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.
9 * See ../COPYING for licensing terms.
11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id
; /* kernel internal index number */
54 unsigned nr
; /* number of io_events */
55 unsigned head
; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
60 unsigned compat_features
;
61 unsigned incompat_features
;
62 unsigned header_length
; /* size of aio_ring */
65 struct io_event io_events
[0];
66 }; /* 128 bytes + ring size */
68 #define AIO_RING_PAGES 8
73 struct kioctx
*table
[];
77 unsigned reqs_available
;
81 struct percpu_ref users
;
84 struct percpu_ref reqs
;
86 unsigned long user_id
;
88 struct __percpu kioctx_cpu
*cpu
;
91 * For percpu reqs_available, number of slots we move to/from global
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
99 * The real limit is nr_events - 1, which will be larger (see
104 /* Size of ringbuffer, in units of struct io_event */
107 unsigned long mmap_base
;
108 unsigned long mmap_size
;
110 struct page
**ring_pages
;
113 struct work_struct free_work
;
116 * signals when all in-flight requests are done
118 struct completion
*requests_done
;
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
127 * We batch accesses to it with a percpu version.
129 atomic_t reqs_available
;
130 } ____cacheline_aligned_in_smp
;
134 struct list_head active_reqs
; /* used for cancellation */
135 } ____cacheline_aligned_in_smp
;
138 struct mutex ring_lock
;
139 wait_queue_head_t wait
;
140 } ____cacheline_aligned_in_smp
;
144 unsigned completed_events
;
145 spinlock_t completion_lock
;
146 } ____cacheline_aligned_in_smp
;
148 struct page
*internal_pages
[AIO_RING_PAGES
];
149 struct file
*aio_ring_file
;
155 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
156 * cancelled or completed (this makes a certain amount of sense because
157 * successful cancellation - io_cancel() - does deliver the completion to
160 * And since most things don't implement kiocb cancellation and we'd really like
161 * kiocb completion to be lockless when possible, we use ki_cancel to
162 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
163 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
165 #define KIOCB_CANCELLED ((void *) (~0ULL))
170 struct kioctx
*ki_ctx
;
171 kiocb_cancel_fn
*ki_cancel
;
173 struct iocb __user
*ki_user_iocb
; /* user's aiocb */
174 __u64 ki_user_data
; /* user's data for completion */
176 struct list_head ki_list
; /* the aio core uses this
177 * for cancellation */
180 * If the aio_resfd field of the userspace iocb is not zero,
181 * this is the underlying eventfd context to deliver events to.
183 struct eventfd_ctx
*ki_eventfd
;
186 /*------ sysctl variables----*/
187 static DEFINE_SPINLOCK(aio_nr_lock
);
188 unsigned long aio_nr
; /* current system wide number of aio requests */
189 unsigned long aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
190 /*----end sysctl variables---*/
192 static struct kmem_cache
*kiocb_cachep
;
193 static struct kmem_cache
*kioctx_cachep
;
195 static struct vfsmount
*aio_mnt
;
197 static const struct file_operations aio_ring_fops
;
198 static const struct address_space_operations aio_ctx_aops
;
200 static struct file
*aio_private_file(struct kioctx
*ctx
, loff_t nr_pages
)
202 struct qstr
this = QSTR_INIT("[aio]", 5);
205 struct inode
*inode
= alloc_anon_inode(aio_mnt
->mnt_sb
);
207 return ERR_CAST(inode
);
209 inode
->i_mapping
->a_ops
= &aio_ctx_aops
;
210 inode
->i_mapping
->private_data
= ctx
;
211 inode
->i_size
= PAGE_SIZE
* nr_pages
;
213 path
.dentry
= d_alloc_pseudo(aio_mnt
->mnt_sb
, &this);
216 return ERR_PTR(-ENOMEM
);
218 path
.mnt
= mntget(aio_mnt
);
220 d_instantiate(path
.dentry
, inode
);
221 file
= alloc_file(&path
, FMODE_READ
| FMODE_WRITE
, &aio_ring_fops
);
227 file
->f_flags
= O_RDWR
;
231 static struct dentry
*aio_mount(struct file_system_type
*fs_type
,
232 int flags
, const char *dev_name
, void *data
)
234 static const struct dentry_operations ops
= {
235 .d_dname
= simple_dname
,
237 return mount_pseudo(fs_type
, "aio:", NULL
, &ops
, AIO_RING_MAGIC
);
241 * Creates the slab caches used by the aio routines, panic on
242 * failure as this is done early during the boot sequence.
244 static int __init
aio_setup(void)
246 static struct file_system_type aio_fs
= {
249 .kill_sb
= kill_anon_super
,
251 aio_mnt
= kern_mount(&aio_fs
);
253 panic("Failed to create aio fs mount.");
255 kiocb_cachep
= KMEM_CACHE(aio_kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
256 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
258 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page
));
262 __initcall(aio_setup
);
264 static void put_aio_ring_file(struct kioctx
*ctx
)
266 struct file
*aio_ring_file
= ctx
->aio_ring_file
;
268 truncate_setsize(aio_ring_file
->f_inode
, 0);
270 /* Prevent further access to the kioctx from migratepages */
271 spin_lock(&aio_ring_file
->f_inode
->i_mapping
->private_lock
);
272 aio_ring_file
->f_inode
->i_mapping
->private_data
= NULL
;
273 ctx
->aio_ring_file
= NULL
;
274 spin_unlock(&aio_ring_file
->f_inode
->i_mapping
->private_lock
);
280 static void aio_free_ring(struct kioctx
*ctx
)
284 /* Disconnect the kiotx from the ring file. This prevents future
285 * accesses to the kioctx from page migration.
