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/kmap_types.h>
33 #include <asm/uaccess.h>
34 #include <asm/mmu_context.h>
37 #define dprintk printk
39 #define dprintk(x...) do { ; } while (0)
42 /*------ sysctl variables----*/
43 atomic_t aio_nr
= ATOMIC_INIT(0); /* current system wide number of aio requests */
44 unsigned aio_max_nr
= 0x10000; /* system wide maximum number of aio requests */
45 /*----end sysctl variables---*/
47 static kmem_cache_t
*kiocb_cachep
;
48 static kmem_cache_t
*kioctx_cachep
;
50 static struct workqueue_struct
*aio_wq
;
52 /* Used for rare fput completion. */
53 static void aio_fput_routine(void *);
54 static DECLARE_WORK(fput_work
, aio_fput_routine
, NULL
);
56 static spinlock_t fput_lock
= SPIN_LOCK_UNLOCKED
;
59 static void aio_kick_handler(void *);
62 * Creates the slab caches used by the aio routines, panic on
63 * failure as this is done early during the boot sequence.
65 static int __init
aio_setup(void)
67 kiocb_cachep
= kmem_cache_create("kiocb", sizeof(struct kiocb
),
68 0, SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
70 panic("unable to create kiocb cache\n");
72 kioctx_cachep
= kmem_cache_create("kioctx", sizeof(struct kioctx
),
73 0, SLAB_HWCACHE_ALIGN
, NULL
, NULL
);
75 panic("unable to create kioctx cache");
77 aio_wq
= create_workqueue("aio");
79 printk(KERN_NOTICE
"aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page
));
84 static void aio_free_ring(struct kioctx
*ctx
)
86 struct aio_ring_info
*info
= &ctx
->ring_info
;
89 for (i
=0; i
<info
->nr_pages
; i
++)
90 put_page(info
->ring_pages
[i
]);
92 if (info
->mmap_size
) {
93 down_write(&ctx
->mm
->mmap_sem
);
94 do_munmap(ctx
->mm
, info
->mmap_base
, info
->mmap_size
);
95 up_write(&ctx
->mm
->mmap_sem
);
98 if (info
->ring_pages
&& info
->ring_pages
!= info
->internal_pages
)
99 kfree(info
->ring_pages
);
100 info
->ring_pages
= NULL
;
104 static int aio_setup_ring(struct kioctx
*ctx
)
106 struct aio_ring
*ring
;
107 struct aio_ring_info
*info
= &ctx
->ring_info
;
108 unsigned nr_events
= ctx
->max_reqs
;
112 /* Compensate for the ring buffer's head/tail overlap entry */
113 nr_events
+= 2; /* 1 is required, 2 for good luck */
115 size
= sizeof(struct aio_ring
);
116 size
+= sizeof(struct io_event
) * nr_events
;
117 nr_pages
= (size
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
122 info
->nr_pages
= nr_pages
;
124 nr_events
= (PAGE_SIZE
* nr_pages
- sizeof(struct aio_ring
)) / sizeof(struct io_event
);
127 info
->ring_pages
= info
->internal_pages
;
128 if (nr_pages
> AIO_RING_PAGES
) {
129 info
->ring_pages
= kmalloc(sizeof(struct page
*) * nr_pages
, GFP_KERNEL
);
130 if (!info
->ring_pages
)
132 memset(info
->ring_pages
, 0, sizeof(struct page
*) * nr_pages
);
135 info
->mmap_size
= nr_pages
* PAGE_SIZE
;
136 dprintk("attempting mmap of %lu bytes\n", info
->mmap_size
);
137 down_write(&ctx
->mm
->mmap_sem
);
138 info
->mmap_base
= do_mmap(NULL
, 0, info
->mmap_size
,
139 PROT_READ
|PROT_WRITE
, MAP_ANON
|MAP_PRIVATE
,
141 if (IS_ERR((void *)info
->mmap_base
)) {
142 up_write(&ctx
->mm
->mmap_sem
);
143 printk("mmap err: %ld\n", -info
->mmap_base
);
149 dprintk("mmap address: 0x%08lx\n", info
->mmap_base
);
150 info
->nr_pages
= get_user_pages(current
, ctx
->mm
,
151 info
->mmap_base
, info
->mmap_size
,
