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/mmu_context.h>
31 #include <linux/percpu.h>
32 #include <linux/slab.h>
33 #include <linux/timer.h>
34 #include <linux/aio.h>
35 #include <linux/highmem.h>
36 #include <linux/workqueue.h>
37 #include <linux/security.h>
38 #include <linux/eventfd.h>
39 #include <linux/blkdev.h>
40 #include <linux/compat.h>
41 #include <linux/migrate.h>
42 #include <linux/ramfs.h>
43 #include <linux/percpu-refcount.h>
44 #include <linux/mount.h>
46 #include <asm/kmap_types.h>
47 #include <linux/uaccess.h>
48 #include <linux/nospec.h>
54 #define AIO_RING_MAGIC 0xa10a10a1
55 #define AIO_RING_COMPAT_FEATURES 1
56 #define AIO_RING_INCOMPAT_FEATURES 0
58 unsigned id
; /* kernel internal index number */
59 unsigned nr
; /* number of io_events */
60 unsigned head
; /* Written to by userland or under ring_lock
61 * mutex by aio_read_events_ring(). */
65 unsigned compat_features
;
66 unsigned incompat_features
;
67 unsigned header_length
; /* size of aio_ring */
70 struct io_event io_events
[0];
71 }; /* 128 bytes + ring size */
74 * Plugging is meant to work with larger batches of IOs. If we don't
75 * have more than the below, then don't bother setting up a plug.
77 #define AIO_PLUG_THRESHOLD 2
79 #define AIO_RING_PAGES 8
84 struct kioctx __rcu
*table
[];
88 unsigned reqs_available
;
92 struct completion comp
;
97 struct percpu_ref users
;
100 struct percpu_ref reqs
;
102 unsigned long user_id
;
104 struct __percpu kioctx_cpu
*cpu
;
107 * For percpu reqs_available, number of slots we move to/from global
112 * This is what userspace passed to io_setup(), it's not used for
113 * anything but counting against the global max_reqs quota.
115 * The real limit is nr_events - 1, which will be larger (see
120 /* Size of ringbuffer, in units of struct io_event */
123 unsigned long mmap_base
;
124 unsigned long mmap_size
;
126 struct page
**ring_pages
;
129 struct rcu_work free_rwork
; /* see free_ioctx() */
132 * signals when all in-flight requests are done
134 struct ctx_rq_wait
*rq_wait
;
138 * This counts the number of available slots in the ringbuffer,
139 * so we avoid overflowing it: it's decremented (if positive)
140 * when allocating a kiocb and incremented when the resulting
141 * io_event is pulled off the ringbuffer.
143 * We batch accesses to it with a percpu version.
145 atomic_t reqs_available
;
146 } ____cacheline_aligned_in_smp
;
150 struct list_head active_reqs
; /* used for cancellation */
151 } ____cacheline_aligned_in_smp
;
154 struct mutex ring_lock
;
155 wait_queue_head_t wait
;
156 } ____cacheline_aligned_in_smp
;
160 unsigned completed_events
;
161 spinlock_t completion_lock
;
162 } ____cacheline_aligned_in_smp
;
164 struct page
*internal_pages
[AIO_RING_PAGES
];
165 struct file
*aio_ring_file
;
171 * First field must be the file pointer in all the
172 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
176 struct work_struct work
;
182 struct wait_queue_head
*head
;
186 struct wait_queue_entry wait
;
187 struct work_struct work
;
191 * NOTE! Each of the iocb union members has the file pointer
192 * as the first entry in their struct definition. So you can
193 * access the file pointer through any of the sub-structs,
194 * or directly as just 'ki_filp' in this struct.
198 struct file
*ki_filp
;
200 struct fsync_iocb fsync
;
201 struct poll_iocb poll
;
204 struct kioctx
*ki_ctx
;
205 kiocb_cancel_fn
*ki_cancel
;
207 struct iocb __user
*ki_user_iocb
; /* user's aiocb */
208 __u64 ki_user_data
; /* user's data for completion */
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 unsigned long aio_nr
; /* current system wide number of aio requests */
224 unsigned long aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct kmem_cache
*kiocb_cachep
;
228 static struct kmem_cache
*kioctx_cachep
;
230 static struct vfsmount
*aio_mnt
;
232 static const struct file_operations aio_ring_fops
;
233 static const struct address_space_operations aio_ctx_aops
;
235 static struct file
*aio_private_file(struct kioctx
*ctx
, loff_t nr_pages
)
238 struct inode
*inode
= alloc_anon_inode(aio_mnt
->mnt_sb
);
240 return ERR_CAST(inode
);
242 inode
->i_mapping
->a_ops
= &aio_ctx_aops
;
243 inode
->i_mapping
->private_data
= ctx
;
244 inode
->i_size
= PAGE_SIZE
* nr_pages
;
246 file
= alloc_file_pseudo(inode
, aio_mnt
, "[aio]",
247 O_RDWR
, &aio_ring_fops
);
253 static struct dentry
*aio_mount(struct file_system_type
*fs_type
,
254 int flags
, const char *dev_name
, void *data
)
256 struct dentry
*root
= mount_pseudo(fs_type
, "aio:", NULL
, NULL
,
260 root
->d_sb
->s_iflags
|= SB_I_NOEXEC
;
265 * Creates the slab caches used by the aio routines, panic on
266 * failure as this is done early during the boot sequence.
268 static int __init
aio_setup(void)
270 static struct file_system_type aio_fs
= {
273 .kill_sb
= kill_anon_super
,
275 aio_mnt
= kern_mount(&aio_fs
);
277 panic("Failed to create aio fs mount.");
279 kiocb_cachep
= KMEM_CACHE(aio_kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
280 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
283 __initcall(aio_setup
);
285 static void put_aio_ring_file(struct kioctx
*ctx
)
287 struct file
*aio_ring_file
= ctx
->aio_ring_file
;
288 struct address_space
*i_mapping
;
291 truncate_setsize(file_inode(aio_ring_file
), 0);
293 /* Prevent further access to the kioctx from migratepages */
294 i_mapping
= aio_ring_file
->f_mapping
;
295 spin_lock(&i_mapping
->private_lock
);
296 i_mapping
->private_data
= NULL
;
297 ctx
->aio_ring_file
= NULL
;
298 spin_unlock(&i_mapping
->private_lock
);
304 static void aio_free_ring(struct kioctx
*ctx
)
308 /* Disconnect the kiotx from the ring file. This prevents future
309 * accesses to the kioctx from page migration.
311 put_aio_ring_file(ctx
);
313 for (i
= 0; i
< ctx
->nr_pages
; i
++) {
315 pr_debug("pid(%d) [%d] page->count=%d\n", current
->pid
, i
,
316 page_count(ctx
->ring_pages
[i
]));
317 page
= ctx
->ring_pages
[i
];
320 ctx
->ring_pages
[i
] = NULL
;
324 if (ctx
->ring_pages
&& ctx
->ring_pages
!= ctx
->internal_pages
) {
325 kfree(ctx
->ring_pages
);
326 ctx
->ring_pages
= NULL
;
330 static int aio_ring_mremap(struct vm_area_struct
*vma
)
332 struct file
*file
= vma
->vm_file
;
333 struct mm_struct
*mm
= vma
->vm_mm
;
334 struct kioctx_table
*table
;
335 int i
, res
= -EINVAL
;
337 spin_lock(&mm
->ioctx_lock
);
339 table
= rcu_dereference(mm
->ioctx_table
);
340 for (i
= 0; i
< table
->nr
; i
++) {
343 ctx
= rcu_dereference(table
->table
[i
]);
344 if (ctx
&& ctx
->aio_ring_file
== file
) {
345 if (!atomic_read(&ctx
->dead
)) {
346 ctx
->user_id
= ctx
->mmap_base
= vma
->vm_start
;
354 spin_unlock(&mm
->ioctx_lock
);
358 static const struct vm_operations_struct aio_ring_vm_ops
= {
359 .mremap
= aio_ring_mremap
,
360 #if IS_ENABLED(CONFIG_MMU)
361 .fault
= filemap_fault
,
362 .map_pages
= filemap_map_pages
,
363 .page_mkwrite
= filemap_page_mkwrite
,
367 static int aio_ring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
369 vma
->vm_flags
|= VM_DONTEXPAND
;
370 vma
->vm_ops
= &aio_ring_vm_ops
;
374 static const struct file_operations aio_ring_fops
= {
375 .mmap
= aio_ring_mmap
,
378 #if IS_ENABLED(CONFIG_MIGRATION)
379 static int aio_migratepage(struct address_space
*mapping
, struct page
*new,
380 struct page
*old
, enum migrate_mode mode
)
388 * We cannot support the _NO_COPY case here, because copy needs to
389 * happen under the ctx->completion_lock. That does not work with the
390 * migration workflow of MIGRATE_SYNC_NO_COPY.
