2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig.
10 * See ../COPYING for licensing terms.
12 #define pr_fmt(fmt) "%s: " fmt, __func__
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
25 #include <linux/sched/signal.h>
27 #include <linux/file.h>
29 #include <linux/mman.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h>
46 #include <linux/uaccess.h>
47 #include <linux/nospec.h>
53 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0
57 unsigned id
; /* kernel internal index number */
58 unsigned nr
; /* number of io_events */
59 unsigned head
; /* Written to by userland or under ring_lock
60 * mutex by aio_read_events_ring(). */
64 unsigned compat_features
;
65 unsigned incompat_features
;
66 unsigned header_length
; /* size of aio_ring */
69 struct io_event io_events
[];
70 }; /* 128 bytes + ring size */
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
76 #define AIO_PLUG_THRESHOLD 2
78 #define AIO_RING_PAGES 8
83 struct kioctx __rcu
*table
[] __counted_by(nr
);
87 unsigned reqs_available
;
91 struct completion comp
;
96 struct percpu_ref users
;
99 struct percpu_ref reqs
;
101 unsigned long user_id
;
103 struct __percpu kioctx_cpu
*cpu
;
106 * For percpu reqs_available, number of slots we move to/from global
111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the global max_reqs quota.
114 * The real limit is nr_events - 1, which will be larger (see
119 /* Size of ringbuffer, in units of struct io_event */
122 unsigned long mmap_base
;
123 unsigned long mmap_size
;
125 struct page
**ring_pages
;
128 struct rcu_work free_rwork
; /* see free_ioctx() */
131 * signals when all in-flight requests are done
133 struct ctx_rq_wait
*rq_wait
;
137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the ringbuffer.
142 * We batch accesses to it with a percpu version.
144 atomic_t reqs_available
;
145 } ____cacheline_aligned_in_smp
;
149 struct list_head active_reqs
; /* used for cancellation */
150 } ____cacheline_aligned_in_smp
;
153 struct mutex ring_lock
;
154 wait_queue_head_t wait
;
155 } ____cacheline_aligned_in_smp
;
159 unsigned completed_events
;
160 spinlock_t completion_lock
;
161 } ____cacheline_aligned_in_smp
;
163 struct page
*internal_pages
[AIO_RING_PAGES
];
164 struct file
*aio_ring_file
;
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
175 struct work_struct work
;
182 struct wait_queue_head
*head
;
186 bool work_need_resched
;
187 struct wait_queue_entry wait
;
188 struct work_struct work
;
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
199 struct file
*ki_filp
;
201 struct fsync_iocb fsync
;
202 struct poll_iocb poll
;
205 struct kioctx
*ki_ctx
;
206 kiocb_cancel_fn
*ki_cancel
;
208 struct io_event ki_res
;
210 struct list_head ki_list
; /* the aio core uses this
211 * for cancellation */
212 refcount_t ki_refcnt
;
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
218 struct eventfd_ctx
*ki_eventfd
;
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock
);
223 static unsigned long aio_nr
; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct ctl_table aio_sysctls
[] = {
229 .procname
= "aio-nr",
231 .maxlen
= sizeof(aio_nr
),
233 .proc_handler
= proc_doulongvec_minmax
,
236 .procname
= "aio-max-nr",
238 .maxlen
= sizeof(aio_max_nr
),
240 .proc_handler
= proc_doulongvec_minmax
,
244 static void __init
aio_sysctl_init(void)
246 register_sysctl_init("fs", aio_sysctls
);
249 #define aio_sysctl_init() do { } while (0)
252 static struct kmem_cache
*kiocb_cachep
;
253 static struct kmem_cache
*kioctx_cachep
;
255 static struct vfsmount
*aio_mnt
;
257 static const struct file_operations aio_ring_fops
;
258 static const struct address_space_operations aio_ctx_aops
;
260 static struct file
*aio_private_file(struct kioctx
*ctx
, loff_t nr_pages
)
263 struct inode
*inode
= alloc_anon_inode(aio_mnt
->mnt_sb
);
265 return ERR_CAST(inode
);
267 inode
->i_mapping
->a_ops
= &aio_ctx_aops
;
268 inode
->i_mapping
->i_private_data
= ctx
;
269 inode
->i_size
= PAGE_SIZE
* nr_pages
;
271 file
= alloc_file_pseudo(inode
, aio_mnt
, "[aio]",
272 O_RDWR
, &aio_ring_fops
);
278 static int aio_init_fs_context(struct fs_context
*fc
)
280 if (!init_pseudo(fc
, AIO_RING_MAGIC
))
282 fc
->s_iflags
|= SB_I_NOEXEC
;
287 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during the boot sequence.
290 static int __init
aio_setup(void)
292 static struct file_system_type aio_fs
= {
294 .init_fs_context
= aio_init_fs_context
,
295 .kill_sb
= kill_anon_super
,
297 aio_mnt
= kern_mount(&aio_fs
);
299 panic("Failed to create aio fs mount.");
301 kiocb_cachep
= KMEM_CACHE(aio_kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
302 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
306 __initcall(aio_setup
);
308 static void put_aio_ring_file(struct kioctx
*ctx
)
310 struct file
*aio_ring_file
= ctx
->aio_ring_file
;
311 struct address_space
*i_mapping
;
314 truncate_setsize(file_inode(aio_ring_file
), 0);
316 /* Prevent further access to the kioctx from migratepages */
317 i_mapping
= aio_ring_file
->f_mapping
;
318 spin_lock(&i_mapping
->i_private_lock
);
319 i_mapping
->i_private_data
= NULL
;
320 ctx
->aio_ring_file
= NULL
;
321 spin_unlock(&i_mapping
->i_private_lock
);
327 static void aio_free_ring(struct kioctx
*ctx
)
331 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page migration.
334 put_aio_ring_file(ctx
);
336 for (i
= 0; i
< ctx
->nr_pages
; i
++) {
338 pr_debug("pid(%d) [%d] page->count=%d\n", current
->pid
, i
,
339 page_count(ctx
->ring_pages
[i
]));
340 page
= ctx
->ring_pages
[i
];
343 ctx
->ring_pages
[i
] = NULL
;
347 if (ctx
->ring_pages
&& ctx
->ring_pages
!= ctx
->internal_pages
) {
348 kfree(ctx
->ring_pages
);
349 ctx
->ring_pages
= NULL
;
353 static int aio_ring_mremap(struct vm_area_struct
*vma
)
355 struct file
*file
= vma
->vm_file
;
356 struct mm_struct
*mm
= vma
->vm_mm
;
357 struct kioctx_table
*table
;
358 int i
, res
= -EINVAL
;
360 spin_lock(&mm
->ioctx_lock
);
362 table
= rcu_dereference(mm
->ioctx_table
);
366 for (i
= 0; i
< table
->nr
; i
++) {
369 ctx
= rcu_dereference(table
->table
[i
]);
370 if (ctx
&& ctx
->aio_ring_file
== file
) {
371 if (!atomic_read(&ctx
->dead
)) {
372 ctx
->user_id
= ctx
->mmap_base
= vma
->vm_start
;
381 spin_unlock(&mm
->ioctx_lock
);
385 static const struct vm_operations_struct aio_ring_vm_ops
= {
386 .mremap
= aio_ring_mremap
,
387 #if IS_ENABLED(CONFIG_MMU)
388 .fault
= filemap_fault
,
389 .map_pages
= filemap_map_pages
,
390 .page_mkwrite
= filemap_page_mkwrite
,
394 static int aio_ring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
396 vm_flags_set(vma
, VM_DONTEXPAND
);
397 vma
->vm_ops
= &aio_ring_vm_ops
;
401 static const struct file_operations aio_ring_fops
= {
402 .mmap
= aio_ring_mmap
,
405 #if IS_ENABLED(CONFIG_MIGRATION)
406 static int aio_migrate_folio(struct address_space
*mapping
, struct folio
*dst
,
407 struct folio
*src
, enum migrate_mode mode
)
415 * We cannot support the _NO_COPY case here, because copy needs to
416 * happen under the ctx->completion_lock. That does not work with the
417 * migration workflow of MIGRATE_SYNC_NO_COPY.
