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;
66 struct mutex ring_lock;
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 spinlock_t completion_lock;
102 struct rcu_head rcu_head;
103 struct work_struct rcu_work;
106 /*------ sysctl variables----*/
107 static DEFINE_SPINLOCK(aio_nr_lock);
108 unsigned long aio_nr; /* current system wide number of aio requests */
109 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
110 /*----end sysctl variables---*/
112 static struct kmem_cache *kiocb_cachep;
113 static struct kmem_cache *kioctx_cachep;
116 * Creates the slab caches used by the aio routines, panic on
117 * failure as this is done early during the boot sequence.
119 static int __init aio_setup(void)
121 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
122 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
124 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
128 __initcall(aio_setup);
130 static void aio_free_ring(struct kioctx *ctx)
132 struct aio_ring_info *info = &ctx->ring_info;
135 for (i=0; i<info->nr_pages; i++)
136 put_page(info->ring_pages[i]);
138 if (info->mmap_size) {
139 vm_munmap(info->mmap_base, info->mmap_size);
142 if (info->ring_pages && info->ring_pages != info->internal_pages)
143 kfree(info->ring_pages);
144 info->ring_pages = NULL;
148 static int aio_setup_ring(struct kioctx *ctx)
150 struct aio_ring *ring;
151 struct aio_ring_info *info = &ctx->ring_info;
152 unsigned nr_events = ctx->max_reqs;
153 struct mm_struct *mm = current->mm;
154 unsigned long size, populate;
157 /* Compensate for the ring buffer's head/tail overlap entry */
158 nr_events += 2; /* 1 is required, 2 for good luck */
160 size = sizeof(struct aio_ring);
161 size += sizeof(struct io_event) * nr_events;
162 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
167 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
170 info->ring_pages = info->internal_pages;
171 if (nr_pages > AIO_RING_PAGES) {
172 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
173 if (!info->ring_pages)
177 info->mmap_size = nr_pages * PAGE_SIZE;
178 pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
179 down_write(&mm->mmap_sem);
180 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
181 PROT_READ|PROT_WRITE,
182 MAP_ANONYMOUS|MAP_PRIVATE, 0,
184 if (IS_ERR((void *)info->mmap_base)) {
185 up_write(&mm->mmap_sem);
191 pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
192 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
193 1, 0, info->ring_pages, NULL);
194 up_write(&mm->mmap_sem);
196 if (unlikely(info->nr_pages != nr_pages)) {
201 mm_populate(info->mmap_base, populate);
203 ctx->user_id = info->mmap_base;
205 info->nr = nr_events; /* trusted copy */
207 ring = kmap_atomic(info->ring_pages[0]);
208 ring->nr = nr_events; /* user copy */
209 ring->id = ctx->user_id;
210 ring->head = ring->tail = 0;
211 ring->magic = AIO_RING_MAGIC;
212 ring->compat_features = AIO_RING_COMPAT_FEATURES;
213 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
214 ring->header_length = sizeof(struct aio_ring);
216 flush_dcache_page(info->ring_pages[0]);
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 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
227 struct kioctx *ctx = req->ki_ctx;
230 spin_lock_irqsave(&ctx->ctx_lock, flags);
232 if (!req->ki_list.next)
233 list_add(&req->ki_list, &ctx->active_reqs);
235 req->ki_cancel = cancel;
237 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
239 EXPORT_SYMBOL(kiocb_set_cancel_fn);
241 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
242 struct io_event *res)
244 kiocb_cancel_fn *old, *cancel;
248 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
249 * actually has a cancel function, hence the cmpxchg()
252 cancel = ACCESS_ONCE(kiocb->ki_cancel);
254 if (!cancel || cancel == KIOCB_CANCELLED)
258 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
259 } while (cancel != old);
261 atomic_inc(&kiocb->ki_users);
262 spin_unlock_irq(&ctx->ctx_lock);
264 memset(res, 0, sizeof(*res));
265 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
266 res->data = kiocb->ki_user_data;
267 ret = cancel(kiocb, res);
269 spin_lock_irq(&ctx->ctx_lock);
274 static void free_ioctx_rcu(struct rcu_head *head)
276 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
277 kmem_cache_free(kioctx_cachep, ctx);
281 * When this function runs, the kioctx has been removed from the "hash table"
282 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
283 * now it's safe to cancel any that need to be.
