2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/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>
42 #define AIO_RING_MAGIC 0xa10a10a1
43 #define AIO_RING_COMPAT_FEATURES 1
44 #define AIO_RING_INCOMPAT_FEATURES 0
46 unsigned id; /* kernel internal index number */
47 unsigned nr; /* number of io_events */
52 unsigned compat_features;
53 unsigned incompat_features;
54 unsigned header_length; /* size of aio_ring */
57 struct io_event io_events[0];
58 }; /* 128 bytes + ring size */
60 #define AIO_RING_PAGES 8
61 struct aio_ring_info {
62 unsigned long mmap_base;
63 unsigned long mmap_size;
65 struct page **ring_pages;
71 struct page *internal_pages[AIO_RING_PAGES];
74 static inline unsigned aio_ring_avail(struct aio_ring_info *info,
75 struct aio_ring *ring)
77 return (ring->head + info->nr - 1 - ring->tail) % info->nr;
84 /* This needs improving */
85 unsigned long user_id;
86 struct hlist_node list;
88 wait_queue_head_t wait;
93 struct list_head active_reqs; /* used for cancellation */
95 /* sys_io_setup currently limits this to an unsigned int */
98 struct aio_ring_info ring_info;
100 struct rcu_head rcu_head;
103 /*------ sysctl variables----*/
104 static DEFINE_SPINLOCK(aio_nr_lock);
105 unsigned long aio_nr; /* current system wide number of aio requests */
106 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
107 /*----end sysctl variables---*/
109 static struct kmem_cache *kiocb_cachep;
110 static struct kmem_cache *kioctx_cachep;
113 * Creates the slab caches used by the aio routines, panic on
114 * failure as this is done early during the boot sequence.
116 static int __init aio_setup(void)
118 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
119 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
121 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
125 __initcall(aio_setup);
127 static void aio_free_ring(struct kioctx *ctx)
129 struct aio_ring_info *info = &ctx->ring_info;
132 for (i=0; i<info->nr_pages; i++)
133 put_page(info->ring_pages[i]);
135 if (info->mmap_size) {
136 vm_munmap(info->mmap_base, info->mmap_size);
139 if (info->ring_pages && info->ring_pages != info->internal_pages)
140 kfree(info->ring_pages);
141 info->ring_pages = NULL;
145 static int aio_setup_ring(struct kioctx *ctx)
147 struct aio_ring *ring;
148 struct aio_ring_info *info = &ctx->ring_info;
149 unsigned nr_events = ctx->max_reqs;
150 struct mm_struct *mm = current->mm;
151 unsigned long size, populate;
154 /* Compensate for the ring buffer's head/tail overlap entry */
155 nr_events += 2; /* 1 is required, 2 for good luck */
157 size = sizeof(struct aio_ring);
158 size += sizeof(struct io_event) * nr_events;
159 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
164 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
167 info->ring_pages = info->internal_pages;
168 if (nr_pages > AIO_RING_PAGES) {
169 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
170 if (!info->ring_pages)
174 info->mmap_size = nr_pages * PAGE_SIZE;
175 pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
176 down_write(&mm->mmap_sem);
177 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
178 PROT_READ|PROT_WRITE,
179 MAP_ANONYMOUS|MAP_PRIVATE, 0,
181 if (IS_ERR((void *)info->mmap_base)) {
182 up_write(&mm->mmap_sem);
188 pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
189 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
190 1, 0, info->ring_pages, NULL);
191 up_write(&mm->mmap_sem);
193 if (unlikely(info->nr_pages != nr_pages)) {
198 mm_populate(info->mmap_base, populate);
200 ctx->user_id = info->mmap_base;
202 info->nr = nr_events; /* trusted copy */
204 ring = kmap_atomic(info->ring_pages[0]);
205 ring->nr = nr_events; /* user copy */
206 ring->id = ctx->user_id;
207 ring->head = ring->tail = 0;
208 ring->magic = AIO_RING_MAGIC;
209 ring->compat_features = AIO_RING_COMPAT_FEATURES;
210 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
211 ring->header_length = sizeof(struct aio_ring);
218 /* aio_ring_event: returns a pointer to the event at the given index from
219 * kmap_atomic(). Release the pointer with put_aio_ring_event();
221 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
222 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
223 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
225 #define aio_ring_event(info, nr) ({ \
226 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
227 struct io_event *__event; \
228 __event = kmap_atomic( \
229 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
230 __event += pos % AIO_EVENTS_PER_PAGE; \
234 #define put_aio_ring_event(event) do { \
235 struct io_event *__event = (event); \
237 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
240 static void ctx_rcu_free(struct rcu_head *head)
242 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
243 kmem_cache_free(kioctx_cachep, ctx);
247 * Called when the last user of an aio context has gone away,
248 * and the struct needs to be freed.
