2 * An async IO implementation for Linux
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;
101 struct work_struct rcu_work;
104 /*------ sysctl variables----*/
105 static DEFINE_SPINLOCK(aio_nr_lock);
106 unsigned long aio_nr; /* current system wide number of aio requests */
107 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
108 /*----end sysctl variables---*/
110 static struct kmem_cache *kiocb_cachep;
111 static struct kmem_cache *kioctx_cachep;
114 * Creates the slab caches used by the aio routines, panic on
115 * failure as this is done early during the boot sequence.
117 static int __init aio_setup(void)
119 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
120 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
122 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
126 __initcall(aio_setup);
128 static void aio_free_ring(struct kioctx *ctx)
130 struct aio_ring_info *info = &ctx->ring_info;
133 for (i=0; i<info->nr_pages; i++)
134 put_page(info->ring_pages[i]);
136 if (info->mmap_size) {
137 vm_munmap(info->mmap_base, info->mmap_size);
140 if (info->ring_pages && info->ring_pages != info->internal_pages)
141 kfree(info->ring_pages);
142 info->ring_pages = NULL;
146 static int aio_setup_ring(struct kioctx *ctx)
148 struct aio_ring *ring;
149 struct aio_ring_info *info = &ctx->ring_info;
150 unsigned nr_events = ctx->max_reqs;
151 struct mm_struct *mm = current->mm;
152 unsigned long size, populate;
155 /* Compensate for the ring buffer's head/tail overlap entry */
156 nr_events += 2; /* 1 is required, 2 for good luck */
158 size = sizeof(struct aio_ring);
159 size += sizeof(struct io_event) * nr_events;
160 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
165 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
168 info->ring_pages = info->internal_pages;
169 if (nr_pages > AIO_RING_PAGES) {
170 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
171 if (!info->ring_pages)
175 info->mmap_size = nr_pages * PAGE_SIZE;
176 pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
177 down_write(&mm->mmap_sem);
178 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
179 PROT_READ|PROT_WRITE,
180 MAP_ANONYMOUS|MAP_PRIVATE, 0,
182 if (IS_ERR((void *)info->mmap_base)) {
183 up_write(&mm->mmap_sem);
189 pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
190 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
191 1, 0, info->ring_pages, NULL);
192 up_write(&mm->mmap_sem);
194 if (unlikely(info->nr_pages != nr_pages)) {
199 mm_populate(info->mmap_base, populate);
201 ctx->user_id = info->mmap_base;
203 info->nr = nr_events; /* trusted copy */
205 ring = kmap_atomic(info->ring_pages[0]);
206 ring->nr = nr_events; /* user copy */
207 ring->id = ctx->user_id;
208 ring->head = ring->tail = 0;
209 ring->magic = AIO_RING_MAGIC;
210 ring->compat_features = AIO_RING_COMPAT_FEATURES;
211 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
212 ring->header_length = sizeof(struct aio_ring);
219 /* aio_ring_event: returns a pointer to the event at the given index from
220 * kmap_atomic(). Release the pointer with put_aio_ring_event();
222 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
223 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
224 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
226 #define aio_ring_event(info, nr) ({ \
227 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
228 struct io_event *__event; \
229 __event = kmap_atomic( \
230 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
231 __event += pos % AIO_EVENTS_PER_PAGE; \
235 #define put_aio_ring_event(event) do { \
236 struct io_event *__event = (event); \
238 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
241 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
242 struct io_event *res)
244 int (*cancel)(struct kiocb *, struct io_event *);
247 cancel = kiocb->ki_cancel;
248 kiocbSetCancelled(kiocb);
250 atomic_inc(&kiocb->ki_users);
251 spin_unlock_irq(&ctx->ctx_lock);
253 memset(res, 0, sizeof(*res));
254 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
255 res->data = kiocb->ki_user_data;
256 ret = cancel(kiocb, res);
258 spin_lock_irq(&ctx->ctx_lock);
264 static void free_ioctx_rcu(struct rcu_head *head)
266 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
267 kmem_cache_free(kioctx_cachep, ctx);
271 * When this function runs, the kioctx has been removed from the "hash table"
272 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
273 * now it's safe to cancel any that need to be.
