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 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/export.h>
17 #include <linux/syscalls.h>
18 #include <linux/backing-dev.h>
19 #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/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/compat.h>
39 #include <asm/kmap_types.h>
40 #include <asm/uaccess.h>
43 #define dprintk printk
45 #define dprintk(x...) do { ; } while (0)
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
67 struct aio_ring_info
{
68 unsigned long mmap_base
;
69 unsigned long mmap_size
;
71 struct page
**ring_pages
;
77 struct page
*internal_pages
[AIO_RING_PAGES
];
80 static inline unsigned aio_ring_avail(struct aio_ring_info
*info
,
81 struct aio_ring
*ring
)
83 return (ring
->head
+ info
->nr
- 1 - ring
->tail
) % info
->nr
;
90 /* This needs improving */
91 unsigned long user_id
;
92 struct hlist_node list
;
94 wait_queue_head_t wait
;
99 struct list_head active_reqs
; /* used for cancellation */
101 /* sys_io_setup currently limits this to an unsigned int */
104 struct aio_ring_info ring_info
;
106 struct rcu_head rcu_head
;
109 /*------ sysctl variables----*/
110 static DEFINE_SPINLOCK(aio_nr_lock
);
111 unsigned long aio_nr
; /* current system wide number of aio requests */
112 unsigned long aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
113 /*----end sysctl variables---*/
115 static struct kmem_cache
*kiocb_cachep
;
116 static struct kmem_cache
*kioctx_cachep
;
119 * Creates the slab caches used by the aio routines, panic on
120 * failure as this is done early during the boot sequence.
122 static int __init
aio_setup(void)
124 kiocb_cachep
= KMEM_CACHE(kiocb
, SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
125 kioctx_cachep
= KMEM_CACHE(kioctx
,SLAB_HWCACHE_ALIGN
|SLAB_PANIC
);
127 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page
));
131 __initcall(aio_setup
);
133 static void aio_free_ring(struct kioctx
*ctx
)
135 struct aio_ring_info
*info
= &ctx
->ring_info
;
138 for (i
=0; i
<info
->nr_pages
; i
++)
139 put_page(info
->ring_pages
[i
]);
141 if (info
->mmap_size
) {
142 vm_munmap(info
->mmap_base
, info
->mmap_size
);
145 if (info
->ring_pages
&& info
->ring_pages
!= info
->internal_pages
)
146 kfree(info
->ring_pages
);
147 info
->ring_pages
= NULL
;
151 static int aio_setup_ring(struct kioctx
*ctx
)
153 struct aio_ring
*ring
;
154 struct aio_ring_info
*info
= &ctx
->ring_info
;
155 unsigned nr_events
= ctx
->max_reqs
;
156 struct mm_struct
*mm
= current
->mm
;
157 unsigned long size
, populate
;
160 /* Compensate for the ring buffer's head/tail overlap entry */
161 nr_events
+= 2; /* 1 is required, 2 for good luck */
163 size
= sizeof(struct aio_ring
);
164 size
+= sizeof(struct io_event
) * nr_events
;
165 nr_pages
= (size
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
170 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
)) / sizeof(struct io_event
);
173 info
->ring_pages
= info
->internal_pages
;
174 if (nr_pages
> AIO_RING_PAGES
) {
175 info
->ring_pages
= kcalloc(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
176 if (!info
->ring_pages
)
180 info
->mmap_size
= nr_pages
* PAGE_SIZE
;
181 dprintk("attempting mmap of %lu bytes\n", info
->mmap_size
);
182 down_write(&mm
->mmap_sem
);
183 info
->mmap_base
= do_mmap_pgoff(NULL
, 0, info
->mmap_size
,
184 PROT_READ
|PROT_WRITE
,
185 MAP_ANONYMOUS
|MAP_PRIVATE
, 0,
187 if (IS_ERR((void *)info
->mmap_base
)) {
188 up_write(&mm
->mmap_sem
);
194 dprintk("mmap address: 0x%08lx\n", info
->mmap_base
);
195 info
->nr_pages
= get_user_pages(current
, mm
, info
->mmap_base
, nr_pages
,
196 1, 0, info
->ring_pages
, NULL
);
197 up_write(&mm
->mmap_sem
);
199 if (unlikely(info
->nr_pages
!= nr_pages
)) {
204 mm_populate(info
->mmap_base
, populate
);
206 ctx
->user_id
= info
->mmap_base
;
208 info
->nr
= nr_events
; /* trusted copy */
210 ring
= kmap_atomic(info
->ring_pages
[0]);
211 ring
->nr
= nr_events
; /* user copy */
212 ring
->id
= ctx
->user_id
;
213 ring
->head
= ring
->tail
= 0;
214 ring
->magic
= AIO_RING_MAGIC
;
215 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
216 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
217 ring
->header_length
= sizeof(struct aio_ring
);
224 /* aio_ring_event: returns a pointer to the event at the given index from
225 * kmap_atomic(). Release the pointer with put_aio_ring_event();
227 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
228 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
229 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
231 #define aio_ring_event(info, nr) ({ \
232 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
233 struct io_event *__event; \
234 __event = kmap_atomic( \
235 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
236 __event += pos % AIO_EVENTS_PER_PAGE; \
240 #define put_aio_ring_event(event) do { \
241 struct io_event *__event = (event); \
243 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
246 static void ctx_rcu_free(struct rcu_head
*head
)
248 struct kioctx
*ctx
= container_of(head
, struct kioctx
, rcu_head
);
249 kmem_cache_free(kioctx_cachep
, ctx
);
253 * Called when the last user of an aio context has gone away,
254 * and the struct needs to be freed.