250 static void __put_ioctx(struct kioctx *ctx)
252 unsigned nr_events = ctx->max_reqs;
253 BUG_ON(ctx->reqs_active);
257 spin_lock(&aio_nr_lock);
258 BUG_ON(aio_nr - nr_events > aio_nr);
260 spin_unlock(&aio_nr_lock);
262 pr_debug("freeing %p\n", ctx);
263 call_rcu(&ctx->rcu_head, ctx_rcu_free);
266 static inline int try_get_ioctx(struct kioctx *kioctx)
268 return atomic_inc_not_zero(&kioctx->users);
271 static inline void put_ioctx(struct kioctx *kioctx)
273 BUG_ON(atomic_read(&kioctx->users) <= 0);
274 if (unlikely(atomic_dec_and_test(&kioctx->users)))
278 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
279 struct io_event *res)
281 int (*cancel)(struct kiocb *, struct io_event *);
284 cancel = kiocb->ki_cancel;
285 kiocbSetCancelled(kiocb);
288 spin_unlock_irq(&ctx->ctx_lock);
290 memset(res, 0, sizeof(*res));
291 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
292 res->data = kiocb->ki_user_data;
293 ret = cancel(kiocb, res);
295 spin_lock_irq(&ctx->ctx_lock);
302 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
304 static struct kioctx *ioctx_alloc(unsigned nr_events)
306 struct mm_struct *mm = current->mm;
310 /* Prevent overflows */
311 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
312 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
313 pr_debug("ENOMEM: nr_events too high\n");
314 return ERR_PTR(-EINVAL);
317 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
318 return ERR_PTR(-EAGAIN);
320 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
322 return ERR_PTR(-ENOMEM);
324 ctx->max_reqs = nr_events;
326 atomic_set(&ctx->users, 2);
327 spin_lock_init(&ctx->ctx_lock);
328 spin_lock_init(&ctx->ring_info.ring_lock);
329 init_waitqueue_head(&ctx->wait);
331 INIT_LIST_HEAD(&ctx->active_reqs);
333 if (aio_setup_ring(ctx) < 0)
336 /* limit the number of system wide aios */
337 spin_lock(&aio_nr_lock);
338 if (aio_nr + nr_events > aio_max_nr ||
339 aio_nr + nr_events < aio_nr) {
340 spin_unlock(&aio_nr_lock);
343 aio_nr += ctx->max_reqs;
344 spin_unlock(&aio_nr_lock);
346 /* now link into global list. */
347 spin_lock(&mm->ioctx_lock);
348 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
349 spin_unlock(&mm->ioctx_lock);
351 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
352 ctx, ctx->user_id, mm, ctx->ring_info.nr);
359 kmem_cache_free(kioctx_cachep, ctx);
360 pr_debug("error allocating ioctx %d\n", err);
365 * Cancels all outstanding aio requests on an aio context. Used
366 * when the processes owning a context have all exited to encourage
367 * the rapid destruction of the kioctx.
369 static void kill_ctx(struct kioctx *ctx)
371 struct task_struct *tsk = current;
372 DECLARE_WAITQUEUE(wait, tsk);
376 spin_lock_irq(&ctx->ctx_lock);
378 while (!list_empty(&ctx->active_reqs)) {
379 req = list_first_entry(&ctx->active_reqs,
380 struct kiocb, ki_list);
382 list_del_init(&req->ki_list);
383 kiocb_cancel(ctx, req, &res);
386 if (!ctx->reqs_active)
389 add_wait_queue(&ctx->wait, &wait);
390 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
391 while (ctx->reqs_active) {
392 spin_unlock_irq(&ctx->ctx_lock);
394 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
395 spin_lock_irq(&ctx->ctx_lock);
397 __set_task_state(tsk, TASK_RUNNING);
398 remove_wait_queue(&ctx->wait, &wait);
401 spin_unlock_irq(&ctx->ctx_lock);
404 /* wait_on_sync_kiocb:
405 * Waits on the given sync kiocb to complete.