287 put_aio_ring_file(ctx
);
289 for (i
= 0; i
< ctx
->nr_pages
; i
++) {
291 pr_debug("pid(%d) [%d] page->count=%d\n", current
->pid
, i
,
292 page_count(ctx
->ring_pages
[i
]));
293 page
= ctx
->ring_pages
[i
];
296 ctx
->ring_pages
[i
] = NULL
;
300 if (ctx
->ring_pages
&& ctx
->ring_pages
!= ctx
->internal_pages
) {
301 kfree(ctx
->ring_pages
);
302 ctx
->ring_pages
= NULL
;
306 static int aio_ring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
308 vma
->vm_flags
|= VM_DONTEXPAND
;
309 vma
->vm_ops
= &generic_file_vm_ops
;
313 static void aio_ring_remap(struct file
*file
, struct vm_area_struct
*vma
)
315 struct mm_struct
*mm
= vma
->vm_mm
;
316 struct kioctx_table
*table
;
319 spin_lock(&mm
->ioctx_lock
);
321 table
= rcu_dereference(mm
->ioctx_table
);
322 for (i
= 0; i
< table
->nr
; i
++) {
325 ctx
= table
->table
[i
];
326 if (ctx
&& ctx
->aio_ring_file
== file
) {
327 ctx
->user_id
= ctx
->mmap_base
= vma
->vm_start
;
333 spin_unlock(&mm
->ioctx_lock
);
336 static const struct file_operations aio_ring_fops
= {
337 .mmap
= aio_ring_mmap
,
338 .mremap
= aio_ring_remap
,
341 #if IS_ENABLED(CONFIG_MIGRATION)
342 static int aio_migratepage(struct address_space
*mapping
, struct page
*new,
343 struct page
*old
, enum migrate_mode mode
)
352 /* mapping->private_lock here protects against the kioctx teardown. */
353 spin_lock(&mapping
->private_lock
);
354 ctx
= mapping
->private_data
;
360 /* The ring_lock mutex. The prevents aio_read_events() from writing
361 * to the ring's head, and prevents page migration from mucking in
362 * a partially initialized kiotx.
364 if (!mutex_trylock(&ctx
->ring_lock
)) {
370 if (idx
< (pgoff_t
)ctx
->nr_pages
) {
371 /* Make sure the old page hasn't already been changed */
372 if (ctx
->ring_pages
[idx
] != old
)
380 /* Writeback must be complete */
381 BUG_ON(PageWriteback(old
));
384 rc
= migrate_page_move_mapping(mapping
, new, old
, NULL
, mode
, 1);
385 if (rc
!= MIGRATEPAGE_SUCCESS
) {
390 /* Take completion_lock to prevent other writes to the ring buffer
391 * while the old page is copied to the new. This prevents new
392 * events from being lost.
394 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
395 migrate_page_copy(new, old
);
396 BUG_ON(ctx
->ring_pages
[idx
] != old
);
397 ctx
->ring_pages
[idx
] = new;
398 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
400 /* The old page is no longer accessible. */
404 mutex_unlock(&ctx
->ring_lock
);
406 spin_unlock(&mapping
->private_lock
);
411 static const struct address_space_operations aio_ctx_aops
= {
412 .set_page_dirty
= __set_page_dirty_no_writeback
,
413 #if IS_ENABLED(CONFIG_MIGRATION)
414 .migratepage
= aio_migratepage
,
418 static int aio_setup_ring(struct kioctx
*ctx
)
420 struct aio_ring
*ring
;
421 unsigned nr_events
= ctx
->max_reqs
;
422 struct mm_struct
*mm
= current
->mm
;
423 unsigned long size
, unused
;
428 /* Compensate for the ring buffer's head/tail overlap entry */
429 nr_events
+= 2; /* 1 is required, 2 for good luck */
431 size
= sizeof(struct aio_ring
);
432 size
+= sizeof(struct io_event
) * nr_events
;
434 nr_pages
= PFN_UP(size
);
438 file
= aio_private_file(ctx
, nr_pages
);
440 ctx
->aio_ring_file
= NULL
;
444 ctx
->aio_ring_file
= file
;
445 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
))
446 / sizeof(struct io_event
);
448 ctx
->ring_pages
= ctx
->internal_pages
;
449 if (nr_pages
> AIO_RING_PAGES
) {
450 ctx
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
452 if (!ctx
->ring_pages
) {
453 put_aio_ring_file(ctx
);
458 for (i
= 0; i
< nr_pages
; i
++) {
460 page
= find_or_create_page(file
->f_inode
->i_mapping
,
461 i
, GFP_HIGHUSER
| __GFP_ZERO
);
464 pr_debug("pid(%d) page[%d]->count=%d\n",
465 current
->pid
, i
, page_count(page
));
466 SetPageUptodate(page
);
469 ctx
->ring_pages
[i
] = page
;
473 if (unlikely(i
!= nr_pages
)) {
478 ctx
->mmap_size
= nr_pages
* PAGE_SIZE
;
479 pr_debug("attempting mmap of %lu bytes\n", ctx
->mmap_size
);
481 down_write(&mm
->mmap_sem
);
482 ctx
->mmap_base
= do_mmap_pgoff(ctx
->aio_ring_file
, 0, ctx
->mmap_size
,
483 PROT_READ
| PROT_WRITE
,
484 MAP_SHARED
, 0, &unused
);
485 up_write(&mm
->mmap_sem
);
486 if (IS_ERR((void *)ctx
->mmap_base
)) {
492 pr_debug("mmap address: 0x%08lx\n", ctx
->mmap_base
);
494 ctx