392 if (mode
== MIGRATE_SYNC_NO_COPY
)
397 /* mapping->private_lock here protects against the kioctx teardown. */
398 spin_lock(&mapping
->private_lock
);
399 ctx
= mapping
->private_data
;
405 /* The ring_lock mutex. The prevents aio_read_events() from writing
406 * to the ring's head, and prevents page migration from mucking in
407 * a partially initialized kiotx.
409 if (!mutex_trylock(&ctx
->ring_lock
)) {
415 if (idx
< (pgoff_t
)ctx
->nr_pages
) {
416 /* Make sure the old page hasn't already been changed */
417 if (ctx
->ring_pages
[idx
] != old
)
425 /* Writeback must be complete */
426 BUG_ON(PageWriteback(old
));
429 rc
= migrate_page_move_mapping(mapping
, new, old
, mode
, 1);
430 if (rc
!= MIGRATEPAGE_SUCCESS
) {
435 /* Take completion_lock to prevent other writes to the ring buffer
436 * while the old page is copied to the new. This prevents new
437 * events from being lost.
439 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
440 migrate_page_copy(new, old
);
441 BUG_ON(ctx
->ring_pages
[idx
] != old
);
442 ctx
->ring_pages
[idx
] = new;
443 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
445 /* The old page is no longer accessible. */
449 mutex_unlock(&ctx
->ring_lock
);
451 spin_unlock(&mapping
->private_lock
);
456 static const struct address_space_operations aio_ctx_aops
= {
457 .set_page_dirty
= __set_page_dirty_no_writeback
,
458 #if IS_ENABLED(CONFIG_MIGRATION)
459 .migratepage
= aio_migratepage
,
463 static int aio_setup_ring(struct kioctx
*ctx
, unsigned int nr_events
)
465 struct aio_ring
*ring
;
466 struct mm_struct
*mm
= current
->mm
;
467 unsigned long size
, unused
;
472 /* Compensate for the ring buffer's head/tail overlap entry */
473 nr_events
+= 2; /* 1 is required, 2 for good luck */
475 size
= sizeof(struct aio_ring
);
476 size
+= sizeof(struct io_event
) * nr_events
;
478 nr_pages
= PFN_UP(size
);
482 file
= aio_private_file(ctx
, nr_pages
);
484 ctx
->aio_ring_file
= NULL
;
488 ctx
->aio_ring_file
= file
;
489 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
))
490 / sizeof(struct io_event
);
492 ctx
->ring_pages
= ctx
->internal_pages
;
493 if (nr_pages
> AIO_RING_PAGES
) {
494 ctx
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
496 if (!ctx
->ring_pages
) {
497 put_aio_ring_file(ctx
);
502 for (i
= 0; i
< nr_pages
; i
++) {
504 page
= find_or_create_page(file
->f_mapping
,
505 i
, GFP_HIGHUSER
| __GFP_ZERO
);
508 pr_debug("pid(%d) page[%d]->count=%d\n",
509 current
->pid
, i
, page_count(page
));
510 SetPageUptodate(page
);
513 ctx
->ring_pages
[i
] = page
;
517 if (unlikely(i
!= nr_pages
)) {
522 ctx
->mmap_size
= nr_pages
* PAGE_SIZE
;
523 pr_debug("attempting mmap of %lu bytes\n", ctx
->mmap_size
);
525 if (down_write_killable(&mm
->mmap_sem
)) {
531 ctx
->mmap_base
= do_mmap_pgoff(ctx
->aio_ring_file
, 0, ctx
->mmap_size
,
532 PROT_READ
| PROT_WRITE
,
533 MAP_SHARED
, 0, &unused
, NULL
);
534 up_write(&mm
->mmap_sem
);
535 if (IS_ERR((void *)ctx
->mmap_base
)) {
541 pr_debug("mmap address: 0x%08lx\n", ctx
->mmap_base
);
543 ctx
->user_id
= ctx
->mmap_base
;
544 ctx
->nr_events
= nr_events
; /* trusted copy */
546 ring
= kmap_atomic(ctx
->ring_pages
[0]);
547 ring
->nr
= nr_events
; /* user copy */
549 ring
->head
= ring
->tail
= 0;
550 ring
->magic
= AIO_RING_MAGIC
;
551 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
552 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
553 ring
->header_length
= sizeof(struct aio_ring
);
555 flush_dcache_page(ctx
->ring_pages
[0]);
560 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
561 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
562 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
564 void kiocb_set_cancel_fn(struct kiocb
*iocb
, kiocb_cancel_fn
*cancel
)
566 struct aio_kiocb
*req
= container_of(iocb
, struct aio_kiocb
, rw
);
567 struct kioctx
*ctx
= req
->ki_ctx
;
570 if (WARN_ON_ONCE(!list_empty(&req
->ki_list
)))
573 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
574 list_add_tail(&req
->ki_list
, &ctx
->active_reqs
);
575 req
->ki_cancel
= cancel
;
576 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
578 EXPORT_SYMBOL(kiocb_set_cancel_fn
);
581 * free_ioctx() should be RCU delayed to synchronize against the RCU
582 * protected lookup_ioctx() and also needs process context to call
583 * aio_free_ring(). Use rcu_work.
585 static void free_ioctx(struct work_struct
*work
)
587 struct kioctx
*ctx
= container_of(to_rcu_work(work
), struct kioctx
,
589 pr_debug("freeing %p\n", ctx
);
592 free_percpu(ctx
->cpu
);
593 percpu_ref_exit(&ctx
->reqs
);
594 percpu_ref_exit(&ctx
->users
);
595 kmem_cache_free(kioctx_cachep
, ctx
);
598 static void free_ioctx_reqs(struct percpu_ref
*ref
)
600 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, reqs
);
602 /* At this point we know that there are no any in-flight requests */
603 if (ctx
->rq_wait
&& atomic_dec_and_test(&ctx
->rq_wait
->count
))
604 complete(&ctx
->rq_wait
->comp
);
606 /* Synchronize against RCU protected table->table[] dereferences */
607 INIT_RCU_WORK(&ctx
->free_rwork
, free_ioctx
);
608 queue_rcu_work(system_wq
, &ctx
->free_rwork
);
612 * When this function runs, the kioctx has been removed from the "hash table"
613 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
614 * now it's safe to cancel any that need to be.
