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
,
245 static void __init
aio_sysctl_init(void)
247 register_sysctl_init("fs", aio_sysctls
);
250 #define aio_sysctl_init() do { } while (0)
253 static struct kmem_cache
*kiocb_cachep
;
254 static struct kmem_cache
*kioctx_cachep
;
256 static struct vfsmount
*aio_mnt
;
258 static const struct file_operations aio_ring_fops
;
259 static const struct address_space_operations aio_ctx_aops
;
261 static struct file
*aio_private_file(struct kioctx
*ctx
, loff_t nr_pages
)
264 struct inode
*inode
= alloc_anon_inode(aio_mnt
->mnt_sb
);
266 return ERR_CAST(inode
);
268 inode
->i_mapping
->a_ops
= &aio_ctx_aops
;
269 inode
->i_mapping
->private_data
= ctx
;
270 inode
->i_size
= PAGE_SIZE
* nr_pages
;
272 file
= alloc_file_pseudo(inode
, aio_mnt
, "[aio]",
273 O_RDWR
, &aio_ring_fops
);
279 static int aio_init_fs_context(struct fs_context
*fc
)
281 if (!init_pseudo(fc
, AIO_RING_MAGIC
))
283 fc
->s_iflags
|= SB_I_NOEXEC
;
288 * Creates the slab caches used by the aio routines, panic on
289 * failure as this is done early during the boot sequence.
291 static int __init
aio_setup(void)
293 static struct file_system_type aio_fs
= {
295 .init_fs_context
= aio_init_fs_context
,
296 .kill_sb
= kill_anon_super
,
298 aio_mnt
= kern_mount(&aio_fs
);
300 panic("Failed to create aio fs mount.");
302 kiocb_cachep
= KMEM_CACHE(aio_kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
303 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
307 __initcall(aio_setup
);
309 static void put_aio_ring_file(struct kioctx
*ctx
)
311 struct file
*aio_ring_file
= ctx
->aio_ring_file
;
312 struct address_space
*i_mapping
;
315 truncate_setsize(file_inode(aio_ring_file
), 0);
317 /* Prevent further access to the kioctx from migratepages */
318 i_mapping
= aio_ring_file
->f_mapping
;
319 spin_lock(&i_mapping
->private_lock
);
320 i_mapping
->private_data
= NULL
;
321 ctx
->aio_ring_file
= NULL
;
322 spin_unlock(&i_mapping
->private_lock
);
328 static void aio_free_ring(struct kioctx
*ctx
)
332 /* Disconnect the kiotx from the ring file. This prevents future
333 * accesses to the kioctx from page migration.
335 put_aio_ring_file(ctx
);
337 for (i
= 0; i
< ctx
->nr_pages
; i
++) {
339 pr_debug("pid(%d) [%d] page->count=%d\n", current
->pid
, i
,
340 page_count(ctx
->ring_pages
[i
]));
341 page
= ctx
->ring_pages
[i
];
344 ctx
->ring_pages
[i
] = NULL
;
348 if (ctx
->ring_pages
&& ctx
->ring_pages
!= ctx
->internal_pages
) {
349 kfree(ctx
->ring_pages
);
350 ctx
->ring_pages
= NULL
;
354 static int aio_ring_mremap(struct vm_area_struct
*vma
)
356 struct file
*file
= vma
->vm_file
;
357 struct mm_struct
*mm
= vma
->vm_mm
;
358 struct kioctx_table
*table
;
359 int i
, res
= -EINVAL
;
361 spin_lock(&mm
->ioctx_lock
);
363 table
= rcu_dereference(mm
->ioctx_table
);
367 for (i
= 0; i
< table
->nr
; i
++) {
370 ctx
= rcu_dereference(table
->table
[i
]);
371 if (ctx
&& ctx
->aio_ring_file
== file
) {
372 if (!atomic_read(&ctx
->dead
)) {
373 ctx
->user_id
= ctx
->mmap_base
= vma
->vm_start
;
382 spin_unlock(&mm
->ioctx_lock
);
386 static const struct vm_operations_struct aio_ring_vm_ops
= {
387 .mremap
= aio_ring_mremap
,
388 #if IS_ENABLED(CONFIG_MMU)
389 .fault
= filemap_fault
,
390 .map_pages
= filemap_map_pages
,
391 .page_mkwrite
= filemap_page_mkwrite
,
395 static int aio_ring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
397 vm_flags_set(vma
, VM_DONTEXPAND
);
398 vma
->vm_ops
= &aio_ring_vm_ops
;
402 static const struct file_operations aio_ring_fops
= {
403 .mmap
= aio_ring_mmap
,
406 #if IS_ENABLED(CONFIG_MIGRATION)
407 static int aio_migrate_folio(struct address_space
*mapping
, struct folio
*dst
,
408 struct folio
*src
, enum migrate_mode mode
)
416 * We cannot support the _NO_COPY case here, because copy needs to
417 * happen under the ctx->completion_lock. That does not work with the
418 * migration workflow of MIGRATE_SYNC_NO_COPY.
420 if (mode
== MIGRATE_SYNC_NO_COPY
)
425 /* mapping->private_lock here protects against the kioctx teardown. */
426 spin_lock(&mapping
->private_lock
);
427 ctx
= mapping
->private_data
;
433 /* The ring_lock mutex. The prevents aio_read_events() from writing
434 * to the ring's head, and prevents page migration from mucking in
435 * a partially initialized kiotx.
437 if (!mutex_trylock(&ctx
->ring_lock
)) {
443 if (idx
< (pgoff_t
)ctx
->nr_pages
) {
444 /* Make sure the old folio hasn't already been changed */
445 if (ctx
->ring_pages
[idx
] != &src
->page
)
453 /* Writeback must be complete */
454 BUG_ON(folio_test_writeback(src
));
457 rc
= folio_migrate_mapping(mapping
, dst
, src
, 1);
458 if (rc
!= MIGRATEPAGE_SUCCESS
) {
463 /* Take completion_lock to prevent other writes to the ring buffer
464 * while the old folio is copied to the new. This prevents new
465 * events from being lost.