->user_id
= ctx
->mmap_base
;
495 ctx
->nr_events
= nr_events
; /* trusted copy */
497 ring
= kmap_atomic(ctx
->ring_pages
[0]);
498 ring
->nr
= nr_events
; /* user copy */
500 ring
->head
= ring
->tail
= 0;
501 ring
->magic
= AIO_RING_MAGIC
;
502 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
503 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
504 ring
->header_length
= sizeof(struct aio_ring
);
506 flush_dcache_page(ctx
->ring_pages
[0]);
511 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
512 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
513 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
515 void kiocb_set_cancel_fn(struct kiocb
*iocb
, kiocb_cancel_fn
*cancel
)
517 struct aio_kiocb
*req
= container_of(iocb
, struct aio_kiocb
, common
);
518 struct kioctx
*ctx
= req
->ki_ctx
;
521 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
523 if (!req
->ki_list
.next
)
524 list_add(&req
->ki_list
, &ctx
->active_reqs
);
526 req
->ki_cancel
= cancel
;
528 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
530 EXPORT_SYMBOL(kiocb_set_cancel_fn
);
532 static int kiocb_cancel(struct aio_kiocb
*kiocb
)
534 kiocb_cancel_fn
*old
, *cancel
;
537 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
538 * actually has a cancel function, hence the cmpxchg()
541 cancel
= ACCESS_ONCE(kiocb
->ki_cancel
);
543 if (!cancel
|| cancel
== KIOCB_CANCELLED
)
547 cancel
= cmpxchg(&kiocb
->ki_cancel
, old
, KIOCB_CANCELLED
);
548 } while (cancel
!= old
);
550 return cancel(&kiocb
->common
);
553 static void free_ioctx(struct work_struct
*work
)
555 struct kioctx
*ctx
= container_of(work
, struct kioctx
, free_work
);
557 pr_debug("freeing %p\n", ctx
);
560 free_percpu(ctx
->cpu
);
561 percpu_ref_exit(&ctx
->reqs
);
562 percpu_ref_exit(&ctx
->users
);
563 kmem_cache_free(kioctx_cachep
, ctx
);
566 static void free_ioctx_reqs(struct percpu_ref
*ref
)
568 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, reqs
);
570 /* At this point we know that there are no any in-flight requests */
571 if (ctx
->requests_done
)
572 complete(ctx
->requests_done
);
574 INIT_WORK(&ctx
->free_work
, free_ioctx
);
575 schedule_work(&ctx
->free_work
);
579 * When this function runs, the kioctx has been removed from the "hash table"
580 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
581 * now it's safe to cancel any that need to be.
583 static void free_ioctx_users(struct percpu_ref
*ref
)
585 struct kioctx
*ctx
= container_of(ref
, struct kioctx
, users
);
586 struct aio_kiocb
*req
;
588 spin_lock_irq(&ctx
->ctx_lock
);
590 while (!list_empty(&ctx
->active_reqs
)) {
591 req
= list_first_entry(&ctx
->active_reqs
,
592 struct aio_kiocb
, ki_list
);
594 list_del_init(&req
->ki_list
);
598 spin_unlock_irq(&ctx
->ctx_lock
);
600 percpu_ref_kill(&ctx
->reqs
);
601 percpu_ref_put(&ctx
->reqs
);
604 static int ioctx_add_table(struct kioctx
*ctx
, struct mm_struct
*mm
)
607 struct kioctx_table
*table
, *old
;
608 struct aio_ring
*ring
;
610 spin_lock(&mm
->ioctx_lock
);
611 table
= rcu_dereference_raw(mm
->ioctx_table
);
615 for (i
= 0; i
< table
->nr
; i
++)
616 if (!table
->table
[i
]) {
618 table
->table
[i
] = ctx
;
619 spin_unlock(&mm
->ioctx_lock
);
621 /* While kioctx setup is in progress,
622 * we are protected from page migration
623 * changes ring_pages by ->ring_lock.
625 ring
= kmap_atomic(ctx
->ring_pages
[0]);
631 new_nr
= (table
? table
->nr
: 1) * 4;
632 spin_unlock(&mm
->ioctx_lock
);
634 table
= kzalloc(sizeof(*table
) + sizeof(struct kioctx
*) *
641 spin_lock(&mm
->ioctx_lock
);
642 old
= rcu_dereference_raw(mm
->ioctx_table
);
645 rcu_assign_pointer(mm
->ioctx_table
, table
);
646 } else if (table
->nr
> old
->nr
) {
647 memcpy(table
->table
, old
->table
,
648 old
->nr
* sizeof(struct kioctx
*));
650 rcu_assign_pointer(mm
->ioctx_table
, table
);
659 static void aio_nr_sub(unsigned nr
)
661 spin_lock(&aio_nr_lock
);
662 if (WARN_ON(aio_nr
- nr
> aio_nr
))
666 spin_unlock(&aio_nr_lock
);
670 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
672 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
674 struct mm_struct
*mm
= current
->mm
;
679 * We keep track of the number of available ringbuffer slots, to prevent
680 * overflow (reqs_available), and we also use percpu counters for this.