616 static void free_ioctx_users(struct percpu_ref
*ref
)
618 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, users
);
619 struct aio_kiocb
*req
;
621 spin_lock_irq(&ctx
->ctx_lock
);
623 while (!list_empty(&ctx
->active_reqs
)) {
624 req
= list_first_entry(&ctx
->active_reqs
,
625 struct aio_kiocb
, ki_list
);
626 req
->ki_cancel(&req
->rw
);
627 list_del_init(&req
->ki_list
);
630 spin_unlock_irq(&ctx
->ctx_lock
);
632 percpu_ref_kill(&ctx
->reqs
);
633 percpu_ref_put(&ctx
->reqs
);
636 static int ioctx_add_table(struct kioctx
*ctx
, struct mm_struct
*mm
)
639 struct kioctx_table
*table
, *old
;
640 struct aio_ring
*ring
;
642 spin_lock(&mm
->ioctx_lock
);
643 table
= rcu_dereference_raw(mm
->ioctx_table
);
647 for (i
= 0; i
< table
->nr
; i
++)
648 if (!rcu_access_pointer(table
->table
[i
])) {
650 rcu_assign_pointer(table
->table
[i
], ctx
);
651 spin_unlock(&mm
->ioctx_lock
);
653 /* While kioctx setup is in progress,
654 * we are protected from page migration
655 * changes ring_pages by ->ring_lock.
657 ring
= kmap_atomic(ctx
->ring_pages
[0]);
663 new_nr
= (table
? table
->nr
: 1) * 4;
664 spin_unlock(&mm
->ioctx_lock
);
666 table
= kzalloc(sizeof(*table
) + sizeof(struct kioctx
*) *
673 spin_lock(&mm
->ioctx_lock
);
674 old
= rcu_dereference_raw(mm
->ioctx_table
);
677 rcu_assign_pointer(mm
->ioctx_table
, table
);
678 } else if (table
->nr
> old
->nr
) {
679 memcpy(table
->table
, old
->table
,
680 old
->nr
* sizeof(struct kioctx
*));
682 rcu_assign_pointer(mm
->ioctx_table
, table
);
691 static void aio_nr_sub(unsigned nr
)
693 spin_lock(&aio_nr_lock
);
694 if (WARN_ON(aio_nr
- nr
> aio_nr
))
698 spin_unlock(&aio_nr_lock
);
702 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
704 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
706 struct mm_struct
*mm
= current
->mm
;
711 * Store the original nr_events -- what userspace passed to io_setup(),
712 * for counting against the global limit -- before it changes.
714 unsigned int max_reqs
= nr_events
;
717 * We keep track of the number of available ringbuffer slots, to prevent
718 * overflow (reqs_available), and we also use percpu counters for this.
720 * So since up to half the slots might be on other cpu's percpu counters
721 * and unavailable, double nr_events so userspace sees what they
722 * expected: additionally, we move req_batch slots to/from percpu
723 * counters at a time, so make sure that isn't 0:
725 nr_events
= max(nr_events
, num_possible_cpus() * 4);
728 /* Prevent overflows */
729 if (nr_events
> (0x10000000U
/ sizeof(struct io_event
))) {
730 pr_debug("ENOMEM: nr_events too high\n");
731 return ERR_PTR(-EINVAL
);
734 if (!nr_events
|| (unsigned long)max_reqs
> aio_max_nr
)
735 return ERR_PTR(-EAGAIN
);
737 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
739 return ERR_PTR(-ENOMEM
);
741 ctx
->max_reqs
= max_reqs
;
743 spin_lock_init(&ctx
->ctx_lock
);
744 spin_lock_init(&ctx
->completion_lock
);
745 mutex_init(&ctx
->ring_lock
);
746 /* Protect against page migration throughout kiotx setup by keeping
747 * the ring_lock mutex held until setup is complete. */
748 mutex_lock(&ctx
->ring_lock
);
749 init_waitqueue_head(&ctx
->wait
);
751 INIT_LIST_HEAD(&ctx
->active_reqs
);
753 if (percpu_ref_init(&ctx
->users
, free_ioctx_users
, 0, GFP_KERNEL
))
756 if (percpu_ref_init(&ctx
->reqs
, free_ioctx_reqs
, 0, GFP_KERNEL
))
759 ctx
->cpu
= alloc_percpu(struct kioctx_cpu
);
763 err
= aio_setup_ring(ctx
, nr_events
);
767 atomic_set(&ctx
->reqs_available
, ctx
->nr_events
- 1);
768 ctx
->req_batch
= (ctx
->nr_events
- 1) / (num_possible_cpus() * 4);
769 if (ctx
->req_batch
< 1)
772 /* limit the number of system wide aios */
773 spin_lock(&aio_nr_lock
);
774 if (aio_nr
+ ctx
->max_reqs
> aio_max_nr
||
775 aio_nr
+ ctx
->max_reqs
< aio_nr
) {
776 spin_unlock(&aio_nr_lock
);
780 aio_nr
+= ctx
->max_reqs
;
781 spin_unlock(&aio_nr_lock
);
783 percpu_ref_get(&ctx
->users
); /* io_setup() will drop this ref */
784 percpu_ref_get(&ctx
->reqs
); /* free_ioctx_users() will drop this */
786 err
= ioctx_add_table(ctx
, mm
);
790 /* Release the ring_lock mutex now that all setup is complete. */
791 mutex_unlock(&ctx
->ring_lock
);
793 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
794 ctx
, ctx
->user_id
, mm
, ctx
->nr_events
);
798 aio_nr_sub(ctx
->max_reqs
);
800 atomic_set(&ctx
->dead
, 1);
802 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
805 mutex_unlock(&ctx
->ring_lock
);
806 free_percpu(ctx
->cpu
);
807 percpu_ref_exit(&ctx
->reqs
);
808 percpu_ref_exit(&ctx
->users
);
809 kmem_cache_free(kioctx_cachep
, ctx
);
810 pr_debug("error allocating ioctx %d\n", err
);
815 * Cancels all outstanding aio requests on an aio context. Used
816 * when the processes owning a context have all exited to encourage
817 * the rapid destruction of the kioctx.
819 static int kill_ioctx(struct mm_struct
*mm
, struct kioctx
*ctx
,
820 struct ctx_rq_wait
*wait
)
822 struct kioctx_table
*table
;
824 spin_lock(&mm
->ioctx_lock
);
825 if (atomic_xchg(&ctx
->dead
, 1)) {
826 spin_unlock(&mm
->ioctx_lock
);
830 table
= rcu_dereference_raw(mm
->ioctx_table
);
831 WARN_ON(ctx
!= rcu_access_pointer(table
->table
[ctx
->id
]));
832 RCU_INIT_POINTER(table
->table
[ctx
->id
], NULL
);
833 spin_unlock(&mm
->ioctx_lock
);
835 /* free_ioctx_reqs() will do the necessary RCU synchronization */
836 wake_up_all(&ctx
->wait
);
839 * It'd be more correct to do this in free_ioctx(), after all
840 * the outstanding kiocbs have finished - but by then io_destroy
841 * has already returned, so io_setup() could potentially return
842 * -EAGAIN with no ioctxs actually in use (as far as userspace
845 aio_nr_sub(ctx
->max_reqs
);
848 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
851 percpu_ref_kill(&ctx
->users
);
856 * exit_aio: called when the last user of mm goes away. At this point, there is
857 * no way for any new requests to be submited or any of the io_* syscalls to be
858 * called on the context.
860 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
863 void exit_aio(struct mm_struct
*mm
)
865 struct kioctx_table
*table
= rcu_dereference_raw(mm
->ioctx_table
);
866 struct ctx_rq_wait wait
;
872 atomic_set(&wait
.count
, table
->nr
);
873 init_completion(&wait
.comp
);
876 for (i
= 0; i
< table
->nr
; ++i
) {
878 rcu_dereference_protected(table
->table
[i
], true);
886 * We don't need to bother with munmap() here - exit_mmap(mm)
887 * is coming and it'll unmap everything. And we simply can't,
888 * this is not necessarily our ->mm.
889 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
890 * that it needs to unmap the area, just set it to 0.