407 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
409 while (iocb->ki_users) {
410 set_current_state(TASK_UNINTERRUPTIBLE);
415 __set_current_state(TASK_RUNNING);
416 return iocb->ki_user_data;
418 EXPORT_SYMBOL(wait_on_sync_kiocb);
420 /* exit_aio: called when the last user of mm goes away. At this point,
421 * there is no way for any new requests to be submited or any of the
422 * io_* syscalls to be called on the context. However, there may be
423 * outstanding requests which hold references to the context; as they
424 * go away, they will call put_ioctx and release any pinned memory
425 * associated with the request (held via struct page * references).
427 void exit_aio(struct mm_struct *mm)
431 while (!hlist_empty(&mm->ioctx_list)) {
432 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
433 hlist_del_rcu(&ctx->list);
437 if (1 != atomic_read(&ctx->users))
439 "exit_aio:ioctx still alive: %d %d %d\n",
440 atomic_read(&ctx->users), ctx->dead,
443 * We don't need to bother with munmap() here -
444 * exit_mmap(mm) is coming and it'll unmap everything.
445 * Since aio_free_ring() uses non-zero ->mmap_size
446 * as indicator that it needs to unmap the area,
447 * just set it to 0; aio_free_ring() is the only
448 * place that uses ->mmap_size, so it's safe.
450 ctx->ring_info.mmap_size = 0;
456 * Allocate a slot for an aio request. Increments the users count
457 * of the kioctx so that the kioctx stays around until all requests are
458 * complete. Returns NULL if no requests are free.
460 * Returns with kiocb->users set to 2. The io submit code path holds
461 * an extra reference while submitting the i/o.
462 * This prevents races between the aio code path referencing the
463 * req (after submitting it) and aio_complete() freeing the req.
465 static struct kiocb *__aio_get_req(struct kioctx *ctx)
467 struct kiocb *req = NULL;
469 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
477 req->ki_cancel = NULL;
478 req->ki_retry = NULL;
481 req->ki_iovec = NULL;
482 req->ki_eventfd = NULL;
488 * struct kiocb's are allocated in batches to reduce the number of
489 * times the ctx lock is acquired and released.
491 #define KIOCB_BATCH_SIZE 32L
493 struct list_head head;
494 long count; /* number of requests left to allocate */
497 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
499 INIT_LIST_HEAD(&batch->head);
500 batch->count = total;
503 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
505 struct kiocb *req, *n;
507 if (list_empty(&batch->head))
510 spin_lock_irq(&ctx->ctx_lock);
511 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
512 list_del(&req->ki_batch);
513 list_del(&req->ki_list);
514 kmem_cache_free(kiocb_cachep, req);
517 if (unlikely(!ctx->reqs_active && ctx->dead))
518 wake_up_all(&ctx->wait);
519 spin_unlock_irq(&ctx->ctx_lock);
523 * Allocate a batch of kiocbs. This avoids taking and dropping the
524 * context lock a lot during setup.
526 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
528 unsigned short allocated, to_alloc;
530 struct kiocb *req, *n;
531 struct aio_ring *ring;
533 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
534 for (allocated = 0; allocated < to_alloc; allocated++) {
535 req = __aio_get_req(ctx);
537 /* allocation failed, go with what we've got */
539 list_add(&req->ki_batch, &batch->head);
545 spin_lock_irq(&ctx->ctx_lock);
546 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
548 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
550 if (avail < allocated) {
551 /* Trim back the number of requests. */
552 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
553 list_del(&req->ki_batch);
554 kmem_cache_free(kiocb_cachep, req);
555 if (--allocated <= avail)
560 batch->count -= allocated;
561 list_for_each_entry(req, &batch->head, ki_batch) {
562 list_add(&req->ki_list, &ctx->active_reqs);
567 spin_unlock_irq(&ctx->ctx_lock);
573 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
574 struct kiocb_batch *batch)
578 if (list_empty(&batch->head))
579 if (kiocb_batch_refill(ctx, batch) == 0)
581 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
582 list_del(&req->ki_batch);
586 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
588 assert_spin_locked(&ctx->ctx_lock);
592 if (req->ki_eventfd != NULL)
593 eventfd_ctx_put(req->ki_eventfd);
596 if (req->ki_iovec != &req->ki_inline_vec)
597 kfree(req->ki_iovec);
598 kmem_cache_free(kiocb_cachep, req);
601 if (unlikely(!ctx->reqs_active && ctx->dead))
602 wake_up_all(&ctx->wait);
606 * Returns true if this put was the last user of the request.