256 static void __put_ioctx(struct kioctx
*ctx
)
258 unsigned nr_events
= ctx
->max_reqs
;
259 BUG_ON(ctx
->reqs_active
);
263 spin_lock(&aio_nr_lock
);
264 BUG_ON(aio_nr
- nr_events
> aio_nr
);
266 spin_unlock(&aio_nr_lock
);
268 pr_debug("__put_ioctx: freeing %p\n", ctx
);
269 call_rcu(&ctx
->rcu_head
, ctx_rcu_free
);
272 static inline int try_get_ioctx(struct kioctx
*kioctx
)
274 return atomic_inc_not_zero(&kioctx
->users
);
277 static inline void put_ioctx(struct kioctx
*kioctx
)
279 BUG_ON(atomic_read(&kioctx
->users
) <= 0);
280 if (unlikely(atomic_dec_and_test(&kioctx
->users
)))
284 static int kiocb_cancel(struct kioctx
*ctx
, struct kiocb
*kiocb
,
285 struct io_event
*res
)
287 int (*cancel
)(struct kiocb
*, struct io_event
*);
290 cancel
= kiocb
->ki_cancel
;
291 kiocbSetCancelled(kiocb
);
294 spin_unlock_irq(&ctx
->ctx_lock
);
296 memset(res
, 0, sizeof(*res
));
297 res
->obj
= (u64
)(unsigned long)kiocb
->ki_obj
.user
;
298 res
->data
= kiocb
->ki_user_data
;
299 ret
= cancel(kiocb
, res
);
301 spin_lock_irq(&ctx
->ctx_lock
);
308 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
310 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
312 struct mm_struct
*mm
= current
->mm
;
316 /* Prevent overflows */
317 if ((nr_events
> (0x10000000U
/ sizeof(struct io_event
))) ||
318 (nr_events
> (0x10000000U
/ sizeof(struct kiocb
)))) {
319 pr_debug("ENOMEM: nr_events too high\n");
320 return ERR_PTR(-EINVAL
);
323 if (!nr_events
|| (unsigned long)nr_events
> aio_max_nr
)
324 return ERR_PTR(-EAGAIN
);
326 ctx
= kmem_cache_zalloc(kioctx_cachep
, GFP_KERNEL
);
328 return ERR_PTR(-ENOMEM
);
330 ctx
->max_reqs
= nr_events
;
332 atomic_set(&ctx
->users
, 2);
333 spin_lock_init(&ctx
->ctx_lock
);
334 spin_lock_init(&ctx
->ring_info
.ring_lock
);
335 init_waitqueue_head(&ctx
->wait
);
337 INIT_LIST_HEAD(&ctx
->active_reqs
);
339 if (aio_setup_ring(ctx
) < 0)
342 /* limit the number of system wide aios */
343 spin_lock(&aio_nr_lock
);
344 if (aio_nr
+ nr_events
> aio_max_nr
||
345 aio_nr
+ nr_events
< aio_nr
) {
346 spin_unlock(&aio_nr_lock
);
349 aio_nr
+= ctx
->max_reqs
;
350 spin_unlock(&aio_nr_lock
);
352 /* now link into global list. */
353 spin_lock(&mm
->ioctx_lock
);
354 hlist_add_head_rcu(&ctx
->list
, &mm
->ioctx_list
);
355 spin_unlock(&mm
->ioctx_lock
);
357 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
358 ctx
, ctx
->user_id
, mm
, ctx
->ring_info
.nr
);
365 kmem_cache_free(kioctx_cachep
, ctx
);
366 dprintk("aio: error allocating ioctx %d\n", err
);
371 * Cancels all outstanding aio requests on an aio context. Used
372 * when the processes owning a context have all exited to encourage
373 * the rapid destruction of the kioctx.
375 static void kill_ctx(struct kioctx
*ctx
)
377 struct task_struct
*tsk
= current
;
378 DECLARE_WAITQUEUE(wait
, tsk
);
382 spin_lock_irq(&ctx
->ctx_lock
);
384 while (!list_empty(&ctx
->active_reqs
)) {
385 req
= list_first_entry(&ctx
->active_reqs
,
386 struct kiocb
, ki_list
);
388 list_del_init(&req
->ki_list
);
389 kiocb_cancel(ctx
, req
, &res
);
392 if (!ctx
->reqs_active
)
395 add_wait_queue(&ctx
->wait
, &wait
);
396 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
397 while (ctx
->reqs_active
) {
398 spin_unlock_irq(&ctx
->ctx_lock
);
400 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
401 spin_lock_irq(&ctx
->ctx_lock
);
403 __set_task_state(tsk
, TASK_RUNNING
);
404 remove_wait_queue(&ctx
->wait
, &wait
);
407 spin_unlock_irq(&ctx
->ctx_lock
);
410 /* wait_on_sync_kiocb:
411 * Waits on the given sync kiocb to complete.