419 if (mode
== MIGRATE_SYNC_NO_COPY
)
424 /* mapping->i_private_lock here protects against the kioctx teardown. */
425 spin_lock(&mapping
->i_private_lock
);
426 ctx
= mapping
->i_private_data
;
432 /* The ring_lock mutex. The prevents aio_read_events() from writing
433 * to the ring's head, and prevents page migration from mucking in
434 * a partially initialized kiotx.
436 if (!mutex_trylock(&ctx
->ring_lock
)) {
442 if (idx
< (pgoff_t
)ctx
->nr_pages
) {
443 /* Make sure the old folio hasn't already been changed */
444 if (ctx
->ring_pages
[idx
] != &src
->page
)
452 /* Writeback must be complete */
453 BUG_ON(folio_test_writeback(src
));
456 rc
= folio_migrate_mapping(mapping
, dst
, src
, 1);
457 if (rc
!= MIGRATEPAGE_SUCCESS
) {
462 /* Take completion_lock to prevent other writes to the ring buffer
463 * while the old folio is copied to the new. This prevents new
464 * events from being lost.
466 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
467 folio_migrate_copy(dst
, src
);
468 BUG_ON(ctx
->ring_pages
[idx
] != &src
->page
);
469 ctx
->ring_pages
[idx
] = &dst
->page
;
470 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
472 /* The old folio is no longer accessible. */
476 mutex_unlock(&ctx
->ring_lock
);
478 spin_unlock(&mapping
->i_private_lock
);
482 #define aio_migrate_folio NULL
485 static const struct address_space_operations aio_ctx_aops
= {
486 .dirty_folio
= noop_dirty_folio
,
487 .migrate_folio
= aio_migrate_folio
,
490 static int aio_setup_ring(struct kioctx
*ctx
, unsigned int nr_events
)
492 struct aio_ring
*ring
;
493 struct mm_struct
*mm
= current
->mm
;
494 unsigned long size
, unused
;
499 /* Compensate for the ring buffer's head/tail overlap entry */
500 nr_events
+= 2; /* 1 is required, 2 for good luck */
502 size
= sizeof(struct aio_ring
);
503 size
+= sizeof(struct io_event
) * nr_events
;
505 nr_pages
= PFN_UP(size
);
509 file
= aio_private_file(ctx
, nr_pages
);
511 ctx
->aio_ring_file
= NULL
;
515 ctx
->aio_ring_file
= file
;
516 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
))
517 / sizeof(struct io_event
);
519 ctx
->ring_pages
= ctx
->internal_pages
;
520 if (nr_pages
> AIO_RING_PAGES
) {
521 ctx
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
523 if (!ctx
->ring_pages
) {
524 put_aio_ring_file(ctx
);
529 for (i
= 0; i
< nr_pages
; i
++) {
531 page
= find_or_create_page(file
->f_mapping
,
532 i
, GFP_USER
| __GFP_ZERO
);
535 pr_debug("pid(%d) page[%d]->count=%d\n",
536 current
->pid
, i
, page_count(page
));
537 SetPageUptodate(page
);
540 ctx
->ring_pages
[i
] = page
;
544 if (unlikely(i
!= nr_pages
)) {
549 ctx
->mmap_size
= nr_pages
* PAGE_SIZE
;
550 pr_debug("attempting mmap of %lu bytes\n", ctx
->mmap_size
);
552 if (mmap_write_lock_killable(mm
)) {
558 ctx
->mmap_base
= do_mmap(ctx
->aio_ring_file
, 0, ctx
->mmap_size
,
559 PROT_READ
| PROT_WRITE
,
560 MAP_SHARED
, 0, 0, &unused
, NULL
);
561 mmap_write_unlock(mm
);
562 if (IS_ERR((void *)ctx
->mmap_base
)) {
568 pr_debug("mmap address: 0x%08lx\n", ctx
->mmap_base
);
570 ctx
->user_id
= ctx
->mmap_base
;
571 ctx
->nr_events
= nr_events
; /* trusted copy */
573 ring
= page_address(ctx
->ring_pages
[0]);
574 ring
->nr
= nr_events
; /* user copy */
576 ring
->head
= ring
->tail
= 0;
577 ring
->magic
= AIO_RING_MAGIC
;
578 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
579 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
580 ring
->header_length
= sizeof(struct aio_ring
);
581 flush_dcache_page(ctx
->ring_pages
[0]);
586 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
587 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
588 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
590 void kiocb_set_cancel_fn(struct kiocb
*iocb
, kiocb_cancel_fn
*cancel
)
592 struct aio_kiocb
*req
= container_of(iocb
, struct aio_kiocb
, rw
);
593 struct kioctx
*ctx
= req
->ki_ctx
;
596 if (WARN_ON_ONCE(!list_empty(&req
->ki_list
)))
599 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
600 list_add_tail(&req
->ki_list
, &ctx
->active_reqs
);
601 req
->ki_cancel
= cancel
;
602 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
604 EXPORT_SYMBOL(kiocb_set_cancel_fn
);
607 * free_ioctx() should be RCU delayed to synchronize against the RCU
608 * protected lookup_ioctx() and also needs process context to call
609 * aio_free_ring(). Use rcu_work.
611 static void free_ioctx(struct work_struct
*work
)
613 struct kioctx
*ctx
= container_of(to_rcu_work(work
), struct kioctx
,
615 pr_debug("freeing %p\n", ctx
);
618 free_percpu(ctx
->cpu
);
619 percpu_ref_exit(&ctx
->reqs
);
620 percpu_ref_exit(&ctx
->users
);
621 kmem_cache_free(kioctx_cachep
, ctx
);
624 static void free_ioctx_reqs(struct percpu_ref
*ref
)
626 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, reqs
);
628 /* At this point we know that there are no any in-flight requests */
629 if (ctx
->rq_wait
&& atomic_dec_and_test(&ctx
->rq_wait
->count
))
630 complete(&ctx
->rq_wait
->comp
);
632 /* Synchronize against RCU protected table->table[] dereferences */
633 INIT_RCU_WORK(&ctx
->free_rwork
, free_ioctx
);
634 queue_rcu_work(system_wq
, &ctx
->free_rwork
);
638 * When this function runs, the kioctx has been removed from the "hash table"
639 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
640 * now it's safe to cancel any that need to be.
642 static void free_ioctx_users(struct percpu_ref
*ref
)
644 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, users
);
645 struct aio_kiocb
*req
;
647 spin_lock_irq(&ctx
->ctx_lock
);
649 while (!list_empty(&ctx
->active_reqs
)) {
650 req
= list_first_entry(&ctx
->active_reqs
,
651 struct aio_kiocb
, ki_list
);
652 req
->ki_cancel(&req
->rw
);
653 list_del_init(&req
->ki_list
);
656 spin_unlock_irq(&ctx
->ctx_lock
);
658 percpu_ref_kill(&ctx
->reqs
);
659 percpu_ref_put(&ctx
->reqs
);
662 static int ioctx_add_table(struct kioctx
*ctx
, struct mm_struct
*mm
)
665 struct kioctx_table
*table
, *old
;
666 struct aio_ring
*ring
;
668 spin_lock(&mm
->ioctx_lock
);
669 table
= rcu_dereference_raw(mm
->ioctx_table
);
673 for (i
= 0; i
< table
->nr
; i
++)
674 if (!rcu_access_pointer(table
->table
[i
])) {
676 rcu_assign_pointer(table
->table
[i
], ctx
);
677 spin_unlock(&mm
->ioctx_lock
);
679 /* While kioctx setup is in progress,
680 * we are protected from page migration
681 * changes ring_pages by ->ring_lock.
683 ring
= page_address(ctx
->ring_pages
[0]);
688 new_nr
= (table
? table
->nr
: 1) * 4;
689 spin_unlock(&mm
->ioctx_lock
);
691 table
= kzalloc(struct_size(table
, table
, new_nr
), GFP_KERNEL
);
697 spin_lock(&mm
->ioctx_lock
);
698 old
= rcu_dereference_raw(mm
->ioctx_table
);
701 rcu_assign_pointer(mm
->ioctx_table
, table
);
702 } else if (table
->nr
> old
->nr
) {
703 memcpy(table
->table
, old
->table
,
704 old
->nr
* sizeof(struct kioctx
*));
706 rcu_assign_pointer(mm
->ioctx_table
, table
);
715 static void aio_nr_sub(unsigned nr
)
717 spin_lock(&aio_nr_lock
);
718 if (WARN_ON(aio_nr
- nr
> aio_nr
))
722 spin_unlock(&aio_nr_lock
);
726 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
728 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
730 struct mm_struct
*mm
= current
->mm
;
735 * Store the original nr_events -- what userspace passed to io_setup(),
736 * for counting against the global limit -- before it changes.