152 1, 0, info
->ring_pages
, NULL
);
153 up_write(&ctx
->mm
->mmap_sem
);
155 if (unlikely(info
->nr_pages
!= nr_pages
)) {
160 ctx
->user_id
= info
->mmap_base
;
162 info
->nr
= nr_events
; /* trusted copy */
164 ring
= kmap_atomic(info
->ring_pages
[0], KM_USER0
);
165 ring
->nr
= nr_events
; /* user copy */
166 ring
->id
= ctx
->user_id
;
167 ring
->head
= ring
->tail
= 0;
168 ring
->magic
= AIO_RING_MAGIC
;
169 ring
->compat_features
= AIO_RING_COMPAT_FEATURES
;
170 ring
->incompat_features
= AIO_RING_INCOMPAT_FEATURES
;
171 ring
->header_length
= sizeof(struct aio_ring
);
172 kunmap_atomic(ring
, KM_USER0
);
178 /* aio_ring_event: returns a pointer to the event at the given index from
179 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
181 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
182 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
183 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
185 #define aio_ring_event(info, nr, km) ({ \
186 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
187 struct io_event *__event; \
188 __event = kmap_atomic( \
189 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
190 __event += pos % AIO_EVENTS_PER_PAGE; \
194 #define put_aio_ring_event(event, km) do { \
195 struct io_event *__event = (event); \
197 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
201 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
203 static struct kioctx
*ioctx_alloc(unsigned nr_events
)
205 struct mm_struct
*mm
;
208 /* Prevent overflows */
209 if ((nr_events
> (0x10000000U
/ sizeof(struct io_event
))) ||
210 (nr_events
> (0x10000000U
/ sizeof(struct kiocb
)))) {
211 pr_debug("ENOMEM: nr_events too high\n");
212 return ERR_PTR(-EINVAL
);
215 if (nr_events
> aio_max_nr
)
216 return ERR_PTR(-EAGAIN
);
218 ctx
= kmem_cache_alloc(kioctx_cachep
, GFP_KERNEL
);
220 return ERR_PTR(-ENOMEM
);
222 memset(ctx
, 0, sizeof(*ctx
));
223 ctx
->max_reqs
= nr_events
;
224 mm
= ctx
->mm
= current
->mm
;
225 atomic_inc(&mm
->mm_count
);
227 atomic_set(&ctx
->users
, 1);
228 spin_lock_init(&ctx
->ctx_lock
);
229 spin_lock_init(&ctx
->ring_info
.ring_lock
);
230 init_waitqueue_head(&ctx
->wait
);
232 INIT_LIST_HEAD(&ctx
->active_reqs
);
233 INIT_LIST_HEAD(&ctx
->run_list
);
234 INIT_WORK(&ctx
->wq
, aio_kick_handler
, ctx
);
236 if (aio_setup_ring(ctx
) < 0)
239 /* limit the number of system wide aios */
240 atomic_add(ctx
->max_reqs
, &aio_nr
); /* undone by __put_ioctx */
241 if (unlikely(atomic_read(&aio_nr
) > aio_max_nr
))
244 /* now link into global list. kludge. FIXME */
245 write_lock(&mm
->ioctx_list_lock
);
246 ctx
->next
= mm
->ioctx_list
;
247 mm
->ioctx_list
= ctx
;
248 write_unlock(&mm
->ioctx_list_lock
);
250 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
251 ctx
, ctx
->user_id
, current
->mm
, ctx
->ring_info
.ring
->nr
);
255 atomic_sub(ctx
->max_reqs
, &aio_nr
); /* undone by __put_ioctx */
256 ctx
->max_reqs
= 0; /* prevent __put_ioctx from sub'ing aio_nr */
258 return ERR_PTR(-EAGAIN
);
261 kmem_cache_free(kioctx_cachep
, ctx
);
262 ctx
= ERR_PTR(-ENOMEM
);
264 dprintk("aio: error allocating ioctx %p\n", ctx
);
269 * Cancels all outstanding aio requests on an aio context. Used
270 * when the processes owning a context have all exited to encourage
271 * the rapid destruction of the kioctx.
