linux/fs/aio.c

1854 lines
47 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.
*
* See ../COPYING for licensing terms.
*/
#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/uio.h>
#define DEBUG 0
#include <linux/sched.h>
#include <linux/fs.h>
#include <linux/file.h>
#include <linux/mm.h>
#include <linux/mman.h>
#include <linux/mmu_context.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 <asm/kmap_types.h>
#include <asm/uaccess.h>
#if DEBUG > 1
#define dprintk printk
#else
#define dprintk(x...) do { ; } while (0)
#endif
/*------ 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 workqueue_struct *aio_wq;
/* Used for rare fput completion. */
2006-11-22 14:55:48 +00:00
static void aio_fput_routine(struct work_struct *);
static DECLARE_WORK(fput_work, aio_fput_routine);
static DEFINE_SPINLOCK(fput_lock);
static LIST_HEAD(fput_head);
2006-11-22 14:55:48 +00:00
static void aio_kick_handler(struct work_struct *);
static void aio_queue_work(struct kioctx *);
/* 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)
{
kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
BUG_ON(!aio_wq);
pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
return 0;
}
__initcall(aio_setup);
static void aio_free_ring(struct kioctx *ctx)
{
struct aio_ring_info *info = &ctx->ring_info;
long i;
for (i=0; i<info->nr_pages; i++)
put_page(info->ring_pages[i]);
if (info->mmap_size) {
down_write(&ctx->mm->mmap_sem);
do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
up_write(&ctx->mm->mmap_sem);
}
if (info->ring_pages && info->ring_pages != info->internal_pages)
kfree(info->ring_pages);
info->ring_pages = NULL;
info->nr = 0;
}
static int aio_setup_ring(struct kioctx *ctx)
{
struct aio_ring *ring;
struct aio_ring_info *info = &ctx->ring_info;
unsigned nr_events = ctx->max_reqs;
unsigned long size;
int nr_pages;
/* 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 = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
if (nr_pages < 0)
return -EINVAL;
nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
info->nr = 0;
info->ring_pages = info->internal_pages;
if (nr_pages > AIO_RING_PAGES) {
info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
if (!info->ring_pages)
return -ENOMEM;
}
info->mmap_size = nr_pages * PAGE_SIZE;
dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
down_write(&ctx->mm->mmap_sem);
info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
0);
if (IS_ERR((void *)info->mmap_base)) {
up_write(&ctx->mm->mmap_sem);
info->mmap_size = 0;
aio_free_ring(ctx);
return -EAGAIN;
}
dprintk("mmap address: 0x%08lx\n", info->mmap_base);
info->nr_pages = get_user_pages(current, ctx->mm,
info->mmap_base, nr_pages,
1, 0, info->ring_pages, NULL);
up_write(&ctx->mm->mmap_sem);
if (unlikely(info->nr_pages != nr_pages)) {
aio_free_ring(ctx);
return -EAGAIN;
}
ctx->user_id = info->mmap_base;
info->nr = nr_events; /* trusted copy */
ring = kmap_atomic(info->ring_pages[0]);
ring->nr = nr_events; /* user copy */
ring->id = ctx->user_id;
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);
return 0;
}
/* aio_ring_event: returns a pointer to the event at the given index from
* kmap_atomic(). Release the pointer with put_aio_ring_event();
*/
#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)
#define aio_ring_event(info, nr) ({ \
unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
struct io_event *__event; \
__event = kmap_atomic( \
(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
__event += pos % AIO_EVENTS_PER_PAGE; \
__event; \
})
#define put_aio_ring_event(event) do { \
struct io_event *__event = (event); \
(void)__event; \
kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
} while(0)
static void ctx_rcu_free(struct rcu_head *head)
{
struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
kmem_cache_free(kioctx_cachep, ctx);
}
/* __put_ioctx
* Called when the last user of an aio context has gone away,
* and the struct needs to be freed.
*/
static void __put_ioctx(struct kioctx *ctx)
{
unsigned nr_events = ctx->max_reqs;
BUG_ON(ctx->reqs_active);
cancel_delayed_work_sync(&ctx->wq);
aio_free_ring(ctx);
mmdrop(ctx->mm);
ctx->mm = NULL;
if (nr_events) {
spin_lock(&aio_nr_lock);
BUG_ON(aio_nr - nr_events > aio_nr);
aio_nr -= nr_events;
spin_unlock(&aio_nr_lock);
}
pr_debug("__put_ioctx: freeing %p\n", ctx);
call_rcu(&ctx->rcu_head, ctx_rcu_free);
}
static inline int try_get_ioctx(struct kioctx *kioctx)
{
return atomic_inc_not_zero(&kioctx->users);
}
static inline void put_ioctx(struct kioctx *kioctx)
{
BUG_ON(atomic_read(&kioctx->users) <= 0);
if (unlikely(atomic_dec_and_test(&kioctx->users)))
__put_ioctx(kioctx);
}
/* ioctx_alloc
* Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
*/
static struct kioctx *ioctx_alloc(unsigned nr_events)
{
struct mm_struct *mm;
struct kioctx *ctx;
int err = -ENOMEM;
/* Prevent overflows */
if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
(nr_events > (0x10000000U / sizeof(struct kiocb)))) {
pr_debug("ENOMEM: nr_events too high\n");
return ERR_PTR(-EINVAL);
}
if (!nr_events || (unsigned long)nr_events > aio_max_nr)
return ERR_PTR(-EAGAIN);
ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
if (!ctx)
return ERR_PTR(-ENOMEM);
ctx->max_reqs = nr_events;
mm = ctx->mm = current->mm;
atomic_inc(&mm->mm_count);
atomic_set(&ctx->users, 2);
spin_lock_init(&ctx->ctx_lock);
spin_lock_init(&ctx->ring_info.ring_lock);
init_waitqueue_head(&ctx->wait);
INIT_LIST_HEAD(&ctx->active_reqs);
INIT_LIST_HEAD(&ctx->run_list);
2006-11-22 14:55:48 +00:00
INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
if (aio_setup_ring(ctx) < 0)
goto out_freectx;
/* limit the number of system wide aios */
spin_lock(&aio_nr_lock);
if (aio_nr + nr_events > aio_max_nr ||
aio_nr + nr_events < aio_nr) {
spin_unlock(&aio_nr_lock);
goto out_cleanup;
}
aio_nr += ctx->max_reqs;
spin_unlock(&aio_nr_lock);
/* now link into global list. */
spin_lock(&mm->ioctx_lock);
hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
spin_unlock(&mm->ioctx_lock);
dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
return ctx;
out_cleanup:
err = -EAGAIN;
aio_free_ring(ctx);
out_freectx:
mmdrop(mm);
kmem_cache_free(kioctx_cachep, ctx);
dprintk("aio: error allocating ioctx %d\n", err);
return ERR_PTR(err);
}
/* kill_ctx
* 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 void kill_ctx(struct kioctx *ctx)
{
int (*cancel)(struct kiocb *, struct io_event *);
struct task_struct *tsk = current;
DECLARE_WAITQUEUE(wait, tsk);
struct io_event res;
spin_lock_irq(&ctx->ctx_lock);
ctx->dead = 1;
while (!list_empty(&ctx->active_reqs)) {
struct list_head *pos = ctx->active_reqs.next;
struct kiocb *iocb = list_kiocb(pos);
list_del_init(&iocb->ki_list);
cancel = iocb->ki_cancel;
kiocbSetCancelled(iocb);
if (cancel) {
iocb->ki_users++;
spin_unlock_irq(&ctx->ctx_lock);
cancel(iocb, &res);
spin_lock_irq(&ctx->ctx_lock);
}
}
if (!ctx->reqs_active)
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
goto out;
add_wait_queue(&ctx->wait, &wait);
set_task_state(tsk, TASK_UNINTERRUPTIBLE);
while (ctx->reqs_active) {
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
spin_unlock_irq(&ctx->ctx_lock);
io_schedule();
set_task_state(tsk, TASK_UNINTERRUPTIBLE);
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
spin_lock_irq(&ctx->ctx_lock);
}
__set_task_state(tsk, TASK_RUNNING);
remove_wait_queue(&ctx->wait, &wait);
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
out:
spin_unlock_irq(&ctx->ctx_lock);
}
/* wait_on_sync_kiocb:
* Waits on the given sync kiocb to complete.