608 static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
610 assert_spin_locked(&ctx->ctx_lock);
613 BUG_ON(req->ki_users < 0);
614 if (likely(req->ki_users))
616 list_del(&req->ki_list); /* remove from active_reqs */
617 req->ki_cancel = NULL;
618 req->ki_retry = NULL;
620 really_put_req(ctx, req);
624 * Returns true if this put was the last user of the kiocb,
625 * false if the request is still in use.
627 void aio_put_req(struct kiocb *req)
629 struct kioctx *ctx = req->ki_ctx;
630 spin_lock_irq(&ctx->ctx_lock);
631 __aio_put_req(ctx, req);
632 spin_unlock_irq(&ctx->ctx_lock);
634 EXPORT_SYMBOL(aio_put_req);
636 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
638 struct mm_struct *mm = current->mm;
639 struct kioctx *ctx, *ret = NULL;
643 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
645 * RCU protects us against accessing freed memory but
646 * we have to be careful not to get a reference when the
647 * reference count already dropped to 0 (ctx->dead test
648 * is unreliable because of races).
650 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
661 * Called when the io request on the given iocb is complete.
663 void aio_complete(struct kiocb *iocb, long res, long res2)
665 struct kioctx *ctx = iocb->ki_ctx;
666 struct aio_ring_info *info;
667 struct aio_ring *ring;
668 struct io_event *event;
673 * Special case handling for sync iocbs:
674 * - events go directly into the iocb for fast handling
675 * - the sync task with the iocb in its stack holds the single iocb
676 * ref, no other paths have a way to get another ref
677 * - the sync task helpfully left a reference to itself in the iocb
679 if (is_sync_kiocb(iocb)) {
680 BUG_ON(iocb->ki_users != 1);
681 iocb->ki_user_data = res;
683 wake_up_process(iocb->ki_obj.tsk);
687 info = &ctx->ring_info;
689 /* add a completion event to the ring buffer.
690 * must be done holding ctx->ctx_lock to prevent
691 * other code from messing with the tail
692 * pointer since we might be called from irq
695 spin_lock_irqsave(&ctx->ctx_lock, flags);
698 * cancelled requests don't get events, userland was given one
699 * when the event got cancelled.
701 if (kiocbIsCancelled(iocb))
704 ring = kmap_atomic(info->ring_pages[0]);
707 event = aio_ring_event(info, tail);
708 if (++tail >= info->nr)
711 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
712 event->data = iocb->ki_user_data;
716 pr_debug("%p[%lu]: %p: %p %Lx %lx %lx\n",
717 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
720 /* after flagging the request as done, we
721 * must never even look at it again
723 smp_wmb(); /* make event visible before updating tail */
728 put_aio_ring_event(event);
731 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
734 * Check if the user asked us to deliver the result through an
735 * eventfd. The eventfd_signal() function is safe to be called
738 if (iocb->ki_eventfd != NULL)
739 eventfd_signal(iocb->ki_eventfd, 1);
742 /* everything turned out well, dispose of the aiocb. */
743 __aio_put_req(ctx, iocb);
746 * We have to order our ring_info tail store above and test
747 * of the wait list below outside the wait lock. This is
748 * like in wake_up_bit() where clearing a bit has to be
749 * ordered with the unlocked test.
753 if (waitqueue_active(&ctx->wait))
756 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
758 EXPORT_SYMBOL(aio_complete);
761 * Pull an event off of the ioctx's event ring. Returns the number of
762 * events fetched (0 or 1 ;-)
763 * FIXME: make this use cmpxchg.
764 * TODO: make the ringbuffer user mmap()able (requires FIXME).
