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/anon_inodes.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
43 #include <asm/kmap_types.h>
44 #include <asm/uaccess.h>
48 #define AIO_RING_MAGIC 0xa10a10a1
49 #define AIO_RING_COMPAT_FEATURES 1
50 #define AIO_RING_INCOMPAT_FEATURES 0
52 unsigned id
; /* kernel internal index number */
53 unsigned nr
; /* number of io_events */
58 unsigned compat_features
;
59 unsigned incompat_features
;
60 unsigned header_length
; /* size of aio_ring */
63 struct io_event io_events
[0];
64 }; /* 128 bytes + ring size */
66 #define AIO_RING_PAGES 8
69 unsigned reqs_available
;
76 /* This needs improving */
77 unsigned long user_id
;
78 struct hlist_node list
;
80 struct __percpu kioctx_cpu
*cpu
;
83 * For percpu reqs_available, number of slots we move to/from global
88 * This is what userspace passed to io_setup(), it's not used for
89 * anything but counting against the global max_reqs quota.
91 * The real limit is nr_events - 1, which will be larger (see
96 /* Size of ringbuffer, in units of struct io_event */
99 unsigned long mmap_base
;
100 unsigned long mmap_size
;
102 struct page
**ring_pages
;
105 struct rcu_head rcu_head
;
106 struct work_struct rcu_work
;
110 * This counts the number of available slots in the ringbuffer,
111 * so we avoid overflowing it: it's decremented (if positive)
112 * when allocating a kiocb and incremented when the resulting
113 * io_event is pulled off the ringbuffer.
115 * We batch accesses to it with a percpu version.
117 atomic_t reqs_available
;
118 } ____cacheline_aligned_in_smp
;
122 struct list_head active_reqs
; /* used for cancellation */
123 } ____cacheline_aligned_in_smp
;
126 struct mutex ring_lock
;
127 wait_queue_head_t wait
;
128 } ____cacheline_aligned_in_smp
;
132 spinlock_t completion_lock
;
133 } ____cacheline_aligned_in_smp
;
135 struct page
*internal_pages
[AIO_RING_PAGES
];
136 struct file
*aio_ring_file
;
139 /*------ sysctl variables----*/
140 static DEFINE_SPINLOCK(aio_nr_lock
);
141 unsigned long aio_nr
; /* current system wide number of aio requests */
142 unsigned long aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
143 /*----end sysctl variables---*/
145 static struct kmem_cache
*kiocb_cachep
;
146 static struct kmem_cache
*kioctx_cachep
;
149 * Creates the slab caches used by the aio routines, panic on
150 * failure as this is done early during the boot sequence.
152 static int __init
aio_setup(void)
154 kiocb_cachep
= KMEM_CACHE(kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
155 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
157 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page
));
161 __initcall(aio_setup
);
163 static void aio_free_ring(struct kioctx
*ctx
)
166 struct file
*aio_ring_file
= ctx
->aio_ring_file
;
168 for (i
= 0; i
< ctx
->nr_pages
; i
++) {
169 pr_debug("pid(%d) [%d] page->count=%d\n", current
->pid
, i
,
170 page_count(ctx
->ring_pages
[i
]));
171 put_page(ctx
->ring_pages
[i
]);
174 if (ctx
->ring_pages
&& ctx
->ring_pages
!= ctx
->internal_pages
)
175 kfree(ctx
->ring_pages
);
178 truncate_setsize(aio_ring_file
->f_inode
, 0);
179 pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n",
180 current
->pid
, aio_ring_file
->f_inode
->i_nlink
,
181 aio_ring_file
->f_path
.dentry
->d_count
,
182 d_unhashed(aio_ring_file
->f_path
.dentry
),
183 atomic_read(&aio_ring_file
->f_inode
->i_count
));
185 ctx
->aio_ring_file
= NULL
;
189 static int aio_ring_mmap(struct file
*file
, struct vm_area_struct
*vma
)
191 vma
->vm_ops
= &generic_file_vm_ops
;
195 static const struct file_operations aio_ring_fops
= {
196 .mmap
= aio_ring_mmap
,
199 static int aio_set_page_dirty(struct page
*page
)
204 #if IS_ENABLED(CONFIG_MIGRATION)
205 static int aio_migratepage(struct address_space
*mapping
, struct page
*new,
206 struct page
*old
, enum migrate_mode mode
)
208 struct kioctx
*ctx
= mapping
->private_data
;
210 unsigned idx
= old
->index
;
213 /* Writeback must be complete */
214 BUG_ON(PageWriteback(old
));
217 rc
= migrate_page_move_mapping(mapping
, new, old
, NULL
, mode
);
218 if (rc
!= MIGRATEPAGE_SUCCESS
) {
225 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
226 migrate_page_copy(new, old
);
227 ctx
->ring_pages
[idx
] = new;
228 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
234 static const struct address_space_operations aio_ctx_aops