*/
ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
{
while (iocb->ki_users) {
set_current_state(TASK_UNINTERRUPTIBLE);
if (!iocb->ki_users)
break;
io_schedule();
}
__set_current_state(TASK_RUNNING);
return iocb->ki_user_data;
}
EXPORT_SYMBOL(wait_on_sync_kiocb);
/* 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. However, there may be
* outstanding requests which hold references to the context; as they
* go away, they will call put_ioctx and release any pinned memory
* associated with the request (held via struct page * references).
*/
void exit_aio(struct mm_struct *mm)
{
struct kioctx *ctx;
while (!hlist_empty(&mm->ioctx_list)) {
ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
hlist_del_rcu(&ctx->list);
kill_ctx(ctx);
if (1 != atomic_read(&ctx->users))
printk(KERN_DEBUG
"exit_aio:ioctx still alive: %d %d %d\n",
atomic_read(&ctx->users), ctx->dead,
ctx->reqs_active);
put_ioctx(ctx);
}
}
/* aio_get_req
* Allocate a slot for an aio request. Increments the users count
* of the kioctx so that the kioctx stays around until all requests are
* complete. Returns NULL if no requests are free.
*
* Returns with kiocb->users set to 2. The io submit code path holds
* an extra reference while submitting the i/o.
* This prevents races between the aio code path referencing the
* req (after submitting it) and aio_complete() freeing the req.
*/
static struct kiocb *__aio_get_req(struct kioctx *ctx)
{
struct kiocb *req = NULL;
req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
if (unlikely(!req))
return NULL;
req->ki_flags = 0;
req->ki_users = 2;
req->ki_key = 0;
req->ki_ctx = ctx;
req->ki_cancel = NULL;
req->ki_retry = NULL;
req->ki_dtor = NULL;
req->private = NULL;
req->ki_iovec = NULL;
INIT_LIST_HEAD(&req->ki_run_list);
req->ki_eventfd = NULL;
return req;
}
/*
* struct kiocb's are allocated in batches to reduce the number of
* times the ctx lock is acquired and released.
*/
#define KIOCB_BATCH_SIZE 32L
struct kiocb_batch {
struct list_head head;
long count; /* number of requests left to allocate */
};
static void kiocb_batch_init(struct kiocb_batch *batch, long total)
{
INIT_LIST_HEAD(&batch->head);
batch->count = total;
}
static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
{
struct kiocb *req, *n;
if (list_empty(&batch->head))
return;
spin_lock_irq(&ctx->ctx_lock);
list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
list_del(&req->ki_batch);
list_del(&req->ki_list);
kmem_cache_free(kiocb_cachep, req);
ctx->reqs_active--;
}
aio: wake up waiters when freeing unused kiocbs Bart Van Assche reported a hung fio process when either hot-removing storage or when interrupting the fio process itself. The (pruned) call trace for the latter looks like so: fio D 0000000000000001 0 6849 6848 0x00000004 ffff880092541b88 0000000000000046 ffff880000000000 ffff88012fa11dc0 ffff88012404be70 ffff880092541fd8 ffff880092541fd8 ffff880092541fd8 ffff880128b894d0 ffff88012404be70 ffff880092541b88 000000018106f24d Call Trace: schedule+0x3f/0x60 io_schedule+0x8f/0xd0 wait_for_all_aios+0xc0/0x100 exit_aio+0x55/0xc0 mmput+0x2d/0x110 exit_mm+0x10d/0x130 do_exit+0x671/0x860 do_group_exit+0x44/0xb0 get_signal_to_deliver+0x218/0x5a0 do_signal+0x65/0x700 do_notify_resume+0x65/0x80 int_signal+0x12/0x17 The problem lies with the allocation batching code. It will opportunistically allocate kiocbs, and then trim back the list of iocbs when there is not enough room in the completion ring to hold all of the events. In the case above, what happens is that the pruning back of events ends up freeing up the last active request and the context is marked as dead, so it is thus responsible for waking up waiters. Unfortunately, the code does not check for this condition, so we end up with a hung task. Signed-off-by: Jeff Moyer <jmoyer@redhat.com> Reported-by: Bart Van Assche <bvanassche@acm.org> Tested-by: Bart Van Assche <bvanassche@acm.org> Cc: <stable@kernel.org> [3.2.x only] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-03-05 22:59:12 +00:00
if (unlikely(!ctx->reqs_active && ctx->dead))
wake_up_all(&ctx->wait);
spin_unlock_irq(&ctx->ctx_lock);
}
/*
* Allocate a batch of kiocbs. This avoids taking and dropping the
* context lock a lot during setup.
*/
static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
{
unsigned short allocated, to_alloc;
long avail;
bool called_fput = false;
struct kiocb *req, *n;
struct aio_ring *ring;
to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
for (allocated = 0; allocated < to_alloc; allocated++) {
req = __aio_get_req(ctx);
if (!req)
/* allocation failed, go with what we've got */
break;
list_add(&req->ki_batch, &batch->head);
}
if (allocated == 0)
goto out;
retry:
spin_lock_irq(&ctx->ctx_lock);
ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
BUG_ON(avail < 0);
if (avail == 0 && !called_fput) {
/*
* Handle a potential starvation case. It is possible that
* we hold the last reference on a struct file, causing us
* to delay the final fput to non-irq context. In this case,
* ctx->reqs_active is artificially high. Calling the fput
* routine here may free up a slot in the event completion
* ring, allowing this allocation to succeed.