413 ssize_t
wait_on_sync_kiocb(struct kiocb
*iocb
)
415 while (iocb
->ki_users
) {
416 set_current_state(TASK_UNINTERRUPTIBLE
);
421 __set_current_state(TASK_RUNNING
);
422 return iocb
->ki_user_data
;
424 EXPORT_SYMBOL(wait_on_sync_kiocb
);
426 /* exit_aio: called when the last user of mm goes away. At this point,
427 * there is no way for any new requests to be submited or any of the
428 * io_* syscalls to be called on the context. However, there may be
429 * outstanding requests which hold references to the context; as they
430 * go away, they will call put_ioctx and release any pinned memory
431 * associated with the request (held via struct page * references).
433 void exit_aio(struct mm_struct
*mm
)
437 while (!hlist_empty(&mm
->ioctx_list
)) {
438 ctx
= hlist_entry(mm
->ioctx_list
.first
, struct kioctx
, list
);
439 hlist_del_rcu(&ctx
->list
);
443 if (1 != atomic_read(&ctx
->users
))
445 "exit_aio:ioctx still alive: %d %d %d\n",
446 atomic_read(&ctx
->users
), ctx
->dead
,
449 * We don't need to bother with munmap() here -
450 * exit_mmap(mm) is coming and it'll unmap everything.
451 * Since aio_free_ring() uses non-zero ->mmap_size
452 * as indicator that it needs to unmap the area,
453 * just set it to 0; aio_free_ring() is the only
454 * place that uses ->mmap_size, so it's safe.
456 ctx
->ring_info
.mmap_size
= 0;
462 * Allocate a slot for an aio request. Increments the users count
463 * of the kioctx so that the kioctx stays around until all requests are
464 * complete. Returns NULL if no requests are free.
466 * Returns with kiocb->users set to 2. The io submit code path holds
467 * an extra reference while submitting the i/o.
468 * This prevents races between the aio code path referencing the
469 * req (after submitting it) and aio_complete() freeing the req.
471 static struct kiocb
*__aio_get_req(struct kioctx
*ctx
)
473 struct kiocb
*req
= NULL
;
475 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
483 req
->ki_cancel
= NULL
;
484 req
->ki_retry
= NULL
;
487 req
->ki_iovec
= NULL
;
488 req
->ki_eventfd
= NULL
;
494 * struct kiocb's are allocated in batches to reduce the number of
495 * times the ctx lock is acquired and released.
497 #define KIOCB_BATCH_SIZE 32L
499 struct list_head head
;
500 long count
; /* number of requests left to allocate */
503 static void kiocb_batch_init(struct kiocb_batch
*batch
, long total
)
505 INIT_LIST_HEAD(&batch
->head
);
506 batch
->count
= total
;
509 static void kiocb_batch_free(struct kioctx
*ctx
, struct kiocb_batch
*batch
)
511 struct kiocb
*req
, *n
;
513 if (list_empty(&batch
->head
))
516 spin_lock_irq(&ctx
->ctx_lock
);
517 list_for_each_entry_safe(req
, n
, &batch
->head
, ki_batch
) {
518 list_del(&req
->ki_batch
);
519 list_del(&req
->ki_list
);
520 kmem_cache_free(kiocb_cachep
, req
);
523 if (unlikely(!ctx
->reqs_active
&& ctx
->dead
))
524 wake_up_all(&ctx
->wait
);
525 spin_unlock_irq(&ctx
->ctx_lock
);
529 * Allocate a batch of kiocbs. This avoids taking and dropping the
530 * context lock a lot during setup.
532 static int kiocb_batch_refill(struct kioctx
*ctx
, struct kiocb_batch
*batch
)
534 unsigned short allocated
, to_alloc
;
536 struct kiocb
*req
, *n
;
537 struct aio_ring
*ring
;
539 to_alloc
= min(batch
->count
, KIOCB_BATCH_SIZE
);
540 for (allocated
= 0; allocated
< to_alloc
; allocated
++) {
541 req
= __aio_get_req(ctx
);
543 /* allocation failed, go with what we've got */
545 list_add(&req
->ki_batch
, &batch
->head
);
551 spin_lock_irq(&ctx
->ctx_lock
);
552 ring
= kmap_atomic(ctx
->ring_info
.ring_pages
[0]);
554 avail
= aio_ring_avail(&ctx
->ring_info
, ring
) - ctx
->reqs_active
;
556 if (avail
< allocated
) {
557 /* Trim back the number of requests. */
558 list_for_each_entry_safe(req
, n
, &batch
->head
, ki_batch
) {
559 list_del(&req
->ki_batch
);
560 kmem_cache_free(kiocb_cachep
, req
);
561 if (--allocated
<= avail
)
566 batch
->count
-= allocated
;
567 list_for_each_entry(req
, &batch
->head
, ki_batch
) {
568 list_add(&req
->ki_list
, &ctx
->active_reqs
);
573 spin_unlock_irq(&ctx
->ctx_lock
);
579 static inline struct kiocb
*aio_get_req(struct kioctx
*ctx
,
580 struct kiocb_batch
*batch
)
584 if (list_empty(&batch
->head
))
585 if (kiocb_batch_refill(ctx
, batch
) == 0)
587 req
= list_first_entry(&batch
->head
, struct kiocb
, ki_batch
);
588 list_del(&req
->ki_batch
);
592 static inline void really_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
594 assert_spin_locked(&ctx
->ctx_lock
);
596 if (req
->ki_eventfd
!= NULL
)
597 eventfd_ctx_put(req
->ki_eventfd
);
600 if (req
->ki_iovec
!= &req
->ki_inline_vec
)
601 kfree(req
->ki_iovec
);
602 kmem_cache_free(kiocb_cachep
, req
);
605 if (unlikely(!ctx
->reqs_active
&& ctx
->dead
))
606 wake_up_all(&ctx
->wait
);
610 * Returns true if this put was the last user of the request.
