2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@redhat.com>
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/module.h>
20 #include <linux/sched.h>
22 #include <linux/file.h>
24 #include <linux/mman.h>
25 #include <linux/slab.h>
26 #include <linux/timer.h>
27 #include <linux/aio.h>
28 #include <linux/module.h>
29 #include <linux/highmem.h>
30 #include <linux/workqueue.h>
32 #include <asm/uaccess.h>
33 #include <asm/mmu_context.h>
36 #define dprintk printk
38 #define dprintk(x...) do { ; } while (0)
41 /*------ sysctl variables----*/
42 atomic_t aio_nr
= ATOMIC_INIT(0); /* current system wide number of aio requests */
43 unsigned aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
44 /*----end sysctl variables---*/
46 static kmem_cache_t
*kiocb_cachep
;
47 static kmem_cache_t
*kioctx_cachep
;
49 static struct workqueue_struct
*aio_wq
;
51 /* Used for rare fput completion. */
52 static void aio_fput_routine(void *);
53 static DECLARE_WORK(fput_work
, aio_fput_routine
, NULL
);
55 static spinlock_t fput_lock
= SPIN_LOCK_UNLOCKED
;
58 static void aio_kick_handler(void *);
61 * Creates the slab caches used by the aio routines, panic on
62 * failure as this is done early during the boot sequence.
64 static int __init
aio_setup(void)
66 kiocb_cachep
= kmem_cache_create("kiocb", sizeof(struct kiocb
),
67 0, SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
69 panic("unable to create kiocb cache\n");
71 kioctx_cachep
= kmem_cache_create("kioctx", sizeof(struct kioctx
),
72 0, SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
74 panic("unable to create kioctx cache");
76 aio_wq
= create_workqueue("aio");
78 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page
));
83 static void aio_free_ring(struct kioctx
*ctx
)
85 struct aio_ring_info
*info
= &ctx
->ring_info
;
88 for (i
=0; i
<info
->nr_pages
; i
++)
89 put_page(info
->ring_pages
[i
]);
91 if (info
->mmap_size
) {
92 down_write(&ctx
->mm
->mmap_sem
);
93 do_munmap(ctx
->mm
, info
->mmap_base
, info
->mmap_size
);
94 up_write(&ctx
->mm
->mmap_sem
);
97 if (info
->ring_pages
&& info
->ring_pages
!= info
->internal_pages
)
98 kfree(info
->ring_pages
);
99 info
->ring_pages
= NULL
;
103 static int aio_setup_ring(struct kioctx
*ctx
)
105 struct aio_ring
*ring
;
106 struct aio_ring_info
*info
= &ctx
->ring_info
;
107 unsigned nr_events
= ctx
->max_reqs
;
111 /* Compensate for the ring buffer's head/tail overlap entry */
112 nr_events
+= 2; /* 1 is required, 2 for good luck */
114 size
= sizeof(struct aio_ring
);
115 size
+= sizeof(struct io_event
) * nr_events
;
116 nr_pages
= (size
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
121 info
->nr_pages
= nr_pages
;
123 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
)) / sizeof(struct io_event
);
126 info
->ring_pages
= info
->internal_pages
;
127 if (nr_pages
> AIO_RING_PAGES
) {
128 info
->ring_pages
= kmalloc(sizeof(struct page
*) * nr_pages
, GFP_KERNEL
);
129 if (!info
->ring_pages
)
131 memset(info
->ring_pages
, 0, sizeof(struct page
*) * nr_pages
);
134 info
->mmap_size
= nr_pages
* PAGE_SIZE
;
135 dprintk("attempting mmap of %lu bytes\n", info
->mmap_size
);
136 down_write(&ctx
->mm
->mmap_sem
);
137 info
->mmap_base
= do_mmap(NULL
, 0, info
->mmap_size
,
138 PROT_READ
|PROT_WRITE
, MAP_ANON
|MAP_PRIVATE
,
140 if (IS_ERR((void *)info
->mmap_base
)) {
141 up_write(&ctx
->mm
->mmap_sem
);
142 printk("mmap err: %ld\n", -info
->mmap_base