273 static void aio_cancel_all(struct kioctx
*ctx
)
275 int (*cancel
)(struct kiocb
*, struct io_event
*);
277 spin_lock_irq(&ctx
->ctx_lock
);
279 while (!list_empty(&ctx
->active_reqs
)) {
280 struct list_head
*pos
= ctx
->active_reqs
.next
;
281 struct kiocb
*iocb
= list_kiocb(pos
);
282 list_del_init(&iocb
->ki_list
);
283 cancel
= iocb
->ki_cancel
;
286 spin_unlock_irq(&ctx
->ctx_lock
);
289 spin_lock_irq(&ctx
->ctx_lock
);
291 spin_unlock_irq(&ctx
->ctx_lock
);
294 void wait_for_all_aios(struct kioctx
*ctx
)
296 struct task_struct
*tsk
= current
;
297 DECLARE_WAITQUEUE(wait
, tsk
);
299 if (!ctx
->reqs_active
)
302 add_wait_queue(&ctx
->wait
, &wait
);
303 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
304 while (ctx
->reqs_active
) {
305 printk("ctx->reqs_active = %d\n", ctx
->reqs_active
);
307 set_task_state(tsk
, TASK_UNINTERRUPTIBLE
);
309 __set_task_state(tsk
, TASK_RUNNING
);
310 remove_wait_queue(&ctx
->wait
, &wait
);
313 /* wait_on_sync_kiocb:
314 * Waits on the given sync kiocb to complete.
316 ssize_t
wait_on_sync_kiocb(struct kiocb
*iocb
)
318 while (iocb
->ki_users
) {
319 set_current_state(TASK_UNINTERRUPTIBLE
);
324 __set_current_state(TASK_RUNNING
);
325 return iocb
->ki_user_data
;
328 /* exit_aio: called when the last user of mm goes away. At this point,
329 * there is no way for any new requests to be submited or any of the
330 * io_* syscalls to be called on the context. However, there may be
331 * outstanding requests which hold references to the context; as they
332 * go away, they will call put_ioctx and release any pinned memory
333 * associated with the request (held via struct page * references).
335 void exit_aio(struct mm_struct
*mm
)
337 struct kioctx
*ctx
= mm
->ioctx_list
;
338 mm
->ioctx_list
= NULL
;
340 struct kioctx
*next
= ctx
->next
;
344 wait_for_all_aios(ctx
);
346 if (1 != atomic_read(&ctx
->users
))
348 "exit_aio:ioctx still alive: %d %d %d\n",
349 atomic_read(&ctx
->users
), ctx
->dead
,
357 * Called when the last user of an aio context has gone away,
358 * and the struct needs to be freed.
360 void __put_ioctx(struct kioctx
*ctx
)
362 unsigned nr_events
= ctx
->max_reqs
;
364 if (unlikely(ctx
->reqs_active
))
370 pr_debug("__put_ioctx: freeing %p\n", ctx
);
371 kmem_cache_free(kioctx_cachep
, ctx
);
373 atomic_sub(nr_events
, &aio_nr
);
377 * Allocate a slot for an aio request. Increments the users count
378 * of the kioctx so that the kioctx stays around until all requests are
379 * complete. Returns -EAGAIN if no requests are free.
381 static struct kiocb
*FASTCALL(__aio_get_req(struct kioctx
*ctx
));
382 static struct kiocb
*__aio_get_req(struct kioctx
*ctx
)
384 struct kiocb
*req
= NULL
;
385 struct aio_ring
*ring
;
388 req
= kmem_cache_alloc(kiocb_cachep
, GFP_KERNEL
);
392 req
->ki_flags
= 1 << KIF_LOCKED
;
396 req
->ki_cancel
= NULL
;
397 req
->ki_retry
= NULL
;
398 req
->ki_user_obj
= NULL
;
400 /* Check if the completion queue has enough free space to
401 * accept an event from this io.
403 spin_lock_irq(&ctx
->ctx_lock
);
404 ring
= kmap_atomic(ctx
->ring_info
.ring_pages
[0], KM_USER0
);
405 if (ctx
->reqs_active
< aio_ring_avail(&ctx
->ring_info
, ring
)) {
406 list_add(&req
->ki_list
, &ctx
->active_reqs
);
409 req
->ki_user_obj
= NULL
;
413 kmem_cache_free(kiocb_cachep
, req
);
416 kunmap_atomic(ring
, KM_USER0
);
417 spin_unlock_irq(&ctx
->ctx_lock
);
420 kmem_cache_free(kiocb_cachep
, req
);
427 static inline struct kiocb
*aio_get_req(struct kioctx
*ctx
)
430 /* Handle a potential starvation case -- should be exceedingly rare as
431 * requests will be stuck on fput_head only if the aio_fput_routine is
432 * delayed and the requests were the last user of the struct file.