467 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
468 folio_migrate_copy(dst
, src
);
469 BUG_ON(ctx
->ring_pages
[idx
] != &src
->page
);
470 ctx
->ring_pages
[idx
] = &dst
->page
;
471 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
473 /* The old folio is no longer accessible. */
477 mutex_unlock(&ctx
->ring_lock
);
479 spin_unlock(&mapping
->private_lock
);
483 #define aio_migrate_folio NULL
486 static const struct address_space_operations aio_ctx_aops
= {
487 .dirty_folio
= noop_dirty_folio
,
488 .migrate_folio
= aio_migrate_folio
,
491 static int aio_setup_ring(struct kioctx
*ctx
, unsigned int nr_events
)
493 struct aio_ring
*ring
;
494 struct mm_struct
*mm
= current
->mm
;
495 unsigned long size
, unused
;
500 /* Compensate for the ring buffer's head/tail overlap entry */
501 nr_events
+= 2; /* 1 is required, 2 for good luck */
503 size
= sizeof(struct aio_ring
);
504 size
+= sizeof(struct io_event
) * nr_events
;
506 nr_pages
= PFN_UP(size
);
510 file
= aio_private_file(ctx
, nr_pages
);
512 ctx
->aio_ring_file
= NULL
;
516 ctx
->aio_ring_file
= file
;
517 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
))
518 / sizeof(struct io_event
);
520 ctx
->ring_pages
= ctx
->internal_pages
;
521 if (nr_pages
> AIO_RING_PAGES
) {
522 ctx
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
524 if (!ctx
->ring_pages
) {
525 put_aio_ring_file(ctx
);
530 for (i
= 0; i
< nr_pages
; i
++) {
532 page
= find_or_create_page(file
->f_mapping
,
533 i
, GFP_USER
| __GFP_ZERO
);
536 pr_debug("pid(%d) page[%d]->count=%d\n",
537 current
->pid
, i
, page_count(page
));
538 SetPageUptodate(page
);
541 ctx
->ring_pages
[i
] = page
;
545 if (unlikely(i
!= nr_pages
)) {
550 ctx
->mmap_size
= nr_pages
* PAGE_SIZE
;
551 pr_debug("attempting mmap of %lu bytes\n", ctx
->mmap_size
);
553 if (mmap_write_lock_killable(mm
)) {
559 ctx
->mmap_base
= do_mmap(ctx
->aio_ring_file
, 0, ctx
->mmap_size
,
560 PROT_READ
| PROT_WRITE
,
561 MAP_SHARED
, 0, 0, &unused
, NULL
);
562 mmap_write_unlock(mm
);
563 if (IS_ERR((void *)ctx
->mmap_base
)) {
569 pr_debug("mmap address: 0x%08lx\n", ctx
->mmap_base
);
571 ctx
->user_id
= ctx
->mmap_base
;
572 ctx
->nr_events
= nr_events
; /* trusted copy */
574 ring
= page_address(ctx
->ring_pages
[0]);
575 ring
->nr
= nr_events
; /* user copy */
577 ring
->head
= ring
->tail
= 0;
578 ring
->magic
= AIO_RING_MAGIC
;
579 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
580 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
581 ring
->header_length
= sizeof(struct aio_ring
);
582 flush_dcache_page(ctx
->ring_pages
[0]);
587 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
588 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
589 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
591 void kiocb_set_cancel_fn(struct kiocb
*iocb
, kiocb_cancel_fn
*cancel
)
593 struct aio_kiocb
*req
= container_of(iocb
, struct aio_kiocb
, rw
);
594 struct kioctx
*ctx
= req
->ki_ctx
;
597 if (WARN_ON_ONCE(!list_empty(&req
->ki_list
)))
600 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
601 list_add_tail(&req
->ki_list
, &ctx
->active_reqs
);
602 req
->ki_cancel
= cancel
;
603 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
605 EXPORT_SYMBOL(kiocb_set_cancel_fn
);
608 * free_ioctx() should be RCU delayed to synchronize against the RCU
609 * protected lookup_ioctx() and also needs process context to call
610 * aio_free_ring(). Use rcu_work.
612 static void free_ioctx(struct work_struct
*work
)
614 struct kioctx
*ctx
= container_of(to_rcu_work(work
), struct kioctx
,
616 pr_debug("freeing %p\n", ctx
);
619 free_percpu(ctx
->cpu
);
620 percpu_ref_exit(&ctx
->reqs
);
621 percpu_ref_exit(&ctx
->users
);
622 kmem_cache_free(kioctx_cachep
, ctx
);
625 static void free_ioctx_reqs(struct percpu_ref
*ref
)
627 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, reqs
);
629 /* At this point we know that there are no any in-flight requests */
630 if (ctx
->rq_wait
&& atomic_dec_and_test(&ctx
->rq_wait
->count
))
631 complete(&ctx
->rq_wait
->comp
);
633 /* Synchronize against RCU protected table->table[] dereferences */
634 INIT_RCU_WORK(&ctx
->free_rwork
, free_ioctx
);
635 queue_rcu_work(system_wq
, &ctx
->free_rwork
);
639 * When this function runs, the kioctx has been removed from the "hash table"
640 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
641 * now it's safe to cancel any that need to be.
643 static void free_ioctx_users(struct percpu_ref
*ref
)
645 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, users
);
646 struct aio_kiocb
*req
;
648 spin_lock_irq(&ctx
->ctx_lock
);
650 while (!list_empty(&ctx
->active_reqs
)) {
651 req
= list_first_entry(&ctx
->active_reqs
,
652 struct aio_kiocb
, ki_list
);
653 req
->ki_cancel(&req
->rw
);
654 list_del_init(&req
->ki_list
);
657 spin_unlock_irq(&ctx
->ctx_lock
);
659 percpu_ref_kill(&ctx
->reqs
);
660 percpu_ref_put(&ctx
->reqs
);
663 static int ioctx_add_table(struct kioctx
*ctx
, struct mm_struct
*mm
)
666 struct kioctx_table
*table
, *old
;
667 struct aio_ring
*ring
;
669 spin_lock(&mm
->ioctx_lock
);
670 table
= rcu_dereference_raw(mm
->ioctx_table
);
674 for (i
= 0; i
< table
->nr
; i
++)
675 if (!rcu_access_pointer(table
->table
[i
])) {
677 rcu_assign_pointer(table
->table
[i
], ctx
);
678 spin_unlock(&mm
->ioctx_lock
);
680 /* While kioctx setup is in progress,
681 * we are protected from page migration
682 * changes ring_pages by ->ring_lock.
684 ring
= page_address(ctx
->ring_pages
[0]);
689 new_nr
= (table
? table
->nr
: 1) * 4;
690 spin_unlock(&mm
->ioctx_lock
);
692 table
= kzalloc(struct_size(table
, table
, new_nr
), GFP_KERNEL
);
698 spin_lock(&mm
->ioctx_lock
);
699 old
= rcu_dereference_raw(mm
->ioctx_table
);
702 rcu_assign_pointer(mm
->ioctx_table
, table
);
703 } else if (table
->nr
> old
->nr
) {
704 memcpy(table
->table
, old
->table
,
705 old
->nr
* sizeof(struct kioctx
*));
707 rcu_assign_pointer(mm
->ioctx_table
, table
);
716 static void aio_nr_sub(unsigned nr
)
718 spin_lock(&aio_nr_lock
);
719 if (WARN_ON(aio_nr
- nr
> aio_nr
))
723 spin_unlock(&aio_nr_lock
);
727 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
729 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
731 struct mm_struct
*mm
= current
->mm
;
736 * Store the original nr_events -- what userspace passed to io_setup(),
737 * for counting against the global limit -- before it changes.
739 unsigned int max_reqs
= nr_events
;
742 * We keep track of the number of available ringbuffer slots, to prevent
743 * overflow (reqs_available), and we also use percpu counters for this.