);
148 dprintk("mmap address: 0x%08lx\n", info
->mmap_base
);
149 info
->nr_pages
= get_user_pages(current
, ctx
->mm
,
150 info
->mmap_base
, nr_pages
,
151 1, 0, info
->ring_pages
, NULL
);
152 up_write(&ctx
->mm
->mmap_sem
);
154 if (unlikely(info
->nr_pages
!= nr_pages
)) {
159 ctx
->user_id
= info
->mmap_base
;
161 info
->nr
= nr_events
; /* trusted copy */
163 ring
= kmap_atomic(info
->ring_pages
[0], KM_USER0
);
164 ring
->nr
= nr_events
; /* user copy */
165 ring
->id
= ctx
->user_id
;
166 ring
->head
= ring
->tail
= 0;
167 ring
->magic
= AIO_RING_MAGIC
;
168 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
169 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
170 ring
->header_length
= sizeof(struct aio_ring
);
171 kunmap_atomic(ring
, KM_USER0
);
177 /* aio_ring_event: returns a pointer to the event at the given index from
178 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
180 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
181 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
182 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
184 #define aio_ring_event(info, nr, km) ({ \
185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
186 struct io_event *__event; \
187 __event = kmap_atomic( \
188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
189 __event += pos % AIO_EVENTS_PER_PAGE; \
193 #define put_aio_ring_event(event, km) do { \
194 struct io_event *__event = (event); \
196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
200 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
202 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
204 struct mm_struct
*mm
;
207 /* Prevent overflows */
208 if ((nr_events
> (0x10000000U
/ sizeof(struct io_event
))) ||
209 (nr_events
> (0x10000000U
/ sizeof(struct kiocb
)))) {
210 pr_debug("ENOMEM: nr_events too high\n");
211 return ERR_PTR(-EINVAL
);
214 if (nr_events
> aio_max_nr
)
215 return ERR_PTR(-EAGAIN
);
217 ctx
= kmem_cache_alloc(kioctx_cachep
, GFP_KERNEL
);
219 return ERR_PTR(-ENOMEM
);
221 memset(ctx
, 0, sizeof(*ctx
));
222 ctx
->max_reqs
= nr_events
;
223 mm
= ctx
->mm
= current
->mm
;
224 atomic_inc(&mm
->mm_count
);
226 atomic_set(&ctx
->users
, 1);
227 spin_lock_init(&ctx
->ctx_lock
);
228 spin_lock_init(&ctx
->ring_info
.ring_lock
);
229 init_waitqueue_head(&ctx
->wait
);
231 INIT_LIST_HEAD(&ctx
->active_reqs
);
232 INIT_LIST_HEAD(&ctx
->run_list
);
233 INIT_WORK(&ctx
->wq
, aio_kick_handler
, ctx
);
235 if (aio_setup_ring(ctx
) < 0)
238 /* limit the number of system wide aios */
239 atomic_add(ctx
->max_reqs
, &aio_nr
); /* undone by __put_ioctx */
240 if (unlikely(atomic_read(&aio_nr
) > aio_max_nr
))
243 /* now link into global list. kludge. FIXME */
244 write_lock(&mm
->ioctx_list_lock
);
245 ctx
->next
= mm
->ioctx_list
;
246 mm
->ioctx_list
= ctx
;
247 write_unlock(&mm
->ioctx_list_lock
);
249 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
250 ctx
, ctx
->user_id
, current
->mm
, ctx
->ring_info
.nr
);
254 atomic_sub(ctx
->max_reqs
, &aio_nr
);
255 ctx
->max_reqs
= 0; /* prevent __put_ioctx from sub'ing aio_nr */
257 return ERR_PTR(-EAGAIN
);
260 kmem_cache_free(kioctx_cachep
, ctx
);
261 ctx
= ERR_PTR(-ENOMEM
);
263 dprintk("aio: error allocating ioctx %p\n", ctx
);
268 * Cancels all outstanding aio requests on an aio context. Used
269 * when the processes owning a context have all exited to encourage
270 * the rapid destruction of the kioctx.