893 kill_ioctx(mm
, ctx
, &wait
);
896 if (!atomic_sub_and_test(skipped
, &wait
.count
)) {
897 /* Wait until all IO for the context are done. */
898 wait_for_completion(&wait
.comp
);
901 RCU_INIT_POINTER(mm
->ioctx_table
, NULL
);
905 static void put_reqs_available(struct kioctx
*ctx
, unsigned nr
)
907 struct kioctx_cpu
*kcpu
;
910 local_irq_save(flags
);
911 kcpu
= this_cpu_ptr(ctx
->cpu
);
912 kcpu
->reqs_available
+= nr
;
914 while (kcpu
->reqs_available
>= ctx
->req_batch
* 2) {
915 kcpu
->reqs_available
-= ctx
->req_batch
;
916 atomic_add(ctx
->req_batch
, &ctx
->reqs_available
);
919 local_irq_restore(flags
);
922 static bool __get_reqs_available(struct kioctx
*ctx
)
924 struct kioctx_cpu
*kcpu
;
928 local_irq_save(flags
);
929 kcpu
= this_cpu_ptr(ctx
->cpu
);
930 if (!kcpu
->reqs_available
) {
931 int old
, avail
= atomic_read(&ctx
->reqs_available
);
934 if (avail
< ctx
->req_batch
)
938 avail
= atomic_cmpxchg(&ctx
->reqs_available
,
939 avail
, avail
- ctx
->req_batch
);
940 } while (avail
!= old
);
942 kcpu
->reqs_available
+= ctx
->req_batch
;
946 kcpu
->reqs_available
--;
948 local_irq_restore(flags
);
952 /* refill_reqs_available
953 * Updates the reqs_available reference counts used for tracking the
954 * number of free slots in the completion ring. This can be called
955 * from aio_complete() (to optimistically update reqs_available) or
956 * from aio_get_req() (the we're out of events case). It must be
957 * called holding ctx->completion_lock.
959 static void refill_reqs_available(struct kioctx
*ctx
, unsigned head
,
962 unsigned events_in_ring
, completed
;
964 /* Clamp head since userland can write to it. */
965 head
%= ctx
->nr_events
;
967 events_in_ring
= tail
- head
;
969 events_in_ring
= ctx
->nr_events
- (head
- tail
);
971 completed
= ctx
->completed_events
;
972 if (events_in_ring
< completed
)
973 completed
-= events_in_ring
;
980 ctx
->completed_events
-= completed
;
981 put_reqs_available(ctx
, completed
);
984 /* user_refill_reqs_available
985 * Called to refill reqs_available when aio_get_req() encounters an
986 * out of space in the completion ring.
988 static void user_refill_reqs_available(struct kioctx
*ctx
)
990 spin_lock_irq(&ctx
->completion_lock
);
991 if (ctx
->completed_events
) {
992 struct aio_ring
*ring
;
995 /* Access of ring->head may race with aio_read_events_ring()
996 * here, but that's okay since whether we read the old version
997 * or the new version, and either will be valid. The important
998 * part is that head cannot pass tail since we prevent
999 * aio_complete() from updating tail by holding
1000 * ctx->completion_lock. Even if head is invalid, the check
1001 * against ctx->completed_events below will make sure we do the
1004 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1006 kunmap_atomic(ring
);
1008 refill_reqs_available(ctx
, head
, ctx
->tail
);
1011 spin_unlock_irq(&ctx
->completion_lock
);
1014 static bool get_reqs_available(struct kioctx
*ctx
)
1016 if (__get_reqs_available(ctx
))
1018 user_refill_reqs_available(ctx
);
1019 return __get_reqs_available(ctx
);
1023 * Allocate a slot for an aio request.
1024 * Returns NULL if no requests are free.
1026 * The refcount is initialized to 2 - one for the async op completion,
1027 * one for the synchronous code that does this.
1029 static inline struct aio_kiocb
*aio_get_req(struct kioctx
*ctx
)
1031 struct aio_kiocb
*req
;
1033 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
1037 percpu_ref_get(&ctx
->reqs
);
1039 INIT_LIST_HEAD(&req
->ki_list
);
1040 refcount_set(&req
->ki_refcnt
, 2);
1041 req
->ki_eventfd
= NULL
;
1045 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
1047 struct aio_ring __user
*ring
= (void __user
*)ctx_id
;
1048 struct mm_struct
*mm
= current
->mm
;
1049 struct kioctx
*ctx
, *ret
= NULL
;
1050 struct kioctx_table
*table
;
1053 if (get_user(id
, &ring
->id
))
1057 table
= rcu_dereference(mm
->ioctx_table
);
1059 if (!table
|| id
>= table
->nr
)
1062 id
= array_index_nospec(id
, table
->nr
);
1063 ctx
= rcu_dereference(table
->table
[id
]);
1064 if (ctx
&& ctx
->user_id
== ctx_id
) {
1065 if (percpu_ref_tryget_live(&ctx
->users
))
1073 static inline void iocb_destroy(struct aio_kiocb
*iocb
)
1076 fput(iocb
->ki_filp
);
1077 percpu_ref_put(&iocb
->ki_ctx
->reqs
);
1078 kmem_cache_free(kiocb_cachep
, iocb
);
1081 static inline void iocb_put(struct aio_kiocb
*iocb
)
1083 if (refcount_dec_and_test(&iocb
->ki_refcnt
))
1087 static void aio_fill_event(struct io_event
*ev
, struct aio_kiocb
*iocb
,
1088 long res
, long res2
)
1090 ev
->obj
= (u64
)(unsigned long)iocb
->ki_user_iocb
;
1091 ev
->data
= iocb
->ki_user_data
;
1097 * Called when the io request on the given iocb is complete.
1099 static void aio_complete(struct aio_kiocb
*iocb
, long res
, long res2
)
1101 struct kioctx
*ctx
= iocb
->ki_ctx
;
1102 struct aio_ring
*ring
;
1103 struct io_event
*ev_page
, *event
;
1104 unsigned tail
, pos
, head
;
1105 unsigned long flags
;
1108 * Add a completion event to the ring buffer. Must be done holding
1109 * ctx->completion_lock to prevent other code from messing with the tail
1110 * pointer since we might be called from irq context.
1112 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1115 pos
= tail
+ AIO_EVENTS_OFFSET
;
1117 if (++tail
>= ctx
->nr_events
)
1120 ev_page
= kmap_atomic(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1121 event
= ev_page
+ pos
% AIO_EVENTS_PER_PAGE
;
1123 aio_fill_event(event
, iocb
, res
, res2
);
1125 kunmap_atomic(ev_page
);
1126 flush_dcache_page(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1128 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1129 ctx
, tail
, iocb
, iocb
->ki_user_iocb
, iocb
->ki_user_data
,
1132 /* after flagging the request as done, we
1133 * must never even look at it again
1135 smp_wmb(); /* make event visible before updating tail */
1139 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1142 kunmap_atomic(ring
);
1143 flush_dcache_page(ctx
->ring_pages
[0]);
1145 ctx
->completed_events
++;
1146 if (ctx
->completed_events
> 1)
1147 refill_reqs_available(ctx
, head
, tail
);
1148 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1150 pr_debug("added to ring %p at [%u]\n", iocb
, tail
);
1153 * Check if the user asked us to deliver the result through an
1154 * eventfd. The eventfd_signal() function is safe to be called
1157 if (iocb
->ki_eventfd
) {
1158 eventfd_signal(iocb
->ki_eventfd
, 1);
1159 eventfd_ctx_put(iocb
->ki_eventfd
);
1163 * We have to order our ring_info tail store above and test
1164 * of the wait list below outside the wait lock. This is
1165 * like in wake_up_bit() where clearing a bit has to be
1166 * ordered with the unlocked test.