745 * So since up to half the slots might be on other cpu's percpu counters
746 * and unavailable, double nr_events so userspace sees what they
747 * expected: additionally, we move req_batch slots to/from percpu
748 * counters at a time, so make sure that isn't 0:
750 nr_events
= max(nr_events
, num_possible_cpus() * 4);
753 /* Prevent overflows */
754 if (nr_events
> (0x10000000U
/ sizeof(struct io_event
))) {
755 pr_debug("ENOMEM: nr_events too high\n");
756 return ERR_PTR(-EINVAL
);
759 if (!nr_events
|| (unsigned long)max_reqs
> aio_max_nr
)
760 return ERR_PTR(-EAGAIN
);
762 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
764 return ERR_PTR(-ENOMEM
);
766 ctx
->max_reqs
= max_reqs
;
768 spin_lock_init(&ctx
->ctx_lock
);
769 spin_lock_init(&ctx
->completion_lock
);
770 mutex_init(&ctx
->ring_lock
);
771 /* Protect against page migration throughout kiotx setup by keeping
772 * the ring_lock mutex held until setup is complete. */
773 mutex_lock(&ctx
->ring_lock
);
774 init_waitqueue_head(&ctx
->wait
);
776 INIT_LIST_HEAD(&ctx
->active_reqs
);
778 if (percpu_ref_init(&ctx
->users
, free_ioctx_users
, 0, GFP_KERNEL
))
781 if (percpu_ref_init(&ctx
->reqs
, free_ioctx_reqs
, 0, GFP_KERNEL
))
784 ctx
->cpu
= alloc_percpu(struct kioctx_cpu
);
788 err
= aio_setup_ring(ctx
, nr_events
);
792 atomic_set(&ctx
->reqs_available
, ctx
->nr_events
- 1);
793 ctx
->req_batch
= (ctx
->nr_events
- 1) / (num_possible_cpus() * 4);
794 if (ctx
->req_batch
< 1)
797 /* limit the number of system wide aios */
798 spin_lock(&aio_nr_lock
);
799 if (aio_nr
+ ctx
->max_reqs
> aio_max_nr
||
800 aio_nr
+ ctx
->max_reqs
< aio_nr
) {
801 spin_unlock(&aio_nr_lock
);
805 aio_nr
+= ctx
->max_reqs
;
806 spin_unlock(&aio_nr_lock
);
808 percpu_ref_get(&ctx
->users
); /* io_setup() will drop this ref */
809 percpu_ref_get(&ctx
->reqs
); /* free_ioctx_users() will drop this */
811 err
= ioctx_add_table(ctx
, mm
);
815 /* Release the ring_lock mutex now that all setup is complete. */
816 mutex_unlock(&ctx
->ring_lock
);
818 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
819 ctx
, ctx
->user_id
, mm
, ctx
->nr_events
);
823 aio_nr_sub(ctx
->max_reqs
);
825 atomic_set(&ctx
->dead
, 1);
827 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
830 mutex_unlock(&ctx
->ring_lock
);
831 free_percpu(ctx
->cpu
);
832 percpu_ref_exit(&ctx
->reqs
);
833 percpu_ref_exit(&ctx
->users
);
834 kmem_cache_free(kioctx_cachep
, ctx
);
835 pr_debug("error allocating ioctx %d\n", err
);
840 * Cancels all outstanding aio requests on an aio context. Used
841 * when the processes owning a context have all exited to encourage
842 * the rapid destruction of the kioctx.
844 static int kill_ioctx(struct mm_struct
*mm
, struct kioctx
*ctx
,
845 struct ctx_rq_wait
*wait
)
847 struct kioctx_table
*table
;
849 spin_lock(&mm
->ioctx_lock
);
850 if (atomic_xchg(&ctx
->dead
, 1)) {
851 spin_unlock(&mm
->ioctx_lock
);
855 table
= rcu_dereference_raw(mm
->ioctx_table
);
856 WARN_ON(ctx
!= rcu_access_pointer(table
->table
[ctx
->id
]));
857 RCU_INIT_POINTER(table
->table
[ctx
->id
], NULL
);
858 spin_unlock(&mm
->ioctx_lock
);
860 /* free_ioctx_reqs() will do the necessary RCU synchronization */
861 wake_up_all(&ctx
->wait
);
864 * It'd be more correct to do this in free_ioctx(), after all
865 * the outstanding kiocbs have finished - but by then io_destroy
866 * has already returned, so io_setup() could potentially return
867 * -EAGAIN with no ioctxs actually in use (as far as userspace
870 aio_nr_sub(ctx
->max_reqs
);
873 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
876 percpu_ref_kill(&ctx
->users
);
881 * exit_aio: called when the last user of mm goes away. At this point, there is
882 * no way for any new requests to be submited or any of the io_* syscalls to be
883 * called on the context.
885 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
888 void exit_aio(struct mm_struct
*mm
)
890 struct kioctx_table
*table
= rcu_dereference_raw(mm
->ioctx_table
);
891 struct ctx_rq_wait wait
;
897 atomic_set(&wait
.count
, table
->nr
);
898 init_completion(&wait
.comp
);
901 for (i
= 0; i
< table
->nr
; ++i
) {
903 rcu_dereference_protected(table
->table
[i
], true);
911 * We don't need to bother with munmap() here - exit_mmap(mm)
912 * is coming and it'll unmap everything. And we simply can't,
913 * this is not necessarily our ->mm.
914 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
915 * that it needs to unmap the area, just set it to 0.
918 kill_ioctx(mm
, ctx
, &wait
);
921 if (!atomic_sub_and_test(skipped
, &wait
.count
)) {
922 /* Wait until all IO for the context are done. */
923 wait_for_completion(&wait
.comp
);
926 RCU_INIT_POINTER(mm
->ioctx_table
, NULL
);
930 static void put_reqs_available(struct kioctx
*ctx
, unsigned nr
)
932 struct kioctx_cpu
*kcpu
;
935 local_irq_save(flags
);
936 kcpu
= this_cpu_ptr(ctx
->cpu
);
937 kcpu
->reqs_available
+= nr
;
939 while (kcpu
->reqs_available
>= ctx
->req_batch
* 2) {
940 kcpu
->reqs_available
-= ctx
->req_batch
;
941 atomic_add(ctx
->req_batch
, &ctx
->reqs_available
);
944 local_irq_restore(flags
);
947 static bool __get_reqs_available(struct kioctx
*ctx
)
949 struct kioctx_cpu
*kcpu
;
953 local_irq_save(flags
);
954 kcpu
= this_cpu_ptr(ctx
->cpu
);
955 if (!kcpu
->reqs_available
) {
956 int avail
= atomic_read(&ctx
->reqs_available
);
959 if (avail
< ctx
->req_batch
)
961 } while (!atomic_try_cmpxchg(&ctx
->reqs_available
,
962 &avail
, avail
- ctx
->req_batch
));
964 kcpu
->reqs_available
+= ctx
->req_batch
;
968 kcpu
->reqs_available
--;
970 local_irq_restore(flags
);
974 /* refill_reqs_available
975 * Updates the reqs_available reference counts used for tracking the
976 * number of free slots in the completion ring. This can be called
977 * from aio_complete() (to optimistically update reqs_available) or
978 * from aio_get_req() (the we're out of events case). It must be
979 * called holding ctx->completion_lock.
981 static void refill_reqs_available(struct kioctx
*ctx
, unsigned head
,
984 unsigned events_in_ring
, completed
;
986 /* Clamp head since userland can write to it. */
987 head
%= ctx
->nr_events
;
989 events_in_ring
= tail
- head
;
991 events_in_ring
= ctx
->nr_events
- (head
- tail
);
993 completed
= ctx
->completed_events
;
994 if (events_in_ring
< completed
)
995 completed
-= events_in_ring
;
1002 ctx
->completed_events
-= completed
;
1003 put_reqs_available(ctx
, completed
);
1006 /* user_refill_reqs_available
1007 * Called to refill reqs_available when aio_get_req() encounters an
1008 * out of space in the completion ring.
1010 static void user_refill_reqs_available(struct kioctx
*ctx
)
1012 spin_lock_irq(&ctx
->completion_lock
);
1013 if (ctx
->completed_events
) {
1014 struct aio_ring
*ring
;
1017 /* Access of ring->head may race with aio_read_events_ring()
1018 * here, but that's okay since whether we read the old version
1019 * or the new version, and either will be valid. The important
1020 * part is that head cannot pass tail since we prevent
1021 * aio_complete() from updating tail by holding
1022 * ctx->completion_lock. Even if head is invalid, the check
1023 * against ctx->completed_events below will make sure we do the
1026 ring
= page_address(ctx
->ring_pages
[0]);
1029 refill_reqs_available(ctx
, head
, ctx
->tail
);
1032 spin_unlock_irq(&ctx
->completion_lock
);
1035 static bool get_reqs_available(struct kioctx
*ctx
)
1037 if (__get_reqs_available(ctx
))
1039 user_refill_reqs_available(ctx
);
1040 return __get_reqs_available(ctx
);
1044 * Allocate a slot for an aio request.