= {
235 .set_page_dirty
= aio_set_page_dirty
,
236 #if IS_ENABLED(CONFIG_MIGRATION)
237 .migratepage
= aio_migratepage
,
241 static int aio_setup_ring(struct kioctx
*ctx
)
243 struct aio_ring
*ring
;
244 unsigned nr_events
= ctx
->max_reqs
;
245 struct mm_struct
*mm
= current
->mm
;
246 unsigned long size
, populate
;
251 /* Compensate for the ring buffer's head/tail overlap entry */
252 nr_events
+= 2; /* 1 is required, 2 for good luck */
254 size
= sizeof(struct aio_ring
);
255 size
+= sizeof(struct io_event
) * nr_events
;
257 nr_pages
= PFN_UP(size
);
261 file
= anon_inode_getfile_private("[aio]", &aio_ring_fops
, ctx
, O_RDWR
);
263 ctx
->aio_ring_file
= NULL
;
267 file
->f_inode
->i_mapping
->a_ops
= &aio_ctx_aops
;
268 file
->f_inode
->i_mapping
->private_data
= ctx
;
269 file
->f_inode
->i_size
= PAGE_SIZE
* (loff_t
)nr_pages
;
271 for (i
= 0; i
< nr_pages
; i
++) {
273 page
= find_or_create_page(file
->f_inode
->i_mapping
,
274 i
, GFP_HIGHUSER
| __GFP_ZERO
);
277 pr_debug("pid(%d) page[%d]->count=%d\n",
278 current
->pid
, i
, page_count(page
));
279 SetPageUptodate(page
);
283 ctx
->aio_ring_file
= file
;
284 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
))
285 / sizeof(struct io_event
);
287 ctx
->ring_pages
= ctx
->internal_pages
;
288 if (nr_pages
> AIO_RING_PAGES
) {
289 ctx
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*),
291 if (!ctx
->ring_pages
)
295 ctx
->mmap_size
= nr_pages
* PAGE_SIZE
;
296 pr_debug("attempting mmap of %lu bytes\n", ctx
->mmap_size
);
298 down_write(&mm
->mmap_sem
);
299 ctx
->mmap_base
= do_mmap_pgoff(ctx
->aio_ring_file
, 0, ctx
->mmap_size
,
300 PROT_READ
| PROT_WRITE
,
301 MAP_SHARED
| MAP_POPULATE
, 0, &populate
);
302 if (IS_ERR((void *)ctx
->mmap_base
)) {
303 up_write(&mm
->mmap_sem
);
308 up_write(&mm
->mmap_sem
);
310 mm_populate(ctx
->mmap_base
, populate
);
312 pr_debug("mmap address: 0x%08lx\n", ctx
->mmap_base
);
313 ctx
->nr_pages
= get_user_pages(current
, mm
, ctx
->mmap_base
, nr_pages
,
314 1, 0, ctx
->ring_pages
, NULL
);
315 for (i
= 0; i
< ctx
->nr_pages
; i
++)
316 put_page(ctx
->ring_pages
[i
]);
318 if (unlikely(ctx
->nr_pages
!= nr_pages
)) {
323 ctx
->user_id
= ctx
->mmap_base
;
324 ctx
->nr_events
= nr_events
; /* trusted copy */
326 ring
= kmap_atomic(ctx
->ring_pages
[0]);
327 ring
->nr
= nr_events
; /* user copy */
328 ring
->id
= ctx
->user_id
;
329 ring
->head
= ring
->tail
= 0;
330 ring
->magic
= AIO_RING_MAGIC
;
331 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
332 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
333 ring
->header_length
= sizeof(struct aio_ring
);
335 flush_dcache_page(ctx
->ring_pages
[0]);
340 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
341 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
342 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
344 void kiocb_set_cancel_fn(struct kiocb
*req
, kiocb_cancel_fn
*cancel
)
346 struct kioctx
*ctx
= req
->ki_ctx
;
349 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
351 if (!req
->ki_list
.next
)
352 list_add(&req
->ki_list
, &ctx
->active_reqs
);
354 req
->ki_cancel
= cancel
;
356 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
358 EXPORT_SYMBOL(kiocb_set_cancel_fn
);
360 static int kiocb_cancel(struct kioctx
*ctx
, struct kiocb
*kiocb
,
361 struct io_event
*res
)
363 kiocb_cancel_fn
*old
, *cancel
;
367 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
368 * actually has a cancel function, hence the cmpxchg()
371 cancel
= ACCESS_ONCE(kiocb
->ki_cancel
);
373 if (!cancel
|| cancel
== KIOCB_CANCELLED
)
377 cancel
= cmpxchg(&kiocb
->ki_cancel
, old
, KIOCB_CANCELLED
);
378 } while (cancel
!= old
);
380 atomic_inc(&kiocb
->ki_users
);
381 spin_unlock_irq(&ctx
->ctx_lock
);
383 memset(res
, 0, sizeof(*res
));
384 res
->obj
= (u64
)(unsigned long)kiocb
->ki_obj
.user
;
385 res
->data
= kiocb
->ki_user_data
;
386 ret
= cancel(kiocb
, res
);
388 spin_lock_irq(&ctx
->ctx_lock
);
393 static void free_ioctx_rcu(struct rcu_head
*head
)
395 struct kioctx
*ctx
= container_of(head
, struct kioctx
, rcu_head
);
397 free_percpu(ctx
->cpu
);
398 kmem_cache_free(kioctx_cachep
, ctx
);
402 * When this function runs, the kioctx has been removed from the "hash table"
403 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
404 * now it's safe to cancel any that need to be.
