linux/fs/aio.c

2202 lines
54 KiB
C
Raw Normal View History

/*
* An async IO implementation for Linux
* Written by Benjamin LaHaise <bcrl@kvack.org>
*
* Implements an efficient asynchronous io interface.
*
* Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
* Copyright 2018 Christoph Hellwig.
*
* See ../COPYING for licensing terms.
*/
#define pr_fmt(fmt) "%s: " fmt, __func__
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/errno.h>
#include <linux/time.h>
#include <linux/aio_abi.h>
#include <linux/export.h>
#include <linux/syscalls.h>
#include <linux/backing-dev.h>
#include <linux/refcount.h>
#include <linux/uio.h>
#include <linux/sched/signal.h>
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/mm.h>
#include <linux/mman.h>
#include <linux/mmu_context.h>
#include <linux/percpu.h>
#include <linux/slab.h>
#include <linux/timer.h>
#include <linux/aio.h>
#include <linux/highmem.h>
#include <linux/workqueue.h>
#include <linux/security.h>
signal/timer/event: KAIO eventfd support example This is an example about how to add eventfd support to the current KAIO code, in order to enable KAIO to post readiness events to a pollable fd (hence compatible with POSIX select/poll). The KAIO code simply signals the eventfd fd when events are ready, and this triggers a POLLIN in the fd. This patch uses a reserved for future use member of the struct iocb to pass an eventfd file descriptor, that KAIO will use to post events every time a request completes. At that point, an aio_getevents() will return the completed result to a struct io_event. I made a quick test program to verify the patch, and it runs fine here: http://www.xmailserver.org/eventfd-aio-test.c The test program uses poll(2), but it'd, of course, work with select and epoll too. This can allow to schedule both block I/O and other poll-able devices requests, and wait for results using select/poll/epoll. In a typical scenario, an application would submit KAIO request using aio_submit(), and will also use epoll_ctl() on the whole other class of devices (that with the addition of signals, timers and user events, now it's pretty much complete), and then would: epoll_wait(...); for_each_event { if (curr_event_is_kaiofd) { aio_getevents(); dispatch_aio_events(); } else { dispatch_epoll_event(); } } Signed-off-by: Davide Libenzi <davidel@xmailserver.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-11 05:23:21 +00:00
#include <linux/eventfd.h>
#include <linux/blkdev.h>
#include <linux/compat.h>
#include <linux/migrate.h>
#include <linux/ramfs.h>
#include <linux/percpu-refcount.h>
#include <linux/mount.h>
#include <asm/kmap_types.h>
#include <linux/uaccess.h>
#include "internal.h"
#define KIOCB_KEY 0
#define AIO_RING_MAGIC 0xa10a10a1
#define AIO_RING_COMPAT_FEATURES 1
#define AIO_RING_INCOMPAT_FEATURES 0
struct aio_ring {
unsigned id; /* kernel internal index number */
unsigned nr; /* number of io_events */
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
unsigned head; /* Written to by userland or under ring_lock
* mutex by aio_read_events_ring(). */
unsigned tail;
unsigned magic;
unsigned compat_features;
unsigned incompat_features;
unsigned header_length; /* size of aio_ring */
struct io_event io_events[0];
}; /* 128 bytes + ring size */
#define AIO_RING_PAGES 8
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
struct kioctx_table {
struct rcu_head rcu;
unsigned nr;
struct kioctx __rcu *table[];
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
};
struct kioctx_cpu {
unsigned reqs_available;
};
struct ctx_rq_wait {
struct completion comp;
atomic_t count;
};
struct kioctx {
struct percpu_ref users;
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
atomic_t dead;
struct percpu_ref reqs;
unsigned long user_id;
struct __percpu kioctx_cpu *cpu;
/*
* For percpu reqs_available, number of slots we move to/from global
* counter at a time:
*/
unsigned req_batch;
/*
* This is what userspace passed to io_setup(), it's not used for
* anything but counting against the global max_reqs quota.
*
* The real limit is nr_events - 1, which will be larger (see
* aio_setup_ring())
*/
unsigned max_reqs;
/* Size of ringbuffer, in units of struct io_event */
unsigned nr_events;
unsigned long mmap_base;
unsigned long mmap_size;
struct page **ring_pages;
long nr_pages;
struct rcu_work free_rwork; /* see free_ioctx() */
/*
* signals when all in-flight requests are done
*/
struct ctx_rq_wait *rq_wait;
struct {
/*
* This counts the number of available slots in the ringbuffer,
* so we avoid overflowing it: it's decremented (if positive)
* when allocating a kiocb and incremented when the resulting
* io_event is pulled off the ringbuffer.
*
* We batch accesses to it with a percpu version.
*/
atomic_t reqs_available;
} ____cacheline_aligned_in_smp;
struct {
spinlock_t ctx_lock;
struct list_head active_reqs; /* used for cancellation */
} ____cacheline_aligned_in_smp;
struct {
struct mutex ring_lock;
wait_queue_head_t wait;
} ____cacheline_aligned_in_smp;
struct {
unsigned tail;
unsigned completed_events;
spinlock_t completion_lock;
} ____cacheline_aligned_in_smp;
struct page *internal_pages[AIO_RING_PAGES];
struct file *aio_ring_file;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
unsigned id;
};
struct fsync_iocb {
struct work_struct work;
struct file *file;
bool datasync;
};
struct poll_iocb {
struct file *file;
struct wait_queue_head *head;
__poll_t events;
bool woken;
bool cancelled;
struct wait_queue_entry wait;
struct work_struct work;
};
struct aio_kiocb {
union {
struct kiocb rw;
struct fsync_iocb fsync;
struct poll_iocb poll;
};
struct kioctx *ki_ctx;
kiocb_cancel_fn *ki_cancel;
struct iocb __user *ki_user_iocb; /* user's aiocb */
__u64 ki_user_data; /* user's data for completion */
struct list_head ki_list; /* the aio core uses this
* for cancellation */
refcount_t ki_refcnt;
/*
* If the aio_resfd field of the userspace iocb is not zero,
* this is the underlying eventfd context to deliver events to.
*/
struct eventfd_ctx *ki_eventfd;
};
/*------ sysctl variables----*/
static DEFINE_SPINLOCK(aio_nr_lock);
unsigned long aio_nr; /* current system wide number of aio requests */
unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
/*----end sysctl variables---*/
static struct kmem_cache *kiocb_cachep;
static struct kmem_cache *kioctx_cachep;
static struct vfsmount *aio_mnt;
static const struct file_operations aio_ring_fops;
static const struct address_space_operations aio_ctx_aops;
static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
{
struct file *file;
struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
if (IS_ERR(inode))
return ERR_CAST(inode);
inode->i_mapping->a_ops = &aio_ctx_aops;
inode->i_mapping->private_data = ctx;
inode->i_size = PAGE_SIZE * nr_pages;
file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
O_RDWR, &aio_ring_fops);
if (IS_ERR(file))
iput(inode);
return file;
}
static struct dentry *aio_mount(struct file_system_type *fs_type,
int flags, const char *dev_name, void *data)
{
struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
AIO_RING_MAGIC);
if (!IS_ERR(root))
root->d_sb->s_iflags |= SB_I_NOEXEC;
return root;
}
/* aio_setup
* Creates the slab caches used by the aio routines, panic on
* failure as this is done early during the boot sequence.
*/
static int __init aio_setup(void)
{
static struct file_system_type aio_fs = {
.name = "aio",
.mount = aio_mount,
.kill_sb = kill_anon_super,
};
aio_mnt = kern_mount(&aio_fs);
if (IS_ERR(aio_mnt))
panic("Failed to create aio fs mount.");
kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
return 0;
}
__initcall(aio_setup);
static void put_aio_ring_file(struct kioctx *ctx)
{
struct file *aio_ring_file = ctx->aio_ring_file;
struct address_space *i_mapping;
if (aio_ring_file) {
truncate_setsize(file_inode(aio_ring_file), 0);
/* Prevent further access to the kioctx from migratepages */
i_mapping = aio_ring_file->f_mapping;
spin_lock(&i_mapping->private_lock);
i_mapping->private_data = NULL;
ctx->aio_ring_file = NULL;
spin_unlock(&i_mapping->private_lock);
fput(aio_ring_file);
}
}
static void aio_free_ring(struct kioctx *ctx)
{
int i;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* Disconnect the kiotx from the ring file. This prevents future
* accesses to the kioctx from page migration.