285 static void free_ioctx(struct kioctx *ctx)
290 spin_lock_irq(&ctx->ctx_lock);
292 while (!list_empty(&ctx->active_reqs)) {
293 req = list_first_entry(&ctx->active_reqs,
294 struct kiocb, ki_list);
296 list_del_init(&req->ki_list);
297 kiocb_cancel(ctx, req, &res);
300 spin_unlock_irq(&ctx->ctx_lock);
302 wait_event(ctx->wait, !atomic_read(&ctx->reqs_active));
306 spin_lock(&aio_nr_lock);
307 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
308 aio_nr -= ctx->max_reqs;
309 spin_unlock(&aio_nr_lock);
311 pr_debug("freeing %p\n", ctx);
314 * Here the call_rcu() is between the wait_event() for reqs_active to
315 * hit 0, and freeing the ioctx.
317 * aio_complete() decrements reqs_active, but it has to touch the ioctx
318 * after to issue a wakeup so we use rcu.
320 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
323 static void put_ioctx(struct kioctx *ctx)
325 if (unlikely(atomic_dec_and_test(&ctx->users)))
330 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
332 static struct kioctx *ioctx_alloc(unsigned nr_events)
334 struct mm_struct *mm = current->mm;
338 /* Prevent overflows */
339 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
340 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
341 pr_debug("ENOMEM: nr_events too high\n");
342 return ERR_PTR(-EINVAL);
345 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
346 return ERR_PTR(-EAGAIN);
348 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
350 return ERR_PTR(-ENOMEM);
352 ctx->max_reqs = nr_events;
354 atomic_set(&ctx->users, 2);
355 atomic_set(&ctx->dead, 0);
356 spin_lock_init(&ctx->ctx_lock);
357 spin_lock_init(&ctx->completion_lock);
358 mutex_init(&ctx->ring_info.ring_lock);
359 init_waitqueue_head(&ctx->wait);
361 INIT_LIST_HEAD(&ctx->active_reqs);
363 if (aio_setup_ring(ctx) < 0)
366 /* limit the number of system wide aios */
367 spin_lock(&aio_nr_lock);
368 if (aio_nr + nr_events > aio_max_nr ||
369 aio_nr + nr_events < aio_nr) {
370 spin_unlock(&aio_nr_lock);
373 aio_nr += ctx->max_reqs;
374 spin_unlock(&aio_nr_lock);
376 /* now link into global list. */
377 spin_lock(&mm->ioctx_lock);
378 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
379 spin_unlock(&mm->ioctx_lock);
381 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
382 ctx, ctx->user_id, mm, ctx->ring_info.nr);
389 kmem_cache_free(kioctx_cachep, ctx);
390 pr_debug("error allocating ioctx %d\n", err);
394 static void kill_ioctx_work(struct work_struct *work)
396 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
398 wake_up_all(&ctx->wait);
402 static void kill_ioctx_rcu(struct rcu_head *head)
404 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
406 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
407 schedule_work(&ctx->rcu_work);
411 * Cancels all outstanding aio requests on an aio context. Used
412 * when the processes owning a context have all exited to encourage
413 * the rapid destruction of the kioctx.
415 static void kill_ioctx(struct kioctx *ctx)
417 if (!atomic_xchg(&ctx->dead, 1)) {
418 hlist_del_rcu(&ctx->list);
419 /* Between hlist_del_rcu() and dropping the initial ref */
423 * We can't punt to workqueue here because put_ioctx() ->
424 * free_ioctx() will unmap the ringbuffer, and that has to be
425 * done in the original process's context. kill_ioctx_rcu/work()
426 * exist for exit_aio(), as in that path free_ioctx() won't do
429 kill_ioctx_work(&ctx->rcu_work);
433 /* wait_on_sync_kiocb:
434 * Waits on the given sync kiocb to complete.
436 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
438 while (atomic_read(&iocb->ki_users)) {
439 set_current_state(TASK_UNINTERRUPTIBLE);
440 if (!atomic_read(&iocb->ki_users))
444 __set_current_state(TASK_RUNNING);
445 return iocb->ki_user_data;
447 EXPORT_SYMBOL(wait_on_sync_kiocb);
450 * exit_aio: called when the last user of mm goes away. At this point, there is
451 * no way for any new requests to be submited or any of the io_* syscalls to be
452 * called on the context.