738 unsigned int max_reqs
= nr_events
;
741 * We keep track of the number of available ringbuffer slots, to prevent
742 * overflow (reqs_available), and we also use percpu counters for this.
744 * So since up to half the slots might be on other cpu's percpu counters
745 * and unavailable, double nr_events so userspace sees what they
746 * expected: additionally, we move req_batch slots to/from percpu
747 * counters at a time, so make sure that isn't 0:
749 nr_events
= max(nr_events
, num_possible_cpus() * 4);
752 /* Prevent overflows */
753 if (nr_events
> (0x10000000U
/ sizeof(struct io_event
))) {
754 pr_debug("ENOMEM: nr_events too high\n");
755 return ERR_PTR(-EINVAL
);
758 if (!nr_events
|| (unsigned long)max_reqs
> aio_max_nr
)
759 return ERR_PTR(-EAGAIN
);
761 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
763 return ERR_PTR(-ENOMEM
);
765 ctx
->max_reqs
= max_reqs
;
767 spin_lock_init(&ctx
->ctx_lock
);
768 spin_lock_init(&ctx
->completion_lock
);
769 mutex_init(&ctx
->ring_lock
);
770 /* Protect against page migration throughout kiotx setup by keeping
771 * the ring_lock mutex held until setup is complete. */
772 mutex_lock(&ctx
->ring_lock
);
773 init_waitqueue_head(&ctx
->wait
);
775 INIT_LIST_HEAD(&ctx
->active_reqs
);
777 if (percpu_ref_init(&ctx
->users
, free_ioctx_users
, 0, GFP_KERNEL
))
780 if (percpu_ref_init(&ctx
->reqs
, free_ioctx_reqs
, 0, GFP_KERNEL
))
783 ctx
->cpu
= alloc_percpu(struct kioctx_cpu
);
787 err
= aio_setup_ring(ctx
, nr_events
);
791 atomic_set(&ctx
->reqs_available
, ctx
->nr_events
- 1);
792 ctx
->req_batch
= (ctx
->nr_events
- 1) / (num_possible_cpus() * 4);
793 if (ctx
->req_batch
< 1)
796 /* limit the number of system wide aios */
797 spin_lock(&aio_nr_lock
);
798 if (aio_nr
+ ctx
->max_reqs
> aio_max_nr
||
799 aio_nr
+ ctx
->max_reqs
< aio_nr
) {
800 spin_unlock(&aio_nr_lock
);
804 aio_nr
+= ctx
->max_reqs
;
805 spin_unlock(&aio_nr_lock
);
807 percpu_ref_get(&ctx
->users
); /* io_setup() will drop this ref */
808 percpu_ref_get(&ctx
->reqs
); /* free_ioctx_users() will drop this */
810 err
= ioctx_add_table(ctx
, mm
);
814 /* Release the ring_lock mutex now that all setup is complete. */
815 mutex_unlock(&ctx
->ring_lock
);
817 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
818 ctx
, ctx
->user_id
, mm
, ctx
->nr_events
);
822 aio_nr_sub(ctx
->max_reqs
);
824 atomic_set(&ctx
->dead
, 1);
826 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
829 mutex_unlock(&ctx
->ring_lock
);
830 free_percpu(ctx
->cpu
);
831 percpu_ref_exit(&ctx
->reqs
);
832 percpu_ref_exit(&ctx
->users
);
833 kmem_cache_free(kioctx_cachep
, ctx
);
834 pr_debug("error allocating ioctx %d\n", err
);
839 * Cancels all outstanding aio requests on an aio context. Used
840 * when the processes owning a context have all exited to encourage
841 * the rapid destruction of the kioctx.
843 static int kill_ioctx(struct mm_struct
*mm
, struct kioctx
*ctx
,
844 struct ctx_rq_wait
*wait
)
846 struct kioctx_table
*table
;
848 spin_lock(&mm
->ioctx_lock
);
849 if (atomic_xchg(&ctx
->dead
, 1)) {
850 spin_unlock(&mm
->ioctx_lock
);
854 table
= rcu_dereference_raw(mm
->ioctx_table
);
855 WARN_ON(ctx
!= rcu_access_pointer(table
->table
[ctx
->id
]));
856 RCU_INIT_POINTER(table
->table
[ctx
->id
], NULL
);
857 spin_unlock(&mm
->ioctx_lock
);
859 /* free_ioctx_reqs() will do the necessary RCU synchronization */
860 wake_up_all(&ctx
->wait
);
863 * It'd be more correct to do this in free_ioctx(), after all
864 * the outstanding kiocbs have finished - but by then io_destroy
865 * has already returned, so io_setup() could potentially return
866 * -EAGAIN with no ioctxs actually in use (as far as userspace
869 aio_nr_sub(ctx
->max_reqs
);
872 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
875 percpu_ref_kill(&ctx
->users
);
880 * exit_aio: called when the last user of mm goes away. At this point, there is
881 * no way for any new requests to be submited or any of the io_* syscalls to be
882 * called on the context.
884 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
887 void exit_aio(struct mm_struct
*mm
)
889 struct kioctx_table
*table
= rcu_dereference_raw(mm
->ioctx_table
);
890 struct ctx_rq_wait wait
;
896 atomic_set(&wait
.count
, table
->nr
);
897 init_completion(&wait
.comp
);
900 for (i
= 0; i
< table
->nr
; ++i
) {
902 rcu_dereference_protected(table
->table
[i
], true);
910 * We don't need to bother with munmap() here - exit_mmap(mm)
911 * is coming and it'll unmap everything. And we simply can't,
912 * this is not necessarily our ->mm.
913 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
914 * that it needs to unmap the area, just set it to 0.
917 kill_ioctx(mm
, ctx
, &wait
);
920 if (!atomic_sub_and_test(skipped
, &wait
.count
)) {
921 /* Wait until all IO for the context are done. */
922 wait_for_completion(&wait
.comp
);
925 RCU_INIT_POINTER(mm
->ioctx_table
, NULL
);
929 static void put_reqs_available(struct kioctx
*ctx
, unsigned nr
)
931 struct kioctx_cpu
*kcpu
;
934 local_irq_save(flags
);
935 kcpu
= this_cpu_ptr(ctx
->cpu
);
936 kcpu
->reqs_available
+= nr
;
938 while (kcpu
->reqs_available
>= ctx
->req_batch
* 2) {
939 kcpu
->reqs_available
-= ctx
->req_batch
;
940 atomic_add(ctx
->req_batch
, &ctx
->reqs_available
);
943 local_irq_restore(flags
);
946 static bool __get_reqs_available(struct kioctx
*ctx
)
948 struct kioctx_cpu
*kcpu
;
952 local_irq_save(flags
);
953 kcpu
= this_cpu_ptr(ctx
->cpu
);
954 if (!kcpu
->reqs_available
) {
955 int avail
= atomic_read(&ctx
->reqs_available
);
958 if (avail
< ctx
->req_batch
)
960 } while (!atomic_try_cmpxchg(&ctx
->reqs_available
,
961 &avail
, avail
- ctx
->req_batch
));
963 kcpu
->reqs_available
+= ctx
->req_batch
;
967 kcpu
->reqs_available
--;
969 local_irq_restore(flags
);
973 /* refill_reqs_available
974 * Updates the reqs_available reference counts used for tracking the
975 * number of free slots in the completion ring. This can be called
976 * from aio_complete() (to optimistically update reqs_available) or
977 * from aio_get_req() (the we're out of events case). It must be
978 * called holding ctx->completion_lock.