612 static void __aio_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
614 dprintk(KERN_DEBUG
"aio_put(%p): f_count=%ld\n",
615 req
, atomic_long_read(&req
->ki_filp
->f_count
));
617 assert_spin_locked(&ctx
->ctx_lock
);
620 BUG_ON(req
->ki_users
< 0);
621 if (likely(req
->ki_users
))
623 list_del(&req
->ki_list
); /* remove from active_reqs */
624 req
->ki_cancel
= NULL
;
625 req
->ki_retry
= NULL
;
629 really_put_req(ctx
, req
);
633 * Returns true if this put was the last user of the kiocb,
634 * false if the request is still in use.
636 void aio_put_req(struct kiocb
*req
)
638 struct kioctx
*ctx
= req
->ki_ctx
;
639 spin_lock_irq(&ctx
->ctx_lock
);
640 __aio_put_req(ctx
, req
);
641 spin_unlock_irq(&ctx
->ctx_lock
);
643 EXPORT_SYMBOL(aio_put_req
);
645 static struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
647 struct mm_struct
*mm
= current
->mm
;
648 struct kioctx
*ctx
, *ret
= NULL
;
652 hlist_for_each_entry_rcu(ctx
, &mm
->ioctx_list
, list
) {
654 * RCU protects us against accessing freed memory but
655 * we have to be careful not to get a reference when the
656 * reference count already dropped to 0 (ctx->dead test
657 * is unreliable because of races).
659 if (ctx
->user_id
== ctx_id
&& !ctx
->dead
&& try_get_ioctx(ctx
)){
670 * Called when the io request on the given iocb is complete.
672 void aio_complete(struct kiocb
*iocb
, long res
, long res2
)
674 struct kioctx
*ctx
= iocb
->ki_ctx
;
675 struct aio_ring_info
*info
;
676 struct aio_ring
*ring
;
677 struct io_event
*event
;
682 * Special case handling for sync iocbs:
683 * - events go directly into the iocb for fast handling
684 * - the sync task with the iocb in its stack holds the single iocb
685 * ref, no other paths have a way to get another ref
686 * - the sync task helpfully left a reference to itself in the iocb
688 if (is_sync_kiocb(iocb
)) {
689 BUG_ON(iocb
->ki_users
!= 1);
690 iocb
->ki_user_data
= res
;
692 wake_up_process(iocb
->ki_obj
.tsk
);
696 info
= &ctx
->ring_info
;
698 /* add a completion event to the ring buffer.
699 * must be done holding ctx->ctx_lock to prevent
700 * other code from messing with the tail
701 * pointer since we might be called from irq
704 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
707 * cancelled requests don't get events, userland was given one
708 * when the event got cancelled.
710 if (kiocbIsCancelled(iocb
))
713 ring
= kmap_atomic(info
->ring_pages
[0]);
716 event
= aio_ring_event(info
, tail
);
717 if (++tail
>= info
->nr
)
720 event
->obj
= (u64
)(unsigned long)iocb
->ki_obj
.user
;
721 event
->data
= iocb
->ki_user_data
;
725 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
726 ctx
, tail
, iocb
, iocb
->ki_obj
.user
, iocb
->ki_user_data
,
729 /* after flagging the request as done, we
730 * must never even look at it again
732 smp_wmb(); /* make event visible before updating tail */
737 put_aio_ring_event(event
);
740 pr_debug("added to ring %p at [%lu]\n", iocb
, tail
);
743 * Check if the user asked us to deliver the result through an
744 * eventfd. The eventfd_signal() function is safe to be called
747 if (iocb
->ki_eventfd
!= NULL
)
748 eventfd_signal(iocb
->ki_eventfd
, 1);
751 /* everything turned out well, dispose of the aiocb. */
752 __aio_put_req(ctx
, iocb
);
755 * We have to order our ring_info tail store above and test
756 * of the wait list below outside the wait lock. This is
757 * like in wake_up_bit() where clearing a bit has to be
758 * ordered with the unlocked test.
762 if (waitqueue_active(&ctx
->wait
))
765 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
767 EXPORT_SYMBOL(aio_complete
);
770 * Pull an event off of the ioctx's event ring. Returns the number of
771 * events fetched (0 or 1 ;-)
772 * FIXME: make this use cmpxchg.
773 * TODO: make the ringbuffer user mmap()able (requires FIXME).