*/
put_aio_ring_file(ctx);
for (i = 0; i < ctx->nr_pages; i++) {
struct page *page;
pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
page_count(ctx->ring_pages[i]));
page = ctx->ring_pages[i];
if (!page)
continue;
ctx->ring_pages[i] = NULL;
put_page(page);
}
if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
kfree(ctx->ring_pages);
ctx->ring_pages = NULL;
}
}
static int aio_ring_mremap(struct vm_area_struct *vma)
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
{
struct file *file = vma->vm_file;
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
struct mm_struct *mm = vma->vm_mm;
struct kioctx_table *table;
int i, res = -EINVAL;
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
spin_lock(&mm->ioctx_lock);
rcu_read_lock();
table = rcu_dereference(mm->ioctx_table);
for (i = 0; i < table->nr; i++) {
struct kioctx *ctx;
ctx = rcu_dereference(table->table[i]);
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
if (ctx && ctx->aio_ring_file == file) {
if (!atomic_read(&ctx->dead)) {
ctx->user_id = ctx->mmap_base = vma->vm_start;
res = 0;
}
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
break;
}
}
rcu_read_unlock();
spin_unlock(&mm->ioctx_lock);
return res;
aio: Make it possible to remap aio ring There are actually two issues this patch addresses. Let me start with the one I tried to solve in the beginning. So, in the checkpoint-restore project (criu) we try to dump tasks' state and restore one back exactly as it was. One of the tasks' state bits is rings set up with io_setup() call. There's (almost) no problems in dumping them, there's a problem restoring them -- if I dump a task with aio ring originally mapped at address A, I want to restore one back at exactly the same address A. Unfortunately, the io_setup() does not allow for that -- it mmaps the ring at whatever place mm finds appropriate (it calls do_mmap_pgoff() with zero address and without the MAP_FIXED flag). To make restore possible I'm going to mremap() the freshly created ring into the address A (under which it was seen before dump). The problem is that the ring's virtual address is passed back to the user-space as the context ID and this ID is then used as search key by all the other io_foo() calls. Reworking this ID to be just some integer doesn't seem to work, as this value is already used by libaio as a pointer using which this library accesses memory for aio meta-data. So, to make restore work we need to make sure that a) ring is mapped at desired virtual address b) kioctx->user_id matches this value Having said that, the patch makes mremap() on aio region update the kioctx's user_id and mmap_base values. Here appears the 2nd issue I mentioned in the beginning of this mail. If (regardless of the C/R dances I do) someone creates an io context with io_setup(), then mremap()-s the ring and then destroys the context, the kill_ioctx() routine will call munmap() on wrong (old) address. This will result in a) aio ring remaining in memory and b) some other vma get unexpectedly unmapped. What do you think? Signed-off-by: Pavel Emelyanov <xemul@parallels.com> Acked-by: Dmitry Monakhov <dmonakhov@openvz.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-09-18 15:56:17 +00:00
}
static const struct vm_operations_struct aio_ring_vm_ops = {
.mremap = aio_ring_mremap,
#if IS_ENABLED(CONFIG_MMU)
.fault = filemap_fault,
.map_pages = filemap_map_pages,
.page_mkwrite = filemap_page_mkwrite,
#endif
};
static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
{
vma->vm_flags |= VM_DONTEXPAND;
vma->vm_ops = &aio_ring_vm_ops;
return 0;
}
static const struct file_operations aio_ring_fops = {
.mmap = aio_ring_mmap,
};
#if IS_ENABLED(CONFIG_MIGRATION)
static int aio_migratepage(struct address_space *mapping, struct page *new,
struct page *old, enum migrate_mode mode)
{
struct kioctx *ctx;
unsigned long flags;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
pgoff_t idx;
int rc;
mm/migrate: new migrate mode MIGRATE_SYNC_NO_COPY Introduce a new migration mode that allow to offload the copy to a device DMA engine. This changes the workflow of migration and not all address_space migratepage callback can support this. This is intended to be use by migrate_vma() which itself is use for thing like HMM (see include/linux/hmm.h). No additional per-filesystem migratepage testing is needed. I disables MIGRATE_SYNC_NO_COPY in all problematic migratepage() callback and i added comment in those to explain why (part of this patch). The commit message is unclear it should say that any callback that wish to support this new mode need to be aware of the difference in the migration flow from other mode. Some of these callbacks do extra locking while copying (aio, zsmalloc, balloon, ...) and for DMA to be effective you want to copy multiple pages in one DMA operations. But in the problematic case you can not easily hold the extra lock accross multiple call to this callback. Usual flow is: For each page { 1 - lock page 2 - call migratepage() callback 3 - (extra locking in some migratepage() callback) 4 - migrate page state (freeze refcount, update page cache, buffer head, ...) 5 - copy page 6 - (unlock any extra lock of migratepage() callback) 7 - return from migratepage() callback 8 - unlock page } The new mode MIGRATE_SYNC_NO_COPY: 1 - lock multiple pages For each page { 2 - call migratepage() callback 3 - abort in all problematic migratepage() callback 4 - migrate page state (freeze refcount, update page cache, buffer head, ...) } // finished all calls to migratepage() callback 5 - DMA copy multiple pages 6 - unlock all the pages To support MIGRATE_SYNC_NO_COPY in the problematic case we would need a new callback migratepages() (for instance) that deals with multiple pages in one transaction. Because the problematic cases are not important for current usage I did not wanted to complexify this patchset even more for no good reason. Link: http://lkml.kernel.org/r/20170817000548.32038-14-jglisse@redhat.com Signed-off-by: Jérôme Glisse <jglisse@redhat.com> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Cc: Balbir Singh <bsingharora@gmail.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Dan Williams <dan.j.williams@intel.com> Cc: David Nellans <dnellans@nvidia.com> Cc: Evgeny Baskakov <ebaskakov@nvidia.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: John Hubbard <jhubbard@nvidia.com> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Mark Hairgrove <mhairgrove@nvidia.com> Cc: Michal Hocko <mhocko@kernel.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Ross Zwisler <ross.zwisler@linux.intel.com> Cc: Sherry Cheung <SCheung@nvidia.com> Cc: Subhash Gutti <sgutti@nvidia.com> Cc: Vladimir Davydov <vdavydov.dev@gmail.com> Cc: Bob Liu <liubo95@huawei.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-09-08 23:12:06 +00:00
/*
* We cannot support the _NO_COPY case here, because copy needs to
* happen under the ctx->completion_lock. That does not work with the
* migration workflow of MIGRATE_SYNC_NO_COPY.
*/
if (mode == MIGRATE_SYNC_NO_COPY)
return -EINVAL;
rc = 0;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* mapping->private_lock here protects against the kioctx teardown. */
spin_lock(&mapping->private_lock);
ctx = mapping->private_data;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
if (!ctx) {
rc = -EINVAL;
goto out;
}
/* The ring_lock mutex. The prevents aio_read_events() from writing
* to the ring's head, and prevents page migration from mucking in
* a partially initialized kiotx.
*/
if (!mutex_trylock(&ctx->ring_lock)) {
rc = -EAGAIN;
goto out;
}
idx = old->index;
if (idx < (pgoff_t)ctx->nr_pages) {
/* Make sure the old page hasn't already been changed */
if (ctx->ring_pages[idx] != old)
rc = -EAGAIN;
} else
rc = -EINVAL;
if (rc != 0)
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
goto out_unlock;
/* Writeback must be complete */
BUG_ON(PageWriteback(old));
get_page(new);
rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
if (rc != MIGRATEPAGE_SUCCESS) {
put_page(new);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
goto out_unlock;
}
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* Take completion_lock to prevent other writes to the ring buffer
* while the old page is copied to the new. This prevents new
* events from being lost.
*/
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
spin_lock_irqsave(&ctx->completion_lock, flags);
migrate_page_copy(new, old);
BUG_ON(ctx->ring_pages[idx] != old);
ctx->ring_pages[idx] = new;
spin_unlock_irqrestore(&ctx->completion_lock, flags);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* The old page is no longer accessible. */
put_page(old);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
out_unlock:
mutex_unlock(&ctx->ring_lock);
out:
spin_unlock(&mapping->private_lock);
return rc;
}
#endif
static const struct address_space_operations aio_ctx_aops = {
aio: fix uncorrent dirty pages accouting when truncating AIO ring buffer https://bugzilla.kernel.org/show_bug.cgi?id=86831 Markus reported that when shutting down mysqld (with AIO support, on a ext3 formatted Harddrive) leads to a negative number of dirty pages (underrun to the counter). The negative number results in a drastic reduction of the write performance because the page cache is not used, because the kernel thinks it is still 2 ^ 32 dirty pages open. Add a warn trace in __dec_zone_state will catch this easily: static inline void __dec_zone_state(struct zone *zone, enum zone_stat_item item) { atomic_long_dec(&zone->vm_stat[item]); + WARN_ON_ONCE(item == NR_FILE_DIRTY && atomic_long_read(&zone->vm_stat[item]) < 0); atomic_long_dec(&vm_stat[item]); } [ 21.341632] ------------[ cut here ]------------ [ 21.346294] WARNING: CPU: 0 PID: 309 at include/linux/vmstat.h:242 cancel_dirty_page+0x164/0x224() [ 21.355296] Modules linked in: wutbox_cp sata_mv [ 21.359968] CPU: 0 PID: 309 Comm: kworker/0:1 Not tainted 3.14.21-WuT #80 [ 21.366793] Workqueue: events free_ioctx [ 21.370760] [<c0016a64>] (unwind_backtrace) from [<c0012f88>] (show_stack+0x20/0x24) [ 21.378562] [<c0012f88>] (show_stack) from [<c03f8ccc>] (dump_stack+0x24/0x28) [ 21.385840] [<c03f8ccc>] (dump_stack) from [<c0023ae4>] (warn_slowpath_common+0x84/0x9c) [ 21.393976] [<c0023ae4>] (warn_slowpath_common) from [<c0023bb8>] (warn_slowpath_null+0x2c/0x34) [ 21.402800] [<c0023bb8>] (warn_slowpath_null) from [<c00c0688>] (cancel_dirty_page+0x164/0x224) [ 21.411524] [<c00c0688>] (cancel_dirty_page) from [<c00c080c>] (truncate_inode_page+0x8c/0x158) [ 21.420272] [<c00c080c>] (truncate_inode_page) from [<c00c0a94>] (truncate_inode_pages_range+0x11c/0x53c) [ 21.429890] [<c00c0a94>] (truncate_inode_pages_range) from [<c00c0f6c>] (truncate_pagecache+0x88/0xac) [ 21.439252] [<c00c0f6c>] (truncate_pagecache) from [<c00c0fec>] (truncate_setsize+0x5c/0x74) [ 21.447731] [<c00c0fec>] (truncate_setsize) from [<c013b3a8>] (put_aio_ring_file.isra.14+0x34/0x90) [ 21.456826] [<c013b3a8>] (put_aio_ring_file.isra.14) from [<c013b424>] (aio_free_ring+0x20/0xcc) [ 21.465660] [<c013b424>] (aio_free_ring) from [<c013b4f4>] (free_ioctx+0x24/0x44) [ 21.473190] [<c013b4f4>] (free_ioctx) from [<c003d8d8>] (process_one_work+0x134/0x47c) [ 21.481132] [<c003d8d8>] (process_one_work) from [<c003e988>] (worker_thread+0x130/0x414) [ 21.489350] [<c003e988>] (worker_thread) from [<c00448ac>] (kthread+0xd4/0xec) [ 21.496621] [<c00448ac>] (kthread) from [<c000ec18>] (ret_from_fork+0x14/0x20) [ 21.503884] ---[ end trace 79c4bf42c038c9a1 ]--- The cause is that we set the aio ring file pages as *DIRTY* via SetPageDirty (bypasses the VFS dirty pages increment) when init, and aio fs uses *default_backing_dev_info* as the backing dev, which does not disable the dirty pages accounting capability. So truncating aio ring file will contribute to accounting dirty pages (VFS dirty pages decrement), then error occurs. The original goal is keeping these pages in memory (can not be reclaimed or swapped) in life-time via marking it dirty. But thinking more, we have already pinned pages via elevating the page's refcount, which can already achieve the goal, so the SetPageDirty seems unnecessary. In order to fix the issue, using the __set_page_dirty_no_writeback instead of the nop .set_page_dirty, and dropped the SetPageDirty (don't manually set the dirty flags, don't disable set_page_dirty(), rely on default behaviour). With the above change, the dirty pages accounting can work well. But as we known, aio fs is an anonymous one, which should never cause any real write-back, we can ignore the dirty pages (write back) accounting by disabling the dirty pages (write back) accounting capability. So we introduce an aio private backing dev info (disabled the ACCT_DIRTY/WRITEBACK/ACCT_WB capabilities) to replace the default one. Reported-by: Markus Königshaus <m.koenigshaus@wut.de> Signed-off-by: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable <stable@vger.kernel.org> Acked-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-11-06 09:46:21 +00:00
.set_page_dirty = __set_page_dirty_no_writeback,
#if IS_ENABLED(CONFIG_MIGRATION)
.migratepage = aio_migratepage,
#endif
};
static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
{
struct aio_ring *ring;
2013-05-07 23:18:25 +00:00
struct mm_struct *mm = current->mm;
unsigned long size, unused;
int nr_pages;
int i;
struct file *file;
/* Compensate for the ring buffer's head/tail overlap entry */
nr_events += 2; /* 1 is required, 2 for good luck */
size = sizeof(struct aio_ring);
size += sizeof(struct io_event) * nr_events;
nr_pages = PFN_UP(size);
if (nr_pages < 0)
return -EINVAL;
file = aio_private_file(ctx, nr_pages);
if (IS_ERR(file)) {
ctx->aio_ring_file = NULL;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
return -ENOMEM;
}
ctx->aio_ring_file = file;
nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
/ sizeof(struct io_event);
ctx->ring_pages = ctx->internal_pages;
if (nr_pages > AIO_RING_PAGES) {
ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
GFP_KERNEL);
if (!ctx->ring_pages) {
put_aio_ring_file(ctx);
return -ENOMEM;
}
}
for (i = 0; i < nr_pages; i++) {
struct page *page;
page = find_or_create_page(file->f_mapping,
i, GFP_HIGHUSER | __GFP_ZERO);
if (!page)
break;
pr_debug("pid(%d) page[%d]->count=%d\n",
current->pid, i, page_count(page));
SetPageUptodate(page);
unlock_page(page);
ctx->ring_pages[i] = page;
}
ctx->nr_pages = i;
if (unlikely(i != nr_pages)) {
aio_free_ring(ctx);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
return -ENOMEM;
}
ctx->mmap_size = nr_pages * PAGE_SIZE;
pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
if (down_write_killable(&mm->mmap_sem)) {
ctx->mmap_size = 0;
aio_free_ring(ctx);
return -EINTR;
}
ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
PROT_READ | PROT_WRITE,
MAP_SHARED, 0, &unused, NULL);
up_write(&mm->mmap_sem);
if (IS_ERR((void *)ctx->mmap_base)) {
ctx->mmap_size = 0;
aio_free_ring(ctx);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
return -ENOMEM;
}
pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
ctx->user_id = ctx->mmap_base;
ctx->nr_events = nr_events; /* trusted copy */
ring = kmap_atomic(ctx->ring_pages[0]);
ring->nr = nr_events; /* user copy */
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
ring->id = ~0U;
ring->head = ring->tail = 0;
ring->magic = AIO_RING_MAGIC;
ring->compat_features = AIO_RING_COMPAT_FEATURES;
ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
ring->header_length = sizeof(struct aio_ring);
kunmap_atomic(ring);
flush_dcache_page(ctx->ring_pages[0]);
return 0;
}
#define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
#define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
#define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
{
struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
struct kioctx *ctx = req->ki_ctx;
unsigned long flags;
if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
return;
spin_lock_irqsave(&ctx->ctx_lock, flags);
list_add_tail(&req->ki_list, &ctx->active_reqs);
req->ki_cancel = cancel;
spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
EXPORT_SYMBOL(kiocb_set_cancel_fn);
/*
* free_ioctx() should be RCU delayed to synchronize against the RCU
* protected lookup_ioctx() and also needs process context to call
* aio_free_ring(). Use rcu_work.