406 static void free_ioctx(struct kioctx
*ctx
)
408 struct aio_ring
*ring
;
411 unsigned cpu
, head
, avail
;
413 spin_lock_irq(&ctx
->ctx_lock
);
415 while (!list_empty(&ctx
->active_reqs
)) {
416 req
= list_first_entry(&ctx
->active_reqs
,
417 struct kiocb
, ki_list
);
419 list_del_init(&req
->ki_list
);
420 kiocb_cancel(ctx
, req
, &res
);
423 spin_unlock_irq(&ctx
->ctx_lock
);
425 for_each_possible_cpu(cpu
) {
426 struct kioctx_cpu
*kcpu
= per_cpu_ptr(ctx
->cpu
, cpu
);
428 atomic_add(kcpu
->reqs_available
, &ctx
->reqs_available
);
429 kcpu
->reqs_available
= 0;
432 ring
= kmap_atomic(ctx
->ring_pages
[0]);
436 while (atomic_read(&ctx
->reqs_available
) < ctx
->nr_events
- 1) {
437 wait_event(ctx
->wait
,
438 (head
!= ctx
->tail
) ||
439 (atomic_read(&ctx
->reqs_available
) >=
440 ctx
->nr_events
- 1));
442 avail
= (head
<= ctx
->tail
? ctx
->tail
: ctx
->nr_events
) - head
;
444 atomic_add(avail
, &ctx
->reqs_available
);
446 head
%= ctx
->nr_events
;
449 WARN_ON(atomic_read(&ctx
->reqs_available
) > ctx
->nr_events
- 1);
453 pr_debug("freeing %p\n", ctx
);
456 * Here the call_rcu() is between the wait_event() for reqs_active to
457 * hit 0, and freeing the ioctx.
459 * aio_complete() decrements reqs_active, but it has to touch the ioctx
460 * after to issue a wakeup so we use rcu.
462 call_rcu(&ctx
->rcu_head
, free_ioctx_rcu
);
465 static void put_ioctx(struct kioctx
*ctx
)
467 if (unlikely(atomic_dec_and_test(&ctx
->users
)))
472 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
474 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
476 struct mm_struct
*mm
= current
->mm
;
481 * We keep track of the number of available ringbuffer slots, to prevent
482 * overflow (reqs_available), and we also use percpu counters for this.
484 * So since up to half the slots might be on other cpu's percpu counters
485 * and unavailable, double nr_events so userspace sees what they
486 * expected: additionally, we move req_batch slots to/from percpu
487 * counters at a time, so make sure that isn't 0:
489 nr_events
= max(nr_events
, num_possible_cpus() * 4);
492 /* Prevent overflows */
493 if ((nr_events
> (0x10000000U
/ sizeof(struct io_event
))) ||
494 (nr_events
> (0x10000000U
/ sizeof(struct kiocb
)))) {
495 pr_debug("ENOMEM: nr_events too high\n");
496 return ERR_PTR(-EINVAL
);
499 if (!nr_events
|| (unsigned long)nr_events
> aio_max_nr
)
500 return ERR_PTR(-EAGAIN
);
502 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
504 return ERR_PTR(-ENOMEM
);
506 ctx
->max_reqs
= nr_events
;
508 atomic_set(&ctx
->users
, 2);
509 atomic_set(&ctx
->dead
, 0);
510 spin_lock_init(&ctx
->ctx_lock
);
511 spin_lock_init(&ctx
->completion_lock
);
512 mutex_init(&ctx
->ring_lock
);
513 init_waitqueue_head(&ctx
->wait
);
515 INIT_LIST_HEAD(&ctx
->active_reqs
);
517 ctx
->cpu
= alloc_percpu(struct kioctx_cpu
);
521 if (aio_setup_ring(ctx
) < 0)
524 atomic_set(&ctx
->reqs_available
, ctx
->nr_events
- 1);
525 ctx
->req_batch
= (ctx
->nr_events
- 1) / (num_possible_cpus() * 4);
526 BUG_ON(!ctx
->req_batch
);
528 /* limit the number of system wide aios */
529 spin_lock(&aio_nr_lock
);
530 if (aio_nr
+ nr_events
> aio_max_nr
||
531 aio_nr
+ nr_events
< aio_nr
) {
532 spin_unlock(&aio_nr_lock
);
535 aio_nr
+= ctx
->max_reqs
;
536 spin_unlock(&aio_nr_lock
);
538 /* now link into global list. */
539 spin_lock(&mm
->ioctx_lock
);
540 hlist_add_head_rcu(&ctx
->list
, &mm
->ioctx_list
);
541 spin_unlock(&mm
->ioctx_lock
);
543 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
544 ctx
, ctx
->user_id
, mm
, ctx
->nr_events
);
551 free_percpu(ctx
->cpu
);
553 if (ctx
->aio_ring_file
)
554 fput(ctx
->aio_ring_file
);
555 kmem_cache_free(kioctx_cachep
, ctx
);
556 pr_debug("error allocating ioctx %d\n", err
);
560 static void kill_ioctx_work(struct work_struct
*work
)
562 struct kioctx
*ctx
= container_of(work
, struct kioctx
, rcu_work
);
564 wake_up_all(&ctx
->wait
);
568 static void kill_ioctx_rcu(struct rcu_head
*head
)
570 struct kioctx
*ctx
= container_of(head
, struct kioctx
, rcu_head
);
572 INIT_WORK(&ctx
->rcu_work
, kill_ioctx_work
);
573 schedule_work(&ctx
->rcu_work
);
577 * Cancels all outstanding aio requests on an aio context. Used
578 * when the processes owning a context have all exited to encourage
579 * the rapid destruction of the kioctx.