1045 * Returns NULL if no requests are free.
1047 * The refcount is initialized to 2 - one for the async op completion,
1048 * one for the synchronous code that does this.
1050 static inline struct aio_kiocb
*aio_get_req(struct kioctx
*ctx
)
1052 struct aio_kiocb
*req
;
1054 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
1058 if (unlikely(!get_reqs_available(ctx
))) {
1059 kmem_cache_free(kiocb_cachep
, req
);
1063 percpu_ref_get(&ctx
->reqs
);
1065 INIT_LIST_HEAD(&req
->ki_list
);
1066 refcount_set(&req
->ki_refcnt
, 2);
1067 req
->ki_eventfd
= NULL
;
1071 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
1073 struct aio_ring __user
*ring
= (void __user
*)ctx_id
;
1074 struct mm_struct
*mm
= current
->mm
;
1075 struct kioctx
*ctx
, *ret
= NULL
;
1076 struct kioctx_table
*table
;
1079 if (get_user(id
, &ring
->id
))
1083 table
= rcu_dereference(mm
->ioctx_table
);
1085 if (!table
|| id
>= table
->nr
)
1088 id
= array_index_nospec(id
, table
->nr
);
1089 ctx
= rcu_dereference(table
->table
[id
]);
1090 if (ctx
&& ctx
->user_id
== ctx_id
) {
1091 if (percpu_ref_tryget_live(&ctx
->users
))
1099 static inline void iocb_destroy(struct aio_kiocb
*iocb
)
1101 if (iocb
->ki_eventfd
)
1102 eventfd_ctx_put(iocb
->ki_eventfd
);
1104 fput(iocb
->ki_filp
);
1105 percpu_ref_put(&iocb
->ki_ctx
->reqs
);
1106 kmem_cache_free(kiocb_cachep
, iocb
);
1110 * Called when the io request on the given iocb is complete.
1112 static void aio_complete(struct aio_kiocb
*iocb
)
1114 struct kioctx
*ctx
= iocb
->ki_ctx
;
1115 struct aio_ring
*ring
;
1116 struct io_event
*ev_page
, *event
;
1117 unsigned tail
, pos
, head
;
1118 unsigned long flags
;
1121 * Add a completion event to the ring buffer. Must be done holding
1122 * ctx->completion_lock to prevent other code from messing with the tail
1123 * pointer since we might be called from irq context.
1125 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1128 pos
= tail
+ AIO_EVENTS_OFFSET
;
1130 if (++tail
>= ctx
->nr_events
)
1133 ev_page
= page_address(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1134 event
= ev_page
+ pos
% AIO_EVENTS_PER_PAGE
;
1136 *event
= iocb
->ki_res
;
1138 flush_dcache_page(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1140 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx
, tail
, iocb
,
1141 (void __user
*)(unsigned long)iocb
->ki_res
.obj
,
1142 iocb
->ki_res
.data
, iocb
->ki_res
.res
, iocb
->ki_res
.res2
);
1144 /* after flagging the request as done, we
1145 * must never even look at it again
1147 smp_wmb(); /* make event visible before updating tail */
1151 ring
= page_address(ctx
->ring_pages
[0]);
1154 flush_dcache_page(ctx
->ring_pages
[0]);
1156 ctx
->completed_events
++;
1157 if (ctx
->completed_events
> 1)
1158 refill_reqs_available(ctx
, head
, tail
);
1159 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1161 pr_debug("added to ring %p at [%u]\n", iocb
, tail
);
1164 * Check if the user asked us to deliver the result through an
1165 * eventfd. The eventfd_signal() function is safe to be called
1168 if (iocb
->ki_eventfd
)
1169 eventfd_signal(iocb
->ki_eventfd
, 1);
1172 * We have to order our ring_info tail store above and test
1173 * of the wait list below outside the wait lock. This is
1174 * like in wake_up_bit() where clearing a bit has to be
1175 * ordered with the unlocked test.
1179 if (waitqueue_active(&ctx
->wait
))
1180 wake_up(&ctx
->wait
);
1183 static inline void iocb_put(struct aio_kiocb
*iocb
)
1185 if (refcount_dec_and_test(&iocb
->ki_refcnt
)) {
1191 /* aio_read_events_ring
1192 * Pull an event off of the ioctx's event ring. Returns the number of
1195 static long aio_read_events_ring(struct kioctx
*ctx
,
1196 struct io_event __user
*event
, long nr
)
1198 struct aio_ring
*ring
;
1199 unsigned head
, tail
, pos
;
1204 * The mutex can block and wake us up and that will cause
1205 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1206 * and repeat. This should be rare enough that it doesn't cause
1207 * peformance issues. See the comment in read_events() for more detail.
1209 sched_annotate_sleep();
1210 mutex_lock(&ctx
->ring_lock
);
1212 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1213 ring
= page_address(ctx
->ring_pages
[0]);
1218 * Ensure that once we've read the current tail pointer, that
1219 * we also see the events that were stored up to the tail.
1223 pr_debug("h%u t%u m%u\n", head
, tail
, ctx
->nr_events
);
1228 head
%= ctx
->nr_events
;
1229 tail
%= ctx
->nr_events
;
1233 struct io_event
*ev
;
1236 avail
= (head
<= tail
? tail
: ctx
->nr_events
) - head
;
1240 pos
= head
+ AIO_EVENTS_OFFSET
;
1241 page
= ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
];
1242 pos
%= AIO_EVENTS_PER_PAGE
;
1244 avail
= min(avail
, nr
- ret
);
1245 avail
= min_t(long, avail
, AIO_EVENTS_PER_PAGE
- pos
);
1247 ev
= page_address(page
);
1248 copy_ret
= copy_to_user(event
+ ret
, ev
+ pos
,
1249 sizeof(*ev
) * avail
);
1251 if (unlikely(copy_ret
)) {
1258 head
%= ctx
->nr_events
;
1261 ring
= page_address(ctx
->ring_pages
[0]);
1263 flush_dcache_page(ctx
->ring_pages
[0]);
1265 pr_debug("%li h%u t%u\n", ret
, head
, tail
);
1267 mutex_unlock(&ctx
->ring_lock
);
1272 static bool aio_read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1273 struct io_event __user
*event
, long *i
)
1275 long ret
= aio_read_events_ring(ctx
, event
+ *i
, nr
- *i
);
1280 if (unlikely(atomic_read(&ctx
->dead
)))
1286 return ret
< 0 || *i
>= min_nr
;
1289 static long read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1290 struct io_event __user
*event
,
1296 * Note that aio_read_events() is being called as the conditional - i.e.
1297 * we're calling it after prepare_to_wait() has set task state to
1298 * TASK_INTERRUPTIBLE.
1300 * But aio_read_events() can block, and if it blocks it's going to flip
1301 * the task state back to TASK_RUNNING.
1303 * This should be ok, provided it doesn't flip the state back to
1304 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1305 * will only happen if the mutex_lock() call blocks, and we then find
1306 * the ringbuffer empty. So in practice we should be ok, but it's
1307 * something to be aware of when touching this code.
1310 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
);
1312 wait_event_interruptible_hrtimeout(ctx
->wait
,
1313 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
),
1319 * Create an aio_context capable of receiving at least nr_events.