275 static void free_ioctx(struct kioctx *ctx)
280 spin_lock_irq(&ctx->ctx_lock);
282 while (!list_empty(&ctx->active_reqs)) {
283 req = list_first_entry(&ctx->active_reqs,
284 struct kiocb, ki_list);
286 list_del_init(&req->ki_list);
287 kiocb_cancel(ctx, req, &res);
290 spin_unlock_irq(&ctx->ctx_lock);
292 wait_event(ctx->wait, !atomic_read(&ctx->reqs_active));
296 spin_lock(&aio_nr_lock);
297 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
298 aio_nr -= ctx->max_reqs;
299 spin_unlock(&aio_nr_lock);
301 pr_debug("freeing %p\n", ctx);
304 * Here the call_rcu() is between the wait_event() for reqs_active to
305 * hit 0, and freeing the ioctx.
307 * aio_complete() decrements reqs_active, but it has to touch the ioctx
308 * after to issue a wakeup so we use rcu.
310 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
313 static void put_ioctx(struct kioctx *ctx)
315 if (unlikely(atomic_dec_and_test(&ctx->users)))
320 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
322 static struct kioctx *ioctx_alloc(unsigned nr_events)
324 struct mm_struct *mm = current->mm;
328 /* Prevent overflows */
329 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
330 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
331 pr_debug("ENOMEM: nr_events too high\n");
332 return ERR_PTR(-EINVAL);
335 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
336 return ERR_PTR(-EAGAIN);
338 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
340 return ERR_PTR(-ENOMEM);
342 ctx->max_reqs = nr_events;
344 atomic_set(&ctx->users, 2);
345 atomic_set(&ctx->dead, 0);
346 spin_lock_init(&ctx->ctx_lock);
347 spin_lock_init(&ctx->ring_info.ring_lock);
348 init_waitqueue_head(&ctx->wait);
350 INIT_LIST_HEAD(&ctx->active_reqs);
352 if (aio_setup_ring(ctx) < 0)
355 /* limit the number of system wide aios */
356 spin_lock(&aio_nr_lock);
357 if (aio_nr + nr_events > aio_max_nr ||
358 aio_nr + nr_events < aio_nr) {
359 spin_unlock(&aio_nr_lock);
362 aio_nr += ctx->max_reqs;
363 spin_unlock(&aio_nr_lock);
365 /* now link into global list. */
366 spin_lock(&mm->ioctx_lock);
367 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
368 spin_unlock(&mm->ioctx_lock);
370 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
371 ctx, ctx->user_id, mm, ctx->ring_info.nr);
378 kmem_cache_free(kioctx_cachep, ctx);
379 pr_debug("error allocating ioctx %d\n", err);
383 static void kill_ioctx_work(struct work_struct *work)
385 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
387 wake_up_all(&ctx->wait);
391 static void kill_ioctx_rcu(struct rcu_head *head)
393 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
395 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
396 schedule_work(&ctx->rcu_work);
400 * Cancels all outstanding aio requests on an aio context. Used
401 * when the processes owning a context have all exited to encourage
402 * the rapid destruction of the kioctx.
404 static void kill_ioctx(struct kioctx *ctx)
406 if (!atomic_xchg(&ctx->dead, 1)) {
407 hlist_del_rcu(&ctx->list);
408 /* Between hlist_del_rcu() and dropping the initial ref */
412 * We can't punt to workqueue here because put_ioctx() ->
413 * free_ioctx() will unmap the ringbuffer, and that has to be
414 * done in the original process's context. kill_ioctx_rcu/work()
415 * exist for exit_aio(), as in that path free_ioctx() won't do
418 kill_ioctx_work(&ctx->rcu_work);
422 /* wait_on_sync_kiocb:
423 * Waits on the given sync kiocb to complete.