581 static void kill_ioctx(struct kioctx
*ctx
)
583 if (!atomic_xchg(&ctx
->dead
, 1)) {
584 hlist_del_rcu(&ctx
->list
);
587 * It'd be more correct to do this in free_ioctx(), after all
588 * the outstanding kiocbs have finished - but by then io_destroy
589 * has already returned, so io_setup() could potentially return
590 * -EAGAIN with no ioctxs actually in use (as far as userspace
593 spin_lock(&aio_nr_lock
);
594 BUG_ON(aio_nr
- ctx
->max_reqs
> aio_nr
);
595 aio_nr
-= ctx
->max_reqs
;
596 spin_unlock(&aio_nr_lock
);
599 vm_munmap(ctx
->mmap_base
, ctx
->mmap_size
);
601 /* Between hlist_del_rcu() and dropping the initial ref */
602 call_rcu(&ctx
->rcu_head
, kill_ioctx_rcu
);
606 /* wait_on_sync_kiocb:
607 * Waits on the given sync kiocb to complete.
609 ssize_t
wait_on_sync_kiocb(struct kiocb
*iocb
)
611 while (atomic_read(&iocb
->ki_users
)) {
612 set_current_state(TASK_UNINTERRUPTIBLE
);
613 if (!atomic_read(&iocb
->ki_users
))
617 __set_current_state(TASK_RUNNING
);
618 return iocb
->ki_user_data
;
620 EXPORT_SYMBOL(wait_on_sync_kiocb
);
623 * exit_aio: called when the last user of mm goes away. At this point, there is
624 * no way for any new requests to be submited or any of the io_* syscalls to be
625 * called on the context.
627 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
630 void exit_aio(struct mm_struct
*mm
)
633 struct hlist_node
*n
;
635 hlist_for_each_entry_safe(ctx
, n
, &mm
->ioctx_list
, list
) {
636 if (1 != atomic_read(&ctx
->users
))
638 "exit_aio:ioctx still alive: %d %d %d\n",
639 atomic_read(&ctx
->users
),
640 atomic_read(&ctx
->dead
),
641 atomic_read(&ctx
->reqs_available
));
643 * We don't need to bother with munmap() here -
644 * exit_mmap(mm) is coming and it'll unmap everything.
645 * Since aio_free_ring() uses non-zero ->mmap_size
646 * as indicator that it needs to unmap the area,
647 * just set it to 0; aio_free_ring() is the only
648 * place that uses ->mmap_size, so it's safe.
656 static void put_reqs_available(struct kioctx
*ctx
, unsigned nr
)
658 struct kioctx_cpu
*kcpu
;
661 kcpu
= this_cpu_ptr(ctx
->cpu
);
663 kcpu
->reqs_available
+= nr
;
664 while (kcpu
->reqs_available
>= ctx
->req_batch
* 2) {
665 kcpu
->reqs_available
-= ctx
->req_batch
;
666 atomic_add(ctx
->req_batch
, &ctx
->reqs_available
);
672 static bool get_reqs_available(struct kioctx
*ctx
)
674 struct kioctx_cpu
*kcpu
;
678 kcpu
= this_cpu_ptr(ctx
->cpu
);
680 if (!kcpu
->reqs_available
) {
681 int old
, avail
= atomic_read(&ctx
->reqs_available
);
684 if (avail
< ctx
->req_batch
)
688 avail
= atomic_cmpxchg(&ctx
->reqs_available
,
689 avail
, avail
- ctx
->req_batch
);
690 } while (avail
!= old
);
692 kcpu
->reqs_available
+= ctx
->req_batch
;
696 kcpu
->reqs_available
--;
703 * Allocate a slot for an aio request. Increments the ki_users count
704 * of the kioctx so that the kioctx stays around until all requests are
705 * complete. Returns NULL if no requests are free.
707 * Returns with kiocb->ki_users set to 2. The io submit code path holds
708 * an extra reference while submitting the i/o.
709 * This prevents races between the aio code path referencing the
710 * req (after submitting it) and aio_complete() freeing the req.
712 static inline struct kiocb
*aio_get_req(struct kioctx
*ctx
)
716 if (!get_reqs_available(ctx
))
719 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
|__GFP_ZERO
);
723 atomic_set(&req
->ki_users
, 2);
727 put_reqs_available(ctx
, 1);
731 static void kiocb_free(struct kiocb
*req
)
735 if (req
->ki_eventfd
!= NULL
)
736 eventfd_ctx_put(req
->ki_eventfd
);
739 if (req
->ki_iovec
!= &req
->ki_inline_vec
)
740 kfree(req
->ki_iovec
);
741 kmem_cache_free(kiocb_cachep
, req
);
744 void aio_put_req(struct kiocb
*req
)
746 if (atomic_dec_and_test(&req
->ki_users
))
749 EXPORT_SYMBOL(aio_put_req
);
751 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
753 struct mm_struct
*mm
= current
->mm
;
754 struct kioctx
*ctx
, *ret
= NULL
;
758 hlist_for_each_entry_rcu(ctx
, &mm
->ioctx_list
, list
) {
759 if (ctx
->user_id
== ctx_id
) {
760 atomic_inc(&ctx
->users
);
771 * Called when the io request on the given iocb is complete.