*/
static void free_ioctx(struct work_struct *work)
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
{
struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
free_rwork);
pr_debug("freeing %p\n", ctx);
aio_free_ring(ctx);
free_percpu(ctx->cpu);
percpu_ref_exit(&ctx->reqs);
percpu_ref_exit(&ctx->users);
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
kmem_cache_free(kioctx_cachep, ctx);
}
static void free_ioctx_reqs(struct percpu_ref *ref)
{
struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
/* At this point we know that there are no any in-flight requests */
if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
complete(&ctx->rq_wait->comp);
/* Synchronize against RCU protected table->table[] dereferences */
INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
queue_rcu_work(system_wq, &ctx->free_rwork);
}
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
/*
* When this function runs, the kioctx has been removed from the "hash table"
* and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
* now it's safe to cancel any that need to be.
*/
static void free_ioctx_users(struct percpu_ref *ref)
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
{
struct kioctx *ctx = container_of(ref, struct kioctx, users);
struct aio_kiocb *req;
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
spin_lock_irq(&ctx->ctx_lock);
while (!list_empty(&ctx->active_reqs)) {
req = list_first_entry(&ctx->active_reqs,
struct aio_kiocb, ki_list);
req->ki_cancel(&req->rw);
list_del_init(&req->ki_list);
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
}
spin_unlock_irq(&ctx->ctx_lock);
percpu_ref_kill(&ctx->reqs);
percpu_ref_put(&ctx->reqs);
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
}
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
{
unsigned i, new_nr;
struct kioctx_table *table, *old;
struct aio_ring *ring;
spin_lock(&mm->ioctx_lock);
table = rcu_dereference_raw(mm->ioctx_table);
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
while (1) {
if (table)
for (i = 0; i < table->nr; i++)
if (!rcu_access_pointer(table->table[i])) {
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
ctx->id = i;
rcu_assign_pointer(table->table[i], ctx);
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
spin_unlock(&mm->ioctx_lock);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* While kioctx setup is in progress,
* we are protected from page migration
* changes ring_pages by ->ring_lock.
*/
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
ring = kmap_atomic(ctx->ring_pages[0]);
ring->id = ctx->id;
kunmap_atomic(ring);
return 0;
}
new_nr = (table ? table->nr : 1) * 4;
spin_unlock(&mm->ioctx_lock);
table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
new_nr, GFP_KERNEL);
if (!table)
return -ENOMEM;
table->nr = new_nr;
spin_lock(&mm->ioctx_lock);
old = rcu_dereference_raw(mm->ioctx_table);
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
if (!old) {
rcu_assign_pointer(mm->ioctx_table, table);
} else if (table->nr > old->nr) {
memcpy(table->table, old->table,
old->nr * sizeof(struct kioctx *));
rcu_assign_pointer(mm->ioctx_table, table);
kfree_rcu(old, rcu);
} else {
kfree(table);
table = old;
}
}
}
static void aio_nr_sub(unsigned nr)
{
spin_lock(&aio_nr_lock);
if (WARN_ON(aio_nr - nr > aio_nr))
aio_nr = 0;
else
aio_nr -= nr;
spin_unlock(&aio_nr_lock);
}
/* ioctx_alloc
* Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
*/
static struct kioctx *ioctx_alloc(unsigned nr_events)
{
2013-05-07 23:18:25 +00:00
struct mm_struct *mm = current->mm;
struct kioctx *ctx;
int err = -ENOMEM;
/*
* Store the original nr_events -- what userspace passed to io_setup(),
* for counting against the global limit -- before it changes.
*/
unsigned int max_reqs = nr_events;
/*
* We keep track of the number of available ringbuffer slots, to prevent
* overflow (reqs_available), and we also use percpu counters for this.
*
* So since up to half the slots might be on other cpu's percpu counters
* and unavailable, double nr_events so userspace sees what they
* expected: additionally, we move req_batch slots to/from percpu
* counters at a time, so make sure that isn't 0:
*/
nr_events = max(nr_events, num_possible_cpus() * 4);
nr_events *= 2;
/* Prevent overflows */
if (nr_events > (0x10000000U / sizeof(struct io_event))) {
pr_debug("ENOMEM: nr_events too high\n");
return ERR_PTR(-EINVAL);
}
if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
return ERR_PTR(-EAGAIN);
ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
if (!ctx)
return ERR_PTR(-ENOMEM);
ctx->max_reqs = max_reqs;
spin_lock_init(&ctx->ctx_lock);
spin_lock_init(&ctx->completion_lock);
mutex_init(&ctx->ring_lock);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* Protect against page migration throughout kiotx setup by keeping
* the ring_lock mutex held until setup is complete. */
mutex_lock(&ctx->ring_lock);
init_waitqueue_head(&ctx->wait);
INIT_LIST_HEAD(&ctx->active_reqs);
if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
goto err;
if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
goto err;
ctx->cpu = alloc_percpu(struct kioctx_cpu);
if (!ctx->cpu)
goto err;
err = aio_setup_ring(ctx, nr_events);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
if (err < 0)
goto err;
atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
if (ctx->req_batch < 1)
ctx->req_batch = 1;
/* limit the number of system wide aios */
spin_lock(&aio_nr_lock);
if (aio_nr + ctx->max_reqs > aio_max_nr ||
aio_nr + ctx->max_reqs < aio_nr) {
spin_unlock(&aio_nr_lock);
err = -EAGAIN;
goto err_ctx;
}
aio_nr += ctx->max_reqs;
spin_unlock(&aio_nr_lock);
percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
err = ioctx_add_table(ctx, mm);
if (err)
goto err_cleanup;
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* Release the ring_lock mutex now that all setup is complete. */
mutex_unlock(&ctx->ring_lock);
pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
ctx, ctx->user_id, mm, ctx->nr_events);
return ctx;
err_cleanup:
aio_nr_sub(ctx->max_reqs);
err_ctx:
atomic_set(&ctx->dead, 1);
if (ctx->mmap_size)
vm_munmap(ctx->mmap_base, ctx->mmap_size);
aio_free_ring(ctx);
err:
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
mutex_unlock(&ctx->ring_lock);
free_percpu(ctx->cpu);
percpu_ref_exit(&ctx->reqs);
percpu_ref_exit(&ctx->users);
kmem_cache_free(kioctx_cachep, ctx);
pr_debug("error allocating ioctx %d\n", err);
return ERR_PTR(err);
}
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
/* kill_ioctx
* Cancels all outstanding aio requests on an aio context. Used
* when the processes owning a context have all exited to encourage
* the rapid destruction of the kioctx.
*/
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
struct ctx_rq_wait *wait)
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
{
struct kioctx_table *table;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
spin_lock(&mm->ioctx_lock);
if (atomic_xchg(&ctx->dead, 1)) {
spin_unlock(&mm->ioctx_lock);
return -EINVAL;
}
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
table = rcu_dereference_raw(mm->ioctx_table);
WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
RCU_INIT_POINTER(table->table[ctx->id], NULL);
spin_unlock(&mm->ioctx_lock);
aio: fix io_destroy() regression by using call_rcu() There was a regression introduced by 36f5588905c1 ("aio: refcounting cleanup"), reported by Jens Axboe - the refcounting cleanup switched to using RCU in the shutdown path, but the synchronize_rcu() was done in the context of the io_destroy() syscall greatly increasing the time it could block. This patch switches it to call_rcu() and makes shutdown asynchronous (more asynchronous than it was originally; before the refcount changes io_destroy() would still wait on pending kiocbs). Note that there's a global quota on the max outstanding kiocbs, and that quota must be manipulated synchronously; otherwise io_setup() could return -EAGAIN when there isn't quota available, and userspace won't have any way of waiting until shutdown of the old kioctxs has finished (besides busy looping). So we release our quota before kioctx shutdown has finished, which should be fine since the quota never corresponded to anything real anyways. Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Reported-by: Jens Axboe <axboe@kernel.dk> Tested-by: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Tested-by: Benjamin LaHaise <bcrl@kvack.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-06-12 21:04:59 +00:00
/* free_ioctx_reqs() will do the necessary RCU synchronization */
wake_up_all(&ctx->wait);
aio: fix io_destroy() regression by using call_rcu() There was a regression introduced by 36f5588905c1 ("aio: refcounting cleanup"), reported by Jens Axboe - the refcounting cleanup switched to using RCU in the shutdown path, but the synchronize_rcu() was done in the context of the io_destroy() syscall greatly increasing the time it could block. This patch switches it to call_rcu() and makes shutdown asynchronous (more asynchronous than it was originally; before the refcount changes io_destroy() would still wait on pending kiocbs). Note that there's a global quota on the max outstanding kiocbs, and that quota must be manipulated synchronously; otherwise io_setup() could return -EAGAIN when there isn't quota available, and userspace won't have any way of waiting until shutdown of the old kioctxs has finished (besides busy looping). So we release our quota before kioctx shutdown has finished, which should be fine since the quota never corresponded to anything real anyways. Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Reported-by: Jens Axboe <axboe@kernel.dk> Tested-by: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Tested-by: Benjamin LaHaise <bcrl@kvack.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-06-12 21:04:59 +00:00
/*
* It'd be more correct to do this in free_ioctx(), after all
* the outstanding kiocbs have finished - but by then io_destroy
* has already returned, so io_setup() could potentially return
* -EAGAIN with no ioctxs actually in use (as far as userspace
* could tell).