682 * So since up to half the slots might be on other cpu's percpu counters
683 * and unavailable, double nr_events so userspace sees what they
684 * expected: additionally, we move req_batch slots to/from percpu
685 * counters at a time, so make sure that isn't 0:
687 nr_events
= max(nr_events
, num_possible_cpus() * 4);
690 /* Prevent overflows */
691 if ((nr_events
> (0x10000000U
/ sizeof(struct io_event
))) ||
692 (nr_events
> (0x10000000U
/ sizeof(struct kiocb
)))) {
693 pr_debug("ENOMEM: nr_events too high\n");
694 return ERR_PTR(-EINVAL
);
697 if (!nr_events
|| (unsigned long)nr_events
> (aio_max_nr
* 2UL))
698 return ERR_PTR(-EAGAIN
);
700 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
702 return ERR_PTR(-ENOMEM
);
704 ctx
->max_reqs
= nr_events
;
706 spin_lock_init(&ctx
->ctx_lock
);
707 spin_lock_init(&ctx
->completion_lock
);
708 mutex_init(&ctx
->ring_lock
);
709 /* Protect against page migration throughout kiotx setup by keeping
710 * the ring_lock mutex held until setup is complete. */
711 mutex_lock(&ctx
->ring_lock
);
712 init_waitqueue_head(&ctx
->wait
);
714 INIT_LIST_HEAD(&ctx
->active_reqs
);
716 if (percpu_ref_init(&ctx
->users
, free_ioctx_users
, 0, GFP_KERNEL
))
719 if (percpu_ref_init(&ctx
->reqs
, free_ioctx_reqs
, 0, GFP_KERNEL
))
722 ctx
->cpu
= alloc_percpu(struct kioctx_cpu
);
726 err
= aio_setup_ring(ctx
);
730 atomic_set(&ctx
->reqs_available
, ctx
->nr_events
- 1);
731 ctx
->req_batch
= (ctx
->nr_events
- 1) / (num_possible_cpus() * 4);
732 if (ctx
->req_batch
< 1)
735 /* limit the number of system wide aios */
736 spin_lock(&aio_nr_lock
);
737 if (aio_nr
+ nr_events
> (aio_max_nr
* 2UL) ||
738 aio_nr
+ nr_events
< aio_nr
) {
739 spin_unlock(&aio_nr_lock
);
743 aio_nr
+= ctx
->max_reqs
;
744 spin_unlock(&aio_nr_lock
);
746 percpu_ref_get(&ctx
->users
); /* io_setup() will drop this ref */
747 percpu_ref_get(&ctx
->reqs
); /* free_ioctx_users() will drop this */
749 err
= ioctx_add_table(ctx
, mm
);
753 /* Release the ring_lock mutex now that all setup is complete. */
754 mutex_unlock(&ctx
->ring_lock
);
756 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
757 ctx
, ctx
->user_id
, mm
, ctx
->nr_events
);
761 aio_nr_sub(ctx
->max_reqs
);
765 mutex_unlock(&ctx
->ring_lock
);
766 free_percpu(ctx
->cpu
);
767 percpu_ref_exit(&ctx
->reqs
);
768 percpu_ref_exit(&ctx
->users
);
769 kmem_cache_free(kioctx_cachep
, ctx
);
770 pr_debug("error allocating ioctx %d\n", err
);
775 * Cancels all outstanding aio requests on an aio context. Used
776 * when the processes owning a context have all exited to encourage
777 * the rapid destruction of the kioctx.
779 static int kill_ioctx(struct mm_struct
*mm
, struct kioctx
*ctx
,
780 struct completion
*requests_done
)
782 struct kioctx_table
*table
;
784 if (atomic_xchg(&ctx
->dead
, 1))
788 spin_lock(&mm
->ioctx_lock
);
789 table
= rcu_dereference_raw(mm
->ioctx_table
);
790 WARN_ON(ctx
!= table
->table
[ctx
->id
]);
791 table
->table
[ctx
->id
] = NULL
;
792 spin_unlock(&mm
->ioctx_lock
);
794 /* percpu_ref_kill() will do the necessary call_rcu() */
795 wake_up_all(&ctx
->wait
);
798 * It'd be more correct to do this in free_ioctx(), after all
799 * the outstanding kiocbs have finished - but by then io_destroy
800 * has already returned, so io_setup() could potentially return
801 * -EAGAIN with no ioctxs actually in use (as far as userspace
804 aio_nr_sub(ctx
->max_reqs
);
807 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
809 ctx
->requests_done
= requests_done
;
810 percpu_ref_kill(&ctx
->users
);
815 * exit_aio: called when the last user of mm goes away. At this point, there is
816 * no way for any new requests to be submited or any of the io_* syscalls to be
817 * called on the context.
819 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
822 void exit_aio(struct mm_struct
*mm
)
824 struct kioctx_table
*table
= rcu_dereference_raw(mm
->ioctx_table
);
830 for (i
= 0; i
< table
->nr
; ++i
) {
831 struct kioctx
*ctx
= table
->table
[i
];
832 struct completion requests_done
=
833 COMPLETION_INITIALIZER_ONSTACK(requests_done
);
838 * We don't need to bother with munmap() here - exit_mmap(mm)
839 * is coming and it'll unmap everything. And we simply can't,
840 * this is not necessarily our ->mm.
841 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
842 * that it needs to unmap the area, just set it to 0.
845 kill_ioctx(mm
, ctx
, &requests_done
);
847 /* Wait until all IO for the context are done. */
848 wait_for_completion(&requests_done
);
851 RCU_INIT_POINTER(mm
->ioctx_table
, NULL
);
855 static void put_reqs_available(struct kioctx
*ctx
, unsigned nr
)
857 struct kioctx_cpu
*kcpu
;
860 local_irq_save(flags
);
861 kcpu
= this_cpu_ptr(ctx
->cpu
);
862 kcpu
->reqs_available
+= nr
;
864 while (kcpu
->reqs_available
>= ctx
->req_batch
* 2) {
865 kcpu
->reqs_available
-= ctx
->req_batch
;
866 atomic_add(ctx
->req_batch
, &ctx
->reqs_available
);
869 local_irq_restore(flags
);
872 static bool get_reqs_available(struct kioctx
*ctx
)
874 struct kioctx_cpu
*kcpu
;
878 local_irq_save(flags
);
879 kcpu
= this_cpu_ptr(ctx
->cpu
);
880 if (!kcpu
->reqs_available
) {
881 int old
, avail
= atomic_read(&ctx
->reqs_available
);
884 if (avail
< ctx
->req_batch
)
888 avail
= atomic_cmpxchg(&ctx
->reqs_available
,
889 avail
, avail
- ctx
->req_batch
);
890 } while (avail
!= old
);
892 kcpu
->reqs_available
+= ctx
->req_batch
;
896 kcpu
->reqs_available
--;
898 local_irq_restore(flags
);
902 /* refill_reqs_available
903 * Updates the reqs_available reference counts used for tracking the
904 * number of free slots in the completion ring. This can be called
905 * from aio_complete() (to optimistically update reqs_available) or
906 * from aio_get_req() (the we're out of events case). It must be
907 * called holding ctx->completion_lock.