773 void aio_complete(struct kiocb
*iocb
, long res
, long res2
)
775 struct kioctx
*ctx
= iocb
->ki_ctx
;
776 struct aio_ring
*ring
;
777 struct io_event
*ev_page
, *event
;
782 * Special case handling for sync iocbs:
783 * - events go directly into the iocb for fast handling
784 * - the sync task with the iocb in its stack holds the single iocb
785 * ref, no other paths have a way to get another ref
786 * - the sync task helpfully left a reference to itself in the iocb
788 if (is_sync_kiocb(iocb
)) {
789 BUG_ON(atomic_read(&iocb
->ki_users
) != 1);
790 iocb
->ki_user_data
= res
;
791 atomic_set(&iocb
->ki_users
, 0);
792 wake_up_process(iocb
->ki_obj
.tsk
);
797 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
798 * need to issue a wakeup after incrementing reqs_available.
802 if (iocb
->ki_list
.next
) {
805 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
806 list_del(&iocb
->ki_list
);
807 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
811 * cancelled requests don't get events, userland was given one
812 * when the event got cancelled.
814 if (unlikely(xchg(&iocb
->ki_cancel
,
815 KIOCB_CANCELLED
) == KIOCB_CANCELLED
)) {
817 * Can't use the percpu reqs_available here - could race with
820 atomic_inc(&ctx
->reqs_available
);
821 /* Still need the wake_up in case free_ioctx is waiting */
826 * Add a completion event to the ring buffer. Must be done holding
827 * ctx->completion_lock to prevent other code from messing with the tail
828 * pointer since we might be called from irq context.
830 spin_lock_irqsave(&ctx
->completion_lock
, flags
);
833 pos
= tail
+ AIO_EVENTS_OFFSET
;
835 if (++tail
>= ctx
->nr_events
)
838 ev_page
= kmap_atomic(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
839 event
= ev_page
+ pos
% AIO_EVENTS_PER_PAGE
;
841 event
->obj
= (u64
)(unsigned long)iocb
->ki_obj
.user
;
842 event
->data
= iocb
->ki_user_data
;
846 kunmap_atomic(ev_page
);
847 flush_dcache_page(ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
]);
849 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
850 ctx
, tail
, iocb
, iocb
->ki_obj
.user
, iocb
->ki_user_data
,
853 /* after flagging the request as done, we
854 * must never even look at it again
856 smp_wmb(); /* make event visible before updating tail */
860 ring
= kmap_atomic(ctx
->ring_pages
[0]);
863 flush_dcache_page(ctx
->ring_pages
[0]);
865 spin_unlock_irqrestore(&ctx
->completion_lock
, flags
);
867 pr_debug("added to ring %p at [%u]\n", iocb
, tail
);
870 * Check if the user asked us to deliver the result through an
871 * eventfd. The eventfd_signal() function is safe to be called
874 if (iocb
->ki_eventfd
!= NULL
)
875 eventfd_signal(iocb
->ki_eventfd
, 1);
878 /* everything turned out well, dispose of the aiocb. */
882 * We have to order our ring_info tail store above and test
883 * of the wait list below outside the wait lock. This is
884 * like in wake_up_bit() where clearing a bit has to be
885 * ordered with the unlocked test.
889 if (waitqueue_active(&ctx
->wait
))
894 EXPORT_SYMBOL(aio_complete
);
897 * Pull an event off of the ioctx's event ring. Returns the number of
900 static long aio_read_events_ring(struct kioctx
*ctx
,
901 struct io_event __user
*event
, long nr
)
903 struct aio_ring
*ring
;
908 mutex_lock(&ctx
->ring_lock
);
910 ring
= kmap_atomic(ctx
->ring_pages
[0]);
914 pr_debug("h%u t%u m%u\n", head
, ctx
->tail
, ctx
->nr_events
);
916 if (head
== ctx
->tail
)
924 avail
= (head
<= ctx
->tail
? ctx
->tail
: ctx
->nr_events
) - head
;
925 if (head
== ctx
->tail
)
928 avail
= min(avail
, nr
- ret
);
929 avail
= min_t(long, avail
, AIO_EVENTS_PER_PAGE
-
930 ((head
+ AIO_EVENTS_OFFSET
) % AIO_EVENTS_PER_PAGE
));
932 pos
= head
+ AIO_EVENTS_OFFSET
;
933 page
= ctx
->ring_pages
[pos
/ AIO_EVENTS_PER_PAGE
];
934 pos
%= AIO_EVENTS_PER_PAGE
;
937 copy_ret
= copy_to_user(event
+ ret
, ev
+ pos
,
938 sizeof(*ev
) * avail
);
941 if (unlikely(copy_ret
)) {
948 head
%= ctx
->nr_events
;
951 ring
= kmap_atomic(ctx
->ring_pages
[0]);
954 flush_dcache_page(ctx
->ring_pages
[0]);
956 pr_debug("%li h%u t%u\n", ret
, head
, ctx
->tail
);
958 put_reqs_available(ctx
, ret
);
960 mutex_unlock(&ctx
->ring_lock
);
965 static bool aio_read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
966 struct io_event __user
*event
, long *i
)
968 long ret
= aio_read_events_ring(ctx
, event
+ *i
, nr
- *i
);
973 if (unlikely(atomic_read(&ctx
->dead
)))
979 return ret
< 0 || *i
>= min_nr
;
982 static long read_events(struct kioctx
*ctx
, long min_nr
, long nr
,
983 struct io_event __user
*event
,
984 struct timespec __user
*timeout
)
986 ktime_t until
= { .tv64
= KTIME_MAX
};
992 if (unlikely(copy_from_user(&ts
, timeout
, sizeof(ts
))))
995 until
= timespec_to_ktime(ts
);
999 * Note that aio_read_events() is being called as the conditional - i.e.