425 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
427 while (atomic_read(&iocb->ki_users)) {
428 set_current_state(TASK_UNINTERRUPTIBLE);
429 if (!atomic_read(&iocb->ki_users))
433 __set_current_state(TASK_RUNNING);
434 return iocb->ki_user_data;
436 EXPORT_SYMBOL(wait_on_sync_kiocb);
439 * exit_aio: called when the last user of mm goes away. At this point, there is
440 * no way for any new requests to be submited or any of the io_* syscalls to be
441 * called on the context.
443 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
446 void exit_aio(struct mm_struct *mm)
449 struct hlist_node *n;
451 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
452 if (1 != atomic_read(&ctx->users))
454 "exit_aio:ioctx still alive: %d %d %d\n",
455 atomic_read(&ctx->users),
456 atomic_read(&ctx->dead),
457 atomic_read(&ctx->reqs_active));
459 * We don't need to bother with munmap() here -
460 * exit_mmap(mm) is coming and it'll unmap everything.
461 * Since aio_free_ring() uses non-zero ->mmap_size
462 * as indicator that it needs to unmap the area,
463 * just set it to 0; aio_free_ring() is the only
464 * place that uses ->mmap_size, so it's safe.
466 ctx->ring_info.mmap_size = 0;
468 if (!atomic_xchg(&ctx->dead, 1)) {
469 hlist_del_rcu(&ctx->list);
470 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
476 * Allocate a slot for an aio request. Increments the ki_users count
477 * of the kioctx so that the kioctx stays around until all requests are
478 * complete. Returns NULL if no requests are free.
480 * Returns with kiocb->ki_users set to 2. The io submit code path holds
481 * an extra reference while submitting the i/o.
482 * This prevents races between the aio code path referencing the
483 * req (after submitting it) and aio_complete() freeing the req.
485 static struct kiocb *__aio_get_req(struct kioctx *ctx)
487 struct kiocb *req = NULL;
489 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
494 atomic_set(&req->ki_users, 2);
497 req->ki_cancel = NULL;
498 req->ki_retry = NULL;
501 req->ki_iovec = NULL;
502 req->ki_eventfd = NULL;
508 * struct kiocb's are allocated in batches to reduce the number of
509 * times the ctx lock is acquired and released.
511 #define KIOCB_BATCH_SIZE 32L
513 struct list_head head;
514 long count; /* number of requests left to allocate */
517 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
519 INIT_LIST_HEAD(&batch->head);
520 batch->count = total;
523 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
525 struct kiocb *req, *n;
527 if (list_empty(&batch->head))
530 spin_lock_irq(&ctx->ctx_lock);
531 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
532 list_del(&req->ki_batch);
533 list_del(&req->ki_list);
534 kmem_cache_free(kiocb_cachep, req);
535 atomic_dec(&ctx->reqs_active);
537 spin_unlock_irq(&ctx->ctx_lock);
541 * Allocate a batch of kiocbs. This avoids taking and dropping the
542 * context lock a lot during setup.