909 static void refill_reqs_available(struct kioctx
*ctx
, unsigned head
,
912 unsigned events_in_ring
, completed
;
914 /* Clamp head since userland can write to it. */
915 head
%= ctx
->nr_events
;
917 events_in_ring
= tail
- head
;
919 events_in_ring
= ctx
->nr_events
- (head
- tail
);
921 completed
= ctx
->completed_events
;
922 if (events_in_ring
< completed
)
923 completed
-= events_in_ring
;
930 ctx
->completed_events
-= completed
;
931 put_reqs_available(ctx
, completed
);
934 /* user_refill_reqs_available
935 * Called to refill reqs_available when aio_get_req() encounters an
936 * out of space in the completion ring.
938 static void user_refill_reqs_available(struct kioctx
*ctx
)
940 spin_lock_irq(&ctx
->completion_lock
);
941 if (ctx
->completed_events
) {
942 struct aio_ring
*ring
;
945 /* Access of ring->head may race with aio_read_events_ring()
946 * here, but that's okay since whether we read the old version
947 * or the new version, and either will be valid. The important
948 * part is that head cannot pass tail since we prevent
949 * aio_complete() from updating tail by holding
950 * ctx->completion_lock. Even if head is invalid, the check
951 * against ctx->completed_events below will make sure we do the
954 ring
= kmap_atomic(ctx
->ring_pages
[0]);
958 refill_reqs_available(ctx
, head
, ctx
->tail
);
961 spin_unlock_irq(&ctx
->completion_lock
);
965 * Allocate a slot for an aio request.
966 * Returns NULL if no requests are free.
968 static inline struct aio_kiocb
*aio_get_req(struct kioctx
*ctx
)
970 struct aio_kiocb
*req
;
972 if (!get_reqs_available(ctx
)) {
973 user_refill_reqs_available(ctx
);
974 if (!get_reqs_available(ctx
))
978 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
|__GFP_ZERO
);
982 percpu_ref_get(&ctx
->reqs
);
987 put_reqs_available(ctx
, 1);
991 static void kiocb_free(struct aio_kiocb
*req
)
993 if (req
->common
.ki_filp
)
994 fput(req
->common
.ki_filp
);
995 if (req
->ki_eventfd
!= NULL
)
996 eventfd_ctx_put(req
->ki_eventfd
);
997 kmem_cache_free(kiocb_cachep
, req
);
1000 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
1002 struct aio_ring __user
*ring
= (void __user
*)ctx_id
;
1003 struct mm_struct
*mm
= current
->mm
;
1004 struct kioctx
*ctx
, *ret
= NULL
;
1005 struct kioctx_table
*table
;
1008 if (get_user(id
, &ring
->id
))
1012 table
= rcu_dereference(mm
->ioctx_table
);
1014 if (!table
|| id
>= table
->nr
)
1017 ctx
= table
->table
[id
];
1018 if (ctx
&& ctx
->user_id
== ctx_id
) {
1019 percpu_ref_get(&ctx
->users
);
1028 * Called when the io request on the given iocb is complete.
1030 static void aio_complete(struct kiocb
*kiocb
, long res
, long res2
)
1032 struct aio_kiocb
*iocb
= container_of(kiocb
, struct aio_kiocb
, common
);
1033 struct kioctx
*ctx
= iocb
->ki_ctx
;
1034 struct aio_ring
*ring
;
1035 struct io_event
*ev_page
, *event
;
1036 unsigned tail
, pos
, head
;
1037 unsigned long flags
;
1040 * Special case handling for sync iocbs:
1041 * - events go directly into the iocb for fast handling
1042 * - the sync task with the iocb in its stack holds the single iocb
1043 * ref, no other paths have a way to get another ref
1044 * - the sync task helpfully left a reference to itself in the iocb
1046 BUG_ON(is_sync_kiocb(kiocb
));
1048 if (iocb
->ki_list
.next
) {
1049 unsigned long flags
;
1051 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
1052 list_del(&iocb
->ki_list
);
1053 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
1057 * Add a completion event to the ring buffer. Must be done holding
1058 * ctx->completion_lock to prevent other code from messing with the tail
1059 * pointer since we might be called from irq context.
1061 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
1064 pos
= tail
+ AIO_EVENTS_OFFSET
;
1066 if (++tail
>= ctx
->nr_events
)
1069 ev_page
= kmap_atomic(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1070 event
= ev_page
+ pos
% AIO_EVENTS_PER_PAGE
;
1072 event
->obj
= (u64
)(unsigned long)iocb
->ki_user_iocb
;
1073 event
->data
= iocb
->ki_user_data
;
1077 kunmap_atomic(ev_page
);
1078 flush_dcache_page(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
1080 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1081 ctx
, tail
, iocb
, iocb
->ki_user_iocb
, iocb
->ki_user_data
,
1084 /* after flagging the request as done, we
1085 * must never even look at it again
1087 smp_wmb(); /* make event visible before updating tail */
1091 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1094 kunmap_atomic(ring
);
1095 flush_dcache_page(ctx
->ring_pages
[0]);
1097 ctx
->completed_events
++;
1098 if (ctx
->completed_events
> 1)
1099 refill_reqs_available(ctx
, head
, tail
);
1100 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
1102 pr_debug("added to ring %p at [%u]\n", iocb
, tail
);
1105 * Check if the user asked us to deliver the result through an
1106 * eventfd. The eventfd_signal() function is safe to be called
1109 if (iocb
->ki_eventfd
!= NULL
)
1110 eventfd_signal(iocb
->ki_eventfd
, 1);
1112 /* everything turned out well, dispose of the aiocb. */
1116 * We have to order our ring_info tail store above and test
1117 * of the wait list below outside the wait lock. This is
1118 * like in wake_up_bit() where clearing a bit has to be
1119 * ordered with the unlocked test.