272 static void aio_cancel_all(struct kioctx
*ctx
)
274 int (*cancel
)(struct kiocb
*, struct io_event
*);
276 spin_lock_irq(&ctx
->ctx_lock
);
278 while (!list_empty(&ctx
->active_reqs
)) {
279 struct list_head
*pos
= ctx
->active_reqs
.next
;
280 struct kiocb
*iocb
= list_kiocb(pos
);
281 list_del_init(&iocb
->ki_list
);
282 cancel
= iocb
->ki_cancel
;
285 spin_unlock_irq(&ctx
->ctx_lock
);
287 spin_lock_irq(&ctx
->ctx_lock
);
290 spin_unlock_irq(&ctx
->ctx_lock
);
293 void wait_for_all_aios(struct kioctx
*ctx
)
295 struct task_struct
*tsk
= current
;
296 DECLARE_WAITQUEUE(wait
, tsk
);
298 if (!ctx
->reqs_active
)
301 add_wait_queue(&ctx
->wait
, &wait
);
302 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
303 while (ctx
->reqs_active
) {
305 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
307 __set_task_state(tsk
, TASK_RUNNING
);
308 remove_wait_queue(&ctx
->wait
, &wait
);
311 /* wait_on_sync_kiocb:
312 * Waits on the given sync kiocb to complete.
314 ssize_t
wait_on_sync_kiocb(struct kiocb
*iocb
)
316 while (iocb
->ki_users
) {
317 set_current_state(TASK_UNINTERRUPTIBLE
);
322 __set_current_state(TASK_RUNNING
);
323 return iocb
->ki_user_data
;
326 /* exit_aio: called when the last user of mm goes away. At this point,
327 * there is no way for any new requests to be submited or any of the
328 * io_* syscalls to be called on the context. However, there may be
329 * outstanding requests which hold references to the context; as they
330 * go away, they will call put_ioctx and release any pinned memory
331 * associated with the request (held via struct page * references).
333 void exit_aio(struct mm_struct
*mm
)
335 struct kioctx
*ctx
= mm
->ioctx_list
;
336 mm
->ioctx_list
= NULL
;
338 struct kioctx
*next
= ctx
->next
;
342 wait_for_all_aios(ctx
);
344 if (1 != atomic_read(&ctx
->users
))
346 "exit_aio:ioctx still alive: %d %d %d\n",
347 atomic_read(&ctx
->users
), ctx
->dead
,
355 * Called when the last user of an aio context has gone away,
356 * and the struct needs to be freed.
358 void __put_ioctx(struct kioctx
*ctx
)
360 unsigned nr_events
= ctx
->max_reqs
;
362 if (unlikely(ctx
->reqs_active
))
368 pr_debug("__put_ioctx: freeing %p\n", ctx
);
369 kmem_cache_free(kioctx_cachep
, ctx
);
371 atomic_sub(nr_events
, &aio_nr
);
375 * Allocate a slot for an aio request. Increments the users count
376 * of the kioctx so that the kioctx stays around until all requests are
377 * complete. Returns NULL if no requests are free.
379 static struct kiocb
*FASTCALL(__aio_get_req(struct kioctx
*ctx
));
380 static struct kiocb
*__aio_get_req(struct kioctx
*ctx
)
382 struct kiocb
*req
= NULL
;
383 struct aio_ring
*ring
;
386 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
390 req
->ki_flags
= 1 << KIF_LOCKED
;
394 req
->ki_cancel
= NULL
;
395 req
->ki_retry
= NULL
;
396 req
->ki_user_obj
= NULL
;
398 /* Check if the completion queue has enough free space to
399 * accept an event from this io.
401 spin_lock_irq(&ctx
->ctx_lock
);
402 ring
= kmap_atomic(ctx
->ring_info
.ring_pages
[0], KM_USER0
);
403 if (ctx
->reqs_active
< aio_ring_avail(&ctx
->ring_info
, ring
)) {
404 list_add(&req
->ki_list
, &ctx
->active_reqs
);
409 kunmap_atomic(ring
, KM_USER0
);
410 spin_unlock_irq(&ctx
->ctx_lock
);
413 kmem_cache_free(kiocb_cachep
, req
);
420 static inline struct kiocb
*aio_get_req(struct kioctx
*ctx
)
423 /* Handle a potential starvation case -- should be exceedingly rare as
424 * requests will be stuck on fput_head only if the aio_fput_routine is
425 * delayed and the requests were the last user of the struct file.