544 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
546 unsigned short allocated, to_alloc;
548 struct kiocb *req, *n;
549 struct aio_ring *ring;
551 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
552 for (allocated = 0; allocated < to_alloc; allocated++) {
553 req = __aio_get_req(ctx);
555 /* allocation failed, go with what we've got */
557 list_add(&req->ki_batch, &batch->head);
563 spin_lock_irq(&ctx->ctx_lock);
564 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
566 avail = aio_ring_avail(&ctx->ring_info, ring) -
567 atomic_read(&ctx->reqs_active);
569 if (avail < allocated) {
570 /* Trim back the number of requests. */
571 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
572 list_del(&req->ki_batch);
573 kmem_cache_free(kiocb_cachep, req);
574 if (--allocated <= avail)
579 batch->count -= allocated;
580 list_for_each_entry(req, &batch->head, ki_batch) {
581 list_add(&req->ki_list, &ctx->active_reqs);
582 atomic_inc(&ctx->reqs_active);
586 spin_unlock_irq(&ctx->ctx_lock);
592 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
593 struct kiocb_batch *batch)
597 if (list_empty(&batch->head))
598 if (kiocb_batch_refill(ctx, batch) == 0)
600 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
601 list_del(&req->ki_batch);
605 static void kiocb_free(struct kiocb *req)
609 if (req->ki_eventfd != NULL)
610 eventfd_ctx_put(req->ki_eventfd);
613 if (req->ki_iovec != &req->ki_inline_vec)
614 kfree(req->ki_iovec);
615 kmem_cache_free(kiocb_cachep, req);
618 void aio_put_req(struct kiocb *req)
620 if (atomic_dec_and_test(&req->ki_users))
623 EXPORT_SYMBOL(aio_put_req);
625 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
627 struct mm_struct *mm = current->mm;
628 struct kioctx *ctx, *ret = NULL;
632 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
633 if (ctx->user_id == ctx_id) {
634 atomic_inc(&ctx->users);
645 * Called when the io request on the given iocb is complete.
647 void aio_complete(struct kiocb *iocb, long res, long res2)
649 struct kioctx *ctx = iocb->ki_ctx;
650 struct aio_ring_info *info;
651 struct aio_ring *ring;
652 struct io_event *event;
657 * Special case handling for sync iocbs:
658 * - events go directly into the iocb for fast handling
659 * - the sync task with the iocb in its stack holds the single iocb
660 * ref, no other paths have a way to get another ref
661 * - the sync task helpfully left a reference to itself in the iocb
663 if (is_sync_kiocb(iocb)) {
664 BUG_ON(atomic_read(&iocb->ki_users) != 1);
665 iocb->ki_user_data = res;
666 atomic_set(&iocb->ki_users, 0);
667 wake_up_process(iocb->ki_obj.tsk);
671 info = &ctx->ring_info;
674 * Add a completion event to the ring buffer. Must be done holding
675 * ctx->ctx_lock to prevent other code from messing with the tail
676 * pointer since we might be called from irq context.
678 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
679 * need to issue a wakeup after decrementing reqs_active.
682 spin_lock_irqsave(&ctx->ctx_lock, flags);
684 list_del(&iocb->ki_list); /* remove from active_reqs */
687 * cancelled requests don't get events, userland was given one
688 * when the event got cancelled.
690 if (kiocbIsCancelled(iocb))
693 ring = kmap_atomic(info->ring_pages[0]);
696 event = aio_ring_event(info, tail);
697 if (++tail >= info->nr)
700 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
701 event->data = iocb->ki_user_data;
705 pr_debug("%p[%lu]: %p: %p %Lx %lx %lx\n",
706 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
709 /* after flagging the request as done, we
710 * must never even look at it again
712 smp_wmb(); /* make event visible before updating tail */
717 put_aio_ring_event(event);
720 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
723 * Check if the user asked us to deliver the result through an
724 * eventfd. The eventfd_signal() function is safe to be called
727 if (iocb->ki_eventfd != NULL)
728 eventfd_signal(iocb->ki_eventfd, 1);
731 /* everything turned out well, dispose of the aiocb. */
733 atomic_dec(&ctx->reqs_active);
736 * We have to order our ring_info tail store above and test
737 * of the wait list below outside the wait lock. This is
738 * like in wake_up_bit() where clearing a bit has to be
739 * ordered with the unlocked test.
743 if (waitqueue_active(&ctx->wait))
746 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
749 EXPORT_SYMBOL(aio_complete);
752 * Pull an event off of the ioctx's event ring. Returns the number of
753 * events fetched (0 or 1 ;-)
754 * FIXME: make this use cmpxchg.
755 * TODO: make the ringbuffer user mmap()able (requires FIXME).
