints so the code flow is more easily understood. Having entry and exit tracepoints for faults also allows us to easily see what filesystems functions were called during the fault. These filesystem functions get executed via iomap_begin() and iomap_end() calls, for example, and will have their own tracepoints. For PMD faults we primarily want to understand the type of mapping, the fault flags, the faulting address and whether it fell back to 4k faults. If it fell back to 4k faults the tracepoints should let us understand why. I named the new tracepoint header file "fs_dax.h" to allow for device DAX to have its own separate tracing header in the same directory at some point. Here is an example output for these events from a successful PMD fault: big-1441 [005] .... 32.582758: xfs_filemap_pmd_fault: dev 259:0 ino 0x1003 big-1441 [005] .... 32.582776: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 big-1441 [005] .... 32.583292: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 NOPAGE Link: http://lkml.kernel.org/r/1484085142-2297-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dave Chinner <david@fromorbit.com> Reviewed-by: Jan Kara <jack@suse.cz> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:39:50 +00:00
if (pgoff > max_pgoff) {
result = VM_FAULT_SIGBUS;
goto out;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/* If the PMD would extend beyond the file size */
if ((pgoff | PG_PMD_COLOUR) > max_pgoff)
goto fallback;
/*
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
* grab_mapping_entry() will make sure we get a 2MiB empty entry, a
* 2MiB zero page entry or a DAX PMD. If it can't (because a 4k page
* is already in the tree, for instance), it will return -EEXIST and
* we just fall back to 4k entries.
*/
entry = grab_mapping_entry(mapping, pgoff, RADIX_DAX_PMD);
if (IS_ERR(entry))
goto fallback;
dax: fix race between colliding PMD & PTE entries We currently have two related PMD vs PTE races in the DAX code. These can both be easily triggered by having two threads reading and writing simultaneously to the same private mapping, with the key being that private mapping reads can be handled with PMDs but private mapping writes are always handled with PTEs so that we can COW. Here is the first race: CPU 0 CPU 1 (private mapping write) __handle_mm_fault() create_huge_pmd() - FALLBACK handle_pte_fault() passes check for pmd_devmap() (private mapping read) __handle_mm_fault() create_huge_pmd() dax_iomap_pmd_fault() inserts PMD dax_iomap_pte_fault() does a PTE fault, but we already have a DAX PMD installed in our page tables at this spot. Here's the second race: CPU 0 CPU 1 (private mapping read) __handle_mm_fault() passes check for pmd_none() create_huge_pmd() dax_iomap_pmd_fault() inserts PMD (private mapping write) __handle_mm_fault() create_huge_pmd() - FALLBACK (private mapping read) __handle_mm_fault() passes check for pmd_none() create_huge_pmd() handle_pte_fault() dax_iomap_pte_fault() inserts PTE dax_iomap_pmd_fault() inserts PMD, but we already have a PTE at this spot. The core of the issue is that while there is isolation between faults to the same range in the DAX fault handlers via our DAX entry locking, there is no isolation between faults in the code in mm/memory.c. This means for instance that this code in __handle_mm_fault() can run: if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) { ret = create_huge_pmd(&vmf); But by the time we actually get to run the fault handler called by create_huge_pmd(), the PMD is no longer pmd_none() because a racing PTE fault has installed a normal PMD here as a parent. This is the cause of the 2nd race. The first race is similar - there is the following check in handle_pte_fault(): } else { /* See comment in pte_alloc_one_map() */ if (pmd_devmap(*vmf->pmd) || pmd_trans_unstable(vmf->pmd)) return 0; So if a pmd_devmap() PMD (a DAX PMD) has been installed at vmf->pmd, we will bail and retry the fault. This is correct, but there is nothing preventing the PMD from being installed after this check but before we actually get to the DAX PTE fault handlers. In my testing these races result in the following types of errors: BUG: Bad rss-counter state mm:ffff8800a817d280 idx:1 val:1 BUG: non-zero nr_ptes on freeing mm: 15 Fix this issue by having the DAX fault handlers verify that it is safe to continue their fault after they have taken an entry lock to block other racing faults. [ross.zwisler@linux.intel.com: improve fix for colliding PMD & PTE entries] Link: http://lkml.kernel.org/r/20170526195932.32178-1-ross.zwisler@linux.intel.com Link: http://lkml.kernel.org/r/20170522215749.23516-2-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reported-by: Pawel Lebioda <pawel.lebioda@intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: "Kirill A . Shutemov" <kirill.shutemov@linux.intel.com> Cc: Pawel Lebioda <pawel.lebioda@intel.com> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Xiong Zhou <xzhou@redhat.com> Cc: Eryu Guan <eguan@redhat.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-06-02 21:46:37 +00:00
/*
* It is possible, particularly with mixed reads & writes to private
* mappings, that we have raced with a PTE fault that overlaps with
* the PMD we need to set up. If so just return and the fault will be
* retried.