1170 if (waitqueue_active(&ctx
->wait
))
1171 wake_up(&ctx
->wait
);
1175 /* aio_read_events_ring
1176 * Pull an event off of the ioctx's event ring. Returns the number of
1179 static long aio_read_events_ring(struct kioctx
*ctx
,
1180 struct io_event __user
*event
, long nr
)
1182 struct aio_ring
*ring
;
1183 unsigned head
, tail
, pos
;
1188 * The mutex can block and wake us up and that will cause
1189 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1190 * and repeat. This should be rare enough that it doesn't cause
1191 * peformance issues. See the comment in read_events() for more detail.
1193 sched_annotate_sleep();
1194 mutex_lock(&ctx
->ring_lock
);
1196 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1197 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1200 kunmap_atomic(ring
);
1203 * Ensure that once we've read the current tail pointer, that
1204 * we also see the events that were stored up to the tail.
1208 pr_debug("h%u t%u m%u\n", head
, tail
, ctx
->nr_events
);
1213 head
%= ctx
->nr_events
;
1214 tail
%= ctx
->nr_events
;
1218 struct io_event
*ev
;
1221 avail
= (head
<= tail
? tail
: ctx
->nr_events
) - head
;
1225 pos
= head
+ AIO_EVENTS_OFFSET
;
1226 page
= ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
];
1227 pos
%= AIO_EVENTS_PER_PAGE
;
1229 avail
= min(avail
, nr
- ret
);
1230 avail
= min_t(long, avail
, AIO_EVENTS_PER_PAGE
- pos
);
1233 copy_ret
= copy_to_user(event
+ ret
, ev
+ pos
,
1234 sizeof(*ev
) * avail
);
1237 if (unlikely(copy_ret
)) {
1244 head
%= ctx
->nr_events
;
1247 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1249 kunmap_atomic(ring
);
1250 flush_dcache_page(ctx
->ring_pages
[0]);
1252 pr_debug("%li h%u t%u\n", ret
, head
, tail
);
1254 mutex_unlock(&ctx
->ring_lock
);
1259 static bool aio_read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1260 struct io_event __user
*event
, long *i
)
1262 long ret
= aio_read_events_ring(ctx
, event
+ *i
, nr
- *i
);
1267 if (unlikely(atomic_read(&ctx
->dead
)))
1273 return ret
< 0 || *i
>= min_nr
;
1276 static long read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1277 struct io_event __user
*event
,
1283 * Note that aio_read_events() is being called as the conditional - i.e.
1284 * we're calling it after prepare_to_wait() has set task state to
1285 * TASK_INTERRUPTIBLE.
1287 * But aio_read_events() can block, and if it blocks it's going to flip
1288 * the task state back to TASK_RUNNING.
1290 * This should be ok, provided it doesn't flip the state back to
1291 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1292 * will only happen if the mutex_lock() call blocks, and we then find
1293 * the ringbuffer empty. So in practice we should be ok, but it's
1294 * something to be aware of when touching this code.
1297 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
);
1299 wait_event_interruptible_hrtimeout(ctx
->wait
,
1300 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
),
1306 * Create an aio_context capable of receiving at least nr_events.
1307 * ctxp must not point to an aio_context that already exists, and
1308 * must be initialized to 0 prior to the call. On successful
1309 * creation of the aio_context, *ctxp is filled in with the resulting
1310 * handle. May fail with -EINVAL if *ctxp is not initialized,
1311 * if the specified nr_events exceeds internal limits. May fail
1312 * with -EAGAIN if the specified nr_events exceeds the user's limit
1313 * of available events. May fail with -ENOMEM if insufficient kernel
1314 * resources are available. May fail with -EFAULT if an invalid
1315 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1318 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
1320 struct kioctx
*ioctx
= NULL
;
1324 ret
= get_user(ctx
, ctxp
);
1329 if (unlikely(ctx
|| nr_events
== 0)) {
1330 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1335 ioctx
= ioctx_alloc(nr_events
);
1336 ret
= PTR_ERR(ioctx
);
1337 if (!IS_ERR(ioctx
)) {
1338 ret
= put_user(ioctx
->user_id
, ctxp
);
1340 kill_ioctx(current
->mm
, ioctx
, NULL
);
1341 percpu_ref_put(&ioctx
->users
);
1348 #ifdef CONFIG_COMPAT
1349 COMPAT_SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, u32 __user
*, ctx32p
)
1351 struct kioctx
*ioctx
= NULL
;
1355 ret
= get_user(ctx
, ctx32p
);
1360 if (unlikely(ctx
|| nr_events
== 0)) {
1361 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1366 ioctx
= ioctx_alloc(nr_events
);
1367 ret
= PTR_ERR(ioctx
);
1368 if (!IS_ERR(ioctx
)) {
1369 /* truncating is ok because it's a user address */
1370 ret
= put_user((u32
)ioctx
->user_id
, ctx32p
);
1372 kill_ioctx(current
->mm
, ioctx
, NULL
);
1373 percpu_ref_put(&ioctx
->users
);
1382 * Destroy the aio_context specified. May cancel any outstanding
1383 * AIOs and block on completion. Will fail with -ENOSYS if not
1384 * implemented. May fail with -EINVAL if the context pointed to
1387 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1389 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1390 if (likely(NULL
!= ioctx
)) {
1391 struct ctx_rq_wait wait
;
1394 init_completion(&wait
.comp
);
1395 atomic_set(&wait
.count
, 1);
1397 /* Pass requests_done to kill_ioctx() where it can be set
1398 * in a thread-safe way. If we try to set it here then we have
1399 * a race condition if two io_destroy() called simultaneously.
1401 ret
= kill_ioctx(current
->mm
, ioctx
, &wait
);
1402 percpu_ref_put(&ioctx
->users
);
1404 /* Wait until all IO for the context are done. Otherwise kernel
1405 * keep using user-space buffers even if user thinks the context
1409 wait_for_completion(&wait
.comp
);
1413 pr_debug("EINVAL: invalid context id\n");
1417 static void aio_remove_iocb(struct aio_kiocb
*iocb
)
1419 struct kioctx
*ctx
= iocb
->ki_ctx
;
1420 unsigned long flags
;
1422 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
1423 list_del(&iocb
->ki_list
);
1424 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1427 static void aio_complete_rw(struct kiocb
*kiocb
, long res
, long res2
)
1429 struct aio_kiocb
*iocb
= container_of(kiocb
, struct aio_kiocb
, rw
);
1431 if (!list_empty_careful(&iocb
->ki_list
))
1432 aio_remove_iocb(iocb
);
1434 if (kiocb
->ki_flags
& IOCB_WRITE
) {
1435 struct inode
*inode
= file_inode(kiocb
->ki_filp
);
1438 * Tell lockdep we inherited freeze protection from submission
1441 if (S_ISREG(inode
->i_mode
))
1442 __sb_writers_acquired(inode
->i_sb
, SB_FREEZE_WRITE
);
1443 file_end_write(kiocb
->ki_filp
);
1446 aio_complete(iocb
, res
, res2
);
1449 static int aio_prep_rw(struct kiocb
*req
, const struct iocb
*iocb
)
1453 req
->ki_complete
= aio_complete_rw
;
1454 req
->private = NULL
;
1455 req
->ki_pos
= iocb
->aio_offset
;
1456 req
->ki_flags
= iocb_flags(req
->ki_filp
);
1457 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
)
1458 req
->ki_flags
|= IOCB_EVENTFD
;
1459 req
->ki_hint
= ki_hint_validate(file_write_hint(req
->ki_filp
));
1460 if (iocb
->aio_flags
& IOCB_FLAG_IOPRIO
) {
1462 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1463 * aio_reqprio is interpreted as an I/O scheduling
1464 * class and priority.