*/
kunmap_atomic(ring);
spin_unlock_irq(&ctx->ctx_lock);
aio_fput_routine(NULL);
called_fput = true;
goto retry;
}
if (avail < allocated) {
/* Trim back the number of requests. */
list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
list_del(&req->ki_batch);
kmem_cache_free(kiocb_cachep, req);
if (--allocated <= avail)
break;
}
}
batch->count -= allocated;
list_for_each_entry(req, &batch->head, ki_batch) {
list_add(&req->ki_list, &ctx->active_reqs);
ctx->reqs_active++;
}
kunmap_atomic(ring);
spin_unlock_irq(&ctx->ctx_lock);
out:
return allocated;
}
static inline struct kiocb *aio_get_req(struct kioctx *ctx,
struct kiocb_batch *batch)
{
struct kiocb *req;
if (list_empty(&batch->head))
if (kiocb_batch_refill(ctx, batch) == 0)
return NULL;
req = list_first_entry(&batch->head, struct kiocb, ki_batch);
list_del(&req->ki_batch);
return req;
}
static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
{
assert_spin_locked(&ctx->ctx_lock);
if (req->ki_eventfd != NULL)
eventfd_ctx_put(req->ki_eventfd);
if (req->ki_dtor)
req->ki_dtor(req);
if (req->ki_iovec != &req->ki_inline_vec)
kfree(req->ki_iovec);
kmem_cache_free(kiocb_cachep, req);
ctx->reqs_active--;
if (unlikely(!ctx->reqs_active && ctx->dead))
wake_up_all(&ctx->wait);
}
2006-11-22 14:55:48 +00:00
static void aio_fput_routine(struct work_struct *data)
{
spin_lock_irq(&fput_lock);
while (likely(!list_empty(&fput_head))) {
struct kiocb *req = list_kiocb(fput_head.next);
struct kioctx *ctx = req->ki_ctx;
list_del(&req->ki_list);
spin_unlock_irq(&fput_lock);
/* Complete the fput(s) */
if (req->ki_filp != NULL)
fput(req->ki_filp);
/* Link the iocb into the context's free list */
rcu_read_lock();
spin_lock_irq(&ctx->ctx_lock);
really_put_req(ctx, req);
/*
* at that point ctx might've been killed, but actual
* freeing is RCU'd
*/
spin_unlock_irq(&ctx->ctx_lock);
rcu_read_unlock();
spin_lock_irq(&fput_lock);
}
spin_unlock_irq(&fput_lock);
}
/* __aio_put_req
* Returns true if this put was the last user of the request.
*/
static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
{
dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
req, atomic_long_read(&req->ki_filp->f_count));
assert_spin_locked(&ctx->ctx_lock);
req->ki_users--;
BUG_ON(req->ki_users < 0);
if (likely(req->ki_users))
return 0;
list_del(&req->ki_list); /* remove from active_reqs */
req->ki_cancel = NULL;
req->ki_retry = NULL;
/*
* Try to optimize the aio and eventfd file* puts, by avoiding to
* schedule work in case it is not final fput() time. In normal cases,
* we would not be holding the last reference to the file*, so
* this function will be executed w/out any aio kthread wakeup.
*/
if (unlikely(!fput_atomic(req->ki_filp))) {
spin_lock(&fput_lock);
list_add(&req->ki_list, &fput_head);
spin_unlock(&fput_lock);
schedule_work(&fput_work);
} else {
req->ki_filp = NULL;
really_put_req(ctx, req);
}
return 1;
}
/* aio_put_req
* Returns true if this put was the last user of the kiocb,
* false if the request is still in use.
*/
int aio_put_req(struct kiocb *req)
{
struct kioctx *ctx = req->ki_ctx;
int ret;
spin_lock_irq(&ctx->ctx_lock);
ret = __aio_put_req(ctx, req);
spin_unlock_irq(&ctx->ctx_lock);
return ret;
}
EXPORT_SYMBOL(aio_put_req);
static struct kioctx *lookup_ioctx(unsigned long ctx_id)
{
struct mm_struct *mm = current->mm;
struct kioctx *ctx, *ret = NULL;
struct hlist_node *n;
rcu_read_lock();
hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
/*
* RCU protects us against accessing freed memory but
* we have to be careful not to get a reference when the
* reference count already dropped to 0 (ctx->dead test
* is unreliable because of races).
*/
if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
ret = ctx;
break;
}
}
rcu_read_unlock();
return ret;
}
/*
* Queue up a kiocb to be retried. Assumes that the kiocb
* has already been marked as kicked, and places it on
* the retry run list for the corresponding ioctx, if it
* isn't already queued. Returns 1 if it actually queued
* the kiocb (to tell the caller to activate the work
* queue to process it), or 0, if it found that it was
* already queued.
*/
static inline int __queue_kicked_iocb(struct kiocb *iocb)
{
struct kioctx *ctx = iocb->ki_ctx;
assert_spin_locked(&ctx->ctx_lock);
if (list_empty(&iocb->ki_run_list)) {
list_add_tail(&iocb->ki_run_list,
&ctx->run_list);
return 1;
}
return 0;
}
/* aio_run_iocb
* This is the core aio execution routine. It is
* invoked both for initial i/o submission and
* subsequent retries via the aio_kick_handler.
* Expects to be invoked with iocb->ki_ctx->lock
* already held. The lock is released and reacquired
* as needed during processing.
*
* Calls the iocb retry method (already setup for the
* iocb on initial submission) for operation specific
* handling, but takes care of most of common retry
* execution details for a given iocb. The retry method
* needs to be non-blocking as far as possible, to avoid
* holding up other iocbs waiting to be serviced by the
* retry kernel thread.
*
* The trickier parts in this code have to do with
* ensuring that only one retry instance is in progress
* for a given iocb at any time. Providing that guarantee
* simplifies the coding of individual aio operations as
* it avoids various potential races.
*/
static ssize_t aio_run_iocb(struct kiocb *iocb)
{
struct kioctx *ctx = iocb->ki_ctx;
ssize_t (*retry)(struct kiocb *);
ssize_t ret;
if (!(retry = iocb->ki_retry)) {
printk("aio_run_iocb: iocb->ki_retry = NULL\n");
return 0;
}
/*
* We don't want the next retry iteration for this
* operation to start until this one has returned and
* updated the iocb state. However, wait_queue functions
* can trigger a kick_iocb from interrupt context in the
* meantime, indicating that data is available for the next
* iteration. We want to remember that and enable the
* next retry iteration _after_ we are through with
* this one.
*
* So, in order to be able to register a "kick", but
* prevent it from being queued now, we clear the kick
* flag, but make the kick code *think* that the iocb is
* still on the run list until we are actually done.