766 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
768 struct aio_ring_info *info = &ioctx->ring_info;
769 struct aio_ring *ring;
773 ring = kmap_atomic(info->ring_pages[0]);
774 pr_debug("h%u t%u m%u\n", ring->head, ring->tail, ring->nr);
776 if (ring->head == ring->tail)
779 spin_lock(&info->ring_lock);
781 head = ring->head % info->nr;
782 if (head != ring->tail) {
783 struct io_event *evp = aio_ring_event(info, head);
785 head = (head + 1) % info->nr;
786 smp_mb(); /* finish reading the event before updatng the head */
789 put_aio_ring_event(evp);
791 spin_unlock(&info->ring_lock);
795 pr_debug("%d h%u t%u\n", ret, ring->head, ring->tail);
800 struct timer_list timer;
802 struct task_struct *p;
805 static void timeout_func(unsigned long data)
807 struct aio_timeout *to = (struct aio_timeout *)data;
810 wake_up_process(to->p);
813 static inline void init_timeout(struct aio_timeout *to)
815 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
820 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
821 const struct timespec *ts)
823 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
824 if (time_after(to->timer.expires, jiffies))
825 add_timer(&to->timer);
830 static inline void clear_timeout(struct aio_timeout *to)
832 del_singleshot_timer_sync(&to->timer);
835 static int read_events(struct kioctx *ctx,
836 long min_nr, long nr,
837 struct io_event __user *event,
838 struct timespec __user *timeout)
840 long start_jiffies = jiffies;
841 struct task_struct *tsk = current;
842 DECLARE_WAITQUEUE(wait, tsk);
846 struct aio_timeout to;
848 /* needed to zero any padding within an entry (there shouldn't be
851 memset(&ent, 0, sizeof(ent));
853 while (likely(i < nr)) {
854 ret = aio_read_evt(ctx, &ent);
855 if (unlikely(ret <= 0))
858 pr_debug("%Lx %Lx %Lx %Lx\n",
859 ent.data, ent.obj, ent.res, ent.res2);
861 /* Could we split the check in two? */
863 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
864 pr_debug("lost an event due to EFAULT.\n");
869 /* Good, event copied to userland, update counts. */
885 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
888 set_timeout(start_jiffies, &to, &ts);
891 while (likely(i < nr)) {
892 add_wait_queue_exclusive(&ctx->wait, &wait);
894 set_task_state(tsk, TASK_INTERRUPTIBLE);
895 ret = aio_read_evt(ctx, &ent);
900 if (unlikely(ctx->dead)) {
904 if (to.timed_out) /* Only check after read evt */
906 /* Try to only show up in io wait if there are ops
908 if (ctx->reqs_active)
912 if (signal_pending(tsk)) {
916 /*ret = aio_read_evt(ctx, &ent);*/
919 set_task_state(tsk, TASK_RUNNING);
920 remove_wait_queue(&ctx->wait, &wait);
922 if (unlikely(ret <= 0))
926 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
927 pr_debug("lost an event due to EFAULT.\n");
931 /* Good, event copied to userland, update counts. */
939 destroy_timer_on_stack(&to.timer);
943 /* Take an ioctx and remove it from the list of ioctx's. Protects
944 * against races with itself via ->dead.
946 static void io_destroy(struct kioctx *ioctx)
948 struct mm_struct *mm = current->mm;
951 /* delete the entry from the list is someone else hasn't already */
952 spin_lock(&mm->ioctx_lock);
953 was_dead = ioctx->dead;
955 hlist_del_rcu(&ioctx->list);
956 spin_unlock(&mm->ioctx_lock);
958 pr_debug("(%p)\n", ioctx);
959 if (likely(!was_dead))
960 put_ioctx(ioctx); /* twice for the list */
965 * Wake up any waiters. The setting of ctx->dead must be seen
966 * by other CPUs at this point. Right now, we rely on the
967 * locking done by the above calls to ensure this consistency.
969 wake_up_all(&ioctx->wait);
973 * Create an aio_context capable of receiving at least nr_events.