427 req
= __aio_get_req(ctx
);
428 if (unlikely(NULL
== req
)) {
429 aio_fput_routine(NULL
);
430 req
= __aio_get_req(ctx
);
435 static inline void really_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
439 req
->ki_user_obj
= NULL
;
440 kmem_cache_free(kiocb_cachep
, req
);
443 if (unlikely(!ctx
->reqs_active
&& ctx
->dead
))
447 static void aio_fput_routine(void *data
)
449 spin_lock_irq(&fput_lock
);
450 while (likely(!list_empty(&fput_head
))) {
451 struct kiocb
*req
= list_kiocb(fput_head
.next
);
452 struct kioctx
*ctx
= req
->ki_ctx
;
454 list_del(&req
->ki_list
);
455 spin_unlock_irq(&fput_lock
);
457 /* Complete the fput */
458 __fput(req
->ki_filp
);
460 /* Link the iocb into the context's free list */
461 spin_lock_irq(&ctx
->ctx_lock
);
462 really_put_req(ctx
, req
);
463 spin_unlock_irq(&ctx
->ctx_lock
);
466 spin_lock_irq(&fput_lock
);
468 spin_unlock_irq(&fput_lock
);
472 * Returns true if this put was the last user of the request.
474 static int __aio_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
476 dprintk(KERN_DEBUG
"aio_put(%p): f_count=%d\n",
477 req
, atomic_read(&req
->ki_filp
->f_count
));
480 if (unlikely(req
->ki_users
< 0))
482 if (likely(req
->ki_users
))
484 list_del(&req
->ki_list
); /* remove from active_reqs */
485 req
->ki_cancel
= NULL
;
486 req
->ki_retry
= NULL
;
488 /* Must be done under the lock to serialise against cancellation.
489 * Call this aio_fput as it duplicates fput via the fput_work.
491 if (unlikely(atomic_dec_and_test(&req
->ki_filp
->f_count
))) {
493 spin_lock(&fput_lock
);
494 list_add(&req
->ki_list
, &fput_head
);
495 spin_unlock(&fput_lock
);
496 queue_work(aio_wq
, &fput_work
);
498 really_put_req(ctx
, req
);
503 * Returns true if this put was the last user of the kiocb,
504 * false if the request is still in use.
506 int aio_put_req(struct kiocb
*req
)
508 struct kioctx
*ctx
= req
->ki_ctx
;
510 spin_lock_irq(&ctx
->ctx_lock
);
511 ret
= __aio_put_req(ctx
, req
);
512 spin_unlock_irq(&ctx
->ctx_lock
);
518 /* Lookup an ioctx id. ioctx_list is lockless for reads.
519 * FIXME: this is O(n) and is only suitable for development.
521 struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
523 struct kioctx
*ioctx
;
524 struct mm_struct
*mm
;
527 read_lock(&mm
->ioctx_list_lock
);
528 for (ioctx
= mm
->ioctx_list
; ioctx
; ioctx
= ioctx
->next
)
529 if (likely(ioctx
->user_id
== ctx_id
&& !ioctx
->dead
)) {
533 read_unlock(&mm
->ioctx_list_lock
);
538 static void use_mm(struct mm_struct
*mm
)
540 struct mm_struct
*active_mm
= current
->active_mm
;
541 atomic_inc(&mm
->mm_count
);
543 if (mm
!= active_mm
) {
544 current
->active_mm
= mm
;
545 activate_mm(active_mm
, mm
);
550 static void unuse_mm(struct mm_struct
*mm
)
553 /* active_mm is still 'mm' */
554 enter_lazy_tlb(mm
, current
);
557 /* Run on kevent's context. FIXME: needs to be per-cpu and warn if an
560 static void aio_kick_handler(void *data
)
562 struct kioctx
*ctx
= data
;
566 spin_lock_irq(&ctx
->ctx_lock
);
567 while (!list_empty(&ctx
->run_list
)) {
571 iocb
= list_entry(ctx
->run_list
.next
, struct kiocb
,
573 list_del(&iocb
->ki_run_list
);
575 spin_unlock_irq(&ctx
->ctx_lock
);
577 kiocbClearKicked(iocb
);
578 ret
= iocb
->ki_retry(iocb
);
579 if (-EIOCBQUEUED
!= ret
) {
580 aio_complete(iocb
, ret
, 0);
584 spin_lock_irq(&ctx
->ctx_lock
);
586 __aio_put_req(ctx
, iocb
);
588 spin_unlock_irq(&ctx
->ctx_lock
);
593 void kick_iocb(struct kiocb
*iocb
)
595 struct kioctx
*ctx
= iocb
->ki_ctx
;
597 /* sync iocbs are easy: they can only ever be executing from a
599 if (is_sync_kiocb(iocb
)) {
600 kiocbSetKicked(iocb
);
601 wake_up_process(iocb
->ki_user_obj
);
605 if (!kiocbTryKick(iocb
)) {
607 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
608 list_add_tail(&iocb
->ki_run_list
, &ctx
->run_list
);
609 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
610 schedule_work(&ctx
->wq
);
615 * Called when the io request on the given iocb is complete.