454 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
457 void exit_aio(struct mm_struct *mm)
460 struct hlist_node *n;
462 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
463 if (1 != atomic_read(&ctx->users))
465 "exit_aio:ioctx still alive: %d %d %d\n",
466 atomic_read(&ctx->users),
467 atomic_read(&ctx->dead),
468 atomic_read(&ctx->reqs_active));
470 * We don't need to bother with munmap() here -
471 * exit_mmap(mm) is coming and it'll unmap everything.
472 * Since aio_free_ring() uses non-zero ->mmap_size
473 * as indicator that it needs to unmap the area,
474 * just set it to 0; aio_free_ring() is the only
475 * place that uses ->mmap_size, so it's safe.
477 ctx->ring_info.mmap_size = 0;
479 if (!atomic_xchg(&ctx->dead, 1)) {
480 hlist_del_rcu(&ctx->list);
481 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
487 * Allocate a slot for an aio request. Increments the ki_users count
488 * of the kioctx so that the kioctx stays around until all requests are
489 * complete. Returns NULL if no requests are free.
491 * Returns with kiocb->ki_users set to 2. The io submit code path holds
492 * an extra reference while submitting the i/o.
493 * This prevents races between the aio code path referencing the
494 * req (after submitting it) and aio_complete() freeing the req.
496 static struct kiocb *__aio_get_req(struct kioctx *ctx)
498 struct kiocb *req = NULL;
500 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
504 atomic_set(&req->ki_users, 2);
511 * struct kiocb's are allocated in batches to reduce the number of
512 * times the ctx lock is acquired and released.
514 #define KIOCB_BATCH_SIZE 32L
516 struct list_head head;
517 long count; /* number of requests left to allocate */
520 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
522 INIT_LIST_HEAD(&batch->head);
523 batch->count = total;
526 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
528 struct kiocb *req, *n;
530 if (list_empty(&batch->head))
533 spin_lock_irq(&ctx->ctx_lock);
534 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
535 list_del(&req->ki_batch);
536 kmem_cache_free(kiocb_cachep, req);
537 atomic_dec(&ctx->reqs_active);
539 spin_unlock_irq(&ctx->ctx_lock);
543 * Allocate a batch of kiocbs. This avoids taking and dropping the
544 * context lock a lot during setup.
546 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
548 unsigned short allocated, to_alloc;
550 struct kiocb *req, *n;
551 struct aio_ring *ring;
553 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
554 for (allocated = 0; allocated < to_alloc; allocated++) {
555 req = __aio_get_req(ctx);
557 /* allocation failed, go with what we've got */
559 list_add(&req->ki_batch, &batch->head);
565 spin_lock_irq(&ctx->ctx_lock);
566 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
568 avail = aio_ring_avail(&ctx->ring_info, ring) -
569 atomic_read(&ctx->reqs_active);
571 if (avail < allocated) {
572 /* Trim back the number of requests. */
573 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
574 list_del(&req->ki_batch);
575 kmem_cache_free(kiocb_cachep, req);
576 if (--allocated <= avail)
581 batch->count -= allocated;
582 atomic_add(allocated, &ctx->reqs_active);
585 spin_unlock_irq(&ctx->ctx_lock);
591 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
592 struct kiocb_batch *batch)
596 if (list_empty(&batch->head))
597 if (kiocb_batch_refill(ctx, batch) == 0)
599 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
600 list_del(&req->ki_batch);
604 static void kiocb_free(struct kiocb *req)
608 if (req->ki_eventfd != NULL)
609 eventfd_ctx_put(req->ki_eventfd);
612 if (req->ki_iovec != &req->ki_inline_vec)
613 kfree(req->ki_iovec);
614 kmem_cache_free(kiocb_cachep, req);
617 void aio_put_req(struct kiocb *req)
619 if (atomic_dec_and_test(&req->ki_users))
622 EXPORT_SYMBOL(aio_put_req);
624 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
626 struct mm_struct *mm = current->mm;
627 struct kioctx *ctx, *ret = NULL;
631 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
632 if (ctx->user_id == ctx_id) {
633 atomic_inc(&ctx->users);
644 * Called when the io request on the given iocb is complete.