1320 * ctxp must not point to an aio_context that already exists, and
1321 * must be initialized to 0 prior to the call. On successful
1322 * creation of the aio_context, *ctxp is filled in with the resulting
1323 * handle. May fail with -EINVAL if *ctxp is not initialized,
1324 * if the specified nr_events exceeds internal limits. May fail
1325 * with -EAGAIN if the specified nr_events exceeds the user's limit
1326 * of available events. May fail with -ENOMEM if insufficient kernel
1327 * resources are available. May fail with -EFAULT if an invalid
1328 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1331 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
1333 struct kioctx
*ioctx
= NULL
;
1337 ret
= get_user(ctx
, ctxp
);
1342 if (unlikely(ctx
|| nr_events
== 0)) {
1343 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1348 ioctx
= ioctx_alloc(nr_events
);
1349 ret
= PTR_ERR(ioctx
);
1350 if (!IS_ERR(ioctx
)) {
1351 ret
= put_user(ioctx
->user_id
, ctxp
);
1353 kill_ioctx(current
->mm
, ioctx
, NULL
);
1354 percpu_ref_put(&ioctx
->users
);
1361 #ifdef CONFIG_COMPAT
1362 COMPAT_SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, u32 __user
*, ctx32p
)
1364 struct kioctx
*ioctx
= NULL
;
1368 ret
= get_user(ctx
, ctx32p
);
1373 if (unlikely(ctx
|| nr_events
== 0)) {
1374 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1379 ioctx
= ioctx_alloc(nr_events
);
1380 ret
= PTR_ERR(ioctx
);
1381 if (!IS_ERR(ioctx
)) {
1382 /* truncating is ok because it's a user address */
1383 ret
= put_user((u32
)ioctx
->user_id
, ctx32p
);
1385 kill_ioctx(current
->mm
, ioctx
, NULL
);
1386 percpu_ref_put(&ioctx
->users
);
1395 * Destroy the aio_context specified. May cancel any outstanding
1396 * AIOs and block on completion. Will fail with -ENOSYS if not
1397 * implemented. May fail with -EINVAL if the context pointed to
1400 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1402 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1403 if (likely(NULL
!= ioctx
)) {
1404 struct ctx_rq_wait wait
;
1407 init_completion(&wait
.comp
);
1408 atomic_set(&wait
.count
, 1);
1410 /* Pass requests_done to kill_ioctx() where it can be set
1411 * in a thread-safe way. If we try to set it here then we have
1412 * a race condition if two io_destroy() called simultaneously.
1414 ret
= kill_ioctx(current
->mm
, ioctx
, &wait
);
1415 percpu_ref_put(&ioctx
->users
);
1417 /* Wait until all IO for the context are done. Otherwise kernel
1418 * keep using user-space buffers even if user thinks the context
1422 wait_for_completion(&wait
.comp
);
1426 pr_debug("EINVAL: invalid context id\n");
1430 static void aio_remove_iocb(struct aio_kiocb
*iocb
)
1432 struct kioctx
*ctx
= iocb
->ki_ctx
;
1433 unsigned long flags
;
1435 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
1436 list_del(&iocb
->ki_list
);
1437 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1440 static void aio_complete_rw(struct kiocb
*kiocb
, long res
)
1442 struct aio_kiocb
*iocb
= container_of(kiocb
, struct aio_kiocb
, rw
);
1444 if (!list_empty_careful(&iocb
->ki_list
))
1445 aio_remove_iocb(iocb
);
1447 if (kiocb
->ki_flags
& IOCB_WRITE
) {
1448 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
1450 if (S_ISREG(inode
->i_mode
))
1451 kiocb_end_write(kiocb
);
1454 iocb
->ki_res
.res
= res
;
1455 iocb
->ki_res
.res2
= 0;
1459 static int aio_prep_rw(struct kiocb
*req
, const struct iocb
*iocb
)
1463 req
->ki_complete
= aio_complete_rw
;
1464 req
->private = NULL
;
1465 req
->ki_pos
= iocb
->aio_offset
;
1466 req
->ki_flags
= req
->ki_filp
->f_iocb_flags
;
1467 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
)
1468 req
->ki_flags
|= IOCB_EVENTFD
;
1469 if (iocb
->aio_flags
& IOCB_FLAG_IOPRIO
) {
1471 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1472 * aio_reqprio is interpreted as an I/O scheduling
1473 * class and priority.
1475 ret
= ioprio_check_cap(iocb
->aio_reqprio
);
1477 pr_debug("aio ioprio check cap error: %d\n", ret
);
1481 req
->ki_ioprio
= iocb
->aio_reqprio
;
1483 req
->ki_ioprio
= get_current_ioprio();
1485 ret
= kiocb_set_rw_flags(req
, iocb
->aio_rw_flags
);
1489 req
->ki_flags
&= ~IOCB_HIPRI
; /* no one is going to poll for this I/O */
1493 static ssize_t
aio_setup_rw(int rw
, const struct iocb
*iocb
,
1494 struct iovec
**iovec
, bool vectored
, bool compat
,
1495 struct iov_iter
*iter
)
1497 void __user
*buf
= (void __user
*)(uintptr_t)iocb
->aio_buf
;
1498 size_t len
= iocb
->aio_nbytes
;
1501 ssize_t ret
= import_single_range(rw
, buf
, len
, *iovec
, iter
);
1506 return __import_iovec(rw
, buf
, len
, UIO_FASTIOV
, iovec
, iter
, compat
);
1509 static inline void aio_rw_done(struct kiocb
*req
, ssize_t ret
)
1515 case -ERESTARTNOINTR
:
1516 case -ERESTARTNOHAND
:
1517 case -ERESTART_RESTARTBLOCK
:
1519 * There's no easy way to restart the syscall since other AIO's
1520 * may be already running. Just fail this IO with EINTR.
1525 req
->ki_complete(req
, ret
);
1529 static int aio_read(struct kiocb
*req
, const struct iocb
*iocb
,
1530 bool vectored
, bool compat
)
1532 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1533 struct iov_iter iter
;
1537 ret
= aio_prep_rw(req
, iocb
);
1540 file
= req
->ki_filp
;
1541 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1543 if (unlikely(!file
->f_op
->read_iter
))
1546 ret
= aio_setup_rw(ITER_DEST
, iocb
, &iovec
, vectored
, compat
, &iter
);
1549 ret
= rw_verify_area(READ
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1551 aio_rw_done(req
, call_read_iter(file
, req
, &iter
));
1556 static int aio_write(struct kiocb
*req
, const struct iocb
*iocb
,
1557 bool vectored
, bool compat
)
1559 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1560 struct iov_iter iter
;
1564 ret
= aio_prep_rw(req
, iocb
);
1567 file
= req
->ki_filp
;
1569 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1571 if (unlikely(!file
->f_op
->write_iter
))
1574 ret
= aio_setup_rw(ITER_SOURCE
, iocb
, &iovec
, vectored
, compat
, &iter
);
1577 ret
= rw_verify_area(WRITE
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1579 if (S_ISREG(file_inode(file
)->i_mode
))
1580 kiocb_start_write(req
);
1581 req
->ki_flags
|= IOCB_WRITE
;
1582 aio_rw_done(req
, call_write_iter(file
, req
, &iter
));
1588 static void aio_fsync_work(struct work_struct
*work
)
1590 struct aio_kiocb
*iocb
= container_of(work
, struct aio_kiocb
, fsync
.work
);
1591 const struct cred
*old_cred
= override_creds(iocb
->fsync
.creds
);
1593 iocb
->ki_res
.res
= vfs_fsync(iocb
->fsync
.file
, iocb
->fsync
.datasync
);
1594 revert_creds(old_cred
);
1595 put_cred(iocb
->fsync
.creds
);
1599 static int aio_fsync(struct fsync_iocb
*req
, const struct iocb
*iocb
,
1602 if (unlikely(iocb
->aio_buf
|| iocb
->aio_offset
|| iocb
->aio_nbytes
||
1603 iocb
->aio_rw_flags
))
1606 if (unlikely(!req
->file
->f_op
->fsync
))
1609 req
->creds
= prepare_creds();
1613 req
->datasync
= datasync
;
1614 INIT_WORK(&req
->work
, aio_fsync_work
);
1615 schedule_work(&req
->work
);
1619 static void aio_poll_put_work(struct work_struct
*work
)
1621 struct poll_iocb
*req
= container_of(work
, struct poll_iocb
, work
);
1622 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1628 * Safely lock the waitqueue which the request is on, synchronizing with the
1629 * case where the ->poll() provider decides to free its waitqueue early.