616 * Returns true if this is the last user of the request. The
617 * only other user of the request can be the cancellation code.
619 int aio_complete(struct kiocb
*iocb
, long res
, long res2
)
621 struct kioctx
*ctx
= iocb
->ki_ctx
;
622 struct aio_ring_info
*info
;
623 struct aio_ring
*ring
;
624 struct io_event
*event
;
629 /* Special case handling for sync iocbs: events go directly
630 * into the iocb for fast handling. Note that this will not
631 * work if we allow sync kiocbs to be cancelled. in which
632 * case the usage count checks will have to move under ctx_lock
635 if (is_sync_kiocb(iocb
)) {
638 iocb
->ki_user_data
= res
;
639 if (iocb
->ki_users
== 1) {
643 spin_lock_irq(&ctx
->ctx_lock
);
645 ret
= (0 == iocb
->ki_users
);
646 spin_unlock_irq(&ctx
->ctx_lock
);
648 /* sync iocbs put the task here for us */
649 wake_up_process(iocb
->ki_user_obj
);
653 info
= &ctx
->ring_info
;
655 /* add a completion event to the ring buffer.
656 * must be done holding ctx->ctx_lock to prevent
657 * other code from messing with the tail
658 * pointer since we might be called from irq
661 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
663 ring
= kmap_atomic(info
->ring_pages
[0], KM_IRQ1
);
666 event
= aio_ring_event(info
, tail
, KM_IRQ0
);
667 tail
= (tail
+ 1) % info
->nr
;
669 event
->obj
= (u64
)(unsigned long)iocb
->ki_user_obj
;
670 event
->data
= iocb
->ki_user_data
;
674 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
675 ctx
, tail
, iocb
, iocb
->ki_user_obj
, iocb
->ki_user_data
,
678 /* after flagging the request as done, we
679 * must never even look at it again
681 smp_wmb(); /* make event visible before updating tail */
686 put_aio_ring_event(event
, KM_IRQ0
);
687 kunmap_atomic(ring
, KM_IRQ1
);
689 pr_debug("added to ring %p at [%lu]\n", iocb
, tail
);
691 /* everything turned out well, dispose of the aiocb. */
692 ret
= __aio_put_req(ctx
, iocb
);
694 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
696 if (waitqueue_active(&ctx
->wait
))
706 * Pull an event off of the ioctx's event ring. Returns the number of
707 * events fetched (0 or 1 ;-)
708 * FIXME: make this use cmpxchg.
709 * TODO: make the ringbuffer user mmap()able (requires FIXME).
711 static int aio_read_evt(struct kioctx
*ioctx
, struct io_event
*ent
)
713 struct aio_ring_info
*info
= &ioctx
->ring_info
;
714 struct aio_ring
*ring
;
718 ring
= kmap_atomic(info
->ring_pages
[0], KM_USER0
);
719 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
720 (unsigned long)ring
->head
, (unsigned long)ring
->tail
,
721 (unsigned long)ring
->nr
);
723 if (ring
->head
== ring
->tail
)
726 spin_lock(&info
->ring_lock
);
728 head
= ring
->head
% info
->nr
;
729 if (head
!= ring
->tail
) {
730 struct io_event
*evp
= aio_ring_event(info
, head
, KM_USER1
);
732 head
= (head
+ 1) % info
->nr
;
733 smp_mb(); /* finish reading the event before updatng the head */
736 put_aio_ring_event(evp
, KM_USER1
);
738 spin_unlock(&info
->ring_lock
);
741 kunmap_atomic(ring
, KM_USER0
);
742 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret
,