646 void aio_complete(struct kiocb *iocb, long res, long res2)
648 struct kioctx *ctx = iocb->ki_ctx;
649 struct aio_ring_info *info;
650 struct aio_ring *ring;
651 struct io_event *ev_page, *event;
656 * Special case handling for sync iocbs:
657 * - events go directly into the iocb for fast handling
658 * - the sync task with the iocb in its stack holds the single iocb
659 * ref, no other paths have a way to get another ref
660 * - the sync task helpfully left a reference to itself in the iocb
662 if (is_sync_kiocb(iocb)) {
663 BUG_ON(atomic_read(&iocb->ki_users) != 1);
664 iocb->ki_user_data = res;
665 atomic_set(&iocb->ki_users, 0);
666 wake_up_process(iocb->ki_obj.tsk);
670 info = &ctx->ring_info;
673 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
674 * need to issue a wakeup after decrementing reqs_active.
678 if (iocb->ki_list.next) {
681 spin_lock_irqsave(&ctx->ctx_lock, flags);
682 list_del(&iocb->ki_list);
683 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
687 * cancelled requests don't get events, userland was given one
688 * when the event got cancelled.
690 if (unlikely(xchg(&iocb->ki_cancel,
691 KIOCB_CANCELLED) == KIOCB_CANCELLED))
695 * Add a completion event to the ring buffer. Must be done holding
696 * ctx->ctx_lock to prevent other code from messing with the tail
697 * pointer since we might be called from irq context.
699 spin_lock_irqsave(&ctx->completion_lock, flags);
702 pos = tail + AIO_EVENTS_OFFSET;
704 if (++tail >= info->nr)
707 ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
708 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
710 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
711 event->data = iocb->ki_user_data;
715 kunmap_atomic(ev_page);
716 flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
718 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
719 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
722 /* after flagging the request as done, we
723 * must never even look at it again
725 smp_wmb(); /* make event visible before updating tail */
729 ring = kmap_atomic(info->ring_pages[0]);
732 flush_dcache_page(info->ring_pages[0]);
734 spin_unlock_irqrestore(&ctx->completion_lock, flags);
736 pr_debug("added to ring %p at [%u]\n", iocb, tail);
739 * Check if the user asked us to deliver the result through an
740 * eventfd. The eventfd_signal() function is safe to be called
743 if (iocb->ki_eventfd != NULL)
744 eventfd_signal(iocb->ki_eventfd, 1);
747 /* everything turned out well, dispose of the aiocb. */
749 atomic_dec(&ctx->reqs_active);
752 * We have to order our ring_info tail store above and test
753 * of the wait list below outside the wait lock. This is
754 * like in wake_up_bit() where clearing a bit has to be
755 * ordered with the unlocked test.
759 if (waitqueue_active(&ctx->wait))
764 EXPORT_SYMBOL(aio_complete);
767 * Pull an event off of the ioctx's event ring. Returns the number of
770 static long aio_read_events_ring(struct kioctx *ctx,
771 struct io_event __user *event, long nr)
773 struct aio_ring_info *info = &ctx->ring_info;
774 struct aio_ring *ring;
779 mutex_lock(&info->ring_lock);
781 ring = kmap_atomic(info->ring_pages[0]);
785 pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
787 if (head == info->tail)
795 avail = (head <= info->tail ? info->tail : info->nr) - head;
796 if (head == info->tail)
799 avail = min(avail, nr - ret);
800 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
801 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
803 pos = head + AIO_EVENTS_OFFSET;
804 page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
805 pos %= AIO_EVENTS_PER_PAGE;
808 copy_ret = copy_to_user(event + ret, ev + pos,
809 sizeof(*ev) * avail);
812 if (unlikely(copy_ret)) {
822 ring = kmap_atomic(info->ring_pages[0]);
825 flush_dcache_page(info->ring_pages[0]);
827 pr_debug("%li h%u t%u\n", ret, head, info->tail);
829 mutex_unlock(&info->ring_lock);
834 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
835 struct io_event __user *event, long *i)
837 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
842 if (unlikely(atomic_read(&ctx->dead)))
848 return ret < 0 || *i >= min_nr;
851 static long read_events(struct kioctx *ctx, long min_nr, long nr,
852 struct io_event __user *event,
853 struct timespec __user *timeout)
855 ktime_t until = { .tv64 = KTIME_MAX };
861 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
864 until = timespec_to_ktime(ts);
868 * Note that aio_read_events() is being called as the conditional - i.e.
869 * we're calling it after prepare_to_wait() has set task state to
870 * TASK_INTERRUPTIBLE.