1631 * Returns true on success, meaning that req->head->lock was locked, req->wait
1632 * is on req->head, and an RCU read lock was taken. Returns false if the
1633 * request was already removed from its waitqueue (which might no longer exist).
1635 static bool poll_iocb_lock_wq(struct poll_iocb
*req
)
1637 wait_queue_head_t
*head
;
1640 * While we hold the waitqueue lock and the waitqueue is nonempty,
1641 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1642 * lock in the first place can race with the waitqueue being freed.
1644 * We solve this as eventpoll does: by taking advantage of the fact that
1645 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1646 * we enter rcu_read_lock() and see that the pointer to the queue is
1647 * non-NULL, we can then lock it without the memory being freed out from
1648 * under us, then check whether the request is still on the queue.
1650 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1651 * case the caller deletes the entry from the queue, leaving it empty.
1652 * In that case, only RCU prevents the queue memory from being freed.
1655 head
= smp_load_acquire(&req
->head
);
1657 spin_lock(&head
->lock
);
1658 if (!list_empty(&req
->wait
.entry
))
1660 spin_unlock(&head
->lock
);
1666 static void poll_iocb_unlock_wq(struct poll_iocb
*req
)
1668 spin_unlock(&req
->head
->lock
);
1672 static void aio_poll_complete_work(struct work_struct
*work
)
1674 struct poll_iocb
*req
= container_of(work
, struct poll_iocb
, work
);
1675 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1676 struct poll_table_struct pt
= { ._key
= req
->events
};
1677 struct kioctx
*ctx
= iocb
->ki_ctx
;
1680 if (!READ_ONCE(req
->cancelled
))
1681 mask
= vfs_poll(req
->file
, &pt
) & req
->events
;
1684 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1685 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1686 * synchronize with them. In the cancellation case the list_del_init
1687 * itself is not actually needed, but harmless so we keep it in to
1688 * avoid further branches in the fast path.
1690 spin_lock_irq(&ctx
->ctx_lock
);
1691 if (poll_iocb_lock_wq(req
)) {
1692 if (!mask
&& !READ_ONCE(req
->cancelled
)) {
1694 * The request isn't actually ready to be completed yet.
1695 * Reschedule completion if another wakeup came in.
1697 if (req
->work_need_resched
) {
1698 schedule_work(&req
->work
);
1699 req
->work_need_resched
= false;
1701 req
->work_scheduled
= false;
1703 poll_iocb_unlock_wq(req
);
1704 spin_unlock_irq(&ctx
->ctx_lock
);
1707 list_del_init(&req
->wait
.entry
);
1708 poll_iocb_unlock_wq(req
);
1709 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1710 list_del_init(&iocb
->ki_list
);
1711 iocb
->ki_res
.res
= mangle_poll(mask
);
1712 spin_unlock_irq(&ctx
->ctx_lock
);
1717 /* assumes we are called with irqs disabled */
1718 static int aio_poll_cancel(struct kiocb
*iocb
)
1720 struct aio_kiocb
*aiocb
= container_of(iocb
, struct aio_kiocb
, rw
);
1721 struct poll_iocb
*req
= &aiocb
->poll
;
1723 if (poll_iocb_lock_wq(req
)) {
1724 WRITE_ONCE(req
->cancelled
, true);
1725 if (!req
->work_scheduled
) {
1726 schedule_work(&aiocb
->poll
.work
);
1727 req
->work_scheduled
= true;
1729 poll_iocb_unlock_wq(req
);
1730 } /* else, the request was force-cancelled by POLLFREE already */
1735 static int aio_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1738 struct poll_iocb
*req
= container_of(wait
, struct poll_iocb
, wait
);
1739 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1740 __poll_t mask
= key_to_poll(key
);
1741 unsigned long flags
;
1743 /* for instances that support it check for an event match first: */
1744 if (mask
&& !(mask
& req
->events
))
1748 * Complete the request inline if possible. This requires that three
1749 * conditions be met:
1750 * 1. An event mask must have been passed. If a plain wakeup was done
1751 * instead, then mask == 0 and we have to call vfs_poll() to get
1752 * the events, so inline completion isn't possible.
1753 * 2. The completion work must not have already been scheduled.
1754 * 3. ctx_lock must not be busy. We have to use trylock because we
1755 * already hold the waitqueue lock, so this inverts the normal
1756 * locking order. Use irqsave/irqrestore because not all
1757 * filesystems (e.g. fuse) call this function with IRQs disabled,
1758 * yet IRQs have to be disabled before ctx_lock is obtained.
1760 if (mask
&& !req
->work_scheduled
&&
1761 spin_trylock_irqsave(&iocb
->ki_ctx
->ctx_lock
, flags
)) {
1762 struct kioctx
*ctx
= iocb
->ki_ctx
;
1764 list_del_init(&req
->wait
.entry
);
1765 list_del(&iocb
->ki_list
);
1766 iocb
->ki_res
.res
= mangle_poll(mask
);
1767 if (iocb
->ki_eventfd
&& !eventfd_signal_allowed()) {
1769 INIT_WORK(&req
->work
, aio_poll_put_work
);
1770 schedule_work(&req
->work
);
1772 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1777 * Schedule the completion work if needed. If it was already
1778 * scheduled, record that another wakeup came in.
1780 * Don't remove the request from the waitqueue here, as it might
1781 * not actually be complete yet (we won't know until vfs_poll()
1782 * is called), and we must not miss any wakeups. POLLFREE is an
1783 * exception to this; see below.
1785 if (req
->work_scheduled
) {
1786 req
->work_need_resched
= true;
1788 schedule_work(&req
->work
);
1789 req
->work_scheduled
= true;
1793 * If the waitqueue is being freed early but we can't complete
1794 * the request inline, we have to tear down the request as best
1795 * we can. That means immediately removing the request from its
1796 * waitqueue and preventing all further accesses to the
1797 * waitqueue via the request. We also need to schedule the
1798 * completion work (done above). Also mark the request as
1799 * cancelled, to potentially skip an unneeded call to ->poll().
1801 if (mask
& POLLFREE
) {
1802 WRITE_ONCE(req
->cancelled
, true);
1803 list_del_init(&req
->wait
.entry
);
1806 * Careful: this *must* be the last step, since as soon
1807 * as req->head is NULL'ed out, the request can be
1808 * completed and freed, since aio_poll_complete_work()
1809 * will no longer need to take the waitqueue lock.