743 (unsigned long)ring
->head
, (unsigned long)ring
->tail
);
748 struct timer_list timer
;
750 struct task_struct
*p
;
753 static void timeout_func(unsigned long data
)
755 struct timeout
*to
= (struct timeout
*)data
;
758 wake_up_process(to
->p
);
761 static inline void init_timeout(struct timeout
*to
)
763 init_timer(&to
->timer
);
764 to
->timer
.data
= (unsigned long)to
;
765 to
->timer
.function
= timeout_func
;
770 static inline void set_timeout(long start_jiffies
, struct timeout
*to
,
771 const struct timespec
*ts
)
773 unsigned long how_long
;
775 if (ts
->tv_sec
< 0 || (!ts
->tv_sec
&& !ts
->tv_nsec
)) {
780 how_long
= ts
->tv_sec
* HZ
;
781 #define HZ_NS (1000000000 / HZ)
782 how_long
+= (ts
->tv_nsec
+ HZ_NS
- 1) / HZ_NS
;
784 to
->timer
.expires
= jiffies
+ how_long
;
785 add_timer(&to
->timer
);
788 static inline void clear_timeout(struct timeout
*to
)
790 del_timer_sync(&to
->timer
);
793 static int read_events(struct kioctx
*ctx
,
794 long min_nr
, long nr
,
795 struct io_event
*event
,
796 struct timespec
*timeout
)
798 long start_jiffies
= jiffies
;
799 struct task_struct
*tsk
= current
;
800 DECLARE_WAITQUEUE(wait
, tsk
);
806 /* needed to zero any padding within an entry (there shouldn't be
809 memset(&ent
, 0, sizeof(ent
));
812 while (likely(i
< nr
)) {
813 ret
= aio_read_evt(ctx
, &ent
);
814 if (unlikely(ret
<= 0))
817 dprintk("read event: %Lx %Lx %Lx %Lx\n",
818 ent
.data
, ent
.obj
, ent
.res
, ent
.res2
);
820 /* Could we split the check in two? */
822 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
823 dprintk("aio: lost an event due to EFAULT.\n");
828 /* Good, event copied to userland, update counts. */
844 if (unlikely(copy_from_user(&ts
, timeout
, sizeof(ts
))))
847 set_timeout(start_jiffies
, &to
, &ts
);
850 while (likely(i
< nr
)) {
851 add_wait_queue_exclusive(&ctx
->wait
, &wait
);
853 set_task_state(tsk
, TASK_INTERRUPTIBLE
);
855 ret
= aio_read_evt(ctx
, &ent
);
861 if (to
.timed_out
) /* Only check after read evt */
864 if (signal_pending(tsk
)) {
868 /*ret = aio_read_evt(ctx, &ent);*/
871 set_task_state(tsk
, TASK_RUNNING
);
872 remove_wait_queue(&ctx
->wait
, &wait
);
874 if (unlikely(ret
<= 0))
878 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
879 dprintk("aio: lost an event due to EFAULT.\n");
883 /* Good, event copied to userland, update counts. */
894 /* Take an ioctx and remove it from the list of ioctx's. Protects
895 * against races with itself via ->dead.
897 static void io_destroy(struct kioctx
*ioctx
)
899 struct mm_struct
*mm
= current
->mm
;
903 /* delete the entry from the list is someone else hasn't already */
904 write_lock(&mm
->ioctx_list_lock
);
905 was_dead
= ioctx
->dead
;
907 for (tmp
= &mm
->ioctx_list
; *tmp
&& *tmp
!= ioctx
;
912 write_unlock(&mm
->ioctx_list_lock
);
914 dprintk("aio_release(%p)\n", ioctx
);
915 if (likely(!was_dead
))
916 put_ioctx(ioctx
); /* twice for the list */
918 aio_cancel_all(ioctx
);
919 wait_for_all_aios(ioctx
);
920 put_ioctx(ioctx
); /* once for the lookup */
924 * Create an aio_context capable of receiving at least nr_events.