*/
aio_nr_sub(ctx->max_reqs);
aio: fix io_destroy() regression by using call_rcu() There was a regression introduced by 36f5588905c1 ("aio: refcounting cleanup"), reported by Jens Axboe - the refcounting cleanup switched to using RCU in the shutdown path, but the synchronize_rcu() was done in the context of the io_destroy() syscall greatly increasing the time it could block. This patch switches it to call_rcu() and makes shutdown asynchronous (more asynchronous than it was originally; before the refcount changes io_destroy() would still wait on pending kiocbs). Note that there's a global quota on the max outstanding kiocbs, and that quota must be manipulated synchronously; otherwise io_setup() could return -EAGAIN when there isn't quota available, and userspace won't have any way of waiting until shutdown of the old kioctxs has finished (besides busy looping). So we release our quota before kioctx shutdown has finished, which should be fine since the quota never corresponded to anything real anyways. Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Reported-by: Jens Axboe <axboe@kernel.dk> Tested-by: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Tested-by: Benjamin LaHaise <bcrl@kvack.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-06-12 21:04:59 +00:00
if (ctx->mmap_size)
vm_munmap(ctx->mmap_base, ctx->mmap_size);
ctx->rq_wait = wait;
percpu_ref_kill(&ctx->users);
return 0;
}
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
/*
* exit_aio: called when the last user of mm goes away. At this point, there is
* no way for any new requests to be submited or any of the io_* syscalls to be
* called on the context.
*
* There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
* them.
*/
void exit_aio(struct mm_struct *mm)
{
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
struct ctx_rq_wait wait;
int i, skipped;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
if (!table)
return;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
atomic_set(&wait.count, table->nr);
init_completion(&wait.comp);
skipped = 0;
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
for (i = 0; i < table->nr; ++i) {
struct kioctx *ctx =
rcu_dereference_protected(table->table[i], true);
if (!ctx) {
skipped++;
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
continue;
}
/*
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
* We don't need to bother with munmap() here - exit_mmap(mm)
* is coming and it'll unmap everything. And we simply can't,
* this is not necessarily our ->mm.
* Since kill_ioctx() uses non-zero ->mmap_size as indicator
* that it needs to unmap the area, just set it to 0.
*/
ctx->mmap_size = 0;
kill_ioctx(mm, ctx, &wait);
}
aio: refcounting cleanup The usage of ctx->dead was fubar - it makes no sense to explicitly check it all over the place, especially when we're already using RCU. Now, ctx->dead only indicates whether we've dropped the initial refcount. The new teardown sequence is: set ctx->dead hlist_del_rcu(); synchronize_rcu(); Now we know no system calls can take a new ref, and it's safe to drop the initial ref: put_ioctx(); We also need to ensure there are no more outstanding kiocbs. This was done incorrectly - it was being done in kill_ctx(), and before dropping the initial refcount. At this point, other syscalls may still be submitting kiocbs! Now, we cancel and wait for outstanding kiocbs in free_ioctx(), after kioctx->users has dropped to 0 and we know no more iocbs could be submitted. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Kent Overstreet <koverstreet@google.com> Cc: Zach Brown <zab@redhat.com> Cc: Felipe Balbi <balbi@ti.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Mark Fasheh <mfasheh@suse.com> Cc: Joel Becker <jlbec@evilplan.org> Cc: Rusty Russell <rusty@rustcorp.com.au> Cc: Jens Axboe <axboe@kernel.dk> Cc: Asai Thambi S P <asamymuthupa@micron.com> Cc: Selvan Mani <smani@micron.com> Cc: Sam Bradshaw <sbradshaw@micron.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Benjamin LaHaise <bcrl@kvack.org> Reviewed-by: "Theodore Ts'o" <tytso@mit.edu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-07 23:18:41 +00:00
if (!atomic_sub_and_test(skipped, &wait.count)) {
/* Wait until all IO for the context are done. */
wait_for_completion(&wait.comp);
}
aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() On 04/30, Benjamin LaHaise wrote: > > > - ctx->mmap_size = 0; > > - > > - kill_ioctx(mm, ctx, NULL); > > + if (ctx) { > > + ctx->mmap_size = 0; > > + kill_ioctx(mm, ctx, NULL); > > + } > > Rather than indenting and moving the two lines changing mmap_size and the > kill_ioctx() call, why not just do "if (!ctx) ... continue;"? That reduces > the number of lines changed and avoid excessive indentation. OK. To me the code looks better/simpler with "if (ctx)", but this is subjective of course, I won't argue. The patch still removes the empty line between mmap_size = 0 and kill_ioctx(), we reset mmap_size only for kill_ioctx(). But feel free to remove this change. ------------------------------------------------------------------------------- Subject: [PATCH v3 1/2] aio: change exit_aio() to load mm->ioctx_table once and avoid rcu_read_lock() 1. We can read ->ioctx_table only once and we do not read rcu_read_lock() or even rcu_dereference(). This mm has no users, nobody else can play with ->ioctx_table. Otherwise the code is buggy anyway, if we need rcu_read_lock() in a loop because ->ioctx_table can be updated then kfree(table) is obviously wrong. 2. Update the comment. "exit_mmap(mm) is coming" is the good reason to avoid munmap(), but another reason is that we simply can't do vm_munmap() unless current->mm == mm and this is not true in general, the caller is mmput(). 3. We do not really need to nullify mm->ioctx_table before return, probably the current code does this to catch the potential problems. But in this case RCU_INIT_POINTER(NULL) looks better. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2014-04-30 17:02:48 +00:00
RCU_INIT_POINTER(mm->ioctx_table, NULL);
kfree(table);
}
static void put_reqs_available(struct kioctx *ctx, unsigned nr)
{
struct kioctx_cpu *kcpu;
unsigned long flags;
local_irq_save(flags);
kcpu = this_cpu_ptr(ctx->cpu);
kcpu->reqs_available += nr;
while (kcpu->reqs_available >= ctx->req_batch * 2) {
kcpu->reqs_available -= ctx->req_batch;
atomic_add(ctx->req_batch, &ctx->reqs_available);
}
local_irq_restore(flags);
}
static bool get_reqs_available(struct kioctx *ctx)
{
struct kioctx_cpu *kcpu;
bool ret = false;
unsigned long flags;
local_irq_save(flags);
kcpu = this_cpu_ptr(ctx->cpu);
if (!kcpu->reqs_available) {
int old, avail = atomic_read(&ctx->reqs_available);
do {
if (avail < ctx->req_batch)
goto out;
old = avail;
avail = atomic_cmpxchg(&ctx->reqs_available,
avail, avail - ctx->req_batch);
} while (avail != old);
kcpu->reqs_available += ctx->req_batch;
}
ret = true;
kcpu->reqs_available--;
out:
local_irq_restore(flags);
return ret;
}
/* refill_reqs_available
* Updates the reqs_available reference counts used for tracking the
* number of free slots in the completion ring. This can be called
* from aio_complete() (to optimistically update reqs_available) or
* from aio_get_req() (the we're out of events case). It must be
* called holding ctx->completion_lock.
*/
static void refill_reqs_available(struct kioctx *ctx, unsigned head,
unsigned tail)
{
unsigned events_in_ring, completed;
/* Clamp head since userland can write to it. */
head %= ctx->nr_events;
if (head <= tail)
events_in_ring = tail - head;
else
events_in_ring = ctx->nr_events - (head - tail);
completed = ctx->completed_events;
if (events_in_ring < completed)
completed -= events_in_ring;
else
completed = 0;
if (!completed)
return;
ctx->completed_events -= completed;
put_reqs_available(ctx, completed);
}
/* user_refill_reqs_available
* Called to refill reqs_available when aio_get_req() encounters an
* out of space in the completion ring.
*/
static void user_refill_reqs_available(struct kioctx *ctx)
{
spin_lock_irq(&ctx->completion_lock);
if (ctx->completed_events) {
struct aio_ring *ring;
unsigned head;
/* Access of ring->head may race with aio_read_events_ring()
* here, but that's okay since whether we read the old version
* or the new version, and either will be valid. The important
* part is that head cannot pass tail since we prevent
* aio_complete() from updating tail by holding
* ctx->completion_lock. Even if head is invalid, the check
* against ctx->completed_events below will make sure we do the
* safe/right thing.
*/
ring = kmap_atomic(ctx->ring_pages[0]);
head = ring->head;
kunmap_atomic(ring);
refill_reqs_available(ctx, head, ctx->tail);
}
spin_unlock_irq(&ctx->completion_lock);
}
/* aio_get_req
* Allocate a slot for an aio request.
* Returns NULL if no requests are free.
*/
static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
{
struct aio_kiocb *req;
if (!get_reqs_available(ctx)) {
user_refill_reqs_available(ctx);
if (!get_reqs_available(ctx))
return NULL;
}
req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
if (unlikely(!req))
goto out_put;
percpu_ref_get(&ctx->reqs);
INIT_LIST_HEAD(&req->ki_list);
refcount_set(&req->ki_refcnt, 0);
req->ki_ctx = ctx;
return req;
out_put:
put_reqs_available(ctx, 1);
return NULL;
}
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
{
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
struct aio_ring __user *ring = (void __user *)ctx_id;
struct mm_struct *mm = current->mm;
struct kioctx *ctx, *ret = NULL;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
struct kioctx_table *table;
unsigned id;
if (get_user(id, &ring->id))
return NULL;
rcu_read_lock();
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
table = rcu_dereference(mm->ioctx_table);
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
if (!table || id >= table->nr)
goto out;
ctx = rcu_dereference(table->table[id]);
if (ctx && ctx->user_id == ctx_id) {
aio: fix io_destroy(2) vs. lookup_ioctx() race kill_ioctx() used to have an explicit RCU delay between removing the reference from ->ioctx_table and percpu_ref_kill() dropping the refcount. At some point that delay had been removed, on the theory that percpu_ref_kill() itself contained an RCU delay. Unfortunately, that was the wrong kind of RCU delay and it didn't care about rcu_read_lock() used by lookup_ioctx(). As the result, we could get ctx freed right under lookup_ioctx(). Tejun has fixed that in a6d7cff472e ("fs/aio: Add explicit RCU grace period when freeing kioctx"); however, that fix is not enough. Suppose io_destroy() from one thread races with e.g. io_setup() from another; CPU1 removes the reference from current->mm->ioctx_table[...] just as CPU2 has picked it (under rcu_read_lock()). Then CPU1 proceeds to drop the refcount, getting it to 0 and triggering a call of free_ioctx_users(), which proceeds to drop the secondary refcount and once that reaches zero calls free_ioctx_reqs(). That does INIT_RCU_WORK(&ctx->free_rwork, free_ioctx); queue_rcu_work(system_wq, &ctx->free_rwork); and schedules freeing the whole thing after RCU delay. In the meanwhile CPU2 has gotten around to percpu_ref_get(), bumping the refcount from 0 to 1 and returned the reference to io_setup(). Tejun's fix (that queue_rcu_work() in there) guarantees that ctx won't get freed until after percpu_ref_get(). Sure, we'd increment the counter before ctx can be freed. Now we are out of rcu_read_lock() and there's nothing to stop freeing of the whole thing. Unfortunately, CPU2 assumes that since it has grabbed the reference, ctx is *NOT* going away until it gets around to dropping that reference. The fix is obvious - use percpu_ref_tryget_live() and treat failure as miss. It's not costlier than what we currently do in normal case, it's safe to call since freeing *is* delayed and it closes the race window - either lookup_ioctx() comes before percpu_ref_kill() (in which case ctx->users won't reach 0 until the caller of lookup_ioctx() drops it) or lookup_ioctx() fails, ctx->users is unaffected and caller of lookup_ioctx() doesn't see the object in question at all. Cc: stable@kernel.org Fixes: a6d7cff472e "fs/aio: Add explicit RCU grace period when freeing kioctx" Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-05-20 20:46:23 +00:00