757 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
759 struct aio_ring_info *info = &ioctx->ring_info;
760 struct aio_ring *ring;
764 ring = kmap_atomic(info->ring_pages[0]);
765 pr_debug("h%u t%u m%u\n", ring->head, ring->tail, ring->nr);
767 if (ring->head == ring->tail)
770 spin_lock(&info->ring_lock);
772 head = ring->head % info->nr;
773 if (head != ring->tail) {
774 struct io_event *evp = aio_ring_event(info, head);
776 head = (head + 1) % info->nr;
777 smp_mb(); /* finish reading the event before updatng the head */
780 put_aio_ring_event(evp);
782 spin_unlock(&info->ring_lock);
786 pr_debug("%d h%u t%u\n", ret, ring->head, ring->tail);
791 struct timer_list timer;
793 struct task_struct *p;
796 static void timeout_func(unsigned long data)
798 struct aio_timeout *to = (struct aio_timeout *)data;
801 wake_up_process(to->p);
804 static inline void init_timeout(struct aio_timeout *to)
806 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
811 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
812 const struct timespec *ts)
814 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
815 if (time_after(to->timer.expires, jiffies))
816 add_timer(&to->timer);
821 static inline void clear_timeout(struct aio_timeout *to)
823 del_singleshot_timer_sync(&to->timer);
826 static int read_events(struct kioctx *ctx,
827 long min_nr, long nr,
828 struct io_event __user *event,
829 struct timespec __user *timeout)
831 long start_jiffies = jiffies;
832 struct task_struct *tsk = current;
833 DECLARE_WAITQUEUE(wait, tsk);
837 struct aio_timeout to;
839 /* needed to zero any padding within an entry (there shouldn't be
842 memset(&ent, 0, sizeof(ent));
844 while (likely(i < nr)) {
845 ret = aio_read_evt(ctx, &ent);
846 if (unlikely(ret <= 0))
849 pr_debug("%Lx %Lx %Lx %Lx\n",
850 ent.data, ent.obj, ent.res, ent.res2);
852 /* Could we split the check in two? */
854 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
855 pr_debug("lost an event due to EFAULT.\n");
860 /* Good, event copied to userland, update counts. */
876 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
879 set_timeout(start_jiffies, &to, &ts);
882 while (likely(i < nr)) {
883 add_wait_queue_exclusive(&ctx->wait, &wait);
885 set_task_state(tsk, TASK_INTERRUPTIBLE);
886 ret = aio_read_evt(ctx, &ent);
891 if (unlikely(atomic_read(&ctx->dead))) {
895 if (to.timed_out) /* Only check after read evt */
897 /* Try to only show up in io wait if there are ops
899 if (atomic_read(&ctx->reqs_active))
903 if (signal_pending(tsk)) {
907 /*ret = aio_read_evt(ctx, &ent);*/
910 set_task_state(tsk, TASK_RUNNING);
911 remove_wait_queue(&ctx->wait, &wait);
913 if (unlikely(ret <= 0))
917 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
918 pr_debug("lost an event due to EFAULT.\n");
922 /* Good, event copied to userland, update counts. */
930 destroy_timer_on_stack(&to.timer);
935 * Create an aio_context capable of receiving at least nr_events.