1466 ret
= ioprio_check_cap(iocb
->aio_reqprio
);
1468 pr_debug("aio ioprio check cap error: %d\n", ret
);
1472 req
->ki_ioprio
= iocb
->aio_reqprio
;
1474 req
->ki_ioprio
= get_current_ioprio();
1476 ret
= kiocb_set_rw_flags(req
, iocb
->aio_rw_flags
);
1480 req
->ki_flags
&= ~IOCB_HIPRI
; /* no one is going to poll for this I/O */
1484 static int aio_setup_rw(int rw
, const struct iocb
*iocb
, struct iovec
**iovec
,
1485 bool vectored
, bool compat
, struct iov_iter
*iter
)
1487 void __user
*buf
= (void __user
*)(uintptr_t)iocb
->aio_buf
;
1488 size_t len
= iocb
->aio_nbytes
;
1491 ssize_t ret
= import_single_range(rw
, buf
, len
, *iovec
, iter
);
1495 #ifdef CONFIG_COMPAT
1497 return compat_import_iovec(rw
, buf
, len
, UIO_FASTIOV
, iovec
,
1500 return import_iovec(rw
, buf
, len
, UIO_FASTIOV
, iovec
, iter
);
1503 static inline void aio_rw_done(struct kiocb
*req
, ssize_t ret
)
1509 case -ERESTARTNOINTR
:
1510 case -ERESTARTNOHAND
:
1511 case -ERESTART_RESTARTBLOCK
:
1513 * There's no easy way to restart the syscall since other AIO's
1514 * may be already running. Just fail this IO with EINTR.
1519 req
->ki_complete(req
, ret
, 0);
1523 static ssize_t
aio_read(struct kiocb
*req
, const struct iocb
*iocb
,
1524 bool vectored
, bool compat
)
1526 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1527 struct iov_iter iter
;
1531 ret
= aio_prep_rw(req
, iocb
);
1534 file
= req
->ki_filp
;
1535 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1538 if (unlikely(!file
->f_op
->read_iter
))
1541 ret
= aio_setup_rw(READ
, iocb
, &iovec
, vectored
, compat
, &iter
);
1544 ret
= rw_verify_area(READ
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1546 aio_rw_done(req
, call_read_iter(file
, req
, &iter
));
1551 static ssize_t
aio_write(struct kiocb
*req
, const struct iocb
*iocb
,
1552 bool vectored
, bool compat
)
1554 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1555 struct iov_iter iter
;
1559 ret
= aio_prep_rw(req
, iocb
);
1562 file
= req
->ki_filp
;
1564 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1566 if (unlikely(!file
->f_op
->write_iter
))
1569 ret
= aio_setup_rw(WRITE
, iocb
, &iovec
, vectored
, compat
, &iter
);
1572 ret
= rw_verify_area(WRITE
, file
, &req
->ki_pos
, iov_iter_count(&iter
));
1575 * Open-code file_start_write here to grab freeze protection,
1576 * which will be released by another thread in
1577 * aio_complete_rw(). Fool lockdep by telling it the lock got
1578 * released so that it doesn't complain about the held lock when
1579 * we return to userspace.
1581 if (S_ISREG(file_inode(file
)->i_mode
)) {
1582 __sb_start_write(file_inode(file
)->i_sb
, SB_FREEZE_WRITE
, true);
1583 __sb_writers_release(file_inode(file
)->i_sb
, SB_FREEZE_WRITE
);
1585 req
->ki_flags
|= IOCB_WRITE
;
1586 aio_rw_done(req
, call_write_iter(file
, req
, &iter
));
1592 static void aio_fsync_work(struct work_struct
*work
)
1594 struct fsync_iocb
*req
= container_of(work
, struct fsync_iocb
, work
);
1597 ret
= vfs_fsync(req
->file
, req
->datasync
);
1598 aio_complete(container_of(req
, struct aio_kiocb
, fsync
), ret
, 0);
1601 static int aio_fsync(struct fsync_iocb
*req
, const struct iocb
*iocb
,
1604 if (unlikely(iocb
->aio_buf
|| iocb
->aio_offset
|| iocb
->aio_nbytes
||
1605 iocb
->aio_rw_flags
))
1608 if (unlikely(!req
->file
->f_op
->fsync
))
1611 req
->datasync
= datasync
;
1612 INIT_WORK(&req
->work
, aio_fsync_work
);
1613 schedule_work(&req
->work
);
1617 static inline void aio_poll_complete(struct aio_kiocb
*iocb
, __poll_t mask
)
1619 aio_complete(iocb
, mangle_poll(mask
), 0);
1622 static void aio_poll_complete_work(struct work_struct
*work
)
1624 struct poll_iocb
*req
= container_of(work
, struct poll_iocb
, work
);
1625 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1626 struct poll_table_struct pt
= { ._key
= req
->events
};
1627 struct kioctx
*ctx
= iocb
->ki_ctx
;
1630 if (!READ_ONCE(req
->cancelled
))
1631 mask
= vfs_poll(req
->file
, &pt
) & req
->events
;
1634 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1635 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1636 * synchronize with them. In the cancellation case the list_del_init
1637 * itself is not actually needed, but harmless so we keep it in to
1638 * avoid further branches in the fast path.
1640 spin_lock_irq(&ctx
->ctx_lock
);
1641 if (!mask
&& !READ_ONCE(req
->cancelled
)) {
1642 add_wait_queue(req
->head
, &req
->wait
);
1643 spin_unlock_irq(&ctx
->ctx_lock
);
1646 list_del_init(&iocb
->ki_list
);
1647 spin_unlock_irq(&ctx
->ctx_lock
);
1649 aio_poll_complete(iocb
, mask
);
1652 /* assumes we are called with irqs disabled */
1653 static int aio_poll_cancel(struct kiocb
*iocb
)
1655 struct aio_kiocb
*aiocb
= container_of(iocb
, struct aio_kiocb
, rw
);
1656 struct poll_iocb
*req
= &aiocb
->poll
;
1658 spin_lock(&req
->head
->lock
);
1659 WRITE_ONCE(req
->cancelled
, true);
1660 if (!list_empty(&req
->wait
.entry
)) {
1661 list_del_init(&req
->wait
.entry
);
1662 schedule_work(&aiocb
->poll
.work
);
1664 spin_unlock(&req
->head
->lock
);
1669 static int aio_poll_wake(struct wait_queue_entry
*wait
, unsigned mode
, int sync
,
1672 struct poll_iocb
*req
= container_of(wait
, struct poll_iocb
, wait
);
1673 struct aio_kiocb
*iocb
= container_of(req
, struct aio_kiocb
, poll
);
1674 __poll_t mask
= key_to_poll(key
);
1675 unsigned long flags
;
1679 /* for instances that support it check for an event match first: */
1681 if (!(mask
& req
->events
))
1685 * Try to complete the iocb inline if we can. Use
1686 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1687 * call this function with IRQs disabled and because IRQs
1688 * have to be disabled before ctx_lock is obtained.