980 static void refill_reqs_available(struct kioctx
*ctx
, unsigned head
,
983 unsigned events_in_ring
, completed
;
985 /* Clamp head since userland can write to it. */
986 head
%= ctx
->nr_events
;
988 events_in_ring
= tail
- head
;
990 events_in_ring
= ctx
->nr_events
- (head
- tail
);
992 completed
= ctx
->completed_events
;
993 if (events_in_ring
< completed
)
994 completed
-= events_in_ring
;
1001 ctx
->completed_events
-= completed
;
1002 put_reqs_available(ctx
, completed
);
1005 /* user_refill_reqs_available
1006 * Called to refill reqs_available when aio_get_req() encounters an
1007 * out of space in the completion ring.
1009 static void user_refill_reqs_available(struct kioctx
*ctx
)
1011 spin_lock_irq(&ctx
->completion_lock
);
1012 if (ctx
->completed_events
) {
1013 struct aio_ring
*ring
;
1016 /* Access of ring->head may race with aio_read_events_ring()
1017 * here, but that's okay since whether we read the old version
1018 * or the new version, and either will be valid. The important
1019 * part is that head cannot pass tail since we prevent
1020 * aio_complete() from updating tail by holding
1021 * ctx->completion_lock. Even if head is invalid, the check
1022 * against ctx->completed_events below will make sure we do the
1025 ring
= page_address(ctx
->ring_pages
[0]);
1028 refill_reqs_available(ctx
, head
, ctx
->tail
);
1031 spin_unlock_irq(&ctx
->completion_lock
);
1034 static bool get_reqs_available(struct kioctx
*ctx
)
1036 if (__get_reqs_available(ctx
))
1038 user_refill_reqs_available(ctx
);
1039 return __get_reqs_available(ctx
);
1043 * Allocate a slot for an aio request.
1044 * Returns NULL if no requests are free.
1046 * The refcount is initialized to 2 - one for the async op completion,
1047 * one for the synchronous code that does this.
1049 static inline struct aio_kiocb
*aio_get_req(struct kioctx
*ctx
)
1051 struct aio_kiocb
*req
;
1053 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
1057 if (unlikely(!get_reqs_available(ctx
))) {
1058 kmem_cache_free(kiocb_cachep
, req
);
1062 percpu_ref_get(&ctx
->reqs
);
1064 INIT_LIST_HEAD(&req
->ki_list
);
1065 refcount_set(&req
->ki_refcnt
, 2);
1066 req
->ki_eventfd
= NULL
;
1070 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
1072 struct aio_ring __user
*ring
= (void __user
*)ctx_id
;
1073 struct mm_struct
*mm
= current
->mm
;
1074 struct kioctx
*ctx
, *ret
= NULL
;
1075 struct kioctx_table
*table
;
1078 if (get_user(id
, &ring
->id
))
1082 table
= rcu_dereference(mm
->ioctx_table
);
1084 if (!table
|| id
>= table
->nr
)
1087 id
= array_index_nospec(id
, table
->nr
);
1088 ctx
= rcu_dereference(table
->table
[id
]);
1089 if (ctx
&& ctx
->user_id
== ctx_id
) {
1090 if (percpu_ref_tryget_live(&ctx
->users
))
1098 static inline void iocb_destroy(struct aio_kiocb
*iocb
)
1100 if (iocb
->ki_eventfd
)
1101 eventfd_ctx_put(iocb
->ki_eventfd
);
1103 fput(iocb
->ki_filp
);
1104 percpu_ref_put(&iocb
->ki_ctx
->reqs
);
1105 kmem_cache_free(kiocb_cachep
, iocb
);
1109 struct wait_queue_entry w
;
1114 * Called when the io request on the given iocb is complete.
1116 static void aio_complete(struct aio_kiocb
*iocb
)
1118 struct kioctx
*ctx
= iocb
->ki_ctx
;
1119 struct aio_ring
*ring
;
1120 struct io_event
*ev_page
, *event
;
1121 unsigned tail
, pos
, head
, avail
;
1122 unsigned long flags
;
1125 * Add a completion event to the ring buffer. Must be done holding
1126 * ctx->completion_lock to prevent other code from messing with the tail
1127 * pointer since we might be called from irq context.
1129 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1132 pos
= tail
+ AIO_EVENTS_OFFSET
;
1134 if (++tail
>= ctx
->nr_events
)
1137 ev_page
= page_address(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1138 event
= ev_page
+ pos
% AIO_EVENTS_PER_PAGE
;
1140 *event
= iocb
->ki_res
;
1142 flush_dcache_page(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1144 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx
, tail
, iocb
,
1145 (void __user
*)(unsigned long)iocb
->ki_res
.obj
,
1146 iocb
->ki_res
.data
, iocb
->ki_res
.res
, iocb
->ki_res
.res2
);
1148 /* after flagging the request as done, we
1149 * must never even look at it again
1151 smp_wmb(); /* make event visible before updating tail */
1155 ring
= page_address(ctx
->ring_pages
[0]);
1158 flush_dcache_page(ctx
->ring_pages
[0]);
1160 ctx
->completed_events
++;
1161 if (ctx
->completed_events
> 1)
1162 refill_reqs_available(ctx
, head
, tail
);
1166 : tail
+ ctx
->nr_events
- head
;
1167 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1169 pr_debug("added to ring %p at [%u]\n", iocb
, tail
);
1172 * Check if the user asked us to deliver the result through an
1173 * eventfd. The eventfd_signal() function is safe to be called
1176 if (iocb
->ki_eventfd
)
1177 eventfd_signal(iocb
->ki_eventfd
);
1180 * We have to order our ring_info tail store above and test
1181 * of the wait list below outside the wait lock. This is
1182 * like in wake_up_bit() where clearing a bit has to be
1183 * ordered with the unlocked test.
1187 if (waitqueue_active(&ctx
->wait
)) {
1188 struct aio_waiter
*curr
, *next
;
1189 unsigned long flags
;
1191 spin_lock_irqsave(&ctx
->wait
.lock
, flags
);
1192 list_for_each_entry_safe(curr
, next
, &ctx
->wait
.head
, w
.entry
)
1193 if (avail
>= curr
->min_nr
) {
1194 list_del_init_careful(&curr
->w
.entry
);
1195 wake_up_process(curr
->w
.private);
1197 spin_unlock_irqrestore(&ctx
->wait
.lock
, flags
);
1201 static inline void iocb_put(struct aio_kiocb
*iocb
)
1203 if (refcount_dec_and_test(&iocb
->ki_refcnt
)) {
1209 /* aio_read_events_ring
1210 * Pull an event off of the ioctx's event ring. Returns the number of
1213 static long aio_read_events_ring(struct kioctx
*ctx
,
1214 struct io_event __user
*event
, long nr
)
1216 struct aio_ring
*ring
;
1217 unsigned head
, tail
, pos
;
1222 * The mutex can block and wake us up and that will cause
1223 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1224 * and repeat. This should be rare enough that it doesn't cause
1225 * peformance issues. See the comment in read_events() for more detail.
1227 sched_annotate_sleep();
1228 mutex_lock(&ctx
->ring_lock
);
1230 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1231 ring
= page_address(ctx
->ring_pages
[0]);
1236 * Ensure that once we've read the current tail pointer, that
1237 * we also see the events that were stored up to the tail.
1241 pr_debug("h%u t%u m%u\n", head
, tail
, ctx
->nr_events
);
1246 head
%= ctx
->nr_events
;
1247 tail
%= ctx
->nr_events
;
1251 struct io_event
*ev
;
1254 avail
= (head
<= tail
? tail
: ctx
->nr_events
) - head
;
1258 pos
= head
+ AIO_EVENTS_OFFSET
;
1259 page
= ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
];
1260 pos
%= AIO_EVENTS_PER_PAGE
;
1262 avail
= min(avail
, nr
- ret
);
1263 avail
= min_t(long, avail
, AIO_EVENTS_PER_PAGE
- pos
);
1265 ev
= page_address(page
);
1266 copy_ret
= copy_to_user(event
+ ret
, ev
+ pos
,
1267 sizeof(*ev
) * avail
);
1269 if (unlikely(copy_ret
)) {
1276 head
%= ctx
->nr_events
;
1279 ring
= page_address(ctx
->ring_pages
[0]);
1281 flush_dcache_page(ctx
->ring_pages
[0]);
1283 pr_debug("%li h%u t%u\n", ret
, head
, tail
);
1285 mutex_unlock(&ctx
->ring_lock
);
1290 static bool aio_read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1291 struct io_event __user
*event
, long *i
)
1293 long ret
= aio_read_events_ring(ctx
, event
+ *i
, nr
- *i
);
1298 if (unlikely(atomic_read(&ctx
->dead
)))
1304 return ret
< 0 || *i
>= min_nr
;
1307 static long read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1308 struct io_event __user
*event
,
1311 struct hrtimer_sleeper t
;
1312 struct aio_waiter w
;
1313 long ret
= 0, ret2
= 0;
1316 * Note that aio_read_events() is being called as the conditional - i.e.