936 * ctxp must not point to an aio_context that already exists, and
937 * must be initialized to 0 prior to the call. On successful
938 * creation of the aio_context, *ctxp is filled in with the resulting
939 * handle. May fail with -EINVAL if *ctxp is not initialized,
940 * if the specified nr_events exceeds internal limits. May fail
941 * with -EAGAIN if the specified nr_events exceeds the user's limit
942 * of available events. May fail with -ENOMEM if insufficient kernel
943 * resources are available. May fail with -EFAULT if an invalid
944 * pointer is passed for ctxp. Will fail with -ENOSYS if not
947 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
949 struct kioctx *ioctx = NULL;
953 ret = get_user(ctx, ctxp);
958 if (unlikely(ctx || nr_events == 0)) {
959 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
964 ioctx = ioctx_alloc(nr_events);
965 ret = PTR_ERR(ioctx);
966 if (!IS_ERR(ioctx)) {
967 ret = put_user(ioctx->user_id, ctxp);
978 * Destroy the aio_context specified. May cancel any outstanding
979 * AIOs and block on completion. Will fail with -ENOSYS if not
980 * implemented. May fail with -EINVAL if the context pointed to
983 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
985 struct kioctx *ioctx = lookup_ioctx(ctx);
986 if (likely(NULL != ioctx)) {
991 pr_debug("EINVAL: io_destroy: invalid context id\n");
995 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
997 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1001 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1002 ssize_t this = min((ssize_t)iov->iov_len, ret);
1003 iov->iov_base += this;
1004 iov->iov_len -= this;
1005 iocb->ki_left -= this;
1007 if (iov->iov_len == 0) {
1013 /* the caller should not have done more io than what fit in
1014 * the remaining iovecs */
1015 BUG_ON(ret > 0 && iocb->ki_left == 0);
1018 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1020 struct file *file = iocb->ki_filp;
1021 struct address_space *mapping = file->f_mapping;
1022 struct inode *inode = mapping->host;
1023 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1024 unsigned long, loff_t);
1026 unsigned short opcode;
1028 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1029 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1030 rw_op = file->f_op->aio_read;
1031 opcode = IOCB_CMD_PREADV;
1033 rw_op = file->f_op->aio_write;
1034 opcode = IOCB_CMD_PWRITEV;
1037 /* This matches the pread()/pwrite() logic */
1038 if (iocb->ki_pos < 0)
1041 if (opcode == IOCB_CMD_PWRITEV)
1042 file_start_write(file);
1044 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1045 iocb->ki_nr_segs - iocb->ki_cur_seg,
1048 aio_advance_iovec(iocb, ret);
1050 /* retry all partial writes. retry partial reads as long as its a
1052 } while (ret > 0 && iocb->ki_left > 0 &&
1053 (opcode == IOCB_CMD_PWRITEV ||
1054 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1055 if (opcode == IOCB_CMD_PWRITEV)
1056 file_end_write(file);
1058 /* This means we must have transferred all that we could */
1059 /* No need to retry anymore */
1060 if ((ret == 0) || (iocb->ki_left == 0))
1061 ret = iocb->ki_nbytes - iocb->ki_left;
1063 /* If we managed to write some out we return that, rather than
1064 * the eventual error. */
1065 if (opcode == IOCB_CMD_PWRITEV
1066 && ret < 0 && ret != -EIOCBQUEUED
1067 && iocb->ki_nbytes - iocb->ki_left)
1068 ret = iocb->ki_nbytes - iocb->ki_left;
1073 static ssize_t aio_fdsync(struct kiocb *iocb)
1075 struct file *file = iocb->ki_filp;
1076 ssize_t ret = -EINVAL;
1078 if (file->f_op->aio_fsync)
1079 ret = file->f_op->aio_fsync(iocb, 1);
1083 static ssize_t aio_fsync(struct kiocb *iocb)
1085 struct file *file = iocb->ki_filp;
1086 ssize_t ret = -EINVAL;
1088 if (file->f_op->aio_fsync)
1089 ret = file->f_op->aio_fsync(iocb, 0);
1093 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1097 #ifdef CONFIG_COMPAT
1099 ret = compat_rw_copy_check_uvector(type,
1100 (struct compat_iovec __user *)kiocb->ki_buf,
1101 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1105 ret = rw_copy_check_uvector(type,
1106 (struct iovec __user *)kiocb->ki_buf,
1107 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1112 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1116 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1117 kiocb->ki_cur_seg = 0;
1118 /* ki_nbytes/left now reflect bytes instead of segs */
1119 kiocb->ki_nbytes = ret;
1120 kiocb->ki_left = ret;
1127 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1131 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1135 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1136 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1137 kiocb->ki_iovec->iov_len = bytes;
1138 kiocb->ki_nr_segs = 1;
1139 kiocb->ki_cur_seg = 0;
1145 * Performs the initial checks and aio retry method
1146 * setup for the kiocb at the time of io submission.