if (percpu_ref_tryget_live(&ctx->users))
ret = ctx;
aio: convert the ioctx list to table lookup v3 On Wed, Jun 12, 2013 at 11:14:40AM -0700, Kent Overstreet wrote: > On Mon, Apr 15, 2013 at 02:40:55PM +0300, Octavian Purdila wrote: > > When using a large number of threads performing AIO operations the > > IOCTX list may get a significant number of entries which will cause > > significant overhead. For example, when running this fio script: > > > > rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=512; thread; loops=100 > > > > on an EXT2 filesystem mounted on top of a ramdisk we can observe up to > > 30% CPU time spent by lookup_ioctx: > > > > 32.51% [guest.kernel] [g] lookup_ioctx > > 9.19% [guest.kernel] [g] __lock_acquire.isra.28 > > 4.40% [guest.kernel] [g] lock_release > > 4.19% [guest.kernel] [g] sched_clock_local > > 3.86% [guest.kernel] [g] local_clock > > 3.68% [guest.kernel] [g] native_sched_clock > > 3.08% [guest.kernel] [g] sched_clock_cpu > > 2.64% [guest.kernel] [g] lock_release_holdtime.part.11 > > 2.60% [guest.kernel] [g] memcpy > > 2.33% [guest.kernel] [g] lock_acquired > > 2.25% [guest.kernel] [g] lock_acquire > > 1.84% [guest.kernel] [g] do_io_submit > > > > This patchs converts the ioctx list to a radix tree. For a performance > > comparison the above FIO script was run on a 2 sockets 8 core > > machine. This are the results (average and %rsd of 10 runs) for the > > original list based implementation and for the radix tree based > > implementation: > > > > cores 1 2 4 8 16 32 > > list 109376 ms 69119 ms 35682 ms 22671 ms 19724 ms 16408 ms > > %rsd 0.69% 1.15% 1.17% 1.21% 1.71% 1.43% > > radix 73651 ms 41748 ms 23028 ms 16766 ms 15232 ms 13787 ms > > %rsd 1.19% 0.98% 0.69% 1.13% 0.72% 0.75% > > % of radix > > relative 66.12% 65.59% 66.63% 72.31% 77.26% 83.66% > > to list > > > > To consider the impact of the patch on the typical case of having > > only one ctx per process the following FIO script was run: > > > > rw=randrw; size=100m ;directory=/mnt/fio; ioengine=libaio; iodepth=1 > > blocksize=1024; numjobs=1; thread; loops=100 > > > > on the same system and the results are the following: > > > > list 58892 ms > > %rsd 0.91% > > radix 59404 ms > > %rsd 0.81% > > % of radix > > relative 100.87% > > to list > > So, I was just doing some benchmarking/profiling to get ready to send > out the aio patches I've got for 3.11 - and it looks like your patch is > causing a ~1.5% throughput regression in my testing :/ ... <snip> I've got an alternate approach for fixing this wart in lookup_ioctx()... Instead of using an rbtree, just use the reserved id in the ring buffer header to index an array pointing the ioctx. It's not finished yet, and it needs to be tidied up, but is most of the way there. -ben -- "Thought is the essence of where you are now." -- kmo> And, a rework of Ben's code, but this was entirely his idea kmo> -Kent bcrl> And fix the code to use the right mm_struct in kill_ioctx(), actually free memory. Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2013-07-30 16:54:40 +00:00
}
out:
rcu_read_unlock();
return ret;
}
static inline void iocb_put(struct aio_kiocb *iocb)
{
if (refcount_read(&iocb->ki_refcnt) == 0 ||
refcount_dec_and_test(&iocb->ki_refcnt)) {
percpu_ref_put(&iocb->ki_ctx->reqs);
kmem_cache_free(kiocb_cachep, iocb);
}
}
/* aio_complete
* Called when the io request on the given iocb is complete.
*/
static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
{
struct kioctx *ctx = iocb->ki_ctx;
struct aio_ring *ring;
struct io_event *ev_page, *event;
unsigned tail, pos, head;
unsigned long flags;
/*
* Add a completion event to the ring buffer. Must be done holding
* ctx->completion_lock to prevent other code from messing with the tail
* pointer since we might be called from irq context.
*/
spin_lock_irqsave(&ctx->completion_lock, flags);
tail = ctx->tail;
pos = tail + AIO_EVENTS_OFFSET;
if (++tail >= ctx->nr_events)
tail = 0;
ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
event = ev_page + pos % AIO_EVENTS_PER_PAGE;
event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
event->data = iocb->ki_user_data;
event->res = res;
event->res2 = res2;
kunmap_atomic(ev_page);
flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
res, res2);
/* after flagging the request as done, we
* must never even look at it again
*/
smp_wmb(); /* make event visible before updating tail */
ctx->tail = tail;
ring = kmap_atomic(ctx->ring_pages[0]);
head = ring->head;
ring->tail = tail;
kunmap_atomic(ring);
flush_dcache_page(ctx->ring_pages[0]);
ctx->completed_events++;
if (ctx->completed_events > 1)
refill_reqs_available(ctx, head, tail);
spin_unlock_irqrestore(&ctx->completion_lock, flags);
pr_debug("added to ring %p at [%u]\n", iocb, tail);
/*
* Check if the user asked us to deliver the result through an
* eventfd. The eventfd_signal() function is safe to be called
* from IRQ context.
*/
if (iocb->ki_eventfd) {
eventfd_signal(iocb->ki_eventfd, 1);
eventfd_ctx_put(iocb->ki_eventfd);
}
aio: bad AIO race in aio_complete() leads to process hang My group ran into a AIO process hang on a 2.6.24 kernel with the process sleeping indefinitely in io_getevents(2) waiting for the last wakeup to come and it never would. We ran the tests on x86_64 SMP. The hang only occurred on a Xeon box ("Clovertown") but not a Core2Duo ("Conroe"). On the Xeon, the L2 cache isn't shared between all eight processors, but is L2 is shared between between all two processors on the Core2Duo we use. My analysis of the hang is if you go down to the second while-loop in read_events(), what happens on processor #1: 1) add_wait_queue_exclusive() adds thread to ctx->wait 2) aio_read_evt() to check tail 3) if aio_read_evt() returned 0, call [io_]schedule() and sleep In aio_complete() with processor #2: A) info->tail = tail; B) waitqueue_active(&ctx->wait) C) if waitqueue_active() returned non-0, call wake_up() The way the code is written, step 1 must be seen by all other processors before processor 1 checks for pending events in step 2 (that were recorded by step A) and step A by processor 2 must be seen by all other processors (checked in step 2) before step B is done. The race I believed I was seeing is that steps 1 and 2 were effectively swapped due to the __list_add() being delayed by the L2 cache not shared by some of the other processors. Imagine: proc 2: just before step A proc 1, step 1: adds to ctx->wait, but is not visible by other processors yet proc 1, step 2: checks tail and sees no pending events proc 2, step A: updates tail proc 1, step 3: calls [io_]schedule() and sleeps proc 2, step B: checks ctx->wait, but sees no one waiting, skips wakeup so proc 1 sleeps indefinitely My patch adds a memory barrier between steps A and B. It ensures that the update in step 1 gets seen on processor 2 before continuing. If processor 1 was just before step 1, the memory barrier makes sure that step A (update tail) gets seen by the time processor 1 makes it to step 2 (check tail). Before the patch our AIO process would hang virtually 100% of the time. After the patch, we have yet to see the process ever hang. Signed-off-by: Quentin Barnes <qbarnes+linux@yahoo-inc.com> Reviewed-by: Zach Brown <zach.brown@oracle.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: <stable@kernel.org> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> [ We should probably disallow that "if (waitqueue_active()) wake_up()" coding pattern, because it's so often buggy wrt memory ordering ] Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-03-20 00:00:39 +00:00
/*
* We have to order our ring_info tail store above and test
* of the wait list below outside the wait lock. This is
* like in wake_up_bit() where clearing a bit has to be
* ordered with the unlocked test.
*/
smp_mb();
if (waitqueue_active(&ctx->wait))
wake_up(&ctx->wait);
iocb_put(iocb);
}
/* aio_read_events_ring
* Pull an event off of the ioctx's event ring. Returns the number of
* events fetched
*/
static long aio_read_events_ring(struct kioctx *ctx,
struct io_event __user *event, long nr)
{
struct aio_ring *ring;
unsigned head, tail, pos;
long ret = 0;
int copy_ret;
aio: annotate aio_read_event_ring for sleep patterns Under CONFIG_DEBUG_ATOMIC_SLEEP=y, aio_read_event_ring() will throw warnings like the following due to being called from wait_event context: WARNING: CPU: 0 PID: 16006 at kernel/sched/core.c:7300 __might_sleep+0x7f/0x90() do not call blocking ops when !TASK_RUNNING; state=1 set at [<ffffffff810d85a3>] prepare_to_wait_event+0x63/0x110 Modules linked in: CPU: 0 PID: 16006 Comm: aio-dio-fcntl-r Not tainted 3.19.0-rc6-dgc+ #705 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 ffffffff821c0372 ffff88003c117cd8 ffffffff81daf2bd 000000000000d8d8 ffff88003c117d28 ffff88003c117d18 ffffffff8109beda ffff88003c117cf8 ffffffff821c115e 0000000000000061 0000000000000000 00007ffffe4aa300 Call Trace: [<ffffffff81daf2bd>] dump_stack+0x4c/0x65 [<ffffffff8109beda>] warn_slowpath_common+0x8a/0xc0 [<ffffffff8109bf56>] warn_slowpath_fmt+0x46/0x50 [<ffffffff810d85a3>] ? prepare_to_wait_event+0x63/0x110 [<ffffffff810d85a3>] ? prepare_to_wait_event+0x63/0x110 [<ffffffff810bdfcf>] __might_sleep+0x7f/0x90 [<ffffffff81db8344>] mutex_lock+0x24/0x45 [<ffffffff81216b7c>] aio_read_events+0x4c/0x290 [<ffffffff81216fac>] read_events+0x1ec/0x220 [<ffffffff810d8650>] ? prepare_to_wait_event+0x110/0x110 [<ffffffff810fdb10>] ? hrtimer_get_res+0x50/0x50 [<ffffffff8121899d>] SyS_io_getevents+0x4d/0xb0 [<ffffffff81dba5a9>] system_call_fastpath+0x12/0x17 ---[ end trace bde69eaf655a4fea ]--- There is not actually a bug here, so annotate the code to tell the debug logic that everything is just fine and not to fire a false positive. Signed-off-by: Dave Chinner <dchinner@redhat.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org>
2015-02-04 00:29:05 +00:00
/*
* The mutex can block and wake us up and that will cause
* wait_event_interruptible_hrtimeout() to schedule without sleeping
* and repeat. This should be rare enough that it doesn't cause
* peformance issues. See the comment in read_events() for more detail.