1317 * we're calling it after prepare_to_wait() has set task state to
1318 * TASK_INTERRUPTIBLE.
1320 * But aio_read_events() can block, and if it blocks it's going to flip
1321 * the task state back to TASK_RUNNING.
1323 * This should be ok, provided it doesn't flip the state back to
1324 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1325 * will only happen if the mutex_lock() call blocks, and we then find
1326 * the ringbuffer empty. So in practice we should be ok, but it's
1327 * something to be aware of when touching this code.
1329 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
);
1330 if (until
== 0 || ret
< 0 || ret
>= min_nr
)
1333 hrtimer_init_sleeper_on_stack(&t
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1334 if (until
!= KTIME_MAX
) {
1335 hrtimer_set_expires_range_ns(&t
.timer
, until
, current
->timer_slack_ns
);
1336 hrtimer_sleeper_start_expires(&t
, HRTIMER_MODE_REL
);
1342 unsigned long nr_got
= ret
;
1344 w
.min_nr
= min_nr
- ret
;
1346 ret2
= prepare_to_wait_event(&ctx
->wait
, &w
.w
, TASK_INTERRUPTIBLE
);
1347 if (!ret2
&& !t
.task
)
1350 if (aio_read_events(ctx
, min_nr
, nr
, event
, &ret
) || ret2
)
1357 finish_wait(&ctx
->wait
, &w
.w
);
1358 hrtimer_cancel(&t
.timer
);
1359 destroy_hrtimer_on_stack(&t
.timer
);
1365 * Create an aio_context capable of receiving at least nr_events.
1366 * ctxp must not point to an aio_context that already exists, and
1367 * must be initialized to 0 prior to the call. On successful
1368 * creation of the aio_context, *ctxp is filled in with the resulting
1369 * handle. May fail with -EINVAL if *ctxp is not initialized,
1370 * if the specified nr_events exceeds internal limits. May fail
1371 * with -EAGAIN if the specified nr_events exceeds the user's limit
1372 * of available events. May fail with -ENOMEM if insufficient kernel
1373 * resources are available. May fail with -EFAULT if an invalid
1374 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1377 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
1379 struct kioctx
*ioctx
= NULL
;
1383 ret
= get_user(ctx
, ctxp
);
1388 if (unlikely(ctx
|| nr_events
== 0)) {
1389 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1394 ioctx
= ioctx_alloc(nr_events
);
1395 ret
= PTR_ERR(ioctx
);
1396 if (!IS_ERR(ioctx
)) {
1397 ret
= put_user(ioctx
->user_id
, ctxp
);
1399 kill_ioctx(current
->mm
, ioctx
, NULL
);
1400 percpu_ref_put(&ioctx
->users
);
1407 #ifdef CONFIG_COMPAT
1408 COMPAT_SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, u32 __user
*, ctx32p
)
1410 struct kioctx
*ioctx
= NULL
;
1414 ret
= get_user(ctx
, ctx32p
);
1419 if (unlikely(ctx
|| nr_events
== 0)) {
1420 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1425 ioctx
= ioctx_alloc(nr_events
);
1426 ret
= PTR_ERR(ioctx
);
1427 if (!IS_ERR(ioctx
)) {
1428 /* truncating is ok because it's a user address */
1429 ret
= put_user((u32
)ioctx
->user_id
, ctx32p
);
1431 kill_ioctx(current
->mm
, ioctx
, NULL
);
1432 percpu_ref_put(&ioctx
->users
);
1441 * Destroy the aio_context specified. May cancel any outstanding
1442 * AIOs and block on completion. Will fail with -ENOSYS if not
1443 * implemented. May fail with -EINVAL if the context pointed to
1446 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1448 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1449 if (likely(NULL
!= ioctx
)) {
1450 struct ctx_rq_wait wait
;
1453 init_completion(&wait
.comp
);
1454 atomic_set(&wait
.count
, 1);
1456 /* Pass requests_done to kill_ioctx() where it can be set
1457 * in a thread-safe way. If we try to set it here then we have
1458 * a race condition if two io_destroy() called simultaneously.
1460 ret
= kill_ioctx(current
->mm
, ioctx
, &wait
);
1461 percpu_ref_put(&ioctx
->users
);
1463 /* Wait until all IO for the context are done. Otherwise kernel
1464 * keep using user-space buffers even if user thinks the context
1468 wait_for_completion(&wait
.comp
);
1472 pr_debug("EINVAL: invalid context id\n");
1476 static void aio_remove_iocb(struct aio_kiocb
*iocb
)
1478 struct kioctx
*ctx
= iocb
->ki_ctx
;
1479 unsigned long flags
;
1481 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
1482 list_del(&iocb
->ki_list
);
1483 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1486 static void aio_complete_rw(struct kiocb
*kiocb
, long res
)
1488 struct aio_kiocb
*iocb
= container_of(kiocb
, struct aio_kiocb
, rw
);
1490 if (!list_empty_careful(&iocb
->ki_list
))
1491 aio_remove_iocb(iocb
);
1493 if (kiocb
->ki_flags
& IOCB_WRITE
) {
1494 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
1496 if (S_ISREG(inode
->i_mode
))
1497 kiocb_end_write(kiocb
);
1500 iocb
->ki_res
.res
= res
;
1501 iocb
->ki_res
.res2
= 0;
1505 static int aio_prep_rw(struct kiocb
*req
, const struct iocb
*iocb
)
1509 req
->ki_complete
= aio_complete_rw
;
1510 req
->private = NULL
;
1511 req
->ki_pos
= iocb
->aio_offset
;
1512 req
->ki_flags
= req
->ki_filp
->f_iocb_flags
;
1513 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
)
1514 req
->ki_flags
|= IOCB_EVENTFD
;
1515 if (iocb
->aio_flags
& IOCB_FLAG_IOPRIO
) {
1517 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1518 * aio_reqprio is interpreted as an I/O scheduling
1519 * class and priority.
1521 ret
= ioprio_check_cap(iocb
->aio_reqprio
);
1523 pr_debug("aio ioprio check cap error: %d\n", ret
);
1527 req
->ki_ioprio
= iocb
->aio_reqprio
;
1529 req
->ki_ioprio
= get_current_ioprio();
1531 ret
= kiocb_set_rw_flags(req
, iocb
->aio_rw_flags
);
1535 req
->ki_flags
&= ~IOCB_HIPRI
; /* no one is going to poll for this I/O */
1539 static ssize_t
aio_setup_rw(int rw
, const struct iocb
*iocb
,
1540 struct iovec
**iovec
, bool vectored
, bool compat
,
1541 struct iov_iter
*iter
)
1543 void __user
*buf
= (void __user
*)(uintptr_t)iocb
->aio_buf
;
1544 size_t len
= iocb
->aio_nbytes
;
1547 ssize_t ret
= import_ubuf(rw
, buf
, len
, iter
);
1552 return __import_iovec(rw
, buf
, len
, UIO_FASTIOV
, iovec
, iter
, compat
);
1555 static inline void aio_rw_done(struct kiocb
*req
, ssize_t ret
)
1561 case -ERESTARTNOINTR
:
1562 case -ERESTARTNOHAND
:
1563 case -ERESTART_RESTARTBLOCK
:
1565 * There's no easy way to restart the syscall since other AIO's
1566 * may be already running. Just fail this IO with EINTR.