1123 if (waitqueue_active(&ctx
->wait
))
1124 wake_up(&ctx
->wait
);
1126 percpu_ref_put(&ctx
->reqs
);
1129 /* aio_read_events_ring
1130 * Pull an event off of the ioctx's event ring. Returns the number of
1133 static long aio_read_events_ring(struct kioctx
*ctx
,
1134 struct io_event __user
*event
, long nr
)
1136 struct aio_ring
*ring
;
1137 unsigned head
, tail
, pos
;
1142 * The mutex can block and wake us up and that will cause
1143 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1144 * and repeat. This should be rare enough that it doesn't cause
1145 * peformance issues. See the comment in read_events() for more detail.
1147 sched_annotate_sleep();
1148 mutex_lock(&ctx
->ring_lock
);
1150 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1151 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1154 kunmap_atomic(ring
);
1157 * Ensure that once we've read the current tail pointer, that
1158 * we also see the events that were stored up to the tail.
1162 pr_debug("h%u t%u m%u\n", head
, tail
, ctx
->nr_events
);
1167 head
%= ctx
->nr_events
;
1168 tail
%= ctx
->nr_events
;
1172 struct io_event
*ev
;
1175 avail
= (head
<= tail
? tail
: ctx
->nr_events
) - head
;
1179 avail
= min(avail
, nr
- ret
);
1180 avail
= min_t(long, avail
, AIO_EVENTS_PER_PAGE
-
1181 ((head
+ AIO_EVENTS_OFFSET
) % AIO_EVENTS_PER_PAGE
));
1183 pos
= head
+ AIO_EVENTS_OFFSET
;
1184 page
= ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
];
1185 pos
%= AIO_EVENTS_PER_PAGE
;
1188 copy_ret
= copy_to_user(event
+ ret
, ev
+ pos
,
1189 sizeof(*ev
) * avail
);
1192 if (unlikely(copy_ret
)) {
1199 head
%= ctx
->nr_events
;
1202 ring
= kmap_atomic(ctx
->ring_pages
[0]);
1204 kunmap_atomic(ring
);
1205 flush_dcache_page(ctx
->ring_pages
[0]);
1207 pr_debug("%li h%u t%u\n", ret
, head
, tail
);
1209 mutex_unlock(&ctx
->ring_lock
);
1214 static bool aio_read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1215 struct io_event __user
*event
, long *i
)
1217 long ret
= aio_read_events_ring(ctx
, event
+ *i
, nr
- *i
);
1222 if (unlikely(atomic_read(&ctx
->dead
)))
1228 return ret
< 0 || *i
>= min_nr
;
1231 static long read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
1232 struct io_event __user
*event
,
1233 struct timespec __user
*timeout
)
1235 ktime_t until
= { .tv64
= KTIME_MAX
};
1241 if (unlikely(copy_from_user(&ts
, timeout
, sizeof(ts
))))
1244 until
= timespec_to_ktime(ts
);
1248 * Note that aio_read_events() is being called as the conditional - i.e.
1249 * we're calling it after prepare_to_wait() has set task state to
1250 * TASK_INTERRUPTIBLE.
1252 * But aio_read_events() can block, and if it blocks it's going to flip
1253 * the task state back to TASK_RUNNING.
1255 * This should be ok, provided it doesn't flip the state back to
1256 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1257 * will only happen if the mutex_lock() call blocks, and we then find
1258 * the ringbuffer empty. So in practice we should be ok, but it's
1259 * something to be aware of when touching this code.
1261 if (until
.tv64
== 0)
1262 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
);
1264 wait_event_interruptible_hrtimeout(ctx
->wait
,
1265 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
),
1268 if (!ret
&& signal_pending(current
))
1275 * Create an aio_context capable of receiving at least nr_events.
1276 * ctxp must not point to an aio_context that already exists, and
1277 * must be initialized to 0 prior to the call. On successful
1278 * creation of the aio_context, *ctxp is filled in with the resulting
1279 * handle. May fail with -EINVAL if *ctxp is not initialized,
1280 * if the specified nr_events exceeds internal limits. May fail
1281 * with -EAGAIN if the specified nr_events exceeds the user's limit
1282 * of available events. May fail with -ENOMEM if insufficient kernel
1283 * resources are available. May fail with -EFAULT if an invalid
1284 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1287 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
1289 struct kioctx
*ioctx
= NULL
;
1293 ret
= get_user(ctx
, ctxp
);
1298 if (unlikely(ctx
|| nr_events
== 0)) {
1299 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1304 ioctx
= ioctx_alloc(nr_events
);
1305 ret
= PTR_ERR(ioctx
);
1306 if (!IS_ERR(ioctx
)) {
1307 ret
= put_user(ioctx
->user_id
, ctxp
);
1309 kill_ioctx(current
->mm
, ioctx
, NULL
);
1310 percpu_ref_put(&ioctx
->users
);
1318 * Destroy the aio_context specified. May cancel any outstanding
1319 * AIOs and block on completion. Will fail with -ENOSYS if not
1320 * implemented. May fail with -EINVAL if the context pointed to
1323 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1325 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1326 if (likely(NULL
!= ioctx
)) {
1327 struct completion requests_done
=
1328 COMPLETION_INITIALIZER_ONSTACK(requests_done
);
1331 /* Pass requests_done to kill_ioctx() where it can be set
1332 * in a thread-safe way. If we try to set it here then we have
1333 * a race condition if two io_destroy() called simultaneously.