974 * ctxp must not point to an aio_context that already exists, and
975 * must be initialized to 0 prior to the call. On successful
976 * creation of the aio_context, *ctxp is filled in with the resulting
977 * handle. May fail with -EINVAL if *ctxp is not initialized,
978 * if the specified nr_events exceeds internal limits. May fail
979 * with -EAGAIN if the specified nr_events exceeds the user's limit
980 * of available events. May fail with -ENOMEM if insufficient kernel
981 * resources are available. May fail with -EFAULT if an invalid
982 * pointer is passed for ctxp. Will fail with -ENOSYS if not
985 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
987 struct kioctx *ioctx = NULL;
991 ret = get_user(ctx, ctxp);
996 if (unlikely(ctx || nr_events == 0)) {
997 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1002 ioctx = ioctx_alloc(nr_events);
1003 ret = PTR_ERR(ioctx);
1004 if (!IS_ERR(ioctx)) {
1005 ret = put_user(ioctx->user_id, ctxp);
1016 * Destroy the aio_context specified. May cancel any outstanding
1017 * AIOs and block on completion. Will fail with -ENOSYS if not
1018 * implemented. May fail with -EINVAL if the context pointed to
1021 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1023 struct kioctx *ioctx = lookup_ioctx(ctx);
1024 if (likely(NULL != ioctx)) {
1029 pr_debug("EINVAL: io_destroy: invalid context id\n");
1033 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1035 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1039 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1040 ssize_t this = min((ssize_t)iov->iov_len, ret);
1041 iov->iov_base += this;
1042 iov->iov_len -= this;
1043 iocb->ki_left -= this;
1045 if (iov->iov_len == 0) {
1051 /* the caller should not have done more io than what fit in
1052 * the remaining iovecs */
1053 BUG_ON(ret > 0 && iocb->ki_left == 0);
1056 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1058 struct file *file = iocb->ki_filp;
1059 struct address_space *mapping = file->f_mapping;
1060 struct inode *inode = mapping->host;
1061 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1062 unsigned long, loff_t);
1064 unsigned short opcode;
1066 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1067 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1068 rw_op = file->f_op->aio_read;
1069 opcode = IOCB_CMD_PREADV;
1071 rw_op = file->f_op->aio_write;
1072 opcode = IOCB_CMD_PWRITEV;
1075 /* This matches the pread()/pwrite() logic */
1076 if (iocb->ki_pos < 0)
1079 if (opcode == IOCB_CMD_PWRITEV)
1080 file_start_write(file);
1082 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1083 iocb->ki_nr_segs - iocb->ki_cur_seg,
1086 aio_advance_iovec(iocb, ret);
1088 /* retry all partial writes. retry partial reads as long as its a
1090 } while (ret > 0 && iocb->ki_left > 0 &&
1091 (opcode == IOCB_CMD_PWRITEV ||
1092 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1093 if (opcode == IOCB_CMD_PWRITEV)
1094 file_end_write(file);
1096 /* This means we must have transferred all that we could */
1097 /* No need to retry anymore */
1098 if ((ret == 0) || (iocb->ki_left == 0))
1099 ret = iocb->ki_nbytes - iocb->ki_left;
1101 /* If we managed to write some out we return that, rather than
1102 * the eventual error. */
1103 if (opcode == IOCB_CMD_PWRITEV
1104 && ret < 0 && ret != -EIOCBQUEUED
1105 && iocb->ki_nbytes - iocb->ki_left)
1106 ret = iocb->ki_nbytes - iocb->ki_left;
1111 static ssize_t aio_fdsync(struct kiocb *iocb)
1113 struct file *file = iocb->ki_filp;
1114 ssize_t ret = -EINVAL;
1116 if (file->f_op->aio_fsync)
1117 ret = file->f_op->aio_fsync(iocb, 1);
1121 static ssize_t aio_fsync(struct kiocb *iocb)
1123 struct file *file = iocb->ki_filp;
1124 ssize_t ret = -EINVAL;
1126 if (file->f_op->aio_fsync)
1127 ret = file->f_op->aio_fsync(iocb, 0);
1131 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1135 #ifdef CONFIG_COMPAT
1137 ret = compat_rw_copy_check_uvector(type,
1138 (struct compat_iovec __user *)kiocb->ki_buf,
1139 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1143 ret = rw_copy_check_uvector(type,
1144 (struct iovec __user *)kiocb->ki_buf,
1145 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1150 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1154 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1155 kiocb->ki_cur_seg = 0;
1156 /* ki_nbytes/left now reflect bytes instead of segs */
1157 kiocb->ki_nbytes = ret;
1158 kiocb->ki_left = ret;
1165 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1169 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1173 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1174 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1175 kiocb->ki_iovec->iov_len = bytes;
1176 kiocb->ki_nr_segs = 1;
1177 kiocb->ki_cur_seg = 0;
1183 * Performs the initial checks and aio retry method
1184 * setup for the kiocb at the time of io submission.