434 req
= __aio_get_req(ctx
);
435 if (unlikely(NULL
== req
)) {
436 aio_fput_routine(NULL
);
437 req
= __aio_get_req(ctx
);
442 static inline void really_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
446 req
->ki_user_obj
= NULL
;
447 kmem_cache_free(kiocb_cachep
, req
);
450 if (unlikely(!ctx
->reqs_active
&& ctx
->dead
))
454 static void aio_fput_routine(void *data
)
456 spin_lock_irq(&fput_lock
);
457 while (likely(!list_empty(&fput_head
))) {
458 struct kiocb
*req
= list_kiocb(fput_head
.next
);
459 struct kioctx
*ctx
= req
->ki_ctx
;
461 list_del(&req
->ki_list
);
462 spin_unlock_irq(&fput_lock
);
464 /* Complete the fput */
465 __fput(req
->ki_filp
);
467 /* Link the iocb into the context's free list */
468 spin_lock_irq(&ctx
->ctx_lock
);
469 really_put_req(ctx
, req
);
470 spin_unlock_irq(&ctx
->ctx_lock
);
473 spin_lock_irq(&fput_lock
);
475 spin_unlock_irq(&fput_lock
);
479 * Returns true if this put was the last user of the request.
481 static inline int __aio_put_req(struct kioctx
*ctx
, struct kiocb
*req
)
483 dprintk(KERN_DEBUG
"aio_put(%p): f_count=%d\n",
484 req
, atomic_read(&req
->ki_filp
->f_count
));
487 if (unlikely(req
->ki_users
< 0))
489 if (likely(req
->ki_users
))
491 list_del(&req
->ki_list
); /* remove from active_reqs */
492 req
->ki_cancel
= NULL
;
493 req
->ki_retry
= NULL
;
495 /* Must be done under the lock to serialise against cancellation.
496 * Call this aio_fput as it duplicates fput via the fput_work.
498 if (unlikely(atomic_dec_and_test(&req
->ki_filp
->f_count
))) {
500 spin_lock(&fput_lock
);
501 list_add(&req
->ki_list
, &fput_head
);
502 spin_unlock(&fput_lock
);
503 queue_work(aio_wq
, &fput_work
);
505 really_put_req(ctx
, req
);
510 * Returns true if this put was the last user of the kiocb,
511 * false if the request is still in use.
513 int aio_put_req(struct kiocb
*req
)
515 struct kioctx
*ctx
= req
->ki_ctx
;
517 spin_lock_irq(&ctx
->ctx_lock
);
518 ret
= __aio_put_req(ctx
, req
);
519 spin_unlock_irq(&ctx
->ctx_lock
);
525 /* Lookup an ioctx id. ioctx_list is lockless for reads.
526 * FIXME: this is O(n) and is only suitable for development.
528 static inline struct kioctx
*lookup_ioctx(unsigned long ctx_id
)
530 struct kioctx
*ioctx
;
531 struct mm_struct
*mm
;
534 read_lock(&mm
->ioctx_list_lock
);
535 for (ioctx
= mm
->ioctx_list
; ioctx
; ioctx
= ioctx
->next
)
536 if (likely(ioctx
->user_id
== ctx_id
&& !ioctx
->dead
)) {
540 read_unlock(&mm
->ioctx_list_lock
);
545 static void use_mm(struct mm_struct
*mm
)
547 struct mm_struct
*active_mm
= current
->active_mm
;
548 atomic_inc(&mm
->mm_count
);
550 if (mm
!= active_mm
) {
551 current
->active_mm
= mm
;
552 activate_mm(active_mm
, mm
);
557 static void unuse_mm(struct mm_struct
*mm
)
560 /* active_mm is still 'mm' */
561 enter_lazy_tlb(mm
, current
, smp_processor_id());
564 /* Run on kevent's context. FIXME: needs to be per-cpu and warn if an
567 static void aio_kick_handler(void *data
)
569 struct kioctx
*ctx
= data
;
573 spin_lock_irq(&ctx
->ctx_lock
);
574 while (!list_empty(&ctx
->run_list
)) {
578 iocb
= list_entry(ctx
->run_list
.next
, struct kiocb
,
580 list_del(&iocb
->ki_run_list
);
582 spin_unlock_irq(&ctx
->ctx_lock
);
584 kiocbClearKicked(iocb
);
585 ret
= iocb
->ki_retry(iocb
);
586 if (-EIOCBQUEUED
!= ret
) {
587 aio_complete(iocb
, ret
, 0);
591 spin_lock_irq(&ctx
->ctx_lock
);
593 __aio_put_req(ctx
, iocb
);
595 spin_unlock_irq(&ctx
->ctx_lock
);
600 void kick_iocb(struct kiocb
*iocb
)
602 struct kioctx
*ctx
= iocb
->ki_ctx
;
604 /* sync iocbs are easy: they can only ever be executing from a
606 if (is_sync_kiocb(iocb
)) {
607 kiocbSetKicked(iocb
);
608 wake_up_process(iocb
->ki_user_obj
);
612 if (!kiocbTryKick(iocb
)) {
614 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
615 list_add_tail(&iocb
->ki_run_list
, &ctx
->run_list
);
616 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
617 schedule_work(&ctx
->wq
);
622 * Called when the io request on the given iocb is complete.