1811 smp_store_release(&req
->head
, NULL
);
1817 struct aio_poll_table
{
1818 struct poll_table_struct pt
;
1819 struct aio_kiocb
*iocb
;
1825 aio_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1826 struct poll_table_struct
*p
)
1828 struct aio_poll_table
*pt
= container_of(p
, struct aio_poll_table
, pt
);
1830 /* multiple wait queues per file are not supported */
1831 if (unlikely(pt
->queued
)) {
1832 pt
->error
= -EINVAL
;
1838 pt
->iocb
->poll
.head
= head
;
1839 add_wait_queue(head
, &pt
->iocb
->poll
.wait
);
1842 static int aio_poll(struct aio_kiocb
*aiocb
, const struct iocb
*iocb
)
1844 struct kioctx
*ctx
= aiocb
->ki_ctx
;
1845 struct poll_iocb
*req
= &aiocb
->poll
;
1846 struct aio_poll_table apt
;
1847 bool cancel
= false;
1850 /* reject any unknown events outside the normal event mask. */
1851 if ((u16
)iocb
->aio_buf
!= iocb
->aio_buf
)
1853 /* reject fields that are not defined for poll */
1854 if (iocb
->aio_offset
|| iocb
->aio_nbytes
|| iocb
->aio_rw_flags
)
1857 INIT_WORK(&req
->work
, aio_poll_complete_work
);
1858 req
->events
= demangle_poll(iocb
->aio_buf
) | EPOLLERR
| EPOLLHUP
;
1861 req
->cancelled
= false;
1862 req
->work_scheduled
= false;
1863 req
->work_need_resched
= false;
1865 apt
.pt
._qproc
= aio_poll_queue_proc
;
1866 apt
.pt
._key
= req
->events
;
1869 apt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1871 /* initialized the list so that we can do list_empty checks */
1872 INIT_LIST_HEAD(&req
->wait
.entry
);
1873 init_waitqueue_func_entry(&req
->wait
, aio_poll_wake
);
1875 mask
= vfs_poll(req
->file
, &apt
.pt
) & req
->events
;
1876 spin_lock_irq(&ctx
->ctx_lock
);
1877 if (likely(apt
.queued
)) {
1878 bool on_queue
= poll_iocb_lock_wq(req
);
1880 if (!on_queue
|| req
->work_scheduled
) {
1882 * aio_poll_wake() already either scheduled the async
1883 * completion work, or completed the request inline.
1885 if (apt
.error
) /* unsupported case: multiple queues */
1890 if (mask
|| apt
.error
) {
1891 /* Steal to complete synchronously. */
1892 list_del_init(&req
->wait
.entry
);
1893 } else if (cancel
) {
1894 /* Cancel if possible (may be too late though). */
1895 WRITE_ONCE(req
->cancelled
, true);
1896 } else if (on_queue
) {
1898 * Actually waiting for an event, so add the request to
1899 * active_reqs so that it can be cancelled if needed.
1901 list_add_tail(&aiocb
->ki_list
, &ctx
->active_reqs
);
1902 aiocb
->ki_cancel
= aio_poll_cancel
;
1905 poll_iocb_unlock_wq(req
);
1907 if (mask
) { /* no async, we'd stolen it */
1908 aiocb
->ki_res
.res
= mangle_poll(mask
);
1911 spin_unlock_irq(&ctx
->ctx_lock
);
1917 static int __io_submit_one(struct kioctx
*ctx
, const struct iocb
*iocb
,
1918 struct iocb __user
*user_iocb
, struct aio_kiocb
*req
,
1921 req
->ki_filp
= fget(iocb
->aio_fildes
);
1922 if (unlikely(!req
->ki_filp
))
1925 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1926 struct eventfd_ctx
*eventfd
;
1928 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1929 * instance of the file* now. The file descriptor must be
1930 * an eventfd() fd, and will be signaled for each completed
1931 * event using the eventfd_signal() function.
1933 eventfd
= eventfd_ctx_fdget(iocb
->aio_resfd
);
1934 if (IS_ERR(eventfd
))
1935 return PTR_ERR(eventfd
);
1937 req
->ki_eventfd
= eventfd
;
1940 if (unlikely(put_user(KIOCB_KEY
, &user_iocb
->aio_key
))) {
1941 pr_debug("EFAULT: aio_key\n");
1945 req
->ki_res
.obj
= (u64
)(unsigned long)user_iocb
;
1946 req
->ki_res
.data
= iocb
->aio_data
;
1947 req
->ki_res
.res
= 0;
1948 req
->ki_res
.res2
= 0;
1950 switch (iocb
->aio_lio_opcode
) {
1951 case IOCB_CMD_PREAD
:
1952 return aio_read(&req
->rw
, iocb
, false, compat
);
1953 case IOCB_CMD_PWRITE
:
1954 return aio_write(&req
->rw
, iocb
, false, compat
);
1955 case IOCB_CMD_PREADV
:
1956 return aio_read(&req
->rw
, iocb
, true, compat
);
1957 case IOCB_CMD_PWRITEV
:
1958 return aio_write(&req
->rw
, iocb
, true, compat
);
1959 case IOCB_CMD_FSYNC
:
1960 return aio_fsync(&req
->fsync
, iocb
, false);
1961 case IOCB_CMD_FDSYNC
:
1962 return aio_fsync(&req
->fsync
, iocb
, true);
1964 return aio_poll(req
, iocb
);
1966 pr_debug("invalid aio operation %d\n", iocb
->aio_lio_opcode
);
1971 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
1974 struct aio_kiocb
*req
;
1978 if (unlikely(copy_from_user(&iocb
, user_iocb
, sizeof(iocb
))))
1981 /* enforce forwards compatibility on users */
1982 if (unlikely(iocb
.aio_reserved2
)) {
1983 pr_debug("EINVAL: reserve field set\n");
1987 /* prevent overflows */
1989 (iocb
.aio_buf
!= (unsigned long)iocb
.aio_buf
) ||
1990 (iocb
.aio_nbytes
!= (size_t)iocb
.aio_nbytes
) ||
1991 ((ssize_t
)iocb
.aio_nbytes
< 0)
1993 pr_debug("EINVAL: overflow check\n");
1997 req
= aio_get_req(ctx
);
2001 err
= __io_submit_one(ctx
, &iocb
, user_iocb
, req
, compat
);
2003 /* Done with the synchronous reference */
2007 * If err is 0, we'd either done aio_complete() ourselves or have
2008 * arranged for that to be done asynchronously. Anything non-zero
2009 * means that we need to destroy req ourselves.
2011 if (unlikely(err
)) {
2013 put_reqs_available(ctx
, 1);
2019 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2020 * the number of iocbs queued. May return -EINVAL if the aio_context
2021 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2022 * *iocbpp[0] is not properly initialized, if the operation specified
2023 * is invalid for the file descriptor in the iocb. May fail with
2024 * -EFAULT if any of the data structures point to invalid data. May
2025 * fail with -EBADF if the file descriptor specified in the first
2026 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2027 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2028 * fail with -ENOSYS if not implemented.
2030 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
2031 struct iocb __user
* __user
*, iocbpp
)
2036 struct blk_plug plug
;
2038 if (unlikely(nr
< 0))
2041 ctx
= lookup_ioctx(ctx_id
);
2042 if (unlikely(!ctx
)) {
2043 pr_debug("EINVAL: invalid context id\n");
2047 if (nr
> ctx
->nr_events
)
2048 nr
= ctx
->nr_events
;
2050 if (nr
> AIO_PLUG_THRESHOLD
)
2051 blk_start_plug(&plug
);
2052 for (i
= 0; i
< nr
; i
++) {
2053 struct iocb __user
*user_iocb
;
2055 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
2060 ret
= io_submit_one(ctx
, user_iocb
, false);
2064 if (nr
> AIO_PLUG_THRESHOLD
)
2065 blk_finish_plug(&plug
);
2067 percpu_ref_put(&ctx
->users
);
2071 #ifdef CONFIG_COMPAT
2072 COMPAT_SYSCALL_DEFINE3(io_submit
, compat_aio_context_t
, ctx_id
,
2073 int, nr
, compat_uptr_t __user
*, iocbpp
)
2078 struct blk_plug plug
;
2080 if (unlikely(nr
< 0))
2083 ctx
= lookup_ioctx(ctx_id
);
2084 if (unlikely(!ctx
)) {
2085 pr_debug("EINVAL: invalid context id\n");
2089 if (nr
> ctx
->nr_events
)
2090 nr
= ctx
->nr_events
;
2092 if (nr
> AIO_PLUG_THRESHOLD
)
2093 blk_start_plug(&plug
);
2094 for (i
= 0; i
< nr
; i
++) {
2095 compat_uptr_t user_iocb
;
2097 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
2102 ret
= io_submit_one(ctx
, compat_ptr(user_iocb
), true);
2106 if (nr
> AIO_PLUG_THRESHOLD
)
2107 blk_finish_plug(&plug
);
2109 percpu_ref_put(&ctx
->users
);
2115 * Attempts to cancel an iocb previously passed to io_submit. If
2116 * the operation is successfully cancelled, the resulting event is
2117 * copied into the memory pointed to by result without being placed
2118 * into the completion queue and 0 is returned. May fail with
2119 * -EFAULT if any of the data structures pointed to are invalid.