*/
if (!pmd_none(*vmf->pmd) && !pmd_trans_huge(*vmf->pmd) &&
!pmd_devmap(*vmf->pmd)) {
result = 0;
goto unlock_entry;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
/*
* Note that we don't use iomap_apply here. We aren't doing I/O, only
* setting up a mapping, so really we're using iomap_begin() as a way
* to look up our filesystem block.
*/
pos = (loff_t)pgoff << PAGE_SHIFT;
error = ops->iomap_begin(inode, pos, PMD_SIZE, iomap_flags, &iomap);
if (error)
goto unlock_entry;
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
if (iomap.offset + iomap.length < pos + PMD_SIZE)
goto finish_iomap;
switch (iomap.type) {
case IOMAP_MAPPED:
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
result = dax_pmd_insert_mapping(vmf, &iomap, pos, entry);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
break;
case IOMAP_UNWRITTEN:
case IOMAP_HOLE:
if (WARN_ON_ONCE(write))
break;
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
result = dax_pmd_load_hole(vmf, &iomap, entry);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
break;
default:
WARN_ON_ONCE(1);
break;
}
finish_iomap:
if (ops->iomap_end) {
int copied = PMD_SIZE;
if (result == VM_FAULT_FALLBACK)
copied = 0;
/*
* The fault is done by now and there's no way back (other
* thread may be already happily using PMD we have installed).
* Just ignore error from ->iomap_end since we cannot do much
* with it.
*/
ops->iomap_end(inode, pos, PMD_SIZE, copied, iomap_flags,
&iomap);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
}
unlock_entry:
dax: use common 4k zero page for dax mmap reads When servicing mmap() reads from file holes the current DAX code allocates a page cache page of all zeroes and places the struct page pointer in the mapping->page_tree radix tree. This has three major drawbacks: 1) It consumes memory unnecessarily. For every 4k page that is read via a DAX mmap() over a hole, we allocate a new page cache page. This means that if you read 1GiB worth of pages, you end up using 1GiB of zeroed memory. This is easily visible by looking at the overall memory consumption of the system or by looking at /proc/[pid]/smaps: 7f62e72b3000-7f63272b3000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 1048576 kB Pss: 1048576 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 1048576 kB Private_Dirty: 0 kB Referenced: 1048576 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB 2) It is slower than using a common zero page because each page fault has more work to do. Instead of just inserting a common zero page we have to allocate a page cache page, zero it, and then insert it. Here are the average latencies of dax_load_hole() as measured by ftrace on a random test box: Old method, using zeroed page cache pages: 3.4 us New method, using the common 4k zero page: 0.8 us This was the average latency over 1 GiB of sequential reads done by this simple fio script: [global] size=1G filename=/root/dax/data fallocate=none [io] rw=read ioengine=mmap 3) The fact that we had to check for both DAX exceptional entries and for page cache pages in the radix tree made the DAX code more complex. Solve these issues by following the lead of the DAX PMD code and using a common 4k zero page instead. As with the PMD code we will now insert a DAX exceptional entry into the radix tree instead of a struct page pointer which allows us to remove all the special casing in the DAX code. Note that we do still pretty aggressively check for regular pages in the DAX radix tree, especially where we take action based on the bits set in the page. If we ever find a regular page in our radix tree now that most likely means that someone besides DAX is inserting pages (which has happened lots of times in the past), and we want to find that out early and fail loudly. This solution also removes the extra memory consumption. Here is that same /proc/[pid]/smaps after 1GiB of reading from a hole with the new code: 7f2054a74000-7f2094a74000 rw-s 00000000 103:00 12 /root/dax/data Size: 1048576 kB Rss: 0 kB Pss: 0 kB Shared_Clean: 0 kB Shared_Dirty: 0 kB Private_Clean: 0 kB Private_Dirty: 0 kB Referenced: 0 kB Anonymous: 0 kB LazyFree: 0 kB AnonHugePages: 0 kB ShmemPmdMapped: 0 kB Shared_Hugetlb: 0 kB Private_Hugetlb: 0 kB Swap: 0 kB SwapPss: 0 kB KernelPageSize: 4 kB MMUPageSize: 4 kB Locked: 0 kB Overall system memory consumption is similarly improved. Another major change is that we remove dax_pfn_mkwrite() from our fault flow, and instead rely on the page fault itself to make the PTE dirty and writeable. The following description from the patch adding the vm_insert_mixed_mkwrite() call explains this a little more: "To be able to use the common 4k zero page in DAX we need to have our PTE fault path look more like our PMD fault path where a PTE entry can be marked as dirty and writeable as it is first inserted rather than waiting for a follow-up dax_pfn_mkwrite() => finish_mkwrite_fault() call. Right now we can rely on having a dax_pfn_mkwrite() call because we can distinguish between these two cases in do_wp_page(): case 1: 4k zero page => writable DAX storage case 2: read-only DAX storage => writeable DAX storage This distinction is made by via vm_normal_page(). vm_normal_page() returns false for the common 4k zero page, though, just as it does for DAX ptes. Instead of special casing the DAX + 4k zero page case we will simplify our DAX PTE page fault sequence so that it matches our DAX PMD sequence, and get rid of the dax_pfn_mkwrite() helper. We will instead use dax_iomap_fault() to handle write-protection faults. This means that insert_pfn() needs to follow the lead of insert_pfn_pmd() and allow us to pass in a 'mkwrite' flag. If 'mkwrite' is set insert_pfn() will do the work that was previously done by wp_page_reuse() as part of the dax_pfn_mkwrite() call path" Link: http://lkml.kernel.org/r/20170724170616.25810-4-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Cc: "Darrick J. Wong" <darrick.wong@oracle.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Andreas Dilger <adilger.kernel@dilger.ca> Cc: Christoph Hellwig <hch@lst.de> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Matthew Wilcox <mawilcox@microsoft.com> Cc: Steven Rostedt <rostedt@goodmis.org> 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>