1335 ret
= kill_ioctx(current
->mm
, ioctx
, &requests_done
);
1336 percpu_ref_put(&ioctx
->users
);
1338 /* Wait until all IO for the context are done. Otherwise kernel
1339 * keep using user-space buffers even if user thinks the context
1343 wait_for_completion(&requests_done
);
1347 pr_debug("EINVAL: invalid context id\n");
1351 typedef ssize_t (aio_rw_op
)(struct kiocb
*, const struct iovec
*,
1352 unsigned long, loff_t
);
1353 typedef ssize_t (rw_iter_op
)(struct kiocb
*, struct iov_iter
*);
1355 static ssize_t
aio_setup_vectored_rw(struct kiocb
*kiocb
,
1356 int rw
, char __user
*buf
,
1357 unsigned long *nr_segs
,
1359 struct iovec
**iovec
,
1361 struct iov_iter
*iter
)
1367 #ifdef CONFIG_COMPAT
1369 ret
= compat_rw_copy_check_uvector(rw
,
1370 (struct compat_iovec __user
*)buf
,
1371 *nr_segs
, UIO_FASTIOV
, *iovec
, iovec
);
1374 ret
= rw_copy_check_uvector(rw
,
1375 (struct iovec __user
*)buf
,
1376 *nr_segs
, UIO_FASTIOV
, *iovec
, iovec
);
1380 /* len now reflect bytes instead of segs */
1382 iov_iter_init(iter
, rw
, *iovec
, *nr_segs
, *len
);
1386 static ssize_t
aio_setup_single_vector(struct kiocb
*kiocb
,
1387 int rw
, char __user
*buf
,
1388 unsigned long *nr_segs
,
1390 struct iovec
*iovec
,
1391 struct iov_iter
*iter
)
1393 if (len
> MAX_RW_COUNT
)
1395 if (unlikely(!access_ok(!rw
, buf
, len
)))
1398 iovec
->iov_base
= buf
;
1399 iovec
->iov_len
= len
;
1401 iov_iter_init(iter
, rw
, iovec
, *nr_segs
, len
);
1407 * Performs the initial checks and io submission.
1409 static ssize_t
aio_run_iocb(struct kiocb
*req
, unsigned opcode
,
1410 char __user
*buf
, size_t len
, bool compat
)
1412 struct file
*file
= req
->ki_filp
;
1414 unsigned long nr_segs
;
1418 rw_iter_op
*iter_op
;
1419 struct iovec inline_vecs
[UIO_FASTIOV
], *iovec
= inline_vecs
;
1420 struct iov_iter iter
;
1423 case IOCB_CMD_PREAD
:
1424 case IOCB_CMD_PREADV
:
1427 rw_op
= file
->f_op
->aio_read
;
1428 iter_op
= file
->f_op
->read_iter
;
1431 case IOCB_CMD_PWRITE
:
1432 case IOCB_CMD_PWRITEV
:
1435 rw_op
= file
->f_op
->aio_write
;
1436 iter_op
= file
->f_op
->write_iter
;
1439 if (unlikely(!(file
->f_mode
& mode
)))
1442 if (!rw_op
&& !iter_op
)
1445 if (opcode
== IOCB_CMD_PREADV
|| opcode
== IOCB_CMD_PWRITEV
)
1446 ret
= aio_setup_vectored_rw(req
, rw
, buf
, &nr_segs
,
1447 &len
, &iovec
, compat
, &iter
);
1449 ret
= aio_setup_single_vector(req
, rw
, buf
, &nr_segs
,
1452 ret
= rw_verify_area(rw
, file
, &req
->ki_pos
, len
);
1454 if (iovec
!= inline_vecs
)
1461 /* XXX: move/kill - rw_verify_area()? */
1462 /* This matches the pread()/pwrite() logic */
1463 if (req
->ki_pos
< 0) {
1469 file_start_write(file
);
1472 ret
= iter_op(req
, &iter
);
1474 ret
= rw_op(req
, iter
.iov
, iter
.nr_segs
, req
->ki_pos
);
1478 file_end_write(file
);
1481 case IOCB_CMD_FDSYNC
:
1482 if (!file
->f_op
->aio_fsync
)
1485 ret
= file
->f_op
->aio_fsync(req
, 1);
1488 case IOCB_CMD_FSYNC
:
1489 if (!file
->f_op
->aio_fsync
)
1492 ret
= file
->f_op
->aio_fsync(req
, 0);
1496 pr_debug("EINVAL: no operation provided\n");
1500 if (iovec
!= inline_vecs
)
1503 if (ret
!= -EIOCBQUEUED
) {
1505 * There's no easy way to restart the syscall since other AIO's
1506 * may be already running. Just fail this IO with EINTR.
1508 if (unlikely(ret
== -ERESTARTSYS
|| ret
== -ERESTARTNOINTR
||
1509 ret
== -ERESTARTNOHAND
||
1510 ret
== -ERESTART_RESTARTBLOCK
))
1512 aio_complete(req
, ret
, 0);
1518 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
1519 struct iocb
*iocb
, bool compat
)
1521 struct aio_kiocb
*req
;
1524 /* enforce forwards compatibility on users */
1525 if (unlikely(iocb
->aio_reserved1
|| iocb
->aio_reserved2
)) {
1526 pr_debug("EINVAL: reserve field set\n");
1530 /* prevent overflows */
1532 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1533 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1534 ((ssize_t
)iocb
->aio_nbytes
< 0)
1536 pr_debug("EINVAL: overflow check\n");
1540 req
= aio_get_req(ctx
);
1544 req
->common
.ki_filp
= fget(iocb
->aio_fildes
);
1545 if (unlikely(!req
->common
.ki_filp
)) {
1549 req
->common
.ki_pos
= iocb
->aio_offset
;
1550 req
->common
.ki_complete
= aio_complete
;
1551 req
->common
.ki_flags
= 0;
1553 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1555 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1556 * instance of the file* now. The file descriptor must be
1557 * an eventfd() fd, and will be signaled for each completed
1558 * event using the eventfd_signal() function.