623 * Returns true if this is the last user of the request. The
624 * only other user of the request can be the cancellation code.
626 int aio_complete(struct kiocb
*iocb
, long res
, long res2
)
628 struct kioctx
*ctx
= iocb
->ki_ctx
;
629 struct aio_ring_info
*info
;
630 struct aio_ring
*ring
;
631 struct io_event
*event
;
636 /* Special case handling for sync iocbs: events go directly
637 * into the iocb for fast handling. Note that this will not
638 * work if we allow sync kiocbs to be cancelled. in which
639 * case the usage count checks will have to move under ctx_lock
642 if (is_sync_kiocb(iocb
)) {
645 iocb
->ki_user_data
= res
;
646 if (iocb
->ki_users
== 1) {
650 spin_lock_irq(&ctx
->ctx_lock
);
652 ret
= (0 == iocb
->ki_users
);
653 spin_unlock_irq(&ctx
->ctx_lock
);
655 /* sync iocbs put the task here for us */
656 wake_up_process(iocb
->ki_user_obj
);
660 info
= &ctx
->ring_info
;
662 /* add a completion event to the ring buffer.
663 * must be done holding ctx->ctx_lock to prevent
664 * other code from messing with the tail
665 * pointer since we might be called from irq
668 spin_lock_irqsave(&ctx
->ctx_lock
, flags
);
670 ring
= kmap_atomic(info
->ring_pages
[0], KM_IRQ1
);
673 event
= aio_ring_event(info
, tail
, KM_IRQ0
);
674 tail
= (tail
+ 1) % info
->nr
;
676 event
->obj
= (u64
)(unsigned long)iocb
->ki_user_obj
;
677 event
->data
= iocb
->ki_user_data
;
681 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
682 ctx
, tail
, iocb
, iocb
->ki_user_obj
, iocb
->ki_user_data
,
685 /* after flagging the request as done, we
686 * must never even look at it again
694 put_aio_ring_event(event
, KM_IRQ0
);
695 kunmap_atomic(ring
, KM_IRQ1
);
697 pr_debug("added to ring %p at [%lu]\n", iocb
, tail
);
699 /* everything turned out well, dispose of the aiocb. */
700 ret
= __aio_put_req(ctx
, iocb
);
702 spin_unlock_irqrestore(&ctx
->ctx_lock
, flags
);
704 if (waitqueue_active(&ctx
->wait
))
714 * Pull an event off of the ioctx's event ring. Returns the number of
715 * events fetched (0 or 1 ;-)
716 * FIXME: make this use cmpxchg.
717 * TODO: make the ringbuffer user mmap()able (requires FIXME).