1186 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1188 struct file *file = kiocb->ki_filp;
1191 switch (kiocb->ki_opcode) {
1192 case IOCB_CMD_PREAD:
1194 if (unlikely(!(file->f_mode & FMODE_READ)))
1197 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1200 ret = aio_setup_single_vector(READ, file, kiocb);
1204 if (file->f_op->aio_read)
1205 kiocb->ki_retry = aio_rw_vect_retry;
1207 case IOCB_CMD_PWRITE:
1209 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1212 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1215 ret = aio_setup_single_vector(WRITE, file, kiocb);
1219 if (file->f_op->aio_write)
1220 kiocb->ki_retry = aio_rw_vect_retry;
1222 case IOCB_CMD_PREADV:
1224 if (unlikely(!(file->f_mode & FMODE_READ)))
1226 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1230 if (file->f_op->aio_read)
1231 kiocb->ki_retry = aio_rw_vect_retry;
1233 case IOCB_CMD_PWRITEV:
1235 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1237 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1241 if (file->f_op->aio_write)
1242 kiocb->ki_retry = aio_rw_vect_retry;
1244 case IOCB_CMD_FDSYNC:
1246 if (file->f_op->aio_fsync)
1247 kiocb->ki_retry = aio_fdsync;
1249 case IOCB_CMD_FSYNC:
1251 if (file->f_op->aio_fsync)
1252 kiocb->ki_retry = aio_fsync;
1255 pr_debug("EINVAL: no operation provided\n");
1259 if (!kiocb->ki_retry)
1265 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1266 struct iocb *iocb, struct kiocb_batch *batch,
1272 /* enforce forwards compatibility on users */
1273 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1274 pr_debug("EINVAL: reserve field set\n");
1278 /* prevent overflows */
1280 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1281 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1282 ((ssize_t)iocb->aio_nbytes < 0)
1284 pr_debug("EINVAL: io_submit: overflow check\n");
1288 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1292 req->ki_filp = fget(iocb->aio_fildes);
1293 if (unlikely(!req->ki_filp)) {
1298 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1300 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1301 * instance of the file* now. The file descriptor must be
1302 * an eventfd() fd, and will be signaled for each completed
1303 * event using the eventfd_signal() function.
1305 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1306 if (IS_ERR(req->ki_eventfd)) {
1307 ret = PTR_ERR(req->ki_eventfd);
1308 req->ki_eventfd = NULL;
1313 ret = put_user(req->ki_key, &user_iocb->aio_key);
1314 if (unlikely(ret)) {
1315 pr_debug("EFAULT: aio_key\n");
1319 req->ki_obj.user = user_iocb;
1320 req->ki_user_data = iocb->aio_data;
1321 req->ki_pos = iocb->aio_offset;
1323 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1324 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1325 req->ki_opcode = iocb->aio_lio_opcode;
1327 ret = aio_setup_iocb(req, compat);
1332 spin_lock_irq(&ctx->ctx_lock);
1334 * We could have raced with io_destroy() and are currently holding a
1335 * reference to ctx which should be destroyed. We cannot submit IO
1336 * since ctx gets freed as soon as io_submit() puts its reference. The
1337 * check here is reliable: io_destroy() sets ctx->dead before waiting
1338 * for outstanding IO and the barrier between these two is realized by
1339 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1340 * increment ctx->reqs_active before checking for ctx->dead and the
1341 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1342 * don't see ctx->dead set here, io_destroy() waits for our IO to
1347 spin_unlock_irq(&ctx->ctx_lock);
1351 if (unlikely(kiocbIsCancelled(req)))
1354 ret = req->ki_retry(req);
1356 if (ret != -EIOCBQUEUED) {
1358 * There's no easy way to restart the syscall since other AIO's
1359 * may be already running. Just fail this IO with EINTR.