* When we are done with this iteration, we check if
* the iocb was kicked in the meantime and if so, queue
* it up afresh.
*/
kiocbClearKicked(iocb);
/*
* This is so that aio_complete knows it doesn't need to
* pull the iocb off the run list (We can't just call
* INIT_LIST_HEAD because we don't want a kick_iocb to
* queue this on the run list yet)
*/
iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
spin_unlock_irq(&ctx->ctx_lock);
/* Quit retrying if the i/o has been cancelled */
if (kiocbIsCancelled(iocb)) {
ret = -EINTR;
aio_complete(iocb, ret, 0);
/* must not access the iocb after this */
goto out;
}
/*
* Now we are all set to call the retry method in async
* context.
*/
ret = retry(iocb);
if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
/*
* There's no easy way to restart the syscall since other AIO's
* may be already running. Just fail this IO with EINTR.
*/
if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
ret = -EINTR;
aio_complete(iocb, ret, 0);
}
out:
spin_lock_irq(&ctx->ctx_lock);
if (-EIOCBRETRY == ret) {
/*
* OK, now that we are done with this iteration
* and know that there is more left to go,
* this is where we let go so that a subsequent
* "kick" can start the next iteration
*/
/* will make __queue_kicked_iocb succeed from here on */
INIT_LIST_HEAD(&iocb->ki_run_list);
/* we must queue the next iteration ourselves, if it
* has already been kicked */
if (kiocbIsKicked(iocb)) {
__queue_kicked_iocb(iocb);
/*
* __queue_kicked_iocb will always return 1 here, because
* iocb->ki_run_list is empty at this point so it should
* be safe to unconditionally queue the context into the
* work queue.
*/
aio_queue_work(ctx);
}
}
return ret;
}
/*
* __aio_run_iocbs:
* Process all pending retries queued on the ioctx
* run list.
* Assumes it is operating within the aio issuer's mm
* context.
*/
static int __aio_run_iocbs(struct kioctx *ctx)
{
struct kiocb *iocb;
struct list_head run_list;
assert_spin_locked(&ctx->ctx_lock);
list_replace_init(&ctx->run_list, &run_list);
while (!list_empty(&run_list)) {
iocb = list_entry(run_list.next, struct kiocb,
ki_run_list);
list_del(&iocb->ki_run_list);
/*
* Hold an extra reference while retrying i/o.
*/
iocb->ki_users++; /* grab extra reference */
aio_run_iocb(iocb);
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
__aio_put_req(ctx, iocb);
}
if (!list_empty(&ctx->run_list))
return 1;
return 0;
}
static void aio_queue_work(struct kioctx * ctx)
{
unsigned long timeout;
/*
* if someone is waiting, get the work started right
* away, otherwise, use a longer delay
*/
smp_mb();
if (waitqueue_active(&ctx->wait))
timeout = 1;
else
timeout = HZ/10;
queue_delayed_work(aio_wq, &ctx->wq, timeout);
}
/*
* aio_run_all_iocbs:
* Process all pending retries queued on the ioctx
* run list, and keep running them until the list
* stays empty.
* Assumes it is operating within the aio issuer's mm context.
*/
static inline void aio_run_all_iocbs(struct kioctx *ctx)
{
spin_lock_irq(&ctx->ctx_lock);
while (__aio_run_iocbs(ctx))
;
spin_unlock_irq(&ctx->ctx_lock);
}
/*
* aio_kick_handler:
* Work queue handler triggered to process pending
* retries on an ioctx. Takes on the aio issuer's
* mm context before running the iocbs, so that
* copy_xxx_user operates on the issuer's address
* space.
* Run on aiod's context.
*/
2006-11-22 14:55:48 +00:00
static void aio_kick_handler(struct work_struct *work)
{
2006-11-22 14:55:48 +00:00
struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
mm_segment_t oldfs = get_fs();
struct mm_struct *mm;
int requeue;
set_fs(USER_DS);
use_mm(ctx->mm);
spin_lock_irq(&ctx->ctx_lock);
requeue =__aio_run_iocbs(ctx);
mm = ctx->mm;
spin_unlock_irq(&ctx->ctx_lock);
unuse_mm(mm);
set_fs(oldfs);
/*
* we're in a worker thread already; no point using non-zero delay
*/
if (requeue)
queue_delayed_work(aio_wq, &ctx->wq, 0);
}
/*
* Called by kick_iocb to queue the kiocb for retry
* and if required activate the aio work queue to process
* it
*/
static void try_queue_kicked_iocb(struct kiocb *iocb)
{
struct kioctx *ctx = iocb->ki_ctx;
unsigned long flags;
int run = 0;
spin_lock_irqsave(&ctx->ctx_lock, flags);
/* set this inside the lock so that we can't race with aio_run_iocb()
* testing it and putting the iocb on the run list under the lock */
if (!kiocbTryKick(iocb))
run = __queue_kicked_iocb(iocb);
spin_unlock_irqrestore(&ctx->ctx_lock, flags);
if (run)
aio_queue_work(ctx);
}
/*
* kick_iocb:
* Called typically from a wait queue callback context
* to trigger a retry of the iocb.
* The retry is usually executed by aio workqueue
* threads (See aio_kick_handler).
*/
void kick_iocb(struct kiocb *iocb)
{
/* sync iocbs are easy: they can only ever be executing from a
* single context. */
if (is_sync_kiocb(iocb)) {
kiocbSetKicked(iocb);
wake_up_process(iocb->ki_obj.tsk);
return;
}
try_queue_kicked_iocb(iocb);
}
EXPORT_SYMBOL(kick_iocb);
/* aio_complete
* Called when the io request on the given iocb is complete.
* Returns true if this is the last user of the request. The
* only other user of the request can be the cancellation code.
*/
int aio_complete(struct kiocb *iocb, long res, long res2)
{
struct kioctx *ctx = iocb->ki_ctx;
struct aio_ring_info *info;
struct aio_ring *ring;
struct io_event *event;
unsigned long flags;
unsigned long tail;
int ret;
/*
* Special case handling for sync iocbs:
* - events go directly into the iocb for fast handling
* - the sync task with the iocb in its stack holds the single iocb
* ref, no other paths have a way to get another ref
* - the sync task helpfully left a reference to itself in the iocb
*/
if (is_sync_kiocb(iocb)) {
BUG_ON(iocb->ki_users != 1);
iocb->ki_user_data = res;
iocb->ki_users = 0;
wake_up_process(iocb->ki_obj.tsk);
return 1;
}
info = &ctx->ring_info;
/* add a completion event to the ring buffer.
* must be done holding ctx->ctx_lock to prevent
* other code from messing with the tail
* pointer since we might be called from irq
* context.
*/
spin_lock_irqsave(&ctx->ctx_lock, flags);
if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
list_del_init(&iocb->ki_run_list);
/*
* cancelled requests don't get events, userland was given one
* when the event got cancelled.