719 static int aio_read_evt(struct kioctx
*ioctx
, struct io_event
*ent
)
721 struct aio_ring_info
*info
= &ioctx
->ring_info
;
722 struct aio_ring
*ring
;
726 ring
= kmap_atomic(info
->ring_pages
[0], KM_USER0
);
727 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
728 (unsigned long)ring
->head
, (unsigned long)ring
->tail
,
729 (unsigned long)ring
->nr
);
731 if (ring
->head
== ring
->tail
)
734 spin_lock(&info
->ring_lock
);
736 head
= ring
->head
% info
->nr
;
737 if (head
!= ring
->tail
) {
738 struct io_event
*evp
= aio_ring_event(info
, head
, KM_USER1
);
740 head
= (head
+ 1) % info
->nr
;
744 put_aio_ring_event(evp
, KM_USER1
);
746 spin_unlock(&info
->ring_lock
);
749 kunmap_atomic(ring
, KM_USER0
);
750 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret
,
751 (unsigned long)ring
->head
, (unsigned long)ring
->tail
);
756 struct timer_list timer
;
758 struct task_struct
*p
;
761 static void timeout_func(unsigned long data
)
763 struct timeout
*to
= (struct timeout
*)data
;
766 wake_up_process(to
->p
);
769 static inline void init_timeout(struct timeout
*to
)
771 init_timer(&to
->timer
);
772 to
->timer
.data
= (unsigned long)to
;
773 to
->timer
.function
= timeout_func
;
778 static inline void set_timeout(long start_jiffies
, struct timeout
*to
,
779 const struct timespec
*ts
)
781 unsigned long how_long
;
783 if (ts
->tv_sec
< 0 || (!ts
->tv_sec
&& !ts
->tv_nsec
)) {
788 how_long
= ts
->tv_sec
* HZ
;
789 #define HZ_NS (1000000000 / HZ)
790 how_long
+= (ts
->tv_nsec
+ HZ_NS
- 1) / HZ_NS
;
792 to
->timer
.expires
= jiffies
+ how_long
;
793 add_timer(&to
->timer
);
796 static inline void clear_timeout(struct timeout
*to
)
798 del_timer_sync(&to
->timer
);
801 static int read_events(struct kioctx
*ctx
,
802 long min_nr
, long nr
,
803 struct io_event
*event
,
804 struct timespec
*timeout
)
806 long start_jiffies
= jiffies
;
807 struct task_struct
*tsk
= current
;
808 DECLARE_WAITQUEUE(wait
, tsk
);
814 /* needed to zero any padding within an entry (there shouldn't be
817 memset(&ent
, 0, sizeof(ent
));
820 while (likely(i
< nr
)) {
821 ret
= aio_read_evt(ctx
, &ent
);
822 if (unlikely(ret
<= 0))
825 dprintk("read event: %Lx %Lx %Lx %Lx\n",
826 ent
.data
, ent
.obj
, ent
.res
, ent
.res2
);
828 /* Could we split the check in two? */
830 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
831 dprintk("aio: lost an event due to EFAULT.\n");
836 /* Good, event copied to userland, update counts. */
851 if (unlikely(copy_from_user(&ts
, timeout
, sizeof(ts
))))
855 set_timeout(start_jiffies
, &to
, &ts
);
858 while (likely(i
< nr
)) {
859 add_wait_queue_exclusive(&ctx
->wait
, &wait
);
861 set_task_state(tsk
, TASK_INTERRUPTIBLE
);
863 ret
= aio_read_evt(ctx
, &ent
);
869 if (to
.timed_out
) /* Only check after read evt */
872 if (signal_pending(tsk
)) {
876 /*ret = aio_read_evt(ctx, &ent);*/
879 set_task_state(tsk
, TASK_RUNNING
);
880 remove_wait_queue(&ctx
->wait
, &wait
);
882 if (unlikely(ret
<= 0))
886 if (unlikely(copy_to_user(event
, &ent
, sizeof(ent
)))) {
887 dprintk("aio: lost an event due to EFAULT.\n");
891 /* Good, event copied to userland, update counts. */
902 /* Take an ioctx and remove it from the list of ioctx's. Protects
903 * against races with itself via ->dead.
905 static void io_destroy(struct kioctx
*ioctx
)
907 struct mm_struct
*mm
= current
->mm
;
911 /* delete the entry from the list is someone else hasn't already */
912 write_lock(&mm
->ioctx_list_lock
);
913 was_dead
= ioctx
->dead
;
915 for (tmp
= &mm
->ioctx_list
; *tmp
&& *tmp
!= ioctx
;
920 write_unlock(&mm
->ioctx_list_lock
);
922 dprintk("aio_release(%p)\n", ioctx
);
923 if (likely(!was_dead
))
924 put_ioctx(ioctx
); /* twice for the list */
926 aio_cancel_all(ioctx
);
927 wait_for_all_aios(ioctx
);
928 put_ioctx(ioctx
); /* once for the lookup */
932 * Create an aio_context capable of receiving at least nr_events.