*/
sched_annotate_sleep();
mutex_lock(&ctx->ring_lock);
aio: v4 ensure access to ctx->ring_pages is correctly serialised for migration As reported by Tang Chen, Gu Zheng and Yasuaki Isimatsu, the following issues exist in the aio ring page migration support. As a result, for example, we have the following problem: thread 1 | thread 2 | aio_migratepage() | |-> take ctx->completion_lock | |-> migrate_page_copy(new, old) | | *NOW*, ctx->ring_pages[idx] == old | | | *NOW*, ctx->ring_pages[idx] == old | aio_read_events_ring() | |-> ring = kmap_atomic(ctx->ring_pages[0]) | |-> ring->head = head; *HERE, write to the old ring page* | |-> kunmap_atomic(ring); | |-> ctx->ring_pages[idx] = new | | *BUT NOW*, the content of | | ring_pages[idx] is old. | |-> release ctx->completion_lock | As above, the new ring page will not be updated. Fix this issue, as well as prevent races in aio_ring_setup() by holding the ring_lock mutex during kioctx setup and page migration. This avoids the overhead of taking another spinlock in aio_read_events_ring() as Tang's and Gu's original fix did, pushing the overhead into the migration code. Note that to handle the nesting of ring_lock inside of mmap_sem, the migratepage operation uses mutex_trylock(). Page migration is not a 100% critical operation in this case, so the ocassional failure can be tolerated. This issue was reported by Sasha Levin. Based on feedback from Linus, avoid the extra taking of ctx->completion_lock. Instead, make page migration fully serialised by mapping->private_lock, and have aio_free_ring() simply disconnect the kioctx from the mapping by calling put_aio_ring_file() before touching ctx->ring_pages[]. This simplifies the error handling logic in aio_migratepage(), and should improve robustness. v4: always do mutex_unlock() in cases when kioctx setup fails. Reported-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com> Reported-by: Sasha Levin <sasha.levin@oracle.com> Signed-off-by: Benjamin LaHaise <bcrl@kvack.org> Cc: Tang Chen <tangchen@cn.fujitsu.com> Cc: Gu Zheng <guz.fnst@cn.fujitsu.com> Cc: stable@vger.kernel.org
2014-03-28 14:14:45 +00:00
/* Access to ->ring_pages here is protected by ctx->ring_lock. */
ring = kmap_atomic(ctx->ring_pages[0]);
head = ring->head;
tail = ring->tail;
kunmap_atomic(ring);
/*
* Ensure that once we've read the current tail pointer, that
* we also see the events that were stored up to the tail.
*/
smp_rmb();
pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
if (head == tail)
goto out;
head %= ctx->nr_events;
tail %= ctx->nr_events;
while (ret < nr) {
long avail;
struct io_event *ev;
struct page *page;
avail = (head <= tail ? tail : ctx->nr_events) - head;
if (head == tail)
break;
pos = head + AIO_EVENTS_OFFSET;
page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
pos %= AIO_EVENTS_PER_PAGE;
avail = min(avail, nr - ret);
avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
ev = kmap(page);
copy_ret = copy_to_user(event + ret, ev + pos,
sizeof(*ev) * avail);
kunmap(page);
if (unlikely(copy_ret)) {
ret = -EFAULT;
goto out;
}
ret += avail;
head += avail;
head %= ctx->nr_events;
}
ring = kmap_atomic(ctx->ring_pages[0]);
ring->head = head;
kunmap_atomic(ring);
flush_dcache_page(ctx->ring_pages[0]);
pr_debug("%li h%u t%u\n", ret, head, tail);
out:
mutex_unlock(&ctx->ring_lock);
return ret;
}
static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
struct io_event __user *event, long *i)
{
long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
if (ret > 0)
*i += ret;
if (unlikely(atomic_read(&ctx->dead)))
ret = -EINVAL;
if (!*i)
*i = ret;
return ret < 0 || *i >= min_nr;
}
static long read_events(struct kioctx *ctx, long min_nr, long nr,
struct io_event __user *event,
ktime_t until)
{
long ret = 0;
/*
* Note that aio_read_events() is being called as the conditional - i.e.
* we're calling it after prepare_to_wait() has set task state to
* TASK_INTERRUPTIBLE.
*
* But aio_read_events() can block, and if it blocks it's going to flip
* the task state back to TASK_RUNNING.
*
* This should be ok, provided it doesn't flip the state back to
* TASK_RUNNING and return 0 too much - that causes us to spin. That
* will only happen if the mutex_lock() call blocks, and we then find
* the ringbuffer empty. So in practice we should be ok, but it's
* something to be aware of when touching this code.
*/
if (until == 0)
aio_read_events(ctx, min_nr, nr, event, &ret);
else
wait_event_interruptible_hrtimeout(ctx->wait,
aio_read_events(ctx, min_nr, nr, event, &ret),
until);
return ret;
}
/* sys_io_setup:
* Create an aio_context capable of receiving at least nr_events.
* ctxp must not point to an aio_context that already exists, and
* must be initialized to 0 prior to the call. On successful
* creation of the aio_context, *ctxp is filled in with the resulting
* handle. May fail with -EINVAL if *ctxp is not initialized,
* if the specified nr_events exceeds internal limits. May fail
* with -EAGAIN if the specified nr_events exceeds the user's limit
* of available events. May fail with -ENOMEM if insufficient kernel
* resources are available. May fail with -EFAULT if an invalid
* pointer is passed for ctxp. Will fail with -ENOSYS if not
* implemented.
*/
SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
{
struct kioctx *ioctx = NULL;
unsigned long ctx;
long ret;
ret = get_user(ctx, ctxp);
if (unlikely(ret))
goto out;
ret = -EINVAL;
if (unlikely(ctx || nr_events == 0)) {
pr_debug("EINVAL: ctx %lu nr_events %u\n",
ctx, nr_events);
goto out;
}
ioctx = ioctx_alloc(nr_events);
ret = PTR_ERR(ioctx);
if (!IS_ERR(ioctx)) {
ret = put_user(ioctx->user_id, ctxp);
if (ret)
kill_ioctx(current->mm, ioctx, NULL);
percpu_ref_put(&ioctx->users);
}
out:
return ret;
}
#ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
{
struct kioctx *ioctx = NULL;
unsigned long ctx;
long ret;
ret = get_user(ctx, ctx32p);
if (unlikely(ret))
goto out;
ret = -EINVAL;
if (unlikely(ctx || nr_events == 0)) {
pr_debug("EINVAL: ctx %lu nr_events %u\n",
ctx, nr_events);
goto out;
}
ioctx = ioctx_alloc(nr_events);
ret = PTR_ERR(ioctx);
if (!IS_ERR(ioctx)) {
/* truncating is ok because it's a user address */
ret = put_user((u32)ioctx->user_id, ctx32p);
if (ret)
kill_ioctx(current->mm, ioctx, NULL);
percpu_ref_put(&ioctx->users);
}
out:
return ret;
}
#endif
/* sys_io_destroy:
* Destroy the aio_context specified. May cancel any outstanding
* AIOs and block on completion. Will fail with -ENOSYS if not
* implemented. May fail with -EINVAL if the context pointed to
* is invalid.
*/
SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
{
struct kioctx *ioctx = lookup_ioctx(ctx);
if (likely(NULL != ioctx)) {
struct ctx_rq_wait wait;
int ret;
init_completion(&wait.comp);
atomic_set(&wait.count, 1);
/* Pass requests_done to kill_ioctx() where it can be set
* in a thread-safe way. If we try to set it here then we have
* a race condition if two io_destroy() called simultaneously.
*/
ret = kill_ioctx(current->mm, ioctx, &wait);
percpu_ref_put(&ioctx->users);
/* Wait until all IO for the context are done. Otherwise kernel
* keep using user-space buffers even if user thinks the context
* is destroyed.
*/
if (!ret)
wait_for_completion(&wait.comp);
return ret;
}
pr_debug("EINVAL: invalid context id\n");
return -EINVAL;
}
static void aio_remove_iocb(struct aio_kiocb *iocb)
{
struct kioctx *ctx = iocb->ki_ctx;
unsigned long flags;
spin_lock_irqsave(&ctx->ctx_lock, flags);
list_del(&iocb->ki_list);
spin_unlock_irqrestore(&ctx->ctx_lock, flags);
}
static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
{
struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
if (!list_empty_careful(&iocb->ki_list))
aio_remove_iocb(iocb);
if (kiocb->ki_flags & IOCB_WRITE) {
struct inode *inode = file_inode(kiocb->ki_filp);
/*
* Tell lockdep we inherited freeze protection from submission
* thread.
*/
if (S_ISREG(inode->i_mode))
__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
file_end_write(kiocb->ki_filp);
}
fput(kiocb->ki_filp);
aio_complete(iocb, res, res2);
}
static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
{
int ret;
req->ki_filp = fget(iocb->aio_fildes);
if (unlikely(!req->ki_filp))
return -EBADF;
req->ki_complete = aio_complete_rw;
req->ki_pos = iocb->aio_offset;
req->ki_flags = iocb_flags(req->ki_filp);
if (iocb->aio_flags & IOCB_FLAG_RESFD)
req->ki_flags |= IOCB_EVENTFD;
req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
/*
* If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
* aio_reqprio is interpreted as an I/O scheduling
* class and priority.
*/
ret = ioprio_check_cap(iocb->aio_reqprio);
if (ret) {
pr_debug("aio ioprio check cap error: %d\n", ret);
return ret;
}
req->ki_ioprio = iocb->aio_reqprio;
} else
req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0);
ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
if (unlikely(ret))
fput(req->ki_filp);
return ret;
}
static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
bool vectored, bool compat, struct iov_iter *iter)
{
void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
size_t len = iocb->aio_nbytes;
if (!vectored) {
ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
*iovec = NULL;
return ret;
}
#ifdef CONFIG_COMPAT
if (compat)
return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
iter);
#endif
return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
}
static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
{
switch (ret) {
case -EIOCBQUEUED:
break;
case -ERESTARTSYS:
case -ERESTARTNOINTR:
case -ERESTARTNOHAND:
case -ERESTART_RESTARTBLOCK:
/*
* There's no easy way to restart the syscall since other AIO's
* may be already running. Just fail this IO with EINTR.
*/
ret = -EINTR;
/*FALLTHRU*/
default:
aio_complete_rw(req, ret, 0);
}
}
static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
bool compat)
{
struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
struct iov_iter iter;
struct file *file;
ssize_t ret;
ret = aio_prep_rw(req, iocb);
if (ret)
return ret;
file = req->ki_filp;
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_READ)))
goto out_fput;
ret = -EINVAL;
if (unlikely(!file->f_op->read_iter))
goto out_fput;
ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
if (ret)
goto out_fput;
ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
if (!ret)
aio_rw_done(req, call_read_iter(file, req, &iter));
kfree(iovec);
out_fput:
if (unlikely(ret))
fput(file);
return ret;
}
static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
bool compat)
{
struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
struct iov_iter iter;
struct file *file;
ssize_t ret;
ret = aio_prep_rw(req, iocb);
if (ret)
return ret;
file = req->ki_filp;
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_WRITE)))
goto out_fput;
ret = -EINVAL;
if (unlikely(!file->f_op->write_iter))
goto out_fput;
ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
if (ret)
goto out_fput;
ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
if (!ret) {
/*
* Open-code file_start_write here to grab freeze protection,
* which will be released by another thread in
* aio_complete_rw(). Fool lockdep by telling it the lock got
* released so that it doesn't complain about the held lock when
* we return to userspace.