*/
if (kiocbIsCancelled(iocb))
goto put_rq;
ring = kmap_atomic(info->ring_pages[0]);
tail = info->tail;
event = aio_ring_event(info, tail);
if (++tail >= info->nr)
tail = 0;
event->obj = (u64)(unsigned long)iocb->ki_obj.user;
event->data = iocb->ki_user_data;
event->res = res;
event->res2 = res2;
dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
ctx, tail, iocb, iocb->ki_obj.user, 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 */
info->tail = tail;
ring->tail = tail;
put_aio_ring_event(event);
kunmap_atomic(ring);
pr_debug("added to ring %p at [%lu]\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 != NULL)
eventfd_signal(iocb->ki_eventfd, 1);
put_rq:
/* everything turned out well, dispose of the aiocb. */
ret = __aio_put_req(ctx, iocb);
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);
[PATCH] aio: fix buggy put_ioctx call in aio_complete - v2 An AIO bug was reported that sleeping function is being called in softirq context: BUG: warning at kernel/mutex.c:132/__mutex_lock_common() Call Trace: [<a000000100577b00>] __mutex_lock_slowpath+0x640/0x6c0 [<a000000100577ba0>] mutex_lock+0x20/0x40 [<a0000001000a25b0>] flush_workqueue+0xb0/0x1a0 [<a00000010018c0c0>] __put_ioctx+0xc0/0x240 [<a00000010018d470>] aio_complete+0x2f0/0x420 [<a00000010019cc80>] finished_one_bio+0x200/0x2a0 [<a00000010019d1c0>] dio_bio_complete+0x1c0/0x200 [<a00000010019d260>] dio_bio_end_aio+0x60/0x80 [<a00000010014acd0>] bio_endio+0x110/0x1c0 [<a0000001002770e0>] __end_that_request_first+0x180/0xba0 [<a000000100277b90>] end_that_request_chunk+0x30/0x60 [<a0000002073c0c70>] scsi_end_request+0x50/0x300 [scsi_mod] [<a0000002073c1240>] scsi_io_completion+0x200/0x8a0 [scsi_mod] [<a0000002074729b0>] sd_rw_intr+0x330/0x860 [sd_mod] [<a0000002073b3ac0>] scsi_finish_command+0x100/0x1c0 [scsi_mod] [<a0000002073c2910>] scsi_softirq_done+0x230/0x300 [scsi_mod] [<a000000100277d20>] blk_done_softirq+0x160/0x1c0 [<a000000100083e00>] __do_softirq+0x200/0x240 [<a000000100083eb0>] do_softirq+0x70/0xc0 See report: http://marc.theaimsgroup.com/?l=linux-kernel&m=116599593200888&w=2 flush_workqueue() is not allowed to be called in the softirq context. However, aio_complete() called from I/O interrupt can potentially call put_ioctx with last ref count on ioctx and triggers bug. It is simply incorrect to perform ioctx freeing from aio_complete. The bug is trigger-able from a race between io_destroy() and aio_complete(). A possible scenario: cpu0 cpu1 io_destroy aio_complete wait_for_all_aios { __aio_put_req ... ctx->reqs_active--; if (!ctx->reqs_active) return; } ... put_ioctx(ioctx) put_ioctx(ctx); __put_ioctx bam! Bug trigger! The real problem is that the condition check of ctx->reqs_active in wait_for_all_aios() is incorrect that access to reqs_active is not being properly protected by spin lock. This patch adds that protective spin lock, and at the same time removes all duplicate ref counting for each kiocb as reqs_active is already used as a ref count for each active ioctx. This also ensures that buggy call to flush_workqueue() in softirq context is eliminated. Signed-off-by: "Ken Chen" <kenchen@google.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Benjamin LaHaise <bcrl@kvack.org> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: <stable@kernel.org> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-02-03 09:13:45 +00:00
spin_unlock_irqrestore(&ctx->ctx_lock, flags);
return ret;
}
EXPORT_SYMBOL(aio_complete);
/* aio_read_evt
* Pull an event off of the ioctx's event ring. Returns the number of
* events fetched (0 or 1 ;-)
* FIXME: make this use cmpxchg.
* TODO: make the ringbuffer user mmap()able (requires FIXME).
*/
static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
{
struct aio_ring_info *info = &ioctx->ring_info;
struct aio_ring *ring;
unsigned long head;
int ret = 0;
ring = kmap_atomic(info->ring_pages[0]);
dprintk("in aio_read_evt h%lu t%lu m%lu\n",
(unsigned long)ring->head, (unsigned long)ring->tail,
(unsigned long)ring->nr);
if (ring->head == ring->tail)
goto out;
spin_lock(&info->ring_lock);
head = ring->head % info->nr;
if (head != ring->tail) {
struct io_event *evp = aio_ring_event(info, head);
*ent = *evp;
head = (head + 1) % info->nr;
smp_mb(); /* finish reading the event before updatng the head */
ring->head = head;
ret = 1;
put_aio_ring_event(evp);
}
spin_unlock(&info->ring_lock);
out:
kunmap_atomic(ring);
dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
(unsigned long)ring->head, (unsigned long)ring->tail);
return ret;
}
struct aio_timeout {
struct timer_list timer;
int timed_out;
struct task_struct *p;
};
static void timeout_func(unsigned long data)
{
struct aio_timeout *to = (struct aio_timeout *)data;
to->timed_out = 1;
wake_up_process(to->p);
}
static inline void init_timeout(struct aio_timeout *to)
{
setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
to->timed_out = 0;
to->p = current;
}
static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
const struct timespec *ts)
{
to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
if (time_after(to->timer.expires, jiffies))
add_timer(&to->timer);
else
to->timed_out = 1;
}
static inline void clear_timeout(struct aio_timeout *to)
{
del_singleshot_timer_sync(&to->timer);
}
static int read_events(struct kioctx *ctx,
long min_nr, long nr,
struct io_event __user *event,
struct timespec __user *timeout)
{
long start_jiffies = jiffies;
struct task_struct *tsk = current;
DECLARE_WAITQUEUE(wait, tsk);
int ret;
int i = 0;
struct io_event ent;
struct aio_timeout to;
int retry = 0;
/* needed to zero any padding within an entry (there shouldn't be
* any, but C is fun!