775 static int aio_read_evt(struct kioctx
*ioctx
, struct io_event
*ent
)
777 struct aio_ring_info
*info
= &ioctx
->ring_info
;
778 struct aio_ring
*ring
;
782 ring
= kmap_atomic(info
->ring_pages
[0]);
783 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
784 (unsigned long)ring
->head
, (unsigned long)ring
->tail
,
785 (unsigned long)ring
->nr
);
787 if (ring
->head
== ring
->tail
)
790 spin_lock(&info
->ring_lock
);
792 head
= ring
->head
% info
->nr
;
793 if (head
!= ring
->tail
) {
794 struct io_event
*evp
= aio_ring_event(info
, head
);
796 head
= (head
+ 1) % info
->nr
;
797 smp_mb(); /* finish reading the event before updatng the head */
800 put_aio_ring_event(evp
);
802 spin_unlock(&info
->ring_lock
);
805 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret
,
806 (unsigned long)ring
->head
, (unsigned long)ring
->tail
);
812 struct timer_list timer
;
814 struct task_struct
*p
;
817 static void timeout_func(unsigned long data
)
819 struct aio_timeout
*to
= (struct aio_timeout
*)data
;
822 wake_up_process(to
->p
);
825 static inline void init_timeout(struct aio_timeout
*to
)
827 setup_timer_on_stack(&to
->timer
, timeout_func
, (unsigned long) to
);
832 static inline void set_timeout(long start_jiffies
, struct aio_timeout
*to
,
833 const struct timespec
*ts
)
835 to
->timer
.expires
= start_jiffies
+ timespec_to_jiffies(ts
);
836 if (time_after(to
->timer
.expires
, jiffies
))
837 add_timer(&to
->timer
);
842 static inline void clear_timeout(struct aio_timeout
*to
)
844 del_singleshot_timer_sync(&to
->timer
);
847 static int read_events(struct kioctx
*ctx
,
848 long min_nr
, long nr
,
849 struct io_event __user
*event
,
850 struct timespec __user
*timeout
)
852 long start_jiffies
= jiffies
;
853 struct task_struct
*tsk
= current
;
854 DECLARE_WAITQUEUE(wait
, tsk
);
858 struct aio_timeout to
;
860 /* needed to zero any padding within an entry (there shouldn't be
863 memset(&ent
, 0, sizeof(ent
));
865 while (likely(i
< nr
)) {
866 ret
= aio_read_evt(ctx
, &ent
);
867 if (unlikely(ret
<= 0))
870 dprintk("read event: %Lx %Lx %Lx %Lx\n",
871 ent
.data
, ent
.obj
, ent
.res
, ent
.res2
);
873 /* Could we split the check in two? */
875 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
876 dprintk("aio: lost an event due to EFAULT.\n");
881 /* Good, event copied to userland, update counts. */
897 if (unlikely(copy_from_user(&ts
, timeout
, sizeof(ts
))))
900 set_timeout(start_jiffies
, &to
, &ts
);
903 while (likely(i
< nr
)) {
904 add_wait_queue_exclusive(&ctx
->wait
, &wait
);
906 set_task_state(tsk
, TASK_INTERRUPTIBLE
);
907 ret
= aio_read_evt(ctx
, &ent
);
912 if (unlikely(ctx
->dead
)) {
916 if (to
.timed_out
) /* Only check after read evt */
918 /* Try to only show up in io wait if there are ops
920 if (ctx
->reqs_active
)
924 if (signal_pending(tsk
)) {
928 /*ret = aio_read_evt(ctx, &ent);*/
931 set_task_state(tsk
, TASK_RUNNING
);
932 remove_wait_queue(&ctx
->wait
, &wait
);
934 if (unlikely(ret
<= 0))
938 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
939 dprintk("aio: lost an event due to EFAULT.\n");
943 /* Good, event copied to userland, update counts. */
951 destroy_timer_on_stack(&to
.timer
);
955 /* Take an ioctx and remove it from the list of ioctx's. Protects
956 * against races with itself via ->dead.
958 static void io_destroy(struct kioctx
*ioctx
)
960 struct mm_struct
*mm
= current
->mm
;
963 /* delete the entry from the list is someone else hasn't already */
964 spin_lock(&mm
->ioctx_lock
);
965 was_dead
= ioctx
->dead
;
967 hlist_del_rcu(&ioctx
->list
);
968 spin_unlock(&mm
->ioctx_lock
);
970 dprintk("aio_release(%p)\n", ioctx
);
971 if (likely(!was_dead
))
972 put_ioctx(ioctx
); /* twice for the list */
977 * Wake up any waiters. The setting of ctx->dead must be seen
978 * by other CPUs at this point. Right now, we rely on the
979 * locking done by the above calls to ensure this consistency.
981 wake_up_all(&ioctx
->wait
);
985 * Create an aio_context capable of receiving at least nr_events.