2017-09-06 23:18:43 +00:00
put_locked_mapping_entry(mapping, pgoff);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
fallback:
if (result == VM_FAULT_FALLBACK) {
split_huge_pmd(vma, vmf->pmd, vmf->address);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
count_vm_event(THP_FAULT_FALLBACK);
}
dax: add tracepoint infrastructure, PMD tracing Tracepoints are the standard way to capture debugging and tracing information in many parts of the kernel, including the XFS and ext4 filesystems. Create a tracepoint header for FS DAX and add the first DAX tracepoints to the PMD fault handler. This allows the tracing for DAX to be done in the same way as the filesystem tracing so that developers can look at them together and get a coherent idea of what the system is doing. I added both an entry and exit tracepoint because future patches will add tracepoints to child functions of dax_iomap_pmd_fault() like dax_pmd_load_hole() and dax_pmd_insert_mapping(). We want those messages to be wrapped by the parent function tracepoints so the code flow is more easily understood. Having entry and exit tracepoints for faults also allows us to easily see what filesystems functions were called during the fault. These filesystem functions get executed via iomap_begin() and iomap_end() calls, for example, and will have their own tracepoints. For PMD faults we primarily want to understand the type of mapping, the fault flags, the faulting address and whether it fell back to 4k faults. If it fell back to 4k faults the tracepoints should let us understand why. I named the new tracepoint header file "fs_dax.h" to allow for device DAX to have its own separate tracing header in the same directory at some point. Here is an example output for these events from a successful PMD fault: big-1441 [005] .... 32.582758: xfs_filemap_pmd_fault: dev 259:0 ino 0x1003 big-1441 [005] .... 32.582776: dax_pmd_fault: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 big-1441 [005] .... 32.583292: dax_pmd_fault_done: dev 259:0 ino 0x1003 shared WRITE|ALLOW_RETRY|KILLABLE|USER address 0x10505000 vm_start 0x10200000 vm_end 0x10700000 pgoff 0x200 max_pgoff 0x1400 NOPAGE Link: http://lkml.kernel.org/r/1484085142-2297-3-git-send-email-ross.zwisler@linux.intel.com Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dave Chinner <david@fromorbit.com> Reviewed-by: Jan Kara <jack@suse.cz> Acked-by: Steven Rostedt <rostedt@goodmis.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-02-22 23:39:50 +00:00
out:
trace_dax_pmd_fault_done(inode, vmf, max_pgoff, result);
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
return result;
}
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
#else
static int dax_iomap_pmd_fault(struct vm_fault *vmf,
const struct iomap_ops *ops)
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
{
return VM_FAULT_FALLBACK;
}
dax: add struct iomap based DAX PMD support DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based locking. This patch allows DAX PMDs to participate in the DAX radix tree based locking scheme so that they can be re-enabled using the new struct iomap based fault handlers. There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX mappings that have an associated block allocation, and 4k DAX empty entries. The empty entries exist to provide locking for the duration of a given page fault. This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP) entries, PMD DAX entries that have associated block allocations, and 2 MiB DAX empty entries. Unlike the 4k case where we insert a struct page* into the radix tree for 4k zero pages, for HZP we insert a DAX exceptional entry with the new RADIX_DAX_HZP flag set. This is because we use a single 2 MiB zero page in every 2MiB hole mapping, and it doesn't make sense to have that same struct page* with multiple entries in multiple trees. This would cause contention on the single page lock for the one Huge Zero Page, and it would break the page->index and page->mapping associations that are assumed to be valid in many other places in the kernel. One difficult use case is when one thread is trying to use 4k entries in radix tree for a given offset, and another thread is using 2 MiB entries for that same offset. The current code handles this by making the 2 MiB user fall back to 4k entries for most cases. This was done because it is the simplest solution, and because the use of 2MiB pages is already opportunistic. If we were to try to upgrade from 4k pages to 2MiB pages for a given range, we run into the problem of how we lock out 4k page faults for the entire 2MiB range while we clean out the radix tree so we can insert the 2MiB entry. We can solve this problem if we need to, but I think that the cases where both 2MiB entries and 4K entries are being used for the same range will be rare enough and the gain small enough that it probably won't be worth the complexity. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Dave Chinner <david@fromorbit.com>
2016-11-08 00:34:45 +00:00
#endif /* CONFIG_FS_DAX_PMD */
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
/**
* dax_iomap_fault - handle a page fault on a DAX file
* @vmf: The description of the fault
* @ops: iomap ops passed from the file system
*
* When a page fault occurs, filesystems may call this helper in
* their fault handler for DAX files. dax_iomap_fault() assumes the caller
* has done all the necessary locking for page fault to proceed
* successfully.