*/
memset(&ent, 0, sizeof(ent));
retry:
ret = 0;
while (likely(i < nr)) {
ret = aio_read_evt(ctx, &ent);
if (unlikely(ret <= 0))
break;
dprintk("read event: %Lx %Lx %Lx %Lx\n",
ent.data, ent.obj, ent.res, ent.res2);
/* Could we split the check in two? */
ret = -EFAULT;
if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
dprintk("aio: lost an event due to EFAULT.\n");
break;
}
ret = 0;
/* Good, event copied to userland, update counts. */
event ++;
i ++;
}
if (min_nr <= i)
return i;
if (ret)
return ret;
/* End fast path */
/* racey check, but it gets redone */
if (!retry && unlikely(!list_empty(&ctx->run_list))) {
retry = 1;
aio_run_all_iocbs(ctx);
goto retry;
}
init_timeout(&to);
if (timeout) {
struct timespec ts;
ret = -EFAULT;
if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
goto out;
set_timeout(start_jiffies, &to, &ts);
}
while (likely(i < nr)) {
add_wait_queue_exclusive(&ctx->wait, &wait);
do {
set_task_state(tsk, TASK_INTERRUPTIBLE);
ret = aio_read_evt(ctx, &ent);
if (ret)
break;
if (min_nr <= i)
break;
if (unlikely(ctx->dead)) {
ret = -EINVAL;
break;
}
if (to.timed_out) /* Only check after read evt */
break;
/* Try to only show up in io wait if there are ops
* in flight */
if (ctx->reqs_active)
io_schedule();
else
schedule();
if (signal_pending(tsk)) {
ret = -EINTR;
break;
}
/*ret = aio_read_evt(ctx, &ent);*/
} while (1) ;
set_task_state(tsk, TASK_RUNNING);
remove_wait_queue(&ctx->wait, &wait);
if (unlikely(ret <= 0))
break;
ret = -EFAULT;
if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
dprintk("aio: lost an event due to EFAULT.\n");
break;
}
/* Good, event copied to userland, update counts. */
event ++;
i ++;
}
if (timeout)
clear_timeout(&to);
out:
destroy_timer_on_stack(&to.timer);
return i ? i : ret;
}
/* Take an ioctx and remove it from the list of ioctx's. Protects
* against races with itself via ->dead.
*/
static void io_destroy(struct kioctx *ioctx)
{
struct mm_struct *mm = current->mm;
int was_dead;
/* delete the entry from the list is someone else hasn't already */
spin_lock(&mm->ioctx_lock);
was_dead = ioctx->dead;
ioctx->dead = 1;
hlist_del_rcu(&ioctx->list);
spin_unlock(&mm->ioctx_lock);
dprintk("aio_release(%p)\n", ioctx);
if (likely(!was_dead))
put_ioctx(ioctx); /* twice for the list */
kill_ctx(ioctx);
/*
* Wake up any waiters. The setting of ctx->dead must be seen
* by other CPUs at this point. Right now, we rely on the
* locking done by the above calls to ensure this consistency.
*/
wake_up_all(&ioctx->wait);
}
/* 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: io_setup: 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)
io_destroy(ioctx);
put_ioctx(ioctx);
}
out:
return ret;
}
/* 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)) {
io_destroy(ioctx);
put_ioctx(ioctx);
return 0;
}
pr_debug("EINVAL: io_destroy: invalid context id\n");
return -EINVAL;
}
static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
{
struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
BUG_ON(ret <= 0);
while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
ssize_t this = min((ssize_t)iov->iov_len, ret);
iov->iov_base += this;
iov->iov_len -= this;
iocb->ki_left -= this;
ret -= this;
if (iov->iov_len == 0) {
iocb->ki_cur_seg++;
iov++;
}
}
/* the caller should not have done more io than what fit in
* the remaining iovecs */
BUG_ON(ret > 0 && iocb->ki_left == 0);
}
static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
{
struct file *file = iocb->ki_filp;
struct address_space *mapping = file->f_mapping;
struct inode *inode = mapping->host;
ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
unsigned long, loff_t);
ssize_t ret = 0;
unsigned short opcode;
if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
(iocb->ki_opcode == IOCB_CMD_PREAD)) {
rw_op = file->f_op->aio_read;
opcode = IOCB_CMD_PREADV;
} else {
rw_op = file->f_op->aio_write;
opcode = IOCB_CMD_PWRITEV;
}
/* This matches the pread()/pwrite() logic */
if (iocb->ki_pos < 0)
return -EINVAL;
do {
ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
iocb->ki_nr_segs - iocb->ki_cur_seg,
iocb->ki_pos);
if (ret > 0)
aio_advance_iovec(iocb, ret);
/* retry all partial writes. retry partial reads as long as its a
* regular file. */
} while (ret > 0 && iocb->ki_left > 0 &&
(opcode == IOCB_CMD_PWRITEV ||
(!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
/* This means we must have transferred all that we could */
/* No need to retry anymore */
if ((ret == 0) || (iocb->ki_left == 0))
ret = iocb->ki_nbytes - iocb->ki_left;
/* If we managed to write some out we return that, rather than
* the eventual error. */
if (opcode == IOCB_CMD_PWRITEV
&& ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
&& iocb->ki_nbytes - iocb->ki_left)
ret = iocb->ki_nbytes - iocb->ki_left;
return ret;
}
static ssize_t aio_fdsync(struct kiocb *iocb)
{
struct file *file = iocb->ki_filp;
ssize_t ret = -EINVAL;
if (file->f_op->aio_fsync)
ret = file->f_op->aio_fsync(iocb, 1);
return ret;
}
static ssize_t aio_fsync(struct kiocb *iocb)
{
struct file *file = iocb->ki_filp;
ssize_t ret = -EINVAL;
if (file->f_op->aio_fsync)
ret = file->f_op->aio_fsync(iocb, 0);
return ret;
}
static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
{
ssize_t ret;
#ifdef CONFIG_COMPAT
if (compat)
ret = compat_rw_copy_check_uvector(type,
(struct compat_iovec __user *)kiocb->ki_buf,
kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
Cross Memory Attach The basic idea behind cross memory attach is to allow MPI programs doing intra-node communication to do a single copy of the message rather than a double copy of the message via shared memory. The following patch attempts to achieve this by allowing a destination process, given an address and size from a source process, to copy memory directly from the source process into its own address space via a system call. There is also a symmetrical ability to copy from the current process's address space into a destination process's address space. - Use of /proc/pid/mem has been considered, but there are issues with using it: - Does not allow for specifying iovecs for both src and dest, assuming preadv or pwritev was implemented either the area read from or written to would need to be contiguous. - Currently mem_read allows only processes who are currently ptrace'ing the target and are still able to ptrace the target to read from the target. This check could possibly be moved to the open call, but its not clear exactly what race this restriction is stopping (reason appears to have been lost) - Having to send the fd of /proc/self/mem via SCM_RIGHTS on unix domain socket is a bit ugly from a userspace point of view, especially when you may have hundreds if not (eventually) thousands of processes that all need to do this with each other - Doesn't allow for some future use of the interface we would like to consider adding in the future (see below) - Interestingly reading from /proc/pid/mem currently actually involves two copies! (But this could be fixed pretty easily) As mentioned previously use of vmsplice instead was considered, but has problems. Since you need the reader and writer working co-operatively if the pipe is not drained then you block. Which requires some wrapping to do non blocking on the send side or polling on the receive. In