986 * ctxp must not point to an aio_context that already exists, and
987 * must be initialized to 0 prior to the call. On successful
988 * creation of the aio_context, *ctxp is filled in with the resulting
989 * handle. May fail with -EINVAL if *ctxp is not initialized,
990 * if the specified nr_events exceeds internal limits. May fail
991 * with -EAGAIN if the specified nr_events exceeds the user's limit
992 * of available events. May fail with -ENOMEM if insufficient kernel
993 * resources are available. May fail with -EFAULT if an invalid
994 * pointer is passed for ctxp. Will fail with -ENOSYS if not
997 SYSCALL_DEFINE2(io_setup
, unsigned, nr_events
, aio_context_t __user
*, ctxp
)
999 struct kioctx
*ioctx
= NULL
;
1003 ret
= get_user(ctx
, ctxp
);
1008 if (unlikely(ctx
|| nr_events
== 0)) {
1009 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1014 ioctx
= ioctx_alloc(nr_events
);
1015 ret
= PTR_ERR(ioctx
);
1016 if (!IS_ERR(ioctx
)) {
1017 ret
= put_user(ioctx
->user_id
, ctxp
);
1028 * Destroy the aio_context specified. May cancel any outstanding
1029 * AIOs and block on completion. Will fail with -ENOSYS if not
1030 * implemented. May fail with -EINVAL if the context pointed to
1033 SYSCALL_DEFINE1(io_destroy
, aio_context_t
, ctx
)
1035 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
1036 if (likely(NULL
!= ioctx
)) {
1041 pr_debug("EINVAL: io_destroy: invalid context id\n");
1045 static void aio_advance_iovec(struct kiocb
*iocb
, ssize_t ret
)
1047 struct iovec
*iov
= &iocb
->ki_iovec
[iocb
->ki_cur_seg
];
1051 while (iocb
->ki_cur_seg
< iocb
->ki_nr_segs
&& ret
> 0) {
1052 ssize_t
this = min((ssize_t
)iov
->iov_len
, ret
);
1053 iov
->iov_base
+= this;
1054 iov
->iov_len
-= this;
1055 iocb
->ki_left
-= this;
1057 if (iov
->iov_len
== 0) {
1063 /* the caller should not have done more io than what fit in
1064 * the remaining iovecs */
1065 BUG_ON(ret
> 0 && iocb
->ki_left
== 0);
1068 static ssize_t
aio_rw_vect_retry(struct kiocb
*iocb
)
1070 struct file
*file
= iocb
->ki_filp
;
1071 struct address_space
*mapping
= file
->f_mapping
;
1072 struct inode
*inode
= mapping
->host
;
1073 ssize_t (*rw_op
)(struct kiocb
*, const struct iovec
*,
1074 unsigned long, loff_t
);
1076 unsigned short opcode
;
1078 if ((iocb
->ki_opcode
== IOCB_CMD_PREADV
) ||
1079 (iocb
->ki_opcode
== IOCB_CMD_PREAD
)) {
1080 rw_op
= file
->f_op
->aio_read
;
1081 opcode
= IOCB_CMD_PREADV
;
1083 rw_op
= file
->f_op
->aio_write
;
1084 opcode
= IOCB_CMD_PWRITEV
;
1087 /* This matches the pread()/pwrite() logic */
1088 if (iocb
->ki_pos
< 0)
1091 if (opcode
== IOCB_CMD_PWRITEV
)
1092 file_start_write(file
);
1094 ret
= rw_op(iocb
, &iocb
->ki_iovec
[iocb
->ki_cur_seg
],
1095 iocb
->ki_nr_segs
- iocb
->ki_cur_seg
,
1098 aio_advance_iovec(iocb
, ret
);
1100 /* retry all partial writes. retry partial reads as long as its a
1102 } while (ret
> 0 && iocb
->ki_left
> 0 &&
1103 (opcode
== IOCB_CMD_PWRITEV
||
1104 (!S_ISFIFO(inode
->i_mode
) && !S_ISSOCK(inode
->i_mode
))));
1105 if (opcode
== IOCB_CMD_PWRITEV
)
1106 file_end_write(file
);
1108 /* This means we must have transferred all that we could */
1109 /* No need to retry anymore */
1110 if ((ret
== 0) || (iocb
->ki_left
== 0))
1111 ret
= iocb
->ki_nbytes
- iocb
->ki_left
;
1113 /* If we managed to write some out we return that, rather than
1114 * the eventual error. */
1115 if (opcode
== IOCB_CMD_PWRITEV
1116 && ret
< 0 && ret
!= -EIOCBQUEUED
1117 && iocb
->ki_nbytes
- iocb
->ki_left
)
1118 ret
= iocb
->ki_nbytes
- iocb
->ki_left
;
1123 static ssize_t
aio_fdsync(struct kiocb
*iocb
)
1125 struct file
*file
= iocb
->ki_filp
;
1126 ssize_t ret
= -EINVAL
;
1128 if (file
->f_op
->aio_fsync
)
1129 ret
= file
->f_op
->aio_fsync(iocb
, 1);
1133 static ssize_t
aio_fsync(struct kiocb
*iocb
)
1135 struct file
*file
= iocb
->ki_filp
;
1136 ssize_t ret
= -EINVAL
;
1138 if (file
->f_op
->aio_fsync
)
1139 ret
= file
->f_op
->aio_fsync(iocb
, 0);
1143 static ssize_t
aio_setup_vectored_rw(int type
, struct kiocb
*kiocb
, bool compat
)
1147 #ifdef CONFIG_COMPAT
1149 ret
= compat_rw_copy_check_uvector(type
,
1150 (struct compat_iovec __user
*)kiocb
->ki_buf
,
1151 kiocb
->ki_nbytes
, 1, &kiocb
->ki_inline_vec
,
1155 ret
= rw_copy_check_uvector(type
,
1156 (struct iovec __user
*)kiocb
->ki_buf
,
1157 kiocb
->ki_nbytes
, 1, &kiocb
->ki_inline_vec
,
1162 ret
= rw_verify_area(type
, kiocb
->ki_filp
, &kiocb
->ki_pos
, ret
);
1166 kiocb
->ki_nr_segs
= kiocb
->ki_nbytes
;
1167 kiocb
->ki_cur_seg
= 0;
1168 /* ki_nbytes/left now reflect bytes instead of segs */
1169 kiocb
->ki_nbytes
= ret
;
1170 kiocb
->ki_left
= ret
;
1177 static ssize_t
aio_setup_single_vector(int type
, struct file
* file
, struct kiocb
*kiocb
)
1181 bytes
= rw_verify_area(type
, file
, &kiocb
->ki_pos
, kiocb
->ki_left
);
1185 kiocb
->ki_iovec
= &kiocb
->ki_inline_vec
;
1186 kiocb
->ki_iovec
->iov_base
= kiocb
->ki_buf
;
1187 kiocb
->ki_iovec
->iov_len
= bytes
;
1188 kiocb
->ki_nr_segs
= 1;
1189 kiocb
->ki_cur_seg
= 0;
1195 * Performs the initial checks and aio retry method
1196 * setup for the kiocb at the time of io submission.