925 * ctxp must not point to an aio_context that already exists, and
926 * must be initialized to 0 prior to the call. On successful
927 * creation of the aio_context, *ctxp is filled in with the resulting
928 * handle. May fail with -EINVAL if *ctxp is not initialized,
929 * if the specified nr_events exceeds internal limits. May fail
930 * with -EAGAIN if the specified nr_events exceeds the user's limit
931 * of available events. May fail with -ENOMEM if insufficient kernel
932 * resources are available. May fail with -EFAULT if an invalid
933 * pointer is passed for ctxp. Will fail with -ENOSYS if not
936 asmlinkage
long sys_io_setup(unsigned nr_events
, aio_context_t
*ctxp
)
938 struct kioctx
*ioctx
= NULL
;
942 ret
= get_user(ctx
, ctxp
);
947 if (unlikely(ctx
|| (int)nr_events
<= 0)) {
948 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
952 ioctx
= ioctx_alloc(nr_events
);
953 ret
= PTR_ERR(ioctx
);
954 if (!IS_ERR(ioctx
)) {
955 ret
= put_user(ioctx
->user_id
, ctxp
);
966 * Destroy the aio_context specified. May cancel any outstanding
967 * AIOs and block on completion. Will fail with -ENOSYS if not
968 * implemented. May fail with -EFAULT if the context pointed to
971 asmlinkage
long sys_io_destroy(aio_context_t ctx
)
973 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
974 if (likely(NULL
!= ioctx
)) {
978 pr_debug("EINVAL: io_destroy: invalid context id\n");
982 int io_submit_one(struct kioctx
*ctx
, struct iocb __user
*user_iocb
,
990 /* enforce forwards compatibility on users */
991 if (unlikely(iocb
->aio_reserved1
|| iocb
->aio_reserved2
||
992 iocb
->aio_reserved3
)) {
993 pr_debug("EINVAL: io_submit: reserve field set\n");
997 /* prevent overflows */
999 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1000 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1001 ((ssize_t
)iocb
->aio_nbytes
< 0)
1003 pr_debug("EINVAL: io_submit: overflow check\n");
1007 file
= fget(iocb
->aio_fildes
);
1008 if (unlikely(!file
))
1011 req
= aio_get_req(ctx
);
1012 if (unlikely(!req
)) {
1017 req
->ki_filp
= file
;
1018 iocb
->aio_key
= req
->ki_key
;
1019 ret
= put_user(iocb
->aio_key
, &user_iocb
->aio_key
);
1020 if (unlikely(ret
)) {
1021 dprintk("EFAULT: aio_key\n");
1025 req
->ki_user_obj
= user_iocb
;
1026 req
->ki_user_data
= iocb
->aio_data
;
1027 req
->ki_pos
= iocb
->aio_offset
;
1029 buf
= (char *)(unsigned long)iocb
->aio_buf
;
1031 switch (iocb
->aio_lio_opcode
) {
1032 case IOCB_CMD_PREAD
:
1034 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1037 if (unlikely(!access_ok(VERIFY_WRITE
, buf
, iocb
->aio_nbytes
)))
1040 if (file
->f_op
->aio_read
)
1041 ret
= file
->f_op
->aio_read(req
, buf
,
1042 iocb
->aio_nbytes
, req
->ki_pos
);
1044 case IOCB_CMD_PWRITE
:
1046 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1049 if (unlikely(!access_ok(VERIFY_READ
, buf
, iocb
->aio_nbytes
)))
1052 if (file
->f_op
->aio_write
)
1053 ret
= file
->f_op
->aio_write(req
, buf
,
1054 iocb
->aio_nbytes
, req
->ki_pos
);
1056 case IOCB_CMD_FDSYNC
:
1058 if (file
->f_op
->aio_fsync
)
1059 ret
= file
->f_op
->aio_fsync(req
, 1);
1061 case IOCB_CMD_FSYNC
:
1063 if (file
->f_op
->aio_fsync
)
1064 ret
= file
->f_op
->aio_fsync(req
, 0);
1067 dprintk("EINVAL: io_submit: no operation provided\n");
1071 if (likely(-EIOCBQUEUED
== ret
))
1073 aio_complete(req
, ret
, 0);
1082 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1083 * the number of iocbs queued. May return -EINVAL if the aio_context
1084 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1085 * *iocbpp[0] is not properly initialized, if the operation specified
1086 * is invalid for the file descriptor in the iocb. May fail with
1087 * -EFAULT if any of the data structures point to invalid data. May
1088 * fail with -EBADF if the file descriptor specified in the first
1089 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1090 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1091 * fail with -ENOSYS if not implemented.
1093 asmlinkage
long sys_io_submit(aio_context_t ctx_id
, long nr
,
1094 struct iocb __user
**iocbpp
)
1100 if (unlikely(nr
< 0))
1103 if (unlikely(!access_ok(VERIFY_READ
, iocbpp
, (nr
*sizeof(*iocbpp
)))))
1106 ctx
= lookup_ioctx(ctx_id
);
1107 if (unlikely(!ctx
)) {
1108 pr_debug("EINVAL: io_submit: invalid context id\n");
1113 * AKPM: should this return a partial result if some of the IOs were
1114 * successfully submitted?