1690 if (spin_trylock_irqsave(&iocb
->ki_ctx
->ctx_lock
, flags
)) {
1691 list_del(&iocb
->ki_list
);
1692 spin_unlock_irqrestore(&iocb
->ki_ctx
->ctx_lock
, flags
);
1694 list_del_init(&req
->wait
.entry
);
1695 aio_poll_complete(iocb
, mask
);
1700 list_del_init(&req
->wait
.entry
);
1701 schedule_work(&req
->work
);
1705 struct aio_poll_table
{
1706 struct poll_table_struct pt
;
1707 struct aio_kiocb
*iocb
;
1712 aio_poll_queue_proc(struct file
*file
, struct wait_queue_head
*head
,
1713 struct poll_table_struct
*p
)
1715 struct aio_poll_table
*pt
= container_of(p
, struct aio_poll_table
, pt
);
1717 /* multiple wait queues per file are not supported */
1718 if (unlikely(pt
->iocb
->poll
.head
)) {
1719 pt
->error
= -EINVAL
;
1724 pt
->iocb
->poll
.head
= head
;
1725 add_wait_queue(head
, &pt
->iocb
->poll
.wait
);
1728 static ssize_t
aio_poll(struct aio_kiocb
*aiocb
, const struct iocb
*iocb
)
1730 struct kioctx
*ctx
= aiocb
->ki_ctx
;
1731 struct poll_iocb
*req
= &aiocb
->poll
;
1732 struct aio_poll_table apt
;
1735 /* reject any unknown events outside the normal event mask. */
1736 if ((u16
)iocb
->aio_buf
!= iocb
->aio_buf
)
1738 /* reject fields that are not defined for poll */
1739 if (iocb
->aio_offset
|| iocb
->aio_nbytes
|| iocb
->aio_rw_flags
)
1742 INIT_WORK(&req
->work
, aio_poll_complete_work
);
1743 req
->events
= demangle_poll(iocb
->aio_buf
) | EPOLLERR
| EPOLLHUP
;
1747 req
->cancelled
= false;
1749 apt
.pt
._qproc
= aio_poll_queue_proc
;
1750 apt
.pt
._key
= req
->events
;
1752 apt
.error
= -EINVAL
; /* same as no support for IOCB_CMD_POLL */
1754 /* initialized the list so that we can do list_empty checks */
1755 INIT_LIST_HEAD(&req
->wait
.entry
);
1756 init_waitqueue_func_entry(&req
->wait
, aio_poll_wake
);
1758 mask
= vfs_poll(req
->file
, &apt
.pt
) & req
->events
;
1759 if (unlikely(!req
->head
)) {
1760 /* we did not manage to set up a waitqueue, done */
1764 spin_lock_irq(&ctx
->ctx_lock
);
1765 spin_lock(&req
->head
->lock
);
1767 /* wake_up context handles the rest */
1770 } else if (mask
|| apt
.error
) {
1771 /* if we get an error or a mask we are done */
1772 WARN_ON_ONCE(list_empty(&req
->wait
.entry
));
1773 list_del_init(&req
->wait
.entry
);
1775 /* actually waiting for an event */
1776 list_add_tail(&aiocb
->ki_list
, &ctx
->active_reqs
);
1777 aiocb
->ki_cancel
= aio_poll_cancel
;
1779 spin_unlock(&req
->head
->lock
);
1780 spin_unlock_irq(&ctx
->ctx_lock
);
1783 if (unlikely(apt
.error
))
1787 aio_poll_complete(aiocb
, mask
);
1791 static int __io_submit_one(struct kioctx
*ctx
, const struct iocb
*iocb
,
1792 struct iocb __user
*user_iocb
, bool compat
)
1794 struct aio_kiocb
*req
;
1797 /* enforce forwards compatibility on users */
1798 if (unlikely(iocb
->aio_reserved2
)) {
1799 pr_debug("EINVAL: reserve field set\n");
1803 /* prevent overflows */
1805 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1806 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1807 ((ssize_t
)iocb
->aio_nbytes
< 0)
1809 pr_debug("EINVAL: overflow check\n");
1813 if (!get_reqs_available(ctx
))
1817 req
= aio_get_req(ctx
);
1819 goto out_put_reqs_available
;
1821 req
->ki_filp
= fget(iocb
->aio_fildes
);
1823 if (unlikely(!req
->ki_filp
))
1826 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1828 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1829 * instance of the file* now. The file descriptor must be
1830 * an eventfd() fd, and will be signaled for each completed
1831 * event using the eventfd_signal() function.
1833 req
->ki_eventfd
= eventfd_ctx_fdget((int) iocb
->aio_resfd
);
1834 if (IS_ERR(req
->ki_eventfd
)) {
1835 ret
= PTR_ERR(req
->ki_eventfd
);
1836 req
->ki_eventfd
= NULL
;
1841 ret
= put_user(KIOCB_KEY
, &user_iocb
->aio_key
);
1842 if (unlikely(ret
)) {
1843 pr_debug("EFAULT: aio_key\n");
1847 req
->ki_user_iocb
= user_iocb
;
1848 req
->ki_user_data
= iocb
->aio_data
;
1850 switch (iocb
->aio_lio_opcode
) {
1851 case IOCB_CMD_PREAD
:
1852 ret
= aio_read(&req
->rw
, iocb
, false, compat
);
1854 case IOCB_CMD_PWRITE
:
1855 ret
= aio_write(&req
->rw
, iocb
, false, compat
);
1857 case IOCB_CMD_PREADV
:
1858 ret
= aio_read(&req
->rw
, iocb
, true, compat
);
1860 case IOCB_CMD_PWRITEV
:
1861 ret
= aio_write(&req
->rw
, iocb
, true, compat
);
1863 case IOCB_CMD_FSYNC
:
1864 ret
= aio_fsync(&req
->fsync
, iocb
, false);
1866 case IOCB_CMD_FDSYNC
:
1867 ret
= aio_fsync(&req
->fsync
, iocb
, true);
1870 ret
= aio_poll(req
, iocb
);
1873 pr_debug("invalid aio operation %d\n", iocb
->aio_lio_opcode
);
1878 /* Done with the synchronous reference */
1882 * If ret is 0, we'd either done aio_complete() ourselves or have
1883 * arranged for that to be done asynchronously. Anything non-zero
1884 * means that we need to destroy req ourselves.
1890 if (req
->ki_eventfd
)
1891 eventfd_ctx_put(req
->ki_eventfd
);
1893 out_put_reqs_available
:
1894 put_reqs_available(ctx
, 1);
1898 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
1903 if (unlikely(copy_from_user(&iocb
, user_iocb
, sizeof(iocb
))))
1906 return __io_submit_one(ctx
, &iocb
, user_iocb
, compat
);
1910 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1911 * the number of iocbs queued. May return -EINVAL if the aio_context
1912 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1913 * *iocbpp[0] is not properly initialized, if the operation specified
1914 * is invalid for the file descriptor in the iocb. May fail with
1915 * -EFAULT if any of the data structures point to invalid data. May
1916 * fail with -EBADF if the file descriptor specified in the first
1917 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1918 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1919 * fail with -ENOSYS if not implemented.
1921 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
1922 struct iocb __user
* __user
*, iocbpp
)
1927 struct blk_plug plug
;
1929 if (unlikely(nr
< 0))
1932 ctx
= lookup_ioctx(ctx_id
);
1933 if (unlikely(!ctx
)) {
1934 pr_debug("EINVAL: invalid context id\n");
1938 if (nr
> ctx
->nr_events
)
1939 nr
= ctx
->nr_events
;
1941 if (nr
> AIO_PLUG_THRESHOLD
)
1942 blk_start_plug(&plug
);
1943 for (i
= 0; i
< nr
; i
++) {
1944 struct iocb __user
*user_iocb
;
1946 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
1951 ret
= io_submit_one(ctx
, user_iocb
, false);
1955 if (nr
> AIO_PLUG_THRESHOLD
)
1956 blk_finish_plug(&plug
);
1958 percpu_ref_put(&ctx
->users
);
1962 #ifdef CONFIG_COMPAT
1963 COMPAT_SYSCALL_DEFINE3(io_submit
, compat_aio_context_t
, ctx_id
,
1964 int, nr
, compat_uptr_t __user
*, iocbpp
)
1969 struct blk_plug plug
;
1971 if (unlikely(nr
< 0))
1974 ctx
= lookup_ioctx(ctx_id
);
1975 if (unlikely(!ctx
)) {
1976 pr_debug("EINVAL: invalid context id\n");
1980 if (nr
> ctx
->nr_events
)
1981 nr
= ctx
->nr_events
;
1983 if (nr
> AIO_PLUG_THRESHOLD
)
1984 blk_start_plug(&plug
);
1985 for (i
= 0; i
< nr
; i
++) {
1986 compat_uptr_t user_iocb
;
1988 if (unlikely(get_user(user_iocb
, iocbpp
+ i
))) {
1993 ret
= io_submit_one(ctx
, compat_ptr(user_iocb
), true);
1997 if (nr
> AIO_PLUG_THRESHOLD
)
1998 blk_finish_plug(&plug
);
2000 percpu_ref_put(&ctx
->users
);
2006 * Attempts to cancel an iocb previously passed to io_submit. If
2007 * the operation is successfully cancelled, the resulting event is
2008 * copied into the memory pointed to by result without being placed
2009 * into the completion queue and 0 is returned. May fail with
2010 * -EFAULT if any of the data structures pointed to are invalid.