1571 req
->ki_complete(req
, ret
);
1575 static int aio_read(struct kiocb
*req
, const struct iocb
*iocb
,
1576 bool vectored
, bool compat
)
1578 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1579 struct iov_iter iter
;
1583 ret
= aio_prep_rw(req
, iocb
);
1586 file
= req
->ki_filp
;
1587 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1589 if (unlikely(!file
->f_op
->read_iter
))
1592 ret
= aio_setup_rw(ITER_DEST
, iocb
, &iovec
, vectored
, compat
, &iter
);
1595 ret
= rw_verify_area(READ
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1597 aio_rw_done(req
, call_read_iter(file
, req
, &iter
));
1602 static int aio_write(struct kiocb
*req
, const struct iocb
*iocb
,
1603 bool vectored
, bool compat
)
1605 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1606 struct iov_iter iter
;
1610 ret
= aio_prep_rw(req
, iocb
);
1613 file
= req
->ki_filp
;
1615 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1617 if (unlikely(!file
->f_op
->write_iter
))
1620 ret
= aio_setup_rw(ITER_SOURCE
, iocb
, &iovec
, vectored
, compat
, &iter
);
1623 ret
= rw_verify_area(WRITE
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1625 if (S_ISREG(file_inode(file
)->i_mode
))
1626 kiocb_start_write(req
);
1627 req
->ki_flags
|= IOCB_WRITE
;
1628 aio_rw_done(req
, call_write_iter(file
, req
, &iter
));
1634 static void aio_fsync_work(struct work_struct
*work
)
1636 struct aio_kiocb
*iocb
= container_of(work
, struct aio_kiocb
, fsync
.work
);
1637 const struct cred
*old_cred
= override_creds(iocb
->fsync
.creds
);
1639 iocb
->ki_res
.res
= vfs_fsync(iocb
->fsync
.file
, iocb
->fsync
.datasync
);
1640 revert_creds(old_cred
);
1641 put_cred(iocb
->fsync
.creds
);
1645 static int aio_fsync(struct fsync_iocb
*req
, const struct iocb
*iocb
,
1648 if (unlikely(iocb
->aio_buf
|| iocb
->aio_offset
|| iocb
->aio_nbytes
||
1649 iocb
->aio_rw_flags
))
1652 if (unlikely(!req
->file
->f_op
->fsync
))
1655 req
->creds
= prepare_creds();
1659 req
->datasync
= datasync
;
1660 INIT_WORK(&req
->work
, aio_fsync_work
);
1661 schedule_work(&req
->work
);
1665 static void aio_poll_put_work(struct work_struct
*work
)
1667 struct poll_iocb
*req
= container_of(work
, struct poll_iocb
, work
);
1668 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1674 * Safely lock the waitqueue which the request is on, synchronizing with the
1675 * case where the ->poll() provider decides to free its waitqueue early.
1677 * Returns true on success, meaning that req->head->lock was locked, req->wait
1678 * is on req->head, and an RCU read lock was taken. Returns false if the
1679 * request was already removed from its waitqueue (which might no longer exist).
1681 static bool poll_iocb_lock_wq(struct poll_iocb
*req
)
1683 wait_queue_head_t
*head
;
1686 * While we hold the waitqueue lock and the waitqueue is nonempty,
1687 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1688 * lock in the first place can race with the waitqueue being freed.
1690 * We solve this as eventpoll does: by taking advantage of the fact that
1691 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1692 * we enter rcu_read_lock() and see that the pointer to the queue is
1693 * non-NULL, we can then lock it without the memory being freed out from
1694 * under us, then check whether the request is still on the queue.
1696 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1697 * case the caller deletes the entry from the queue, leaving it empty.
1698 * In that case, only RCU prevents the queue memory from being freed.
1701 head
= smp_load_acquire(&req
->head
);
1703 spin_lock(&head
->lock
);
1704 if (!list_empty(&req
->wait
.entry
))
1706 spin_unlock(&head
->lock
);
1712 static void poll_iocb_unlock_wq(struct poll_iocb
*req
)
1714 spin_unlock(&req
->head
->lock
);
1718 static void aio_poll_complete_work(struct work_struct
*work
)
1720 struct poll_iocb
*req
= container_of(work
, struct poll_iocb
, work
);
1721 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1722 struct poll_table_struct pt
= { ._key
= req
->events
};
1723 struct kioctx
*ctx
= iocb
->ki_ctx
;
1726 if (!READ_ONCE(req
->cancelled
))
1727 mask
= vfs_poll(req
->file
, &pt
) & req
->events
;
1730 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1731 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1732 * synchronize with them. In the cancellation case the list_del_init
1733 * itself is not actually needed, but harmless so we keep it in to
1734 * avoid further branches in the fast path.
1736 spin_lock_irq(&ctx
->ctx_lock
);
1737 if (poll_iocb_lock_wq(req
)) {
1738 if (!mask
&& !READ_ONCE(req
->cancelled
)) {
1740 * The request isn't actually ready to be completed yet.
1741 * Reschedule completion if another wakeup came in.
1743 if (req
->work_need_resched
) {
1744 schedule_work(&req
->work
);
1745 req
->work_need_resched
= false;
1747 req
->work_scheduled
= false;
1749 poll_iocb_unlock_wq(req
);
1750 spin_unlock_irq(&ctx
->ctx_lock
);
1753 list_del_init(&req
->wait
.entry
);
1754 poll_iocb_unlock_wq(req
);
1755 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1756 list_del_init(&iocb
->ki_list
);
1757 iocb
->ki_res
.res
= mangle_poll(mask
);
1758 spin_unlock_irq(&ctx
->ctx_lock
);
1763 /* assumes we are called with irqs disabled */
1764 static int aio_poll_cancel(struct kiocb
*iocb
)
1766 struct aio_kiocb
*aiocb
= container_of(iocb
, struct aio_kiocb
, rw
);
1767 struct poll_iocb
*req
= &aiocb
->poll
;
1769 if (poll_iocb_lock_wq(req
)) {
1770 WRITE_ONCE(req
->cancelled
, true);
1771 if (!req
->work_scheduled
) {
1772 schedule_work(&aiocb
->poll
.work
);
1773 req
->work_scheduled
= true;
1775 poll_iocb_unlock_wq(req
);
1776 } /* else, the request was force-cancelled by POLLFREE already */
1781 static int aio_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1784 struct poll_iocb
*req
= container_of(wait
, struct poll_iocb
, wait
);
1785 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1786 __poll_t mask
= key_to_poll(key
);
1787 unsigned long flags
;
1789 /* for instances that support it check for an event match first: */
1790 if (mask
&& !(mask
& req
->events
))
1794 * Complete the request inline if possible. This requires that three
1795 * conditions be met:
1796 * 1. An event mask must have been passed. If a plain wakeup was done
1797 * instead, then mask == 0 and we have to call vfs_poll() to get
1798 * the events, so inline completion isn't possible.
1799 * 2. The completion work must not have already been scheduled.
1800 * 3. ctx_lock must not be busy. We have to use trylock because we
1801 * already hold the waitqueue lock, so this inverts the normal
1802 * locking order. Use irqsave/irqrestore because not all
1803 * filesystems (e.g. fuse) call this function with IRQs disabled,
1804 * yet IRQs have to be disabled before ctx_lock is obtained.
1806 if (mask
&& !req
->work_scheduled
&&
1807 spin_trylock_irqsave(&iocb
->ki_ctx
->ctx_lock
, flags
)) {
1808 struct kioctx
*ctx
= iocb
->ki_ctx
;
1810 list_del_init(&req
->wait
.entry
);
1811 list_del(&iocb
->ki_list
);
1812 iocb
->ki_res
.res
= mangle_poll(mask
);
1813 if (iocb
->ki_eventfd
&& !eventfd_signal_allowed()) {
1815 INIT_WORK(&req
->work
, aio_poll_put_work
);
1816 schedule_work(&req
->work
);
1818 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1823 * Schedule the completion work if needed. If it was already
1824 * scheduled, record that another wakeup came in.
1826 * Don't remove the request from the waitqueue here, as it might
1827 * not actually be complete yet (we won't know until vfs_poll()
1828 * is called), and we must not miss any wakeups. POLLFREE is an
1829 * exception to this; see below.
1831 if (req
->work_scheduled
) {
1832 req
->work_need_resched
= true;
1834 schedule_work(&req
->work
);
1835 req
->work_scheduled
= true;
1839 * If the waitqueue is being freed early but we can't complete
1840 * the request inline, we have to tear down the request as best
1841 * we can. That means immediately removing the request from its
1842 * waitqueue and preventing all further accesses to the
1843 * waitqueue via the request. We also need to schedule the
1844 * completion work (done above). Also mark the request as
1845 * cancelled, to potentially skip an unneeded call to ->poll().