1361 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1362 ret == -ERESTARTNOHAND ||
1363 ret == -ERESTART_RESTARTBLOCK))
1365 aio_complete(req, ret, 0);
1368 aio_put_req(req); /* drop extra ref to req */
1372 aio_put_req(req); /* drop extra ref to req */
1373 aio_put_req(req); /* drop i/o ref to req */
1377 long do_io_submit(aio_context_t ctx_id, long nr,
1378 struct iocb __user *__user *iocbpp, bool compat)
1383 struct blk_plug plug;
1384 struct kiocb_batch batch;
1386 if (unlikely(nr < 0))
1389 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1390 nr = LONG_MAX/sizeof(*iocbpp);
1392 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1395 ctx = lookup_ioctx(ctx_id);
1396 if (unlikely(!ctx)) {
1397 pr_debug("EINVAL: invalid context id\n");
1401 kiocb_batch_init(&batch, nr);
1403 blk_start_plug(&plug);
1406 * AKPM: should this return a partial result if some of the IOs were
1407 * successfully submitted?
1409 for (i=0; i<nr; i++) {
1410 struct iocb __user *user_iocb;
1413 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1418 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1423 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1427 blk_finish_plug(&plug);
1429 kiocb_batch_free(ctx, &batch);
1435 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1436 * the number of iocbs queued. May return -EINVAL if the aio_context
1437 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1438 * *iocbpp[0] is not properly initialized, if the operation specified
1439 * is invalid for the file descriptor in the iocb. May fail with
1440 * -EFAULT if any of the data structures point to invalid data. May
1441 * fail with -EBADF if the file descriptor specified in the first
1442 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1443 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1444 * fail with -ENOSYS if not implemented.
1446 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1447 struct iocb __user * __user *, iocbpp)
1449 return do_io_submit(ctx_id, nr, iocbpp, 0);
1453 * Finds a given iocb for cancellation.
1455 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1458 struct list_head *pos;
1460 assert_spin_locked(&ctx->ctx_lock);
1462 /* TODO: use a hash or array, this sucks. */
1463 list_for_each(pos, &ctx->active_reqs) {
1464 struct kiocb *kiocb = list_kiocb(pos);
1465 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1472 * Attempts to cancel an iocb previously passed to io_submit. If
1473 * the operation is successfully cancelled, the resulting event is
1474 * copied into the memory pointed to by result without being placed
1475 * into the completion queue and 0 is returned. May fail with
1476 * -EFAULT if any of the data structures pointed to are invalid.
1477 * May fail with -EINVAL if aio_context specified by ctx_id is
1478 * invalid. May fail with -EAGAIN if the iocb specified was not
1479 * cancelled. Will fail with -ENOSYS if not implemented.
1481 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1482 struct io_event __user *, result)
1484 struct io_event res;
1486 struct kiocb *kiocb;
1490 ret = get_user(key, &iocb->aio_key);
1494 ctx = lookup_ioctx(ctx_id);
1498 spin_lock_irq(&ctx->ctx_lock);
1500 kiocb = lookup_kiocb(ctx, iocb, key);
1502 ret = kiocb_cancel(ctx, kiocb, &res);
1506 spin_unlock_irq(&ctx->ctx_lock);
1509 /* Cancellation succeeded -- copy the result
1510 * into the user's buffer.
1512 if (copy_to_user(result, &res, sizeof(res)))
1522 * Attempts to read at least min_nr events and up to nr events from
1523 * the completion queue for the aio_context specified by ctx_id. If
1524 * it succeeds, the number of read events is returned. May fail with
1525 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1526 * out of range, if timeout is out of range. May fail with -EFAULT
1527 * if any of the memory specified is invalid. May return 0 or
1528 * < min_nr if the timeout specified by timeout has elapsed
1529 * before sufficient events are available, where timeout == NULL
1530 * specifies an infinite timeout. Note that the timeout pointed to by
1531 * timeout is relative and will be updated if not NULL and the
1532 * operation blocks. Will fail with -ENOSYS if not implemented.
1534 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1537 struct io_event __user *, events,
1538 struct timespec __user *, timeout)
1540 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1543 if (likely(ioctx)) {
1544 if (likely(min_nr <= nr && min_nr >= 0))
1545 ret = read_events(ioctx, min_nr, nr, events, timeout);