all to all communication it requires ordering otherwise you can deadlock. And in the example of many MPI tasks writing to one MPI task vmsplice serialises the copying. There are some cases of MPI collectives where even a single copy interface does not get us the performance gain we could. For example in an MPI_Reduce rather than copy the data from the source we would like to instead use it directly in a mathops (say the reduce is doing a sum) as this would save us doing a copy. We don't need to keep a copy of the data from the source. I haven't implemented this, but I think this interface could in the future do all this through the use of the flags - eg could specify the math operation and type and the kernel rather than just copying the data would apply the specified operation between the source and destination and store it in the destination. Although we don't have a "second user" of the interface (though I've had some nibbles from people who may be interested in using it for intra process messaging which is not MPI). This interface is something which hardware vendors are already doing for their custom drivers to implement fast local communication. And so in addition to this being useful for OpenMPI it would mean the driver maintainers don't have to fix things up when the mm changes. There was some discussion about how much faster a true zero copy would go. Here's a link back to the email with some testing I did on that: http://marc.info/?l=linux-mm&m=130105930902915&w=2 There is a basic man page for the proposed interface here: http://ozlabs.org/~cyeoh/cma/process_vm_readv.txt This has been implemented for x86 and powerpc, other architecture should mainly (I think) just need to add syscall numbers for the process_vm_readv and process_vm_writev. There are 32 bit compatibility versions for 64-bit kernels. For arch maintainers there are some simple tests to be able to quickly verify that the syscalls are working correctly here: http://ozlabs.org/~cyeoh/cma/cma-test-20110718.tgz Signed-off-by: Chris Yeoh <yeohc@au1.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: David Howells <dhowells@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: <linux-man@vger.kernel.org> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 00:06:39 +00:00
&kiocb->ki_iovec, 1);
else
#endif
ret = rw_copy_check_uvector(type,
(struct iovec __user *)kiocb->ki_buf,
kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
Cross Memory Attach The basic idea behind cross memory attach is to allow MPI programs doing intra-node communication to do a single copy of the message rather than a double copy of the message via shared memory. The following patch attempts to achieve this by allowing a destination process, given an address and size from a source process, to copy memory directly from the source process into its own address space via a system call. There is also a symmetrical ability to copy from the current process's address space into a destination process's address space. - Use of /proc/pid/mem has been considered, but there are issues with using it: - Does not allow for specifying iovecs for both src and dest, assuming preadv or pwritev was implemented either the area read from or written to would need to be contiguous. - Currently mem_read allows only processes who are currently ptrace'ing the target and are still able to ptrace the target to read from the target. This check could possibly be moved to the open call, but its not clear exactly what race this restriction is stopping (reason appears to have been lost) - Having to send the fd of /proc/self/mem via SCM_RIGHTS on unix domain socket is a bit ugly from a userspace point of view, especially when you may have hundreds if not (eventually) thousands of processes that all need to do this with each other - Doesn't allow for some future use of the interface we would like to consider adding in the future (see below) - Interestingly reading from /proc/pid/mem currently actually involves two copies! (But this could be fixed pretty easily) As mentioned previously use of vmsplice instead was considered, but has problems. Since you need the reader and writer working co-operatively if the pipe is not drained then you block. Which requires some wrapping to do non blocking on the send side or polling on the receive. In all to all communication it requires ordering otherwise you can deadlock. And in the example of many MPI tasks writing to one MPI task vmsplice serialises the copying. There are some cases of MPI collectives where even a single copy interface does not get us the performance gain we could. For example in an MPI_Reduce rather than copy the data from the source we would like to instead use it directly in a mathops (say the reduce is doing a sum) as this would save us doing a copy. We don't need to keep a copy of the data from the source. I haven't implemented this, but I think this interface could in the future do all this through the use of the flags - eg could specify the math operation and type and the kernel rather than just copying the data would apply the specified operation between the source and destination and store it in the destination. Although we don't have a "second user" of the interface (though I've had some nibbles from people who may be interested in using it for intra process messaging which is not MPI). This interface is something which hardware vendors are already doing for their custom drivers to implement fast local communication. And so in addition to this being useful for OpenMPI it would mean the driver maintainers don't have to fix things up when the mm changes. There was some discussion about how much faster a true zero copy would go. Here's a link back to the email with some testing I did on that: http://marc.info/?l=linux-mm&m=130105930902915&w=2 There is a basic man page for the proposed interface here: http://ozlabs.org/~cyeoh/cma/process_vm_readv.txt This has been implemented for x86 and powerpc, other architecture should mainly (I think) just need to add syscall numbers for the process_vm_readv and process_vm_writev. There are 32 bit compatibility versions for 64-bit kernels. For arch maintainers there are some simple tests to be able to quickly verify that the syscalls are working correctly here: http://ozlabs.org/~cyeoh/cma/cma-test-20110718.tgz Signed-off-by: Chris Yeoh <yeohc@au1.ibm.com> Cc: Ingo Molnar <mingo@elte.hu> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: David Howells <dhowells@redhat.com> Cc: James Morris <jmorris@namei.org> Cc: <linux-man@vger.kernel.org> Cc: <linux-arch@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-01 00:06:39 +00:00
&kiocb->ki_iovec, 1);
if (ret < 0)
goto out;
kiocb->ki_nr_segs = kiocb->ki_nbytes;
kiocb->ki_cur_seg = 0;
/* ki_nbytes/left now reflect bytes instead of segs */
kiocb->ki_nbytes = ret;
kiocb->ki_left = ret;
ret = 0;
out:
return ret;
}
static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
{
kiocb->ki_iovec = &kiocb->ki_inline_vec;
kiocb->ki_iovec->iov_base = kiocb->ki_buf;
kiocb->ki_iovec->iov_len = kiocb->ki_left;
kiocb->ki_nr_segs = 1;
kiocb->ki_cur_seg = 0;
return 0;
}
/*
* aio_setup_iocb:
* Performs the initial checks and aio retry method
* setup for the kiocb at the time of io submission.