1000 * we're calling it after prepare_to_wait() has set task state to
1001 * TASK_INTERRUPTIBLE.
1003 * But aio_read_events() can block, and if it blocks it's going to flip
1004 * the task state back to TASK_RUNNING.
1006 * This should be ok, provided it doesn't flip the state back to
1007 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1008 * will only happen if the mutex_lock() call blocks, and we then find
1009 * the ringbuffer empty. So in practice we should be ok, but it's
1010 * something to be aware of when touching this code.
1012 wait_event_interruptible_hrtimeout(ctx
->wait
,
1013 aio_read_events(ctx
, min_nr
, nr
, event
, &ret
), until
);
1015 if (!ret
&& signal_pending(current
))
1022 * Create an aio_context capable of receiving at least nr_events.
1023 * ctxp must not point to an aio_context that already exists, and
1024 * must be initialized to 0 prior to the call. On successful
1025 * creation of the aio_context, *ctxp is filled in with the resulting
1026 * handle. May fail with -EINVAL if *ctxp is not initialized,
1027 * if the specified nr_events exceeds internal limits. May fail
1028 * with -EAGAIN if the specified nr_events exceeds the user's limit
1029 * of available events. May fail with -ENOMEM if insufficient kernel
1030 * resources are available. May fail with -EFAULT if an invalid
1031 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1034 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
1036 struct kioctx
*ioctx
= NULL
;
1040 ret
= get_user(ctx
, ctxp
);
1045 if (unlikely(ctx
|| nr_events
== 0)) {
1046 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1051 ioctx
= ioctx_alloc(nr_events
);
1052 ret
= PTR_ERR(ioctx
);
1053 if (!IS_ERR(ioctx
)) {
1054 ret
= put_user(ioctx
->user_id
, ctxp
);
1065 * Destroy the aio_context specified. May cancel any outstanding
1066 * AIOs and block on completion. Will fail with -ENOSYS if not
1067 * implemented. May fail with -EINVAL if the context pointed to
1070 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1072 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1073 if (likely(NULL
!= ioctx
)) {
1078 pr_debug("EINVAL: io_destroy: invalid context id\n");
1082 static void aio_advance_iovec(struct kiocb
*iocb
, ssize_t ret
)
1084 struct iovec
*iov
= &iocb
->ki_iovec
[iocb
->ki_cur_seg
];
1088 while (iocb
->ki_cur_seg
< iocb
->ki_nr_segs
&& ret
> 0) {
1089 ssize_t
this = min((ssize_t
)iov
->iov_len
, ret
);
1090 iov
->iov_base
+= this;
1091 iov
->iov_len
-= this;
1092 iocb
->ki_left
-= this;
1094 if (iov
->iov_len
== 0) {
1100 /* the caller should not have done more io than what fit in
1101 * the remaining iovecs */
1102 BUG_ON(ret
> 0 && iocb
->ki_left
== 0);
1105 typedef ssize_t (aio_rw_op
)(struct kiocb
*, const struct iovec
*,
1106 unsigned long, loff_t
);
1108 static ssize_t
aio_rw_vect_retry(struct kiocb
*iocb
, int rw
, aio_rw_op
*rw_op
)
1110 struct file
*file
= iocb
->ki_filp
;
1111 struct address_space
*mapping
= file
->f_mapping
;
1112 struct inode
*inode
= mapping
->host
;
1115 /* This matches the pread()/pwrite() logic */
1116 if (iocb
->ki_pos
< 0)
1120 file_start_write(file
);
1122 ret
= rw_op(iocb
, &iocb
->ki_iovec
[iocb
->ki_cur_seg
],
1123 iocb
->ki_nr_segs
- iocb
->ki_cur_seg
,
1126 aio_advance_iovec(iocb
, ret
);
1128 /* retry all partial writes. retry partial reads as long as its a
1130 } while (ret
> 0 && iocb
->ki_left
> 0 &&
1132 (!S_ISFIFO(inode
->i_mode
) && !S_ISSOCK(inode
->i_mode
))));
1134 file_end_write(file
);
1136 /* This means we must have transferred all that we could */