872 * But aio_read_events() can block, and if it blocks it's going to flip
873 * the task state back to TASK_RUNNING.
875 * This should be ok, provided it doesn't flip the state back to
876 * TASK_RUNNING and return 0 too much - that causes us to spin. That
877 * will only happen if the mutex_lock() call blocks, and we then find
878 * the ringbuffer empty. So in practice we should be ok, but it's
879 * something to be aware of when touching this code.
881 wait_event_interruptible_hrtimeout(ctx->wait,
882 aio_read_events(ctx, min_nr, nr, event, &ret), until);
884 if (!ret && signal_pending(current))
891 * Create an aio_context capable of receiving at least nr_events.
892 * ctxp must not point to an aio_context that already exists, and
893 * must be initialized to 0 prior to the call. On successful
894 * creation of the aio_context, *ctxp is filled in with the resulting
895 * handle. May fail with -EINVAL if *ctxp is not initialized,
896 * if the specified nr_events exceeds internal limits. May fail
897 * with -EAGAIN if the specified nr_events exceeds the user's limit
898 * of available events. May fail with -ENOMEM if insufficient kernel
899 * resources are available. May fail with -EFAULT if an invalid
900 * pointer is passed for ctxp. Will fail with -ENOSYS if not
903 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
905 struct kioctx *ioctx = NULL;
909 ret = get_user(ctx, ctxp);
914 if (unlikely(ctx || nr_events == 0)) {
915 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
920 ioctx = ioctx_alloc(nr_events);
921 ret = PTR_ERR(ioctx);
922 if (!IS_ERR(ioctx)) {
923 ret = put_user(ioctx->user_id, ctxp);
934 * Destroy the aio_context specified. May cancel any outstanding
935 * AIOs and block on completion. Will fail with -ENOSYS if not
936 * implemented. May fail with -EINVAL if the context pointed to
939 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
941 struct kioctx *ioctx = lookup_ioctx(ctx);
942 if (likely(NULL != ioctx)) {
947 pr_debug("EINVAL: io_destroy: invalid context id\n");
951 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
953 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
957 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
958 ssize_t this = min((ssize_t)iov->iov_len, ret);
959 iov->iov_base += this;
960 iov->iov_len -= this;
961 iocb->ki_left -= this;
963 if (iov->iov_len == 0) {
969 /* the caller should not have done more io than what fit in
970 * the remaining iovecs */
971 BUG_ON(ret > 0 && iocb->ki_left == 0);
974 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
976 struct file *file = iocb->ki_filp;
977 struct address_space *mapping = file->f_mapping;
978 struct inode *inode = mapping->host;
979 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
980 unsigned long, loff_t);
982 unsigned short opcode;
984 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
985 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
986 rw_op = file->f_op->aio_read;
987 opcode = IOCB_CMD_PREADV;
989 rw_op = file->f_op->aio_write;
990 opcode = IOCB_CMD_PWRITEV;
993 /* This matches the pread()/pwrite() logic */
994 if (iocb->ki_pos < 0)
997 if (opcode == IOCB_CMD_PWRITEV)
998 file_start_write(file);
1000 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1001 iocb->ki_nr_segs - iocb->ki_cur_seg,
1004 aio_advance_iovec(iocb, ret);
1006 /* retry all partial writes. retry partial reads as long as its a
1008 } while (ret > 0 && iocb->ki_left > 0 &&
1009 (opcode == IOCB_CMD_PWRITEV ||
1010 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1011 if (opcode == IOCB_CMD_PWRITEV)
1012 file_end_write(file);
1014 /* This means we must have transferred all that we could */
1015 /* No need to retry anymore */
1016 if ((ret == 0) || (iocb->ki_left == 0))
1017 ret = iocb->ki_nbytes - iocb->ki_left;
1019 /* If we managed to write some out we return that, rather than
1020 * the eventual error. */
1021 if (opcode == IOCB_CMD_PWRITEV
1022 && ret < 0 && ret != -EIOCBQUEUED
1023 && iocb->ki_nbytes - iocb->ki_left)
1024 ret = iocb->ki_nbytes - iocb->ki_left;
1029 static ssize_t aio_fdsync(struct kiocb *iocb)
1031 struct file *file = iocb->ki_filp;
1032 ssize_t ret = -EINVAL;
1034 if (file->f_op->aio_fsync)
1035 ret = file->f_op->aio_fsync(iocb, 1);
1039 static ssize_t aio_fsync(struct kiocb *iocb)
1041 struct file *file = iocb->ki_filp;
1042 ssize_t ret = -EINVAL;
1044 if (file->f_op->aio_fsync)
1045 ret = file->f_op->aio_fsync(iocb, 0);
1049 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1053 #ifdef CONFIG_COMPAT
1055 ret = compat_rw_copy_check_uvector(type,
1056 (struct compat_iovec __user *)kiocb->ki_buf,
1057 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1061 ret = rw_copy_check_uvector(type,
1062 (struct iovec __user *)kiocb->ki_buf,
1063 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1068 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1072 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1073 kiocb->ki_cur_seg = 0;
1074 /* ki_nbytes/left now reflect bytes instead of segs */
1075 kiocb->ki_nbytes = ret;
1076 kiocb->ki_left = ret;
1083 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1087 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1091 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1092 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1093 kiocb->ki_iovec->iov_len = bytes;
1094 kiocb->ki_nr_segs = 1;
1095 kiocb->ki_cur_seg = 0;
1101 * Performs the initial checks and aio retry method
1102 * setup for the kiocb at the time of io submission.