*/
if (S_ISREG(file_inode(file)->i_mode)) {
__sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
aio: fix lock dep warning lockdep reports a warnning. file_start_write/file_end_write only acquire/release the lock for regular files. So checking the files in aio side too. [ 453.532141] ------------[ cut here ]------------ [ 453.533011] WARNING: CPU: 1 PID: 1298 at ../kernel/locking/lockdep.c:3514 lock_release+0x434/0x670 [ 453.533011] DEBUG_LOCKS_WARN_ON(depth <= 0) [ 453.533011] Modules linked in: [ 453.533011] CPU: 1 PID: 1298 Comm: fio Not tainted 4.9.0+ #964 [ 453.533011] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.9.0-1.fc24 04/01/2014 [ 453.533011] ffff8803a24b7a70 ffffffff8196cffb ffff8803a24b7ae8 0000000000000000 [ 453.533011] ffff8803a24b7ab8 ffffffff81091ee1 ffff8803a5dba700 00000dba00000008 [ 453.533011] ffffed0074496f59 ffff8803a5dbaf54 ffff8803ae0f8488 fffffffffffffdef [ 453.533011] Call Trace: [ 453.533011] [<ffffffff8196cffb>] dump_stack+0x67/0x9c [ 453.533011] [<ffffffff81091ee1>] __warn+0x111/0x130 [ 453.533011] [<ffffffff81091f97>] warn_slowpath_fmt+0x97/0xb0 [ 453.533011] [<ffffffff81091f00>] ? __warn+0x130/0x130 [ 453.533011] [<ffffffff8191b789>] ? blk_finish_plug+0x29/0x60 [ 453.533011] [<ffffffff811205d4>] lock_release+0x434/0x670 [ 453.533011] [<ffffffff8198af94>] ? import_single_range+0xd4/0x110 [ 453.533011] [<ffffffff81322195>] ? rw_verify_area+0x65/0x140 [ 453.533011] [<ffffffff813aa696>] ? aio_write+0x1f6/0x280 [ 453.533011] [<ffffffff813aa6c9>] aio_write+0x229/0x280 [ 453.533011] [<ffffffff813aa4a0>] ? aio_complete+0x640/0x640 [ 453.533011] [<ffffffff8111df20>] ? debug_check_no_locks_freed+0x1a0/0x1a0 [ 453.533011] [<ffffffff8114793a>] ? debug_lockdep_rcu_enabled.part.2+0x1a/0x30 [ 453.533011] [<ffffffff81147985>] ? debug_lockdep_rcu_enabled+0x35/0x40 [ 453.533011] [<ffffffff812a92be>] ? __might_fault+0x7e/0xf0 [ 453.533011] [<ffffffff813ac9bc>] do_io_submit+0x94c/0xb10 [ 453.533011] [<ffffffff813ac2ae>] ? do_io_submit+0x23e/0xb10 [ 453.533011] [<ffffffff813ac070>] ? SyS_io_destroy+0x270/0x270 [ 453.533011] [<ffffffff8111d7b3>] ? mark_held_locks+0x23/0xc0 [ 453.533011] [<ffffffff8100201a>] ? trace_hardirqs_on_thunk+0x1a/0x1c [ 453.533011] [<ffffffff813acb90>] SyS_io_submit+0x10/0x20 [ 453.533011] [<ffffffff824f96aa>] entry_SYSCALL_64_fastpath+0x18/0xad [ 453.533011] [<ffffffff81119190>] ? trace_hardirqs_off_caller+0xc0/0x110 [ 453.533011] ---[ end trace b2fbe664d1cc0082 ]--- Cc: Dmitry Monakhov <dmonakhov@openvz.org> Cc: Jan Kara <jack@suse.cz> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Shaohua Li <shli@fb.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-12-13 20:09:56 +00:00
__sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
}
req->ki_flags |= IOCB_WRITE;
aio_rw_done(req, call_write_iter(file, req, &iter));
}
kfree(iovec);
out_fput:
if (unlikely(ret))
fput(file);
return ret;
}
static void aio_fsync_work(struct work_struct *work)
{
struct fsync_iocb *req = container_of(work, struct fsync_iocb, work);
int ret;
ret = vfs_fsync(req->file, req->datasync);
fput(req->file);
aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0);
}
static int aio_fsync(struct fsync_iocb *req, struct iocb *iocb, bool datasync)
{
if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
iocb->aio_rw_flags))
return -EINVAL;
req->file = fget(iocb->aio_fildes);
if (unlikely(!req->file))
return -EBADF;
if (unlikely(!req->file->f_op->fsync)) {
fput(req->file);
return -EINVAL;
}
req->datasync = datasync;
INIT_WORK(&req->work, aio_fsync_work);
schedule_work(&req->work);
return 0;
}
static inline void aio_poll_complete(struct aio_kiocb *iocb, __poll_t mask)
{
struct file *file = iocb->poll.file;
aio_complete(iocb, mangle_poll(mask), 0);
fput(file);
}
static void aio_poll_complete_work(struct work_struct *work)
{
struct poll_iocb *req = container_of(work, struct poll_iocb, work);
struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
struct poll_table_struct pt = { ._key = req->events };
struct kioctx *ctx = iocb->ki_ctx;
__poll_t mask = 0;
if (!READ_ONCE(req->cancelled))
mask = vfs_poll(req->file, &pt) & req->events;
/*
* Note that ->ki_cancel callers also delete iocb from active_reqs after
* calling ->ki_cancel. We need the ctx_lock roundtrip here to
* synchronize with them. In the cancellation case the list_del_init
* itself is not actually needed, but harmless so we keep it in to
* avoid further branches in the fast path.
*/
spin_lock_irq(&ctx->ctx_lock);
if (!mask && !READ_ONCE(req->cancelled)) {
add_wait_queue(req->head, &req->wait);
spin_unlock_irq(&ctx->ctx_lock);
return;
}
list_del_init(&iocb->ki_list);
spin_unlock_irq(&ctx->ctx_lock);
aio_poll_complete(iocb, mask);
}
/* assumes we are called with irqs disabled */
static int aio_poll_cancel(struct kiocb *iocb)
{
struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
struct poll_iocb *req = &aiocb->poll;
spin_lock(&req->head->lock);
WRITE_ONCE(req->cancelled, true);
if (!list_empty(&req->wait.entry)) {
list_del_init(&req->wait.entry);
schedule_work(&aiocb->poll.work);
}
spin_unlock(&req->head->lock);
return 0;
}
static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
void *key)
{
struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
__poll_t mask = key_to_poll(key);
req->woken = true;
/* for instances that support it check for an event match first: */
if (mask) {
if (!(mask & req->events))
return 0;
/* try to complete the iocb inline if we can: */
if (spin_trylock(&iocb->ki_ctx->ctx_lock)) {
list_del(&iocb->ki_list);
spin_unlock(&iocb->ki_ctx->ctx_lock);
list_del_init(&req->wait.entry);
aio_poll_complete(iocb, mask);
return 1;
}
}
list_del_init(&req->wait.entry);
schedule_work(&req->work);
return 1;
}
struct aio_poll_table {
struct poll_table_struct pt;
struct aio_kiocb *iocb;
int error;
};
static void
aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
struct poll_table_struct *p)
{
struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
/* multiple wait queues per file are not supported */
if (unlikely(pt->iocb->poll.head)) {
pt->error = -EINVAL;
return;
}
pt->error = 0;
pt->iocb->poll.head = head;
add_wait_queue(head, &pt->iocb->poll.wait);
}
static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb)
{
struct kioctx *ctx = aiocb->ki_ctx;
struct poll_iocb *req = &aiocb->poll;
struct aio_poll_table apt;
__poll_t mask;
/* reject any unknown events outside the normal event mask. */
if ((u16)iocb->aio_buf != iocb->aio_buf)
return -EINVAL;
/* reject fields that are not defined for poll */
if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
return -EINVAL;
INIT_WORK(&req->work, aio_poll_complete_work);
req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
req->file = fget(iocb->aio_fildes);
if (unlikely(!req->file))
return -EBADF;
apt.pt._qproc = aio_poll_queue_proc;
apt.pt._key = req->events;
apt.iocb = aiocb;
apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
/* initialized the list so that we can do list_empty checks */
INIT_LIST_HEAD(&req->wait.entry);
init_waitqueue_func_entry(&req->wait, aio_poll_wake);
/* one for removal from waitqueue, one for this function */
refcount_set(&aiocb->ki_refcnt, 2);
mask = vfs_poll(req->file, &apt.pt) & req->events;
if (unlikely(!req->head)) {
/* we did not manage to set up a waitqueue, done */
goto out;
}
spin_lock_irq(&ctx->ctx_lock);
spin_lock(&req->head->lock);
if (req->woken) {
/* wake_up context handles the rest */
mask = 0;
apt.error = 0;
} else if (mask || apt.error) {
/* if we get an error or a mask we are done */
WARN_ON_ONCE(list_empty(&req->wait.entry));
list_del_init(&req->wait.entry);
} else {
/* actually waiting for an event */
list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
aiocb->ki_cancel = aio_poll_cancel;
}
spin_unlock(&req->head->lock);
spin_unlock_irq(&ctx->ctx_lock);
out:
if (unlikely(apt.error)) {
fput(req->file);
return apt.error;
}
if (mask)
aio_poll_complete(aiocb, mask);
iocb_put(aiocb);
return 0;
}
static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
bool compat)
{
struct aio_kiocb *req;
struct iocb iocb;
ssize_t ret;
if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
return -EFAULT;
/* enforce forwards compatibility on users */
if (unlikely(iocb.aio_reserved2)) {
pr_debug("EINVAL: reserve field set\n");
return -EINVAL;
}
/* prevent overflows */
if (unlikely(
(iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
(iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
((ssize_t)iocb.aio_nbytes < 0)
)) {
pr_debug("EINVAL: overflow check\n");
return -EINVAL;
}
req = aio_get_req(ctx);
if (unlikely(!req))
return -EAGAIN;
if (iocb.aio_flags & IOCB_FLAG_RESFD) {
signal/timer/event: KAIO eventfd support example This is an example about how to add eventfd support to the current KAIO code, in order to enable KAIO to post readiness events to a pollable fd (hence compatible with POSIX select/poll). The KAIO code simply signals the eventfd fd when events are ready, and this triggers a POLLIN in the fd. This patch uses a reserved for future use member of the struct iocb to pass an eventfd file descriptor, that KAIO will use to post events every time a request completes. At that point, an aio_getevents() will return the completed result to a struct io_event. I made a quick test program to verify the patch, and it runs fine here: http://www.xmailserver.org/eventfd-aio-test.c The test program uses poll(2), but it'd, of course, work with select and epoll too. This can allow to schedule both block I/O and other poll-able devices requests, and wait for results using select/poll/epoll. In a typical scenario, an application would submit KAIO request using aio_submit(), and will also use epoll_ctl() on the whole other class of devices (that with the addition of signals, timers and user events, now it's pretty much complete), and then would: epoll_wait(...); for_each_event { if (curr_event_is_kaiofd) { aio_getevents(); dispatch_aio_events(); } else { dispatch_epoll_event(); } } Signed-off-by: Davide Libenzi <davidel@xmailserver.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-11 05:23:21 +00:00
/*
* If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
* instance of the file* now. The file descriptor must be
* an eventfd() fd, and will be signaled for each completed
* event using the eventfd_signal() function.