1137 /* No need to retry anymore */
1138 if ((ret
== 0) || (iocb
->ki_left
== 0))
1139 ret
= iocb
->ki_nbytes
- iocb
->ki_left
;
1141 /* If we managed to write some out we return that, rather than
1142 * the eventual error. */
1144 && ret
< 0 && ret
!= -EIOCBQUEUED
1145 && iocb
->ki_nbytes
- iocb
->ki_left
)
1146 ret
= iocb
->ki_nbytes
- iocb
->ki_left
;
1151 static ssize_t
aio_setup_vectored_rw(int rw
, struct kiocb
*kiocb
, bool compat
)
1155 kiocb
->ki_nr_segs
= kiocb
->ki_nbytes
;
1157 #ifdef CONFIG_COMPAT
1159 ret
= compat_rw_copy_check_uvector(rw
,
1160 (struct compat_iovec __user
*)kiocb
->ki_buf
,
1161 kiocb
->ki_nr_segs
, 1, &kiocb
->ki_inline_vec
,
1165 ret
= rw_copy_check_uvector(rw
,
1166 (struct iovec __user
*)kiocb
->ki_buf
,
1167 kiocb
->ki_nr_segs
, 1, &kiocb
->ki_inline_vec
,
1172 /* ki_nbytes now reflect bytes instead of segs */
1173 kiocb
->ki_nbytes
= ret
;
1177 static ssize_t
aio_setup_single_vector(int rw
, struct kiocb
*kiocb
)
1179 if (unlikely(!access_ok(!rw
, kiocb
->ki_buf
, kiocb
->ki_nbytes
)))
1182 kiocb
->ki_iovec
= &kiocb
->ki_inline_vec
;
1183 kiocb
->ki_iovec
->iov_base
= kiocb
->ki_buf
;
1184 kiocb
->ki_iovec
->iov_len
= kiocb
->ki_nbytes
;
1185 kiocb
->ki_nr_segs
= 1;
1191 * Performs the initial checks and aio retry method
1192 * setup for the kiocb at the time of io submission.
1194 static ssize_t
aio_run_iocb(struct kiocb
*req
, bool compat
)
1196 struct file
*file
= req
->ki_filp
;
1202 switch (req
->ki_opcode
) {
1203 case IOCB_CMD_PREAD
:
1204 case IOCB_CMD_PREADV
:
1207 rw_op
= file
->f_op
->aio_read
;
1210 case IOCB_CMD_PWRITE
:
1211 case IOCB_CMD_PWRITEV
:
1214 rw_op
= file
->f_op
->aio_write
;
1217 if (unlikely(!(file
->f_mode
& mode
)))
1223 ret
= (req
->ki_opcode
== IOCB_CMD_PREADV
||
1224 req
->ki_opcode
== IOCB_CMD_PWRITEV
)
1225 ? aio_setup_vectored_rw(rw
, req
, compat
)
1226 : aio_setup_single_vector(rw
, req
);
1230 ret
= rw_verify_area(rw
, file
, &req
->ki_pos
, req
->ki_nbytes
);
1234 req
->ki_nbytes
= ret
;
1237 ret
= aio_rw_vect_retry(req
, rw
, rw_op
);
1240 case IOCB_CMD_FDSYNC
:
1241 if (!file
->f_op
->aio_fsync
)
1244 ret
= file
->f_op
->aio_fsync(req
, 1);
1247 case IOCB_CMD_FSYNC
:
1248 if (!file
->f_op
->aio_fsync
)
1251 ret
= file
->f_op
->aio_fsync(req
, 0);
1255 pr_debug("EINVAL: no operation provided\n");
1259 if (ret
!= -EIOCBQUEUED
) {
1261 * There's no easy way to restart the syscall since other AIO's
1262 * may be already running. Just fail this IO with EINTR.
1264 if (unlikely(ret
== -ERESTARTSYS
|| ret
== -ERESTARTNOINTR
||
1265 ret
== -ERESTARTNOHAND
||
1266 ret
== -ERESTART_RESTARTBLOCK
))
1268 aio_complete(req
, ret
, 0);
1274 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
1275 struct iocb
*iocb
, bool compat
)
1280 /* enforce forwards compatibility on users */
1281 if (unlikely(iocb
->aio_reserved1
|| iocb
->aio_reserved2
)) {
1282 pr_debug("EINVAL: reserve field set\n");
1286 /* prevent overflows */
1288 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1289 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1290 ((ssize_t
)iocb
->aio_nbytes
< 0)
1292 pr_debug("EINVAL: io_submit: overflow check\n");
1296 req
= aio_get_req(ctx
);
1300 req
->ki_filp
= fget(iocb
->aio_fildes
);
1301 if (unlikely(!req
->ki_filp
)) {
1306 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1308 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1309 * instance of the file* now. The file descriptor must be
1310 * an eventfd() fd, and will be signaled for each completed
1311 * event using the eventfd_signal() function.