*/
static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
{
struct file *file = kiocb->ki_filp;
ssize_t ret = 0;
switch (kiocb->ki_opcode) {
case IOCB_CMD_PREAD:
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_READ)))
break;
ret = -EFAULT;
if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
kiocb->ki_left)))
break;
ret = security_file_permission(file, MAY_READ);
if (unlikely(ret))
break;
ret = aio_setup_single_vector(kiocb);
if (ret)
break;
ret = -EINVAL;
if (file->f_op->aio_read)
kiocb->ki_retry = aio_rw_vect_retry;
break;
case IOCB_CMD_PWRITE:
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_WRITE)))
break;
ret = -EFAULT;
if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
kiocb->ki_left)))
break;
ret = security_file_permission(file, MAY_WRITE);
if (unlikely(ret))
break;
ret = aio_setup_single_vector(kiocb);
if (ret)
break;
ret = -EINVAL;
if (file->f_op->aio_write)
kiocb->ki_retry = aio_rw_vect_retry;
break;
case IOCB_CMD_PREADV:
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_READ)))
break;
ret = security_file_permission(file, MAY_READ);
if (unlikely(ret))
break;
ret = aio_setup_vectored_rw(READ, kiocb, compat);
if (ret)
break;
ret = -EINVAL;
if (file->f_op->aio_read)
kiocb->ki_retry = aio_rw_vect_retry;
break;
case IOCB_CMD_PWRITEV:
ret = -EBADF;
if (unlikely(!(file->f_mode & FMODE_WRITE)))
break;
ret = security_file_permission(file, MAY_WRITE);
if (unlikely(ret))
break;
ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
if (ret)
break;
ret = -EINVAL;
if (file->f_op->aio_write)
kiocb->ki_retry = aio_rw_vect_retry;
break;
case IOCB_CMD_FDSYNC:
ret = -EINVAL;
if (file->f_op->aio_fsync)
kiocb->ki_retry = aio_fdsync;
break;
case IOCB_CMD_FSYNC:
ret = -EINVAL;
if (file->f_op->aio_fsync)
kiocb->ki_retry = aio_fsync;
break;
default:
dprintk("EINVAL: io_submit: no operation provided\n");
ret = -EINVAL;
}
if (!kiocb->ki_retry)
return ret;
return 0;
}
static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
struct iocb *iocb, struct kiocb_batch *batch,
bool compat)
{
struct kiocb *req;
struct file *file;
ssize_t ret;
/* enforce forwards compatibility on users */
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 (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
pr_debug("EINVAL: io_submit: 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: io_submit: overflow check\n");
return -EINVAL;
}
file = fget(iocb->aio_fildes);
if (unlikely(!file))
return -EBADF;
req = aio_get_req(ctx, batch); /* returns with 2 references to req */
if (unlikely(!req)) {
fput(file);
return -EAGAIN;
}
req->ki_filp = file;
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 (iocb->aio_flags & IOCB_FLAG_RESFD) {
/*
* 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(req->ki_key, &user_iocb->aio_key);
if (unlikely(ret)) {
dprintk("EFAULT: aio_key\n");
goto out_put_req;
}
req->ki_obj.user = user_iocb;
req->ki_user_data = iocb->aio_data;
req->ki_pos = iocb->aio_offset;
req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
req->ki_opcode = iocb->aio_lio_opcode;
ret = aio_setup_iocb(req, compat);
if (ret)
goto out_put_req;
spin_lock_irq(&ctx->ctx_lock);
/*
* We could have raced with io_destroy() and are currently holding a
* reference to ctx which should be destroyed. We cannot submit IO
* since ctx gets freed as soon as io_submit() puts its reference. The
* check here is reliable: io_destroy() sets ctx->dead before waiting
* for outstanding IO and the barrier between these two is realized by
* unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
* increment ctx->reqs_active before checking for ctx->dead and the
* barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
* don't see ctx->dead set here, io_destroy() waits for our IO to
* finish.
*/
if (ctx->dead) {
spin_unlock_irq(&ctx->ctx_lock);
ret = -EINVAL;
goto out_put_req;
}
aio_run_iocb(req);
if (!list_empty(&ctx->run_list)) {
/* drain the run list */
while (__aio_run_iocbs(ctx))
;
}
spin_unlock_irq(&ctx->ctx_lock);
aio_put_req(req); /* drop extra ref to req */
return 0;
out_put_req:
aio_put_req(req); /* drop extra ref to req */
aio_put_req(req); /* drop i/o ref to req */
return ret;
}
long do_io_submit(aio_context_t ctx_id, long nr,
struct iocb __user *__user *iocbpp, bool compat)
{
struct kioctx *ctx;
long ret = 0;
int i = 0;
struct blk_plug plug;
struct kiocb_batch batch;
if (unlikely(nr < 0))
return -EINVAL;
if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
nr = LONG_MAX/sizeof(*iocbpp);
if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
return -EFAULT;
ctx = lookup_ioctx(ctx_id);
if (unlikely(!ctx)) {
pr_debug("EINVAL: io_submit: invalid context id\n");
return -EINVAL;
}
kiocb_batch_init(&batch, nr);
blk_start_plug(&plug);
/*
* AKPM: should this return a partial result if some of the IOs were
* successfully submitted?
*/
for (i=0; i<nr; i++) {
struct iocb __user *user_iocb;
struct iocb tmp;
if (unlikely(__get_user(user_iocb, iocbpp + i))) {
ret = -EFAULT;
break;
}
if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
ret = -EFAULT;
break;
}
ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
if (ret)
break;
}
blk_finish_plug(&plug);
kiocb_batch_free(ctx, &batch);
put_ioctx(ctx);
return i ? i : 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)
{
return do_io_submit(ctx_id, nr, iocbpp, 0);
}
/* lookup_kiocb
* Finds a given iocb for cancellation.
*/
static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
u32 key)
{
struct list_head *pos;
assert_spin_locked(&ctx->ctx_lock);
/* TODO: use a hash or array, this sucks. */
list_for_each(pos, &ctx->active_reqs) {
struct kiocb *kiocb = list_kiocb(pos);
if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
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)
{
int (*cancel)(struct kiocb *iocb, struct io_event *res);
struct kioctx *ctx;
struct kiocb *kiocb;
u32 key;
int ret;
ret = get_user(key, &iocb->aio_key);
if (unlikely(ret))
return -EFAULT;
ctx = lookup_ioctx(ctx_id);
if (unlikely(!ctx))
return -EINVAL;
spin_lock_irq(&ctx->ctx_lock);
ret = -EAGAIN;
kiocb = lookup_kiocb(ctx, iocb, key);
if (kiocb && kiocb->ki_cancel) {
cancel = kiocb->ki_cancel;
kiocb->ki_users ++;
kiocbSetCancelled(kiocb);
} else
cancel = NULL;
spin_unlock_irq(&ctx->ctx_lock);
if (NULL != cancel) {
struct io_event tmp;
pr_debug("calling cancel\n");
memset(&tmp, 0, sizeof(tmp));
tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
tmp.data = kiocb->ki_user_data;
ret = cancel(kiocb, &tmp);
if (!ret) {
/* Cancellation succeeded -- copy the result
* into the user's buffer.
*/
if (copy_to_user(result, &tmp, sizeof(tmp)))
ret = -EFAULT;
}
} else
ret = -EINVAL;
put_ioctx(ctx);
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 and will be updated if not NULL and the
* operation blocks. 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 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, timeout);
put_ioctx(ioctx);
}
asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
return ret;
}