1560 req
->ki_eventfd
= eventfd_ctx_fdget((int) iocb
->aio_resfd
);
1561 if (IS_ERR(req
->ki_eventfd
)) {
1562 ret
= PTR_ERR(req
->ki_eventfd
);
1563 req
->ki_eventfd
= NULL
;
1567 req
->common
.ki_flags
|= IOCB_EVENTFD
;
1570 ret
= put_user(KIOCB_KEY
, &user_iocb
->aio_key
);
1571 if (unlikely(ret
)) {
1572 pr_debug("EFAULT: aio_key\n");
1576 req
->ki_user_iocb
= user_iocb
;
1577 req
->ki_user_data
= iocb
->aio_data
;
1579 ret
= aio_run_iocb(&req
->common
, iocb
->aio_lio_opcode
,
1580 (char __user
*)(unsigned long)iocb
->aio_buf
,
1588 put_reqs_available(ctx
, 1);
1589 percpu_ref_put(&ctx
->reqs
);
1594 long do_io_submit(aio_context_t ctx_id
, long nr
,
1595 struct iocb __user
*__user
*iocbpp
, bool compat
)
1600 struct blk_plug plug
;
1602 if (unlikely(nr
< 0))
1605 if (unlikely(nr
> LONG_MAX
/sizeof(*iocbpp
)))
1606 nr
= LONG_MAX
/sizeof(*iocbpp
);
1608 if (unlikely(!access_ok(VERIFY_READ
, iocbpp
, (nr
*sizeof(*iocbpp
)))))
1611 ctx
= lookup_ioctx(ctx_id
);
1612 if (unlikely(!ctx
)) {
1613 pr_debug("EINVAL: invalid context id\n");
1617 blk_start_plug(&plug
);
1620 * AKPM: should this return a partial result if some of the IOs were
1621 * successfully submitted?
1623 for (i
=0; i
<nr
; i
++) {
1624 struct iocb __user
*user_iocb
;
1627 if (unlikely(__get_user(user_iocb
, iocbpp
+ i
))) {
1632 if (unlikely(copy_from_user(&tmp
, user_iocb
, sizeof(tmp
)))) {
1637 ret
= io_submit_one(ctx
, user_iocb
, &tmp
, compat
);
1641 blk_finish_plug(&plug
);
1643 percpu_ref_put(&ctx
->users
);
1648 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1649 * the number of iocbs queued. May return -EINVAL if the aio_context
1650 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1651 * *iocbpp[0] is not properly initialized, if the operation specified
1652 * is invalid for the file descriptor in the iocb. May fail with
1653 * -EFAULT if any of the data structures point to invalid data. May
1654 * fail with -EBADF if the file descriptor specified in the first
1655 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1656 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1657 * fail with -ENOSYS if not implemented.
1659 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
1660 struct iocb __user
* __user
*, iocbpp
)
1662 return do_io_submit(ctx_id
, nr
, iocbpp
, 0);
1666 * Finds a given iocb for cancellation.
1668 static struct aio_kiocb
*
1669 lookup_kiocb(struct kioctx
*ctx
, struct iocb __user
*iocb
, u32 key
)
1671 struct aio_kiocb
*kiocb
;
1673 assert_spin_locked(&ctx
->ctx_lock
);
1675 if (key
!= KIOCB_KEY
)
1678 /* TODO: use a hash or array, this sucks. */
1679 list_for_each_entry(kiocb
, &ctx
->active_reqs
, ki_list
) {
1680 if (kiocb
->ki_user_iocb
== iocb
)
1687 * Attempts to cancel an iocb previously passed to io_submit. If
1688 * the operation is successfully cancelled, the resulting event is
1689 * copied into the memory pointed to by result without being placed
1690 * into the completion queue and 0 is returned. May fail with
1691 * -EFAULT if any of the data structures pointed to are invalid.
1692 * May fail with -EINVAL if aio_context specified by ctx_id is
1693 * invalid. May fail with -EAGAIN if the iocb specified was not
1694 * cancelled. Will fail with -ENOSYS if not implemented.
1696 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
1697 struct io_event __user
*, result
)
1700 struct aio_kiocb
*kiocb
;
1704 ret
= get_user(key
, &iocb
->aio_key
);
1708 ctx
= lookup_ioctx(ctx_id
);
1712 spin_lock_irq(&ctx
->ctx_lock
);
1714 kiocb
= lookup_kiocb(ctx
, iocb
, key
);
1716 ret
= kiocb_cancel(kiocb
);
1720 spin_unlock_irq(&ctx
->ctx_lock
);
1724 * The result argument is no longer used - the io_event is
1725 * always delivered via the ring buffer. -EINPROGRESS indicates
1726 * cancellation is progress:
1731 percpu_ref_put(&ctx
->users
);
1737 * Attempts to read at least min_nr events and up to nr events from
1738 * the completion queue for the aio_context specified by ctx_id. If
1739 * it succeeds, the number of read events is returned. May fail with
1740 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1741 * out of range, if timeout is out of range. May fail with -EFAULT
1742 * if any of the memory specified is invalid. May return 0 or
1743 * < min_nr if the timeout specified by timeout has elapsed
1744 * before sufficient events are available, where timeout == NULL
1745 * specifies an infinite timeout. Note that the timeout pointed to by
1746 * timeout is relative. Will fail with -ENOSYS if not implemented.
1748 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
1751 struct io_event __user
*, events
,
1752 struct timespec __user
*, timeout
)
1754 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
1757 if (likely(ioctx
)) {
1758 if (likely(min_nr
<= nr
&& min_nr
>= 0))
1759 ret
= read_events(ioctx
, min_nr
, nr
, events
, timeout
);
1760 percpu_ref_put(&ioctx
->users
);