1104 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1106 struct file *file = kiocb->ki_filp;
1109 switch (kiocb->ki_opcode) {
1110 case IOCB_CMD_PREAD:
1112 if (unlikely(!(file->f_mode & FMODE_READ)))
1115 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1118 ret = aio_setup_single_vector(READ, file, kiocb);
1122 if (file->f_op->aio_read)
1123 kiocb->ki_retry = aio_rw_vect_retry;
1125 case IOCB_CMD_PWRITE:
1127 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1130 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1133 ret = aio_setup_single_vector(WRITE, file, kiocb);
1137 if (file->f_op->aio_write)
1138 kiocb->ki_retry = aio_rw_vect_retry;
1140 case IOCB_CMD_PREADV:
1142 if (unlikely(!(file->f_mode & FMODE_READ)))
1144 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1148 if (file->f_op->aio_read)
1149 kiocb->ki_retry = aio_rw_vect_retry;
1151 case IOCB_CMD_PWRITEV:
1153 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1155 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1159 if (file->f_op->aio_write)
1160 kiocb->ki_retry = aio_rw_vect_retry;
1162 case IOCB_CMD_FDSYNC:
1164 if (file->f_op->aio_fsync)
1165 kiocb->ki_retry = aio_fdsync;
1167 case IOCB_CMD_FSYNC:
1169 if (file->f_op->aio_fsync)
1170 kiocb->ki_retry = aio_fsync;
1173 pr_debug("EINVAL: no operation provided\n");
1177 if (!kiocb->ki_retry)
1183 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1184 struct iocb *iocb, struct kiocb_batch *batch,
1190 /* enforce forwards compatibility on users */
1191 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1192 pr_debug("EINVAL: reserve field set\n");
1196 /* prevent overflows */
1198 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1199 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1200 ((ssize_t)iocb->aio_nbytes < 0)
1202 pr_debug("EINVAL: io_submit: overflow check\n");
1206 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1210 req->ki_filp = fget(iocb->aio_fildes);
1211 if (unlikely(!req->ki_filp)) {
1216 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1218 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1219 * instance of the file* now. The file descriptor must be
1220 * an eventfd() fd, and will be signaled for each completed
1221 * event using the eventfd_signal() function.
1223 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1224 if (IS_ERR(req->ki_eventfd)) {
1225 ret = PTR_ERR(req->ki_eventfd);
1226 req->ki_eventfd = NULL;
1231 ret = put_user(req->ki_key, &user_iocb->aio_key);
1232 if (unlikely(ret)) {
1233 pr_debug("EFAULT: aio_key\n");
1237 req->ki_obj.user = user_iocb;
1238 req->ki_user_data = iocb->aio_data;
1239 req->ki_pos = iocb->aio_offset;
1241 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1242 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1243 req->ki_opcode = iocb->aio_lio_opcode;
1245 ret = aio_setup_iocb(req, compat);
1249 ret = req->ki_retry(req);
1250 if (ret != -EIOCBQUEUED) {
1252 * There's no easy way to restart the syscall since other AIO's
1253 * may be already running. Just fail this IO with EINTR.