1148 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1150 struct file *file = kiocb->ki_filp;
1153 switch (kiocb->ki_opcode) {
1154 case IOCB_CMD_PREAD:
1156 if (unlikely(!(file->f_mode & FMODE_READ)))
1159 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1162 ret = aio_setup_single_vector(READ, file, kiocb);
1166 if (file->f_op->aio_read)
1167 kiocb->ki_retry = aio_rw_vect_retry;
1169 case IOCB_CMD_PWRITE:
1171 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1174 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1177 ret = aio_setup_single_vector(WRITE, file, kiocb);
1181 if (file->f_op->aio_write)
1182 kiocb->ki_retry = aio_rw_vect_retry;
1184 case IOCB_CMD_PREADV:
1186 if (unlikely(!(file->f_mode & FMODE_READ)))
1188 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1192 if (file->f_op->aio_read)
1193 kiocb->ki_retry = aio_rw_vect_retry;
1195 case IOCB_CMD_PWRITEV:
1197 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1199 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1203 if (file->f_op->aio_write)
1204 kiocb->ki_retry = aio_rw_vect_retry;
1206 case IOCB_CMD_FDSYNC:
1208 if (file->f_op->aio_fsync)
1209 kiocb->ki_retry = aio_fdsync;
1211 case IOCB_CMD_FSYNC:
1213 if (file->f_op->aio_fsync)
1214 kiocb->ki_retry = aio_fsync;
1217 pr_debug("EINVAL: no operation provided\n");
1221 if (!kiocb->ki_retry)
1227 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1228 struct iocb *iocb, struct kiocb_batch *batch,
1234 /* enforce forwards compatibility on users */
1235 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1236 pr_debug("EINVAL: reserve field set\n");
1240 /* prevent overflows */
1242 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1243 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1244 ((ssize_t)iocb->aio_nbytes < 0)
1246 pr_debug("EINVAL: io_submit: overflow check\n");
1250 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1254 req->ki_filp = fget(iocb->aio_fildes);
1255 if (unlikely(!req->ki_filp)) {
1260 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1262 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1263 * instance of the file* now. The file descriptor must be
1264 * an eventfd() fd, and will be signaled for each completed
1265 * event using the eventfd_signal() function.
1267 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1268 if (IS_ERR(req->ki_eventfd)) {
1269 ret = PTR_ERR(req->ki_eventfd);
1270 req->ki_eventfd = NULL;
1275 ret = put_user(req->ki_key, &user_iocb->aio_key);
1276 if (unlikely(ret)) {
1277 pr_debug("EFAULT: aio_key\n");
1281 req->ki_obj.user = user_iocb;
1282 req->ki_user_data = iocb->aio_data;
1283 req->ki_pos = iocb->aio_offset;
1285 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1286 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1287 req->ki_opcode = iocb->aio_lio_opcode;
1289 ret = aio_setup_iocb(req, compat);
1294 if (unlikely(kiocbIsCancelled(req)))
1297 ret = req->ki_retry(req);
1299 if (ret != -EIOCBQUEUED) {
1301 * There's no easy way to restart the syscall since other AIO's
1302 * may be already running. Just fail this IO with EINTR.