1847 if (mask
& POLLFREE
) {
1848 WRITE_ONCE(req
->cancelled
, true);
1849 list_del_init(&req
->wait
.entry
);
1852 * Careful: this *must* be the last step, since as soon
1853 * as req->head is NULL'ed out, the request can be
1854 * completed and freed, since aio_poll_complete_work()
1855 * will no longer need to take the waitqueue lock.
1857 smp_store_release(&req
->head
, NULL
);
1863 struct aio_poll_table
{
1864 struct poll_table_struct pt
;
1865 struct aio_kiocb
*iocb
;
1871 aio_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1872 struct poll_table_struct
*p
)
1874 struct aio_poll_table
*pt
= container_of(p
, struct aio_poll_table
, pt
);
1876 /* multiple wait queues per file are not supported */
1877 if (unlikely(pt
->queued
)) {
1878 pt
->error
= -EINVAL
;
1884 pt
->iocb
->poll
.head
= head
;
1885 add_wait_queue(head
, &pt
->iocb
->poll
.wait
);
1888 static int aio_poll(struct aio_kiocb
*aiocb
, const struct iocb
*iocb
)
1890 struct kioctx
*ctx
= aiocb
->ki_ctx
;
1891 struct poll_iocb
*req
= &aiocb
->poll
;
1892 struct aio_poll_table apt
;
1893 bool cancel
= false;
1896 /* reject any unknown events outside the normal event mask. */
1897 if ((u16
)iocb
->aio_buf
!= iocb
->aio_buf
)
1899 /* reject fields that are not defined for poll */
1900 if (iocb
->aio_offset
|| iocb
->aio_nbytes
|| iocb
->aio_rw_flags
)
1903 INIT_WORK(&req
->work
, aio_poll_complete_work
);
1904 req
->events
= demangle_poll(iocb
->aio_buf
) | EPOLLERR
| EPOLLHUP
;
1907 req
->cancelled
= false;
1908 req
->work_scheduled
= false;
1909 req
->work_need_resched
= false;
1911 apt
.pt
._qproc
= aio_poll_queue_proc
;
1912 apt
.pt
._key
= req
->events
;
1915 apt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1917 /* initialized the list so that we can do list_empty checks */
1918 INIT_LIST_HEAD(&req
->wait
.entry
);
1919 init_waitqueue_func_entry(&req
->wait
, aio_poll_wake
);
1921 mask
= vfs_poll(req
->file
, &apt
.pt
) & req
->events
;
1922 spin_lock_irq(&ctx
->ctx_lock
);
1923 if (likely(apt
.queued
)) {
1924 bool on_queue
= poll_iocb_lock_wq(req
);
1926 if (!on_queue
|| req
->work_scheduled
) {
1928 * aio_poll_wake() already either scheduled the async
1929 * completion work, or completed the request inline.
1931 if (apt
.error
) /* unsupported case: multiple queues */
1936 if (mask
|| apt
.error
) {
1937 /* Steal to complete synchronously. */
1938 list_del_init(&req
->wait
.entry
);
1939 } else if (cancel
) {
1940 /* Cancel if possible (may be too late though). */
1941 WRITE_ONCE(req
->cancelled
, true);
1942 } else if (on_queue
) {
1944 * Actually waiting for an event, so add the request to
1945 * active_reqs so that it can be cancelled if needed.
1947 list_add_tail(&aiocb
->ki_list
, &ctx
->active_reqs
);
1948 aiocb
->ki_cancel
= aio_poll_cancel
;
1951 poll_iocb_unlock_wq(req
);
1953 if (mask
) { /* no async, we'd stolen it */
1954 aiocb
->ki_res
.res
= mangle_poll(mask
);
1957 spin_unlock_irq(&ctx
->ctx_lock
);
1963 static int __io_submit_one(struct kioctx
*ctx
, const struct iocb
*iocb
,
1964 struct iocb __user
*user_iocb
, struct aio_kiocb
*req
,
1967 req
->ki_filp
= fget(iocb
->aio_fildes
);
1968 if (unlikely(!req
->ki_filp
))
1971 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1972 struct eventfd_ctx
*eventfd
;
1974 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1975 * instance of the file* now. The file descriptor must be
1976 * an eventfd() fd, and will be signaled for each completed
1977 * event using the eventfd_signal() function.
1979 eventfd
= eventfd_ctx_fdget(iocb
->aio_resfd
);
1980 if (IS_ERR(eventfd
))
1981 return PTR_ERR(eventfd
);
1983 req
->ki_eventfd
= eventfd
;
1986 if (unlikely(put_user(KIOCB_KEY
, &user_iocb
->aio_key
))) {
1987 pr_debug("EFAULT: aio_key\n");
1991 req
->ki_res
.obj
= (u64
)(unsigned long)user_iocb
;
1992 req
->ki_res
.data
= iocb
->aio_data
;
1993 req
->ki_res
.res
= 0;
1994 req
->ki_res
.res2
= 0;
1996 switch (iocb
->aio_lio_opcode
) {
1997 case IOCB_CMD_PREAD
:
1998 return aio_read(&req
->rw
, iocb
, false, compat
);
1999 case IOCB_CMD_PWRITE
:
2000 return aio_write(&req
->rw
, iocb
, false, compat
);
2001 case IOCB_CMD_PREADV
:
2002 return aio_read(&req
->rw
, iocb
, true, compat
);
2003 case IOCB_CMD_PWRITEV
:
2004 return aio_write(&req
->rw
, iocb
, true, compat
);
2005 case IOCB_CMD_FSYNC
:
2006 return aio_fsync(&req
->fsync
, iocb
, false);
2007 case IOCB_CMD_FDSYNC
:
2008 return aio_fsync(&req
->fsync
, iocb
, true);
2010 return aio_poll(req
, iocb
);
2012 pr_debug("invalid aio operation %d\n", iocb
->aio_lio_opcode
);
2017 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
2020 struct aio_kiocb
*req
;
2024 if (unlikely(copy_from_user(&iocb
, user_iocb
, sizeof(iocb
))))
2027 /* enforce forwards compatibility on users */
2028 if (unlikely(iocb
.aio_reserved2
)) {
2029 pr_debug("EINVAL: reserve field set\n");
2033 /* prevent overflows */
2035 (iocb
.aio_buf
!= (unsigned long)iocb
.aio_buf
) ||
2036 (iocb
.aio_nbytes
!= (size_t)iocb
.aio_nbytes
) ||
2037 ((ssize_t
)iocb
.aio_nbytes
< 0)
2039 pr_debug("EINVAL: overflow check\n");
2043 req
= aio_get_req(ctx
);
2047 err
= __io_submit_one(ctx
, &iocb
, user_iocb
, req
, compat
);
2049 /* Done with the synchronous reference */
2053 * If err is 0, we'd either done aio_complete() ourselves or have
2054 * arranged for that to be done asynchronously. Anything non-zero
2055 * means that we need to destroy req ourselves.
2057 if (unlikely(err
)) {
2059 put_reqs_available(ctx
, 1);
2065 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2066 * the number of iocbs queued. May return -EINVAL if the aio_context
2067 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2068 * *iocbpp[0] is not properly initialized, if the operation specified
2069 * is invalid for the file descriptor in the iocb. May fail with
2070 * -EFAULT if any of the data structures point to invalid data. May
2071 * fail with -EBADF if the file descriptor specified in the first
2072 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2073 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2074 * fail with -ENOSYS if not implemented.