1116 for (i
=0; i
<nr
; i
++) {
1117 struct iocb __user
*user_iocb
;
1120 if (unlikely(__get_user(user_iocb
, iocbpp
+ i
))) {
1125 if (unlikely(copy_from_user(&tmp
, user_iocb
, sizeof(tmp
)))) {
1130 ret
= io_submit_one(ctx
, user_iocb
, &tmp
);
1140 * Finds a given iocb for cancellation.
1141 * MUST be called with ctx->ctx_lock held.
1143 struct kiocb
*lookup_kiocb(struct kioctx
*ctx
, struct iocb
*iocb
, u32 key
)
1145 struct list_head
*pos
;
1146 /* TODO: use a hash or array, this sucks. */
1147 list_for_each(pos
, &ctx
->active_reqs
) {
1148 struct kiocb
*kiocb
= list_kiocb(pos
);
1149 if (kiocb
->ki_user_obj
== iocb
&& kiocb
->ki_key
== key
)
1156 * Attempts to cancel an iocb previously passed to io_submit. If
1157 * the operation is successfully cancelled, the resulting event is
1158 * copied into the memory pointed to by result without being placed
1159 * into the completion queue and 0 is returned. May fail with
1160 * -EFAULT if any of the data structures pointed to are invalid.
1161 * May fail with -EINVAL if aio_context specified by ctx_id is
1162 * invalid. May fail with -EAGAIN if the iocb specified was not
1163 * cancelled. Will fail with -ENOSYS if not implemented.
1165 asmlinkage
long sys_io_cancel(aio_context_t ctx_id
, struct iocb
*iocb
,
1166 struct io_event
*result
)
1168 int (*cancel
)(struct kiocb
*iocb
, struct io_event
*res
);
1170 struct kiocb
*kiocb
;
1174 ret
= get_user(key
, &iocb
->aio_key
);
1178 ctx
= lookup_ioctx(ctx_id
);
1182 spin_lock_irq(&ctx
->ctx_lock
);
1184 kiocb
= lookup_kiocb(ctx
, iocb
, key
);
1185 if (kiocb
&& kiocb
->ki_cancel
) {
1186 cancel
= kiocb
->ki_cancel
;
1190 spin_unlock_irq(&ctx
->ctx_lock
);
1192 if (NULL
!= cancel
) {
1193 struct io_event tmp
;
1194 pr_debug("calling cancel\n");
1195 memset(&tmp
, 0, sizeof(tmp
));
1196 tmp
.obj
= (u64
)(unsigned long)kiocb
->ki_user_obj
;
1197 tmp
.data
= kiocb
->ki_user_data
;
1198 ret
= cancel(kiocb
, &tmp
);
1200 /* Cancellation succeeded -- copy the result
1201 * into the user's buffer.
1203 if (copy_to_user(result
, &tmp
, sizeof(tmp
)))
1207 printk(KERN_DEBUG
"iocb has no cancel operation\n");
1215 * Attempts to read at least min_nr events and up to nr events from
1216 * the completion queue for the aio_context specified by ctx_id. May
1217 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1218 * if nr is out of range, if when is out of range. May fail with
1219 * -EFAULT if any of the memory specified to is invalid. May return
1220 * 0 or < min_nr if no events are available and the timeout specified
1221 * by when has elapsed, where when == NULL specifies an infinite
1222 * timeout. Note that the timeout pointed to by when is relative and
1223 * will be updated if not NULL and the operation blocks. Will fail
1224 * with -ENOSYS if not implemented.
1226 asmlinkage
long sys_io_getevents(aio_context_t ctx_id
,
1229 struct io_event
*events
,
1230 struct timespec
*timeout
)
1232 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
1235 if (unlikely(min_nr
> nr
|| min_nr
< 0 || nr
< 0))
1238 if (likely(NULL
!= ioctx
)) {
1239 ret
= read_events(ioctx
, min_nr
, nr
, events
, timeout
);
1246 __initcall(aio_setup
);
1248 EXPORT_SYMBOL(aio_complete
);
1249 EXPORT_SYMBOL(aio_put_req
);
1250 EXPORT_SYMBOL(wait_on_sync_kiocb
);