1313 req
->ki_eventfd
= eventfd_ctx_fdget((int) iocb
->aio_resfd
);
1314 if (IS_ERR(req
->ki_eventfd
)) {
1315 ret
= PTR_ERR(req
->ki_eventfd
);
1316 req
->ki_eventfd
= NULL
;
1321 ret
= put_user(KIOCB_KEY
, &user_iocb
->aio_key
);
1322 if (unlikely(ret
)) {
1323 pr_debug("EFAULT: aio_key\n");
1327 req
->ki_obj
.user
= user_iocb
;
1328 req
->ki_user_data
= iocb
->aio_data
;
1329 req
->ki_pos
= iocb
->aio_offset
;
1331 req
->ki_buf
= (char __user
*)(unsigned long)iocb
->aio_buf
;
1332 req
->ki_left
= req
->ki_nbytes
= iocb
->aio_nbytes
;
1333 req
->ki_opcode
= iocb
->aio_lio_opcode
;
1335 ret
= aio_run_iocb(req
, compat
);
1339 aio_put_req(req
); /* drop extra ref to req */
1342 put_reqs_available(ctx
, 1);
1343 aio_put_req(req
); /* drop extra ref to req */
1344 aio_put_req(req
); /* drop i/o ref to req */
1348 long do_io_submit(aio_context_t ctx_id
, long nr
,
1349 struct iocb __user
*__user
*iocbpp
, bool compat
)
1354 struct blk_plug plug
;
1356 if (unlikely(nr
< 0))
1359 if (unlikely(nr
> LONG_MAX
/sizeof(*iocbpp
)))
1360 nr
= LONG_MAX
/sizeof(*iocbpp
);
1362 if (unlikely(!access_ok(VERIFY_READ
, iocbpp
, (nr
*sizeof(*iocbpp
)))))
1365 ctx
= lookup_ioctx(ctx_id
);
1366 if (unlikely(!ctx
)) {
1367 pr_debug("EINVAL: invalid context id\n");
1371 blk_start_plug(&plug
);
1374 * AKPM: should this return a partial result if some of the IOs were
1375 * successfully submitted?
1377 for (i
=0; i
<nr
; i
++) {
1378 struct iocb __user
*user_iocb
;
1381 if (unlikely(__get_user(user_iocb
, iocbpp
+ i
))) {
1386 if (unlikely(copy_from_user(&tmp
, user_iocb
, sizeof(tmp
)))) {
1391 ret
= io_submit_one(ctx
, user_iocb
, &tmp
, compat
);
1395 blk_finish_plug(&plug
);
1402 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1403 * the number of iocbs queued. May return -EINVAL if the aio_context
1404 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1405 * *iocbpp[0] is not properly initialized, if the operation specified
1406 * is invalid for the file descriptor in the iocb. May fail with
1407 * -EFAULT if any of the data structures point to invalid data. May
1408 * fail with -EBADF if the file descriptor specified in the first
1409 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1410 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1411 * fail with -ENOSYS if not implemented.
1413 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
1414 struct iocb __user
* __user
*, iocbpp
)
1416 return do_io_submit(ctx_id
, nr
, iocbpp
, 0);
1420 * Finds a given iocb for cancellation.
1422 static struct kiocb
*lookup_kiocb(struct kioctx
*ctx
, struct iocb __user
*iocb
,
1425 struct list_head
*pos
;
1427 assert_spin_locked(&ctx
->ctx_lock
);
1429 if (key
!= KIOCB_KEY
)
1432 /* TODO: use a hash or array, this sucks. */
1433 list_for_each(pos
, &ctx
->active_reqs
) {
1434 struct kiocb
*kiocb
= list_kiocb(pos
);
1435 if (kiocb
->ki_obj
.user
== iocb
)
1442 * Attempts to cancel an iocb previously passed to io_submit. If
1443 * the operation is successfully cancelled, the resulting event is
1444 * copied into the memory pointed to by result without being placed
1445 * into the completion queue and 0 is returned. May fail with
1446 * -EFAULT if any of the data structures pointed to are invalid.
1447 * May fail with -EINVAL if aio_context specified by ctx_id is
1448 * invalid. May fail with -EAGAIN if the iocb specified was not
1449 * cancelled. Will fail with -ENOSYS if not implemented.
1451 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
1452 struct io_event __user
*, result
)
1454 struct io_event res
;
1456 struct kiocb
*kiocb
;
1460 ret
= get_user(key
, &iocb
->aio_key
);
1464 ctx
= lookup_ioctx(ctx_id
);
1468 spin_lock_irq(&ctx
->ctx_lock
);
1470 kiocb
= lookup_kiocb(ctx
, iocb
, key
);
1472 ret
= kiocb_cancel(ctx
, kiocb
, &res
);
1476 spin_unlock_irq(&ctx
->ctx_lock
);
1479 /* Cancellation succeeded -- copy the result
1480 * into the user's buffer.
1482 if (copy_to_user(result
, &res
, sizeof(res
)))
1492 * Attempts to read at least min_nr events and up to nr events from
1493 * the completion queue for the aio_context specified by ctx_id. If
1494 * it succeeds, the number of read events is returned. May fail with
1495 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1496 * out of range, if timeout is out of range. May fail with -EFAULT
1497 * if any of the memory specified is invalid. May return 0 or
1498 * < min_nr if the timeout specified by timeout has elapsed
1499 * before sufficient events are available, where timeout == NULL
1500 * specifies an infinite timeout. Note that the timeout pointed to by
1501 * timeout is relative. Will fail with -ENOSYS if not implemented.
1503 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
1506 struct io_event __user
*, events
,
1507 struct timespec __user
*, timeout
)
1509 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
1512 if (likely(ioctx
)) {
1513 if (likely(min_nr
<= nr
&& min_nr
>= 0))
1514 ret
= read_events(ioctx
, min_nr
, nr
, events
, timeout
);