2011 * May fail with -EINVAL if aio_context specified by ctx_id is
2012 * invalid. May fail with -EAGAIN if the iocb specified was not
2013 * cancelled. Will fail with -ENOSYS if not implemented.
2015 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
2016 struct io_event __user
*, result
)
2019 struct aio_kiocb
*kiocb
;
2023 if (unlikely(get_user(key
, &iocb
->aio_key
)))
2025 if (unlikely(key
!= KIOCB_KEY
))
2028 ctx
= lookup_ioctx(ctx_id
);
2032 spin_lock_irq(&ctx
->ctx_lock
);
2033 /* TODO: use a hash or array, this sucks. */
2034 list_for_each_entry(kiocb
, &ctx
->active_reqs
, ki_list
) {
2035 if (kiocb
->ki_user_iocb
== iocb
) {
2036 ret
= kiocb
->ki_cancel(&kiocb
->rw
);
2037 list_del_init(&kiocb
->ki_list
);
2041 spin_unlock_irq(&ctx
->ctx_lock
);
2045 * The result argument is no longer used - the io_event is
2046 * always delivered via the ring buffer. -EINPROGRESS indicates
2047 * cancellation is progress:
2052 percpu_ref_put(&ctx
->users
);
2057 static long do_io_getevents(aio_context_t ctx_id
,
2060 struct io_event __user
*events
,
2061 struct timespec64
*ts
)
2063 ktime_t until
= ts
? timespec64_to_ktime(*ts
) : KTIME_MAX
;
2064 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
2067 if (likely(ioctx
)) {
2068 if (likely(min_nr
<= nr
&& min_nr
>= 0))
2069 ret
= read_events(ioctx
, min_nr
, nr
, events
, until
);
2070 percpu_ref_put(&ioctx
->users
);
2077 * Attempts to read at least min_nr events and up to nr events from
2078 * the completion queue for the aio_context specified by ctx_id. If
2079 * it succeeds, the number of read events is returned. May fail with
2080 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2081 * out of range, if timeout is out of range. May fail with -EFAULT
2082 * if any of the memory specified is invalid. May return 0 or
2083 * < min_nr if the timeout specified by timeout has elapsed
2084 * before sufficient events are available, where timeout == NULL
2085 * specifies an infinite timeout. Note that the timeout pointed to by
2086 * timeout is relative. Will fail with -ENOSYS if not implemented.
2088 #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2090 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
2093 struct io_event __user
*, events
,
2094 struct __kernel_timespec __user
*, timeout
)
2096 struct timespec64 ts
;
2099 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2102 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2103 if (!ret
&& signal_pending(current
))
2110 struct __aio_sigset
{
2111 const sigset_t __user
*sigmask
;
2115 SYSCALL_DEFINE6(io_pgetevents
,
2116 aio_context_t
, ctx_id
,
2119 struct io_event __user
*, events
,
2120 struct __kernel_timespec __user
*, timeout
,
2121 const struct __aio_sigset __user
*, usig
)
2123 struct __aio_sigset ksig
= { NULL
, };
2124 sigset_t ksigmask
, sigsaved
;
2125 struct timespec64 ts
;
2128 if (timeout
&& unlikely(get_timespec64(&ts
, timeout
)))
2131 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2134 ret
= set_user_sigmask(ksig
.sigmask
, &ksigmask
, &sigsaved
, ksig
.sigsetsize
);
2138 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2139 restore_user_sigmask(ksig
.sigmask
, &sigsaved
);
2140 if (signal_pending(current
) && !ret
)
2141 ret
= -ERESTARTNOHAND
;
2146 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2148 SYSCALL_DEFINE6(io_pgetevents_time32
,
2149 aio_context_t
, ctx_id
,
2152 struct io_event __user
*, events
,
2153 struct old_timespec32 __user
*, timeout
,
2154 const struct __aio_sigset __user
*, usig
)
2156 struct __aio_sigset ksig
= { NULL
, };
2157 sigset_t ksigmask
, sigsaved
;
2158 struct timespec64 ts
;
2161 if (timeout
&& unlikely(get_old_timespec32(&ts
, timeout
)))
2164 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2168 ret
= set_user_sigmask(ksig
.sigmask
, &ksigmask
, &sigsaved
, ksig
.sigsetsize
);
2172 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &ts
: NULL
);
2173 restore_user_sigmask(ksig
.sigmask
, &sigsaved
);
2174 if (signal_pending(current
) && !ret
)
2175 ret
= -ERESTARTNOHAND
;
2182 #if defined(CONFIG_COMPAT_32BIT_TIME)
2184 SYSCALL_DEFINE5(io_getevents_time32
, __u32
, ctx_id
,
2187 struct io_event __user
*, events
,
2188 struct old_timespec32 __user
*, timeout
)
2190 struct timespec64 t
;
2193 if (timeout
&& get_old_timespec32(&t
, timeout
))
2196 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2197 if (!ret
&& signal_pending(current
))
2204 #ifdef CONFIG_COMPAT
2206 struct __compat_aio_sigset
{
2207 compat_sigset_t __user
*sigmask
;
2208 compat_size_t sigsetsize
;
2211 #if defined(CONFIG_COMPAT_32BIT_TIME)
2213 COMPAT_SYSCALL_DEFINE6(io_pgetevents
,
2214 compat_aio_context_t
, ctx_id
,
2215 compat_long_t
, min_nr
,
2217 struct io_event __user
*, events
,
2218 struct old_timespec32 __user
*, timeout
,
2219 const struct __compat_aio_sigset __user
*, usig
)
2221 struct __compat_aio_sigset ksig
= { NULL
, };
2222 sigset_t ksigmask
, sigsaved
;
2223 struct timespec64 t
;
2226 if (timeout
&& get_old_timespec32(&t
, timeout
))
2229 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2232 ret
= set_compat_user_sigmask(ksig
.sigmask
, &ksigmask
, &sigsaved
, ksig
.sigsetsize
);
2236 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2237 restore_user_sigmask(ksig
.sigmask
, &sigsaved
);
2238 if (signal_pending(current
) && !ret
)
2239 ret
= -ERESTARTNOHAND
;
2246 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64
,
2247 compat_aio_context_t
, ctx_id
,
2248 compat_long_t
, min_nr
,
2250 struct io_event __user
*, events
,
2251 struct __kernel_timespec __user
*, timeout
,
2252 const struct __compat_aio_sigset __user
*, usig
)
2254 struct __compat_aio_sigset ksig
= { NULL
, };
2255 sigset_t ksigmask
, sigsaved
;
2256 struct timespec64 t
;
2259 if (timeout
&& get_timespec64(&t
, timeout
))
2262 if (usig
&& copy_from_user(&ksig
, usig
, sizeof(ksig
)))
2265 ret
= set_compat_user_sigmask(ksig
.sigmask
, &ksigmask
, &sigsaved
, ksig
.sigsetsize
);
2269 ret
= do_io_getevents(ctx_id
, min_nr
, nr
, events
, timeout
? &t
: NULL
);
2270 restore_user_sigmask(ksig
.sigmask
, &sigsaved
);
2271 if (signal_pending(current
) && !ret
)
2272 ret
= -ERESTARTNOHAND
;