933 * ctxp must not point to an aio_context that already exists, and
934 * must be initialized to 0 prior to the call. On successful
935 * creation of the aio_context, *ctxp is filled in with the resulting
936 * handle. May fail with -EINVAL if *ctxp is not initialized,
937 * if the specified nr_events exceeds internal limits. May fail
938 * with -EAGAIN if the specified nr_events exceeds the user's limit
939 * of available events. May fail with -ENOMEM if insufficient kernel
940 * resources are available. May fail with -EFAULT if an invalid
941 * pointer is passed for ctxp. Will fail with -ENOSYS if not
944 asmlinkage
long sys_io_setup(unsigned nr_events
, aio_context_t
*ctxp
)
946 struct kioctx
*ioctx
= NULL
;
950 ret
= get_user(ctx
, ctxp
);
955 if (unlikely(ctx
|| !nr_events
|| (int)nr_events
< 0)) {
956 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
960 ioctx
= ioctx_alloc(nr_events
);
961 ret
= PTR_ERR(ioctx
);
962 if (!IS_ERR(ioctx
)) {
963 ret
= put_user(ioctx
->user_id
, ctxp
);
974 * Destroy the aio_context specified. May cancel any outstanding
975 * AIOs and block on completion. Will fail with -ENOSYS if not
976 * implemented. May fail with -EFAULT if the context pointed to
979 asmlinkage
long sys_io_destroy(aio_context_t ctx
)
981 struct kioctx
*ioctx
= lookup_ioctx(ctx
);
982 if (likely(NULL
!= ioctx
)) {
986 pr_debug("EINVAL: io_destroy: invalid context id\n");
990 static int FASTCALL(io_submit_one(struct kioctx
*ctx
, struct iocb
*user_iocb
,
992 static int io_submit_one(struct kioctx
*ctx
, struct iocb
*user_iocb
,
1000 /* enforce forwards compatibility on users */
1001 if (unlikely(iocb
->aio_reserved1
|| iocb
->aio_reserved2
||
1002 iocb
->aio_reserved3
)) {
1003 pr_debug("EINVAL: io_submit: reserve field set\n");
1007 /* prevent overflows */
1009 (iocb
->aio_buf
!= (unsigned long)iocb
->aio_buf
) ||
1010 (iocb
->aio_nbytes
!= (size_t)iocb
->aio_nbytes
) ||
1011 ((ssize_t
)iocb
->aio_nbytes
< 0)
1013 pr_debug("EINVAL: io_submit: overflow check\n");
1017 file
= fget(iocb
->aio_fildes
);
1018 if (unlikely(!file
))
1021 req
= aio_get_req(ctx
);
1022 if (unlikely(!req
)) {
1027 req
->ki_filp
= file
;
1028 iocb
->aio_key
= req
->ki_key
;
1029 ret
= put_user(iocb
->aio_key
, &user_iocb
->aio_key
);
1030 if (unlikely(ret
)) {
1031 dprintk("EFAULT: aio_key\n");
1035 req
->ki_user_obj
= user_iocb
;
1036 req
->ki_user_data
= iocb
->aio_data
;
1037 req
->ki_pos
= iocb
->aio_offset
;
1039 buf
= (char *)(unsigned long)iocb
->aio_buf
;
1041 switch (iocb
->aio_lio_opcode
) {
1042 case IOCB_CMD_PREAD
:
1044 if (unlikely(!(file
->f_mode
& FMODE_READ
)))
1047 if (unlikely(!access_ok(VERIFY_WRITE
, buf
, iocb
->aio_nbytes
)))
1050 if (file
->f_op
->aio_read
)
1051 ret
= file
->f_op
->aio_read(req
, buf
,
1052 iocb
->aio_nbytes
, req
->ki_pos
);
1054 case IOCB_CMD_PWRITE
:
1056 if (unlikely(!(file
->f_mode
& FMODE_WRITE
)))
1059 if (unlikely(!access_ok(VERIFY_READ
, buf
, iocb
->aio_nbytes
)))
1062 if (file
->f_op
->aio_write
)
1063 ret
= file
->f_op
->aio_write(req
, buf
,
1064 iocb
->aio_nbytes
, req
->ki_pos
);
1066 case IOCB_CMD_FDSYNC
:
1068 if (file
->f_op
->aio_fsync
)
1069 ret
= file
->f_op
->aio_fsync(req
, 1);
1071 case IOCB_CMD_FSYNC
:
1073 if (file
->f_op
->aio_fsync
)
1074 ret
= file
->f_op
->aio_fsync(req
, 0);
1077 dprintk("EINVAL: io_submit: no operation provided\n");
1081 if (likely(-EIOCBQUEUED
== ret
))
1083 aio_complete(req
, ret
, 0);
1092 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1093 * the number of iocbs queued. May return -EINVAL if the aio_context
1094 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1095 * *iocbpp[0] is not properly initialized, if the operation specified
1096 * is invalid for the file descriptor in the iocb. May fail with
1097 * -EFAULT if any of the data structures point to invalid data. May
1098 * fail with -EBADF if the file descriptor specified in the first
1099 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1100 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1101 * fail with -ENOSYS if not implemented.