1304 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1305 ret == -ERESTARTNOHAND ||
1306 ret == -ERESTART_RESTARTBLOCK))
1308 aio_complete(req, ret, 0);
1311 aio_put_req(req); /* drop extra ref to req */
1315 spin_lock_irq(&ctx->ctx_lock);
1316 list_del(&req->ki_list);
1317 spin_unlock_irq(&ctx->ctx_lock);
1319 atomic_dec(&ctx->reqs_active);
1320 aio_put_req(req); /* drop extra ref to req */
1321 aio_put_req(req); /* drop i/o ref to req */
1325 long do_io_submit(aio_context_t ctx_id, long nr,
1326 struct iocb __user *__user *iocbpp, bool compat)
1331 struct blk_plug plug;
1332 struct kiocb_batch batch;
1334 if (unlikely(nr < 0))
1337 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1338 nr = LONG_MAX/sizeof(*iocbpp);
1340 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1343 ctx = lookup_ioctx(ctx_id);
1344 if (unlikely(!ctx)) {
1345 pr_debug("EINVAL: invalid context id\n");
1349 kiocb_batch_init(&batch, nr);
1351 blk_start_plug(&plug);
1354 * AKPM: should this return a partial result if some of the IOs were
1355 * successfully submitted?
1357 for (i=0; i<nr; i++) {
1358 struct iocb __user *user_iocb;
1361 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1366 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1371 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1375 blk_finish_plug(&plug);
1377 kiocb_batch_free(ctx, &batch);
1383 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1384 * the number of iocbs queued. May return -EINVAL if the aio_context
1385 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1386 * *iocbpp[0] is not properly initialized, if the operation specified
1387 * is invalid for the file descriptor in the iocb. May fail with
1388 * -EFAULT if any of the data structures point to invalid data. May
1389 * fail with -EBADF if the file descriptor specified in the first
1390 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1391 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1392 * fail with -ENOSYS if not implemented.
1394 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1395 struct iocb __user * __user *, iocbpp)
1397 return do_io_submit(ctx_id, nr, iocbpp, 0);
1401 * Finds a given iocb for cancellation.
1403 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1406 struct list_head *pos;
1408 assert_spin_locked(&ctx->ctx_lock);
1410 /* TODO: use a hash or array, this sucks. */
1411 list_for_each(pos, &ctx->active_reqs) {
1412 struct kiocb *kiocb = list_kiocb(pos);
1413 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1420 * Attempts to cancel an iocb previously passed to io_submit. If
1421 * the operation is successfully cancelled, the resulting event is
1422 * copied into the memory pointed to by result without being placed
1423 * into the completion queue and 0 is returned. May fail with
1424 * -EFAULT if any of the data structures pointed to are invalid.
1425 * May fail with -EINVAL if aio_context specified by ctx_id is
1426 * invalid. May fail with -EAGAIN if the iocb specified was not
1427 * cancelled. Will fail with -ENOSYS if not implemented.
1429 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1430 struct io_event __user *, result)
1432 struct io_event res;
1434 struct kiocb *kiocb;
1438 ret = get_user(key, &iocb->aio_key);
1442 ctx = lookup_ioctx(ctx_id);
1446 spin_lock_irq(&ctx->ctx_lock);
1448 kiocb = lookup_kiocb(ctx, iocb, key);
1450 ret = kiocb_cancel(ctx, kiocb, &res);
1454 spin_unlock_irq(&ctx->ctx_lock);
1457 /* Cancellation succeeded -- copy the result
1458 * into the user's buffer.
1460 if (copy_to_user(result, &res, sizeof(res)))
1470 * Attempts to read at least min_nr events and up to nr events from
1471 * the completion queue for the aio_context specified by ctx_id. If
1472 * it succeeds, the number of read events is returned. May fail with
1473 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1474 * out of range, if timeout is out of range. May fail with -EFAULT
1475 * if any of the memory specified is invalid. May return 0 or
1476 * < min_nr if the timeout specified by timeout has elapsed
1477 * before sufficient events are available, where timeout == NULL
1478 * specifies an infinite timeout. Note that the timeout pointed to by
1479 * timeout is relative and will be updated if not NULL and the
1480 * operation blocks. Will fail with -ENOSYS if not implemented.
1482 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1485 struct io_event __user *, events,
1486 struct timespec __user *, timeout)
1488 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1491 if (likely(ioctx)) {
1492 if (likely(min_nr <= nr && min_nr >= 0))
1493 ret = read_events(ioctx, min_nr, nr, events, timeout);