1198 static ssize_t
aio_setup_iocb(struct kiocb
*kiocb
, bool compat
)
1200 struct file
*file
= kiocb
->ki_filp
;
1203 switch (kiocb
->ki_opcode
) {
1204 case IOCB_CMD_PREAD
:
1206 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1209 if (unlikely(!access_ok(VERIFY_WRITE
, kiocb
->ki_buf
,
1212 ret
= aio_setup_single_vector(READ
, file
, kiocb
);
1216 if (file
->f_op
->aio_read
)
1217 kiocb
->ki_retry
= aio_rw_vect_retry
;
1219 case IOCB_CMD_PWRITE
:
1221 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1224 if (unlikely(!access_ok(VERIFY_READ
, kiocb
->ki_buf
,
1227 ret
= aio_setup_single_vector(WRITE
, file
, kiocb
);
1231 if (file
->f_op
->aio_write
)
1232 kiocb
->ki_retry
= aio_rw_vect_retry
;
1234 case IOCB_CMD_PREADV
:
1236 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1238 ret
= aio_setup_vectored_rw(READ
, kiocb
, compat
);
1242 if (file
->f_op
->aio_read
)
1243 kiocb
->ki_retry
= aio_rw_vect_retry
;
1245 case IOCB_CMD_PWRITEV
:
1247 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1249 ret
= aio_setup_vectored_rw(WRITE
, kiocb
, compat
);
1253 if (file
->f_op
->aio_write
)
1254 kiocb
->ki_retry
= aio_rw_vect_retry
;
1256 case IOCB_CMD_FDSYNC
:
1258 if (file
->f_op
->aio_fsync
)
1259 kiocb
->ki_retry
= aio_fdsync
;
1261 case IOCB_CMD_FSYNC
:
1263 if (file
->f_op
->aio_fsync
)
1264 kiocb
->ki_retry
= aio_fsync
;
1267 dprintk("EINVAL: io_submit: no operation provided\n");
1271 if (!kiocb
->ki_retry
)
1277 static int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
1278 struct iocb
*iocb
, struct kiocb_batch
*batch
,
1285 /* enforce forwards compatibility on users */
1286 if (unlikely(iocb
->aio_reserved1
|| iocb
->aio_reserved2
)) {
1287 pr_debug("EINVAL: io_submit: reserve field set\n");
1291 /* prevent overflows */
1293 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1294 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1295 ((ssize_t
)iocb
->aio_nbytes
< 0)
1297 pr_debug("EINVAL: io_submit: overflow check\n");
1301 file
= fget(iocb
->aio_fildes
);
1302 if (unlikely(!file
))
1305 req
= aio_get_req(ctx
, batch
); /* returns with 2 references to req */
1306 if (unlikely(!req
)) {
1310 req
->ki_filp
= file
;
1311 if (iocb
->aio_flags
& IOCB_FLAG_RESFD
) {
1313 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1314 * instance of the file* now. The file descriptor must be
1315 * an eventfd() fd, and will be signaled for each completed
1316 * event using the eventfd_signal() function.
1318 req
->ki_eventfd
= eventfd_ctx_fdget((int) iocb
->aio_resfd
);
1319 if (IS_ERR(req
->ki_eventfd
)) {
1320 ret
= PTR_ERR(req
->ki_eventfd
);
1321 req
->ki_eventfd
= NULL
;
1326 ret
= put_user(req
->ki_key
, &user_iocb
->aio_key
);
1327 if (unlikely(ret
)) {
1328 dprintk("EFAULT: aio_key\n");
1332 req
->ki_obj
.user
= user_iocb
;
1333 req
->ki_user_data
= iocb
->aio_data
;
1334 req
->ki_pos
= iocb
->aio_offset
;
1336 req
->ki_buf
= (char __user
*)(unsigned long)iocb
->aio_buf
;
1337 req
->ki_left
= req
->ki_nbytes
= iocb
->aio_nbytes
;
1338 req
->ki_opcode
= iocb
->aio_lio_opcode
;
1340 ret
= aio_setup_iocb(req
, compat
);
1345 spin_lock_irq(&ctx
->ctx_lock
);
1347 * We could have raced with io_destroy() and are currently holding a
1348 * reference to ctx which should be destroyed. We cannot submit IO
1349 * since ctx gets freed as soon as io_submit() puts its reference. The
1350 * check here is reliable: io_destroy() sets ctx->dead before waiting
1351 * for outstanding IO and the barrier between these two is realized by
1352 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1353 * increment ctx->reqs_active before checking for ctx->dead and the
1354 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1355 * don't see ctx->dead set here, io_destroy() waits for our IO to
1360 spin_unlock_irq(&ctx
->ctx_lock
);
1364 if (unlikely(kiocbIsCancelled(req
)))
1367 ret
= req
->ki_retry(req
);
1369 if (ret
!= -EIOCBQUEUED
) {
1371 * There's no easy way to restart the syscall since other AIO's
1372 * may be already running. Just fail this IO with EINTR.