1103 asmlinkage
long sys_io_submit(aio_context_t ctx_id
, long nr
,
1104 struct iocb
**iocbpp
)
1110 if (unlikely(nr
< 0))
1113 if (unlikely(!access_ok(VERIFY_READ
, iocbpp
, (nr
*sizeof(*iocbpp
)))))
1116 ctx
= lookup_ioctx(ctx_id
);
1117 if (unlikely(!ctx
)) {
1118 pr_debug("EINVAL: io_submit: invalid context id\n");
1122 for (i
=0; i
<nr
; i
++) {
1123 struct iocb
*user_iocb
, tmp
;
1125 if (unlikely(__get_user(user_iocb
, iocbpp
+ i
))) {
1130 if (unlikely(copy_from_user(&tmp
, user_iocb
, sizeof(tmp
)))) {
1135 ret
= io_submit_one(ctx
, user_iocb
, &tmp
);
1145 * Finds a given iocb for cancellation.
1146 * MUST be called with ctx->ctx_lock held.
1148 struct kiocb
*lookup_kiocb(struct kioctx
*ctx
, struct iocb
*iocb
, u32 key
)
1150 struct list_head
*pos
;
1151 /* TODO: use a hash or array, this sucks. */
1152 list_for_each(pos
, &ctx
->active_reqs
) {
1153 struct kiocb
*kiocb
= list_kiocb(pos
);
1154 if (kiocb
->ki_user_obj
== iocb
&& kiocb
->ki_key
== key
)
1161 * Attempts to cancel an iocb previously passed to io_submit. If
1162 * the operation is successfully cancelled, the resulting event is
1163 * copied into the memory pointed to by result without being placed
1164 * into the completion queue and 0 is returned. May fail with
1165 * -EFAULT if any of the data structures pointed to are invalid.
1166 * May fail with -EINVAL if aio_context specified by ctx_id is
1167 * invalid. May fail with -EAGAIN if the iocb specified was not
1168 * cancelled. Will fail with -ENOSYS if not implemented.
1170 asmlinkage
long sys_io_cancel(aio_context_t ctx_id
, struct iocb
*iocb
,
1171 struct io_event
*result
)
1173 int (*cancel
)(struct kiocb
*iocb
, struct io_event
*res
);
1175 struct kiocb
*kiocb
;
1179 ret
= get_user(key
, &iocb
->aio_key
);
1183 ctx
= lookup_ioctx(ctx_id
);
1187 spin_lock_irq(&ctx
->ctx_lock
);
1189 kiocb
= lookup_kiocb(ctx
, iocb
, key
);
1190 if (kiocb
&& kiocb
->ki_cancel
) {
1191 cancel
= kiocb
->ki_cancel
;
1195 spin_unlock_irq(&ctx
->ctx_lock
);
1197 if (NULL
!= cancel
) {
1198 struct io_event tmp
;
1199 printk("calling cancel\n");
1200 ret
= cancel(kiocb
, &tmp
);
1202 /* Cancellation succeeded -- copy the result
1203 * into the user's buffer.
1205 if (copy_to_user(result
, &tmp
, sizeof(tmp
)))
1209 printk(KERN_DEBUG
"iocb has no cancel operation\n");
1217 * Attempts to read at least min_nr events and up to nr events from
1218 * the completion queue for the aio_context specified by ctx_id. May
1219 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1220 * if nr is out of range, if when is out of range. May fail with
1221 * -EFAULT if any of the memory specified to is invalid. May return
1222 * 0 or < min_nr if no events are available and the timeout specified
1223 * by when has elapsed, where when == NULL specifies an infinite
1224 * timeout. Note that the timeout pointed to by when is relative and
1225 * will be updated if not NULL and the operation blocks. Will fail
1226 * with -ENOSYS if not implemented.
1228 asmlinkage
long sys_io_getevents(aio_context_t ctx_id
,
1231 struct io_event
*events
,
1232 struct timespec
*timeout
)
1234 struct kioctx
*ioctx
= lookup_ioctx(ctx_id
);
1237 if (unlikely(min_nr
> nr
|| min_nr
< 0 || nr
< 0))
1240 if (likely(NULL
!= ioctx
)) {
1241 ret
= read_events(ioctx
, min_nr
, nr
, events
, timeout
);
1248 __initcall(aio_setup
);
1250 EXPORT_SYMBOL(aio_complete
);
1251 EXPORT_SYMBOL(aio_put_req
);
1252 EXPORT_SYMBOL(wait_on_sync_kiocb
);