2120 * May fail with -EINVAL if aio_context specified by ctx_id is
2121 * invalid. May fail with -EAGAIN if the iocb specified was not
2122 * cancelled. Will fail with -ENOSYS if not implemented.
2124 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
2125 struct io_event __user
*, result
)
2128 struct aio_kiocb
*kiocb
;
2131 u64 obj
= (u64
)(unsigned long)iocb
;
2133 if (unlikely(get_user(key
, &iocb
->aio_key
)))
2135 if (unlikely(key
!= KIOCB_KEY
))
2138 ctx
= lookup_ioctx(ctx_id
);
2142 spin_lock_irq(&ctx
->ctx_lock
);
2143 /* TODO: use a hash or array, this sucks. */
2144 list_for_each_entry(kiocb
, &ctx
->active_reqs
, ki_list
) {
2145 if (kiocb
->ki_res
.obj
== obj
) {
2146 ret
= kiocb
->ki_cancel(&kiocb
->rw
);
2147 list_del_init(&kiocb
->ki_list
);
2151 spin_unlock_irq(&ctx
->ctx_lock
);
2155 * The result argument is no longer used - the io_event is
2156 * always delivered via the ring buffer. -EINPROGRESS indicates
2157 * cancellation is progress:
2162 percpu_ref_put(&ctx
->users
);
2167 static long do_io_getevents(aio_context_t ctx_id
,
2170 struct io_event __user
*events
,
2171 struct timespec64
*ts
)
2173 ktime_t until
= ts
? timespec64_to_ktime(*ts
) : KTIME_MAX
;
2174 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
2177 if (likely(ioctx
)) {
2178 if (likely(min_nr
<= nr
&& min_nr
>= 0))
2179 ret
= read_events(ioctx
, min_nr
, nr
, events
, until
);
2180 percpu_ref_put(&ioctx
->users
);
2187 * Attempts to read at least min_nr events and up to nr events from
2188 * the completion queue for the aio_context specified by ctx_id. If
2189 * it succeeds, the number of read events is returned. May fail with
2190 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2191 * out of range, if timeout is out of range. May fail with -EFAULT
2192 * if any of the memory specified is invalid. May return 0 or
2193 * < min_nr if the timeout specified by timeout has elapsed
2194 * before sufficient events are available, where timeout == NULL
2195 * specifies an infinite timeout. Note that the timeout pointed to by
2196 * timeout is relative. Will fail with -ENOSYS if not implemented.
2200 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
2203 struct io_event __user
*, events
,
2204 struct __kernel_timespec __user
*, timeout
)
2206 struct timespec64 ts
;
2209 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2212 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2213 if (!ret
&& signal_pending(current
))
2220 struct __aio_sigset
{
2221 const sigset_t __user
*sigmask
;
2225 SYSCALL_DEFINE6(io_pgetevents
,
2226 aio_context_t
, ctx_id
,
2229 struct io_event __user
*, events
,
2230 struct __kernel_timespec __user
*, timeout
,
2231 const struct __aio_sigset __user
*, usig
)
2233 struct __aio_sigset ksig
= { NULL
, };
2234 struct timespec64 ts
;
2238 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2241 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2244 ret
= set_user_sigmask(ksig
.sigmask
, ksig
.sigsetsize
);
2248 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2250 interrupted
= signal_pending(current
);
2251 restore_saved_sigmask_unless(interrupted
);
2252 if (interrupted
&& !ret
)
2253 ret
= -ERESTARTNOHAND
;
2258 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2260 SYSCALL_DEFINE6(io_pgetevents_time32
,
2261 aio_context_t
, ctx_id
,
2264 struct io_event __user
*, events
,
2265 struct old_timespec32 __user
*, timeout
,
2266 const struct __aio_sigset __user
*, usig
)
2268 struct __aio_sigset ksig
= { NULL
, };
2269 struct timespec64 ts
;
2273 if (timeout
&& unlikely(get_old_timespec32(&ts
, timeout
)))
2276 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2280 ret
= set_user_sigmask(ksig
.sigmask
, ksig
.sigsetsize
);
2284 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2286 interrupted
= signal_pending(current
);
2287 restore_saved_sigmask_unless(interrupted
);
2288 if (interrupted
&& !ret
)
2289 ret
= -ERESTARTNOHAND
;
2296 #if defined(CONFIG_COMPAT_32BIT_TIME)
2298 SYSCALL_DEFINE5(io_getevents_time32
, __u32
, ctx_id
,
2301 struct io_event __user
*, events
,
2302 struct old_timespec32 __user
*, timeout
)
2304 struct timespec64 t
;
2307 if (timeout
&& get_old_timespec32(&t
, timeout
))
2310 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2311 if (!ret
&& signal_pending(current
))
2318 #ifdef CONFIG_COMPAT
2320 struct __compat_aio_sigset
{
2321 compat_uptr_t sigmask
;
2322 compat_size_t sigsetsize
;
2325 #if defined(CONFIG_COMPAT_32BIT_TIME)
2327 COMPAT_SYSCALL_DEFINE6(io_pgetevents
,
2328 compat_aio_context_t
, ctx_id
,
2329 compat_long_t
, min_nr
,
2331 struct io_event __user
*, events
,
2332 struct old_timespec32 __user
*, timeout
,
2333 const struct __compat_aio_sigset __user
*, usig
)
2335 struct __compat_aio_sigset ksig
= { 0, };
2336 struct timespec64 t
;
2340 if (timeout
&& get_old_timespec32(&t
, timeout
))
2343 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2346 ret
= set_compat_user_sigmask(compat_ptr(ksig
.sigmask
), ksig
.sigsetsize
);
2350 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2352 interrupted
= signal_pending(current
);
2353 restore_saved_sigmask_unless(interrupted
);
2354 if (interrupted
&& !ret
)
2355 ret
= -ERESTARTNOHAND
;
2362 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64
,
2363 compat_aio_context_t
, ctx_id
,
2364 compat_long_t
, min_nr
,
2366 struct io_event __user
*, events
,
2367 struct __kernel_timespec __user
*, timeout
,
2368 const struct __compat_aio_sigset __user
*, usig
)
2370 struct __compat_aio_sigset ksig
= { 0, };
2371 struct timespec64 t
;
2375 if (timeout
&& get_timespec64(&t
, timeout
))
2378 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2381 ret
= set_compat_user_sigmask(compat_ptr(ksig
.sigmask
), ksig
.sigsetsize
);
2385 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2387 interrupted
= signal_pending(current
);
2388 restore_saved_sigmask_unless(interrupted
);
2389 if (interrupted
&& !ret
)
2390 ret
= -ERESTARTNOHAND
;