1374 if (unlikely(ret
== -ERESTARTSYS
|| ret
== -ERESTARTNOINTR
||
1375 ret
== -ERESTARTNOHAND
||
1376 ret
== -ERESTART_RESTARTBLOCK
))
1378 aio_complete(req
, ret
, 0);
1381 aio_put_req(req
); /* drop extra ref to req */
1385 aio_put_req(req
); /* drop extra ref to req */
1386 aio_put_req(req
); /* drop i/o ref to req */
1390 long do_io_submit(aio_context_t ctx_id
, long nr
,
1391 struct iocb __user
*__user
*iocbpp
, bool compat
)
1396 struct blk_plug plug
;
1397 struct kiocb_batch batch
;
1399 if (unlikely(nr
< 0))
1402 if (unlikely(nr
> LONG_MAX
/sizeof(*iocbpp
)))
1403 nr
= LONG_MAX
/sizeof(*iocbpp
);
1405 if (unlikely(!access_ok(VERIFY_READ
, iocbpp
, (nr
*sizeof(*iocbpp
)))))
1408 ctx
= lookup_ioctx(ctx_id
);
1409 if (unlikely(!ctx
)) {
1410 pr_debug("EINVAL: io_submit: invalid context id\n");
1414 kiocb_batch_init(&batch
, nr
);
1416 blk_start_plug(&plug
);
1419 * AKPM: should this return a partial result if some of the IOs were
1420 * successfully submitted?
1422 for (i
=0; i
<nr
; i
++) {
1423 struct iocb __user
*user_iocb
;
1426 if (unlikely(__get_user(user_iocb
, iocbpp
+ i
))) {
1431 if (unlikely(copy_from_user(&tmp
, user_iocb
, sizeof(tmp
)))) {
1436 ret
= io_submit_one(ctx
, user_iocb
, &tmp
, &batch
, compat
);
1440 blk_finish_plug(&plug
);
1442 kiocb_batch_free(ctx
, &batch
);
1448 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1449 * the number of iocbs queued. May return -EINVAL if the aio_context
1450 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1451 * *iocbpp[0] is not properly initialized, if the operation specified
1452 * is invalid for the file descriptor in the iocb. May fail with
1453 * -EFAULT if any of the data structures point to invalid data. May
1454 * fail with -EBADF if the file descriptor specified in the first
1455 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1456 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1457 * fail with -ENOSYS if not implemented.
1459 SYSCALL_DEFINE3(io_submit
, aio_context_t
, ctx_id
, long, nr
,
1460 struct iocb __user
* __user
*, iocbpp
)
1462 return do_io_submit(ctx_id
, nr
, iocbpp
, 0);
1466 * Finds a given iocb for cancellation.
1468 static struct kiocb
*lookup_kiocb(struct kioctx
*ctx
, struct iocb __user
*iocb
,
1471 struct list_head
*pos
;
1473 assert_spin_locked(&ctx
->ctx_lock
);
1475 /* TODO: use a hash or array, this sucks. */
1476 list_for_each(pos
, &ctx
->active_reqs
) {
1477 struct kiocb
*kiocb
= list_kiocb(pos
);
1478 if (kiocb
->ki_obj
.user
== iocb
&& kiocb
->ki_key
== key
)
1485 * Attempts to cancel an iocb previously passed to io_submit. If
1486 * the operation is successfully cancelled, the resulting event is
1487 * copied into the memory pointed to by result without being placed
1488 * into the completion queue and 0 is returned. May fail with
1489 * -EFAULT if any of the data structures pointed to are invalid.
1490 * May fail with -EINVAL if aio_context specified by ctx_id is
1491 * invalid. May fail with -EAGAIN if the iocb specified was not
1492 * cancelled. Will fail with -ENOSYS if not implemented.
1494 SYSCALL_DEFINE3(io_cancel
, aio_context_t
, ctx_id
, struct iocb __user
*, iocb
,
1495 struct io_event __user
*, result
)
1497 struct io_event res
;
1499 struct kiocb
*kiocb
;
1503 ret
= get_user(key
, &iocb
->aio_key
);
1507 ctx
= lookup_ioctx(ctx_id
);
1511 spin_lock_irq(&ctx
->ctx_lock
);
1513 kiocb
= lookup_kiocb(ctx
, iocb
, key
);
1515 ret
= kiocb_cancel(ctx
, kiocb
, &res
);
1519 spin_unlock_irq(&ctx
->ctx_lock
);
1522 /* Cancellation succeeded -- copy the result
1523 * into the user's buffer.
1525 if (copy_to_user(result
, &res
, sizeof(res
)))
1535 * Attempts to read at least min_nr events and up to nr events from
1536 * the completion queue for the aio_context specified by ctx_id. If
1537 * it succeeds, the number of read events is returned. May fail with
1538 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1539 * out of range, if timeout is out of range. May fail with -EFAULT
1540 * if any of the memory specified is invalid. May return 0 or
1541 * < min_nr if the timeout specified by timeout has elapsed
1542 * before sufficient events are available, where timeout == NULL
1543 * specifies an infinite timeout. Note that the timeout pointed to by
1544 * timeout is relative and will be updated if not NULL and the
1545 * operation blocks. Will fail with -ENOSYS if not implemented.
1547 SYSCALL_DEFINE5(io_getevents
, aio_context_t
, ctx_id
,
1550 struct io_event __user
*, events
,
1551 struct timespec __user
*, timeout
)
1553 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
1556 if (likely(ioctx
)) {
1557 if (likely(min_nr
<= nr
&& min_nr
>= 0))
1558 ret
= read_events(ioctx
, min_nr
, nr
, events
, timeout
);