2076 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
2077 struct iocb __user
* __user
*, iocbpp
)
2082 struct blk_plug plug
;
2084 if (unlikely(nr
< 0))
2087 ctx
= lookup_ioctx(ctx_id
);
2088 if (unlikely(!ctx
)) {
2089 pr_debug("EINVAL: invalid context id\n");
2093 if (nr
> ctx
->nr_events
)
2094 nr
= ctx
->nr_events
;
2096 if (nr
> AIO_PLUG_THRESHOLD
)
2097 blk_start_plug(&plug
);
2098 for (i
= 0; i
< nr
; i
++) {
2099 struct iocb __user
*user_iocb
;
2101 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
2106 ret
= io_submit_one(ctx
, user_iocb
, false);
2110 if (nr
> AIO_PLUG_THRESHOLD
)
2111 blk_finish_plug(&plug
);
2113 percpu_ref_put(&ctx
->users
);
2117 #ifdef CONFIG_COMPAT
2118 COMPAT_SYSCALL_DEFINE3(io_submit
, compat_aio_context_t
, ctx_id
,
2119 int, nr
, compat_uptr_t __user
*, iocbpp
)
2124 struct blk_plug plug
;
2126 if (unlikely(nr
< 0))
2129 ctx
= lookup_ioctx(ctx_id
);
2130 if (unlikely(!ctx
)) {
2131 pr_debug("EINVAL: invalid context id\n");
2135 if (nr
> ctx
->nr_events
)
2136 nr
= ctx
->nr_events
;
2138 if (nr
> AIO_PLUG_THRESHOLD
)
2139 blk_start_plug(&plug
);
2140 for (i
= 0; i
< nr
; i
++) {
2141 compat_uptr_t user_iocb
;
2143 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
2148 ret
= io_submit_one(ctx
, compat_ptr(user_iocb
), true);
2152 if (nr
> AIO_PLUG_THRESHOLD
)
2153 blk_finish_plug(&plug
);
2155 percpu_ref_put(&ctx
->users
);
2161 * Attempts to cancel an iocb previously passed to io_submit. If
2162 * the operation is successfully cancelled, the resulting event is
2163 * copied into the memory pointed to by result without being placed
2164 * into the completion queue and 0 is returned. May fail with
2165 * -EFAULT if any of the data structures pointed to are invalid.
2166 * May fail with -EINVAL if aio_context specified by ctx_id is
2167 * invalid. May fail with -EAGAIN if the iocb specified was not
2168 * cancelled. Will fail with -ENOSYS if not implemented.
2170 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
2171 struct io_event __user
*, result
)
2174 struct aio_kiocb
*kiocb
;
2177 u64 obj
= (u64
)(unsigned long)iocb
;
2179 if (unlikely(get_user(key
, &iocb
->aio_key
)))
2181 if (unlikely(key
!= KIOCB_KEY
))
2184 ctx
= lookup_ioctx(ctx_id
);
2188 spin_lock_irq(&ctx
->ctx_lock
);
2189 /* TODO: use a hash or array, this sucks. */
2190 list_for_each_entry(kiocb
, &ctx
->active_reqs
, ki_list
) {
2191 if (kiocb
->ki_res
.obj
== obj
) {
2192 ret
= kiocb
->ki_cancel(&kiocb
->rw
);
2193 list_del_init(&kiocb
->ki_list
);
2197 spin_unlock_irq(&ctx
->ctx_lock
);
2201 * The result argument is no longer used - the io_event is
2202 * always delivered via the ring buffer. -EINPROGRESS indicates
2203 * cancellation is progress:
2208 percpu_ref_put(&ctx
->users
);
2213 static long do_io_getevents(aio_context_t ctx_id
,
2216 struct io_event __user
*events
,
2217 struct timespec64
*ts
)
2219 ktime_t until
= ts
? timespec64_to_ktime(*ts
) : KTIME_MAX
;
2220 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
2223 if (likely(ioctx
)) {
2224 if (likely(min_nr
<= nr
&& min_nr
>= 0))
2225 ret
= read_events(ioctx
, min_nr
, nr
, events
, until
);
2226 percpu_ref_put(&ioctx
->users
);
2233 * Attempts to read at least min_nr events and up to nr events from
2234 * the completion queue for the aio_context specified by ctx_id. If
2235 * it succeeds, the number of read events is returned. May fail with
2236 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2237 * out of range, if timeout is out of range. May fail with -EFAULT
2238 * if any of the memory specified is invalid. May return 0 or
2239 * < min_nr if the timeout specified by timeout has elapsed
2240 * before sufficient events are available, where timeout == NULL
2241 * specifies an infinite timeout. Note that the timeout pointed to by
2242 * timeout is relative. Will fail with -ENOSYS if not implemented.
2246 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
2249 struct io_event __user
*, events
,
2250 struct __kernel_timespec __user
*, timeout
)
2252 struct timespec64 ts
;
2255 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2258 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2259 if (!ret
&& signal_pending(current
))
2266 struct __aio_sigset
{
2267 const sigset_t __user
*sigmask
;
2271 SYSCALL_DEFINE6(io_pgetevents
,
2272 aio_context_t
, ctx_id
,
2275 struct io_event __user
*, events
,
2276 struct __kernel_timespec __user
*, timeout
,
2277 const struct __aio_sigset __user
*, usig
)
2279 struct __aio_sigset ksig
= { NULL
, };
2280 struct timespec64 ts
;
2284 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2287 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2290 ret
= set_user_sigmask(ksig
.sigmask
, ksig
.sigsetsize
);
2294 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2296 interrupted
= signal_pending(current
);
2297 restore_saved_sigmask_unless(interrupted
);
2298 if (interrupted
&& !ret
)
2299 ret
= -ERESTARTNOHAND
;
2304 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2306 SYSCALL_DEFINE6(io_pgetevents_time32
,
2307 aio_context_t
, ctx_id
,
2310 struct io_event __user
*, events
,
2311 struct old_timespec32 __user
*, timeout
,
2312 const struct __aio_sigset __user
*, usig
)
2314 struct __aio_sigset ksig
= { NULL
, };
2315 struct timespec64 ts
;
2319 if (timeout
&& unlikely(get_old_timespec32(&ts
, timeout
)))
2322 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2326 ret
= set_user_sigmask(ksig
.sigmask
, ksig
.sigsetsize
);
2330 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2332 interrupted
= signal_pending(current
);
2333 restore_saved_sigmask_unless(interrupted
);
2334 if (interrupted
&& !ret
)
2335 ret
= -ERESTARTNOHAND
;
2342 #if defined(CONFIG_COMPAT_32BIT_TIME)
2344 SYSCALL_DEFINE5(io_getevents_time32
, __u32
, ctx_id
,
2347 struct io_event __user
*, events
,
2348 struct old_timespec32 __user
*, timeout
)
2350 struct timespec64 t
;
2353 if (timeout
&& get_old_timespec32(&t
, timeout
))
2356 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2357 if (!ret
&& signal_pending(current
))
2364 #ifdef CONFIG_COMPAT
2366 struct __compat_aio_sigset
{
2367 compat_uptr_t sigmask
;
2368 compat_size_t sigsetsize
;
2371 #if defined(CONFIG_COMPAT_32BIT_TIME)
2373 COMPAT_SYSCALL_DEFINE6(io_pgetevents
,
2374 compat_aio_context_t
, ctx_id
,
2375 compat_long_t
, min_nr
,
2377 struct io_event __user
*, events
,
2378 struct old_timespec32 __user
*, timeout
,
2379 const struct __compat_aio_sigset __user
*, usig
)
2381 struct __compat_aio_sigset ksig
= { 0, };
2382 struct timespec64 t
;
2386 if (timeout
&& get_old_timespec32(&t
, timeout
))
2389 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2392 ret
= set_compat_user_sigmask(compat_ptr(ksig
.sigmask
), ksig
.sigsetsize
);
2396 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2398 interrupted
= signal_pending(current
);
2399 restore_saved_sigmask_unless(interrupted
);
2400 if (interrupted
&& !ret
)
2401 ret
= -ERESTARTNOHAND
;
2408 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64
,
2409 compat_aio_context_t
, ctx_id
,
2410 compat_long_t
, min_nr
,
2412 struct io_event __user
*, events
,
2413 struct __kernel_timespec __user
*, timeout
,
2414 const struct __compat_aio_sigset __user
*, usig
)
2416 struct __compat_aio_sigset ksig
= { 0, };
2417 struct timespec64 t
;
2421 if (timeout
&& get_timespec64(&t
, timeout
))
2424 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2427 ret
= set_compat_user_sigmask(compat_ptr(ksig
.sigmask
), ksig
.sigsetsize
);
2431 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2433 interrupted
= signal_pending(current
);
2434 restore_saved_sigmask_unless(interrupted
);
2435 if (interrupted
&& !ret
)
2436 ret
= -ERESTARTNOHAND
;