*/
req->ki_eventfd = eventfd_ctx_fdget((int) iocb.aio_resfd);
if (IS_ERR(req->ki_eventfd)) {
signal/timer/event: KAIO eventfd support example This is an example about how to add eventfd support to the current KAIO code, in order to enable KAIO to post readiness events to a pollable fd (hence compatible with POSIX select/poll). The KAIO code simply signals the eventfd fd when events are ready, and this triggers a POLLIN in the fd. This patch uses a reserved for future use member of the struct iocb to pass an eventfd file descriptor, that KAIO will use to post events every time a request completes. At that point, an aio_getevents() will return the completed result to a struct io_event. I made a quick test program to verify the patch, and it runs fine here: http://www.xmailserver.org/eventfd-aio-test.c The test program uses poll(2), but it'd, of course, work with select and epoll too. This can allow to schedule both block I/O and other poll-able devices requests, and wait for results using select/poll/epoll. In a typical scenario, an application would submit KAIO request using aio_submit(), and will also use epoll_ctl() on the whole other class of devices (that with the addition of signals, timers and user events, now it's pretty much complete), and then would: epoll_wait(...); for_each_event { if (curr_event_is_kaiofd) { aio_getevents(); dispatch_aio_events(); } else { dispatch_epoll_event(); } } Signed-off-by: Davide Libenzi <davidel@xmailserver.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-11 05:23:21 +00:00
ret = PTR_ERR(req->ki_eventfd);
req->ki_eventfd = NULL;
signal/timer/event: KAIO eventfd support example This is an example about how to add eventfd support to the current KAIO code, in order to enable KAIO to post readiness events to a pollable fd (hence compatible with POSIX select/poll). The KAIO code simply signals the eventfd fd when events are ready, and this triggers a POLLIN in the fd. This patch uses a reserved for future use member of the struct iocb to pass an eventfd file descriptor, that KAIO will use to post events every time a request completes. At that point, an aio_getevents() will return the completed result to a struct io_event. I made a quick test program to verify the patch, and it runs fine here: http://www.xmailserver.org/eventfd-aio-test.c The test program uses poll(2), but it'd, of course, work with select and epoll too. This can allow to schedule both block I/O and other poll-able devices requests, and wait for results using select/poll/epoll. In a typical scenario, an application would submit KAIO request using aio_submit(), and will also use epoll_ctl() on the whole other class of devices (that with the addition of signals, timers and user events, now it's pretty much complete), and then would: epoll_wait(...); for_each_event { if (curr_event_is_kaiofd) { aio_getevents(); dispatch_aio_events(); } else { dispatch_epoll_event(); } } Signed-off-by: Davide Libenzi <davidel@xmailserver.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-05-11 05:23:21 +00:00
goto out_put_req;
}
}
ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
if (unlikely(ret)) {
pr_debug("EFAULT: aio_key\n");
goto out_put_req;
}
req->ki_user_iocb = user_iocb;
req->ki_user_data = iocb.aio_data;
switch (iocb.aio_lio_opcode) {
case IOCB_CMD_PREAD:
ret = aio_read(&req->rw, &iocb, false, compat);
break;
case IOCB_CMD_PWRITE:
ret = aio_write(&req->rw, &iocb, false, compat);
break;
case IOCB_CMD_PREADV:
ret = aio_read(&req->rw, &iocb, true, compat);
break;
case IOCB_CMD_PWRITEV:
ret = aio_write(&req->rw, &iocb, true, compat);
break;
case IOCB_CMD_FSYNC:
ret = aio_fsync(&req->fsync, &iocb, false);
break;
case IOCB_CMD_FDSYNC:
ret = aio_fsync(&req->fsync, &iocb, true);
break;
case IOCB_CMD_POLL:
ret = aio_poll(req, &iocb);
break;
default:
pr_debug("invalid aio operation %d\n", iocb.aio_lio_opcode);
ret = -EINVAL;
break;
}
2013-05-07 23:18:25 +00:00
/*
* If ret is 0, we'd either done aio_complete() ourselves or have
* arranged for that to be done asynchronously. Anything non-zero
* means that we need to destroy req ourselves.
*/
if (ret)
goto out_put_req;
return 0;
out_put_req:
put_reqs_available(ctx, 1);
percpu_ref_put(&ctx->reqs);
if (req->ki_eventfd)
eventfd_ctx_put(req->ki_eventfd);
kmem_cache_free(kiocb_cachep, req);
return ret;
}
/* sys_io_submit:
* Queue the nr iocbs pointed to by iocbpp for processing. Returns
* the number of iocbs queued. May return -EINVAL if the aio_context
* specified by ctx_id is invalid, if nr is < 0, if the iocb at
* *iocbpp[0] is not properly initialized, if the operation specified
* is invalid for the file descriptor in the iocb. May fail with
* -EFAULT if any of the data structures point to invalid data. May
* fail with -EBADF if the file descriptor specified in the first
* iocb is invalid. May fail with -EAGAIN if insufficient resources
* are available to queue any iocbs. Will return 0 if nr is 0. Will
* fail with -ENOSYS if not implemented.
*/
SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
struct iocb __user * __user *, iocbpp)
{
struct kioctx *ctx;
long ret = 0;
int i = 0;
struct blk_plug plug;
if (unlikely(nr < 0))
return -EINVAL;
ctx = lookup_ioctx(ctx_id);
if (unlikely(!ctx)) {
pr_debug("EINVAL: invalid context id\n");
return -EINVAL;
}
if (nr > ctx->nr_events)
nr = ctx->nr_events;
blk_start_plug(&plug);
for (i = 0; i < nr; i++) {
struct iocb __user *user_iocb;
if (unlikely(get_user(user_iocb, iocbpp + i))) {
ret = -EFAULT;
break;
}
ret = io_submit_one(ctx, user_iocb, false);
if (ret)
break;
}
blk_finish_plug(&plug);
percpu_ref_put(&ctx->users);
return i ? i : ret;
}
#ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
int, nr, compat_uptr_t __user *, iocbpp)
{
struct kioctx *ctx;
long ret = 0;
int i = 0;
struct blk_plug plug;
if (unlikely(nr < 0))
return -EINVAL;
ctx = lookup_ioctx(ctx_id);
if (unlikely(!ctx)) {
pr_debug("EINVAL: invalid context id\n");
return -EINVAL;
}
if (nr > ctx->nr_events)
nr = ctx->nr_events;
blk_start_plug(&plug);
for (i = 0; i < nr; i++) {
compat_uptr_t user_iocb;
if (unlikely(get_user(user_iocb, iocbpp + i))) {
ret = -EFAULT;
break;
}
ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
if (ret)
break;
}
blk_finish_plug(&plug);
percpu_ref_put(&ctx->users);
return i ? i : ret;
}
#endif
/* lookup_kiocb
* Finds a given iocb for cancellation.
*/
static struct aio_kiocb *
lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb)
{
struct aio_kiocb *kiocb;
assert_spin_locked(&ctx->ctx_lock);
/* TODO: use a hash or array, this sucks. */
list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
if (kiocb->ki_user_iocb == iocb)
return kiocb;
}
return NULL;
}
/* sys_io_cancel:
* Attempts to cancel an iocb previously passed to io_submit. If
* the operation is successfully cancelled, the resulting event is
* copied into the memory pointed to by result without being placed
* into the completion queue and 0 is returned. May fail with
* -EFAULT if any of the data structures pointed to are invalid.
* May fail with -EINVAL if aio_context specified by ctx_id is
* invalid. May fail with -EAGAIN if the iocb specified was not
* cancelled. Will fail with -ENOSYS if not implemented.
*/
SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
struct io_event __user *, result)
{
struct kioctx *ctx;
struct aio_kiocb *kiocb;
int ret = -EINVAL;
u32 key;
if (unlikely(get_user(key, &iocb->aio_key)))
return -EFAULT;
if (unlikely(key != KIOCB_KEY))
return -EINVAL;
ctx = lookup_ioctx(ctx_id);
if (unlikely(!ctx))
return -EINVAL;
spin_lock_irq(&ctx->ctx_lock);
kiocb = lookup_kiocb(ctx, iocb);
if (kiocb) {
ret = kiocb->ki_cancel(&kiocb->rw);
list_del_init(&kiocb->ki_list);
}
spin_unlock_irq(&ctx->ctx_lock);
if (!ret) {
/*
* The result argument is no longer used - the io_event is
* always delivered via the ring buffer. -EINPROGRESS indicates
* cancellation is progress:
*/
ret = -EINPROGRESS;
}
percpu_ref_put(&ctx->users);
return ret;
}
static long do_io_getevents(aio_context_t ctx_id,
long min_nr,
long nr,
struct io_event __user *events,
struct timespec64 *ts)
{
ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
struct kioctx *ioctx = lookup_ioctx(ctx_id);
long ret = -EINVAL;
if (likely(ioctx)) {
if (likely(min_nr <= nr && min_nr >= 0))
ret = read_events(ioctx, min_nr, nr, events, until);
percpu_ref_put(&ioctx->users);
}
return ret;
}
/* io_getevents:
* Attempts to read at least min_nr events and up to nr events from
* the completion queue for the aio_context specified by ctx_id. If
* it succeeds, the number of read events is returned. May fail with
* -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
* out of range, if timeout is out of range. May fail with -EFAULT
* if any of the memory specified is invalid. May return 0 or
* < min_nr if the timeout specified by timeout has elapsed
* before sufficient events are available, where timeout == NULL
* specifies an infinite timeout. Note that the timeout pointed to by
* timeout is relative. Will fail with -ENOSYS if not implemented.
*/
SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
long, min_nr,
long, nr,
struct io_event __user *, events,
struct timespec __user *, timeout)
{
struct timespec64 ts;
int ret;
if (timeout && unlikely(get_timespec64(&ts, timeout)))
return -EFAULT;
ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
if (!ret && signal_pending(current))
ret = -EINTR;
return ret;
}
struct __aio_sigset {
const sigset_t __user *sigmask;
size_t sigsetsize;
};
SYSCALL_DEFINE6(io_pgetevents,
aio_context_t, ctx_id,
long, min_nr,
long, nr,
struct io_event __user *, events,
struct timespec __user *, timeout,
const struct __aio_sigset __user *, usig)
{
struct __aio_sigset ksig = { NULL, };
sigset_t ksigmask, sigsaved;
struct timespec64 ts;
int ret;
if (timeout && unlikely(get_timespec64(&ts, timeout)))
return -EFAULT;
if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
return -EFAULT;
if (ksig.sigmask) {
if (ksig.sigsetsize != sizeof(sigset_t))
return -EINVAL;
if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask)))
return -EFAULT;
sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
}
ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
if (signal_pending(current)) {
if (ksig.sigmask) {
current->saved_sigmask = sigsaved;
set_restore_sigmask();
}
if (!ret)
ret = -ERESTARTNOHAND;
} else {
if (ksig.sigmask)
sigprocmask(SIG_SETMASK, &sigsaved, NULL);
}
return ret;
}
#ifdef CONFIG_COMPAT
COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
compat_long_t, min_nr,
compat_long_t, nr,
struct io_event __user *, events,
struct compat_timespec __user *, timeout)
{
struct timespec64 t;
int ret;
if (timeout && compat_get_timespec64(&t, timeout))
return -EFAULT;
ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
if (!ret && signal_pending(current))
ret = -EINTR;
return ret;
}
struct __compat_aio_sigset {
compat_sigset_t __user *sigmask;
compat_size_t sigsetsize;
};
COMPAT_SYSCALL_DEFINE6(io_pgetevents,
compat_aio_context_t, ctx_id,
compat_long_t, min_nr,
compat_long_t, nr,
struct io_event __user *, events,
struct compat_timespec __user *, timeout,
const struct __compat_aio_sigset __user *, usig)
{
struct __compat_aio_sigset ksig = { NULL, };
sigset_t ksigmask, sigsaved;
struct timespec64 t;
int ret;
if (timeout && compat_get_timespec64(&t, timeout))
return -EFAULT;
if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
return -EFAULT;
if (ksig.sigmask) {
if (ksig.sigsetsize != sizeof(compat_sigset_t))
return -EINVAL;
if (get_compat_sigset(&ksigmask, ksig.sigmask))
return -EFAULT;
sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
}
ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
if (signal_pending(current)) {
if (ksig.sigmask) {
current->saved_sigmask = sigsaved;
set_restore_sigmask();
}
if (!ret)
ret = -ERESTARTNOHAND;
} else {
if (ksig.sigmask)
sigprocmask(SIG_SETMASK, &sigsaved, NULL);
}
return ret;
}
#endif