1255 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1256 ret == -ERESTARTNOHAND ||
1257 ret == -ERESTART_RESTARTBLOCK))
1259 aio_complete(req, ret, 0);
1262 aio_put_req(req); /* drop extra ref to req */
1266 atomic_dec(&ctx->reqs_active);
1267 aio_put_req(req); /* drop extra ref to req */
1268 aio_put_req(req); /* drop i/o ref to req */
1272 long do_io_submit(aio_context_t ctx_id, long nr,
1273 struct iocb __user *__user *iocbpp, bool compat)
1278 struct blk_plug plug;
1279 struct kiocb_batch batch;
1281 if (unlikely(nr < 0))
1284 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1285 nr = LONG_MAX/sizeof(*iocbpp);
1287 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1290 ctx = lookup_ioctx(ctx_id);
1291 if (unlikely(!ctx)) {
1292 pr_debug("EINVAL: invalid context id\n");
1296 kiocb_batch_init(&batch, nr);
1298 blk_start_plug(&plug);
1301 * AKPM: should this return a partial result if some of the IOs were
1302 * successfully submitted?
1304 for (i=0; i<nr; i++) {
1305 struct iocb __user *user_iocb;
1308 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1313 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1318 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1322 blk_finish_plug(&plug);
1324 kiocb_batch_free(ctx, &batch);
1330 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1331 * the number of iocbs queued. May return -EINVAL if the aio_context
1332 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1333 * *iocbpp[0] is not properly initialized, if the operation specified
1334 * is invalid for the file descriptor in the iocb. May fail with
1335 * -EFAULT if any of the data structures point to invalid data. May
1336 * fail with -EBADF if the file descriptor specified in the first
1337 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1338 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1339 * fail with -ENOSYS if not implemented.
1341 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1342 struct iocb __user * __user *, iocbpp)
1344 return do_io_submit(ctx_id, nr, iocbpp, 0);
1348 * Finds a given iocb for cancellation.
1350 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1353 struct list_head *pos;
1355 assert_spin_locked(&ctx->ctx_lock);
1357 /* TODO: use a hash or array, this sucks. */
1358 list_for_each(pos, &ctx->active_reqs) {
1359 struct kiocb *kiocb = list_kiocb(pos);
1360 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1367 * Attempts to cancel an iocb previously passed to io_submit. If
1368 * the operation is successfully cancelled, the resulting event is
1369 * copied into the memory pointed to by result without being placed
1370 * into the completion queue and 0 is returned. May fail with
1371 * -EFAULT if any of the data structures pointed to are invalid.
1372 * May fail with -EINVAL if aio_context specified by ctx_id is
1373 * invalid. May fail with -EAGAIN if the iocb specified was not
1374 * cancelled. Will fail with -ENOSYS if not implemented.
1376 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1377 struct io_event __user *, result)
1379 struct io_event res;
1381 struct kiocb *kiocb;
1385 ret = get_user(key, &iocb->aio_key);
1389 ctx = lookup_ioctx(ctx_id);
1393 spin_lock_irq(&ctx->ctx_lock);
1395 kiocb = lookup_kiocb(ctx, iocb, key);
1397 ret = kiocb_cancel(ctx, kiocb, &res);
1401 spin_unlock_irq(&ctx->ctx_lock);
1404 /* Cancellation succeeded -- copy the result
1405 * into the user's buffer.
1407 if (copy_to_user(result, &res, sizeof(res)))
1417 * Attempts to read at least min_nr events and up to nr events from
1418 * the completion queue for the aio_context specified by ctx_id. If
1419 * it succeeds, the number of read events is returned. May fail with
1420 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1421 * out of range, if timeout is out of range. May fail with -EFAULT
1422 * if any of the memory specified is invalid. May return 0 or
1423 * < min_nr if the timeout specified by timeout has elapsed
1424 * before sufficient events are available, where timeout == NULL
1425 * specifies an infinite timeout. Note that the timeout pointed to by
1426 * timeout is relative and will be updated if not NULL and the
1427 * operation blocks. Will fail with -ENOSYS if not implemented.
1429 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1432 struct io_event __user *, events,
1433 struct timespec __user *, timeout)
1435 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1438 if (likely(ioctx)) {
1439 if (likely(min_nr <= nr && min_nr >= 0))
1440 ret = read_events(ioctx, min_nr, nr, events, timeout);