*/
int dax_iomap_fault(struct vm_fault *vmf, enum page_entry_size pe_size,
const struct iomap_ops *ops)
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
{
switch (pe_size) {
case PE_SIZE_PTE:
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
return dax_iomap_pte_fault(vmf, ops);
case PE_SIZE_PMD:
mm,fs,dax: change ->pmd_fault to ->huge_fault Patch series "1G transparent hugepage support for device dax", v2. The following series implements support for 1G trasparent hugepage on x86 for device dax. The bulk of the code was written by Mathew Wilcox a while back supporting transparent 1G hugepage for fs DAX. I have forward ported the relevant bits to 4.10-rc. The current submission has only the necessary code to support device DAX. Comments from Dan Williams: So the motivation and intended user of this functionality mirrors the motivation and users of 1GB page support in hugetlbfs. Given expected capacities of persistent memory devices an in-memory database may want to reduce tlb pressure beyond what they can already achieve with 2MB mappings of a device-dax file. We have customer feedback to that effect as Willy mentioned in his previous version of these patches [1]. [1]: https://lkml.org/lkml/2016/1/31/52 Comments from Nilesh @ Oracle: There are applications which have a process model; and if you assume 10,000 processes attempting to mmap all the 6TB memory available on a server; we are looking at the following: processes : 10,000 memory : 6TB pte @ 4k page size: 8 bytes / 4K of memory * #processes = 6TB / 4k * 8 * 10000 = 1.5GB * 80000 = 120,000GB pmd @ 2M page size: 120,000 / 512 = ~240GB pud @ 1G page size: 240GB / 512 = ~480MB As you can see with 2M pages, this system will use up an exorbitant amount of DRAM to hold the page tables; but the 1G pages finally brings it down to a reasonable level. Memory sizes will keep increasing; so this number will keep increasing. An argument can be made to convert the applications from process model to thread model, but in the real world that may not be always practical. Hopefully this helps explain the use case where this is valuable. This patch (of 3): In preparation for adding the ability to handle PUD pages, convert vm_operations_struct.pmd_fault to vm_operations_struct.huge_fault. The vm_fault structure is extended to include a union of the different page table pointers that may be needed, and three flag bits are reserved to indicate which type of pointer is in the union. [ross.zwisler@linux.intel.com: remove unused function ext4_dax_huge_fault()] Link: http://lkml.kernel.org/r/1485813172-7284-1-git-send-email-ross.zwisler@linux.intel.com [dave.jiang@intel.com: clear PMD or PUD size flags when in fall through path] Link: http://lkml.kernel.org/r/148589842696.5820.16078080610311444794.stgit@djiang5-desk3.ch.intel.com Link: http://lkml.kernel.org/r/148545058784.17912.6353162518188733642.stgit@djiang5-desk3.ch.intel.com Signed-off-by: Matthew Wilcox <mawilcox@microsoft.com> Signed-off-by: Dave Jiang <dave.jiang@intel.com> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Jan Kara <jack@suse.com> Cc: Dan Williams <dan.j.williams@intel.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Nilesh Choudhury <nilesh.choudhury@oracle.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> 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>
2017-02-24 22:56:59 +00:00
return dax_iomap_pmd_fault(vmf, ops);
default:
return VM_FAULT_FALLBACK;
}
}
EXPORT_SYMBOL_GPL(dax_iomap_fault);