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];
78 /* This needs improving */
79 unsigned long user_id;
80 struct hlist_node list;
82 wait_queue_head_t wait;
87 struct list_head active_reqs; /* used for cancellation */
90 * This is what userspace passed to io_setup(), it's not used for
91 * anything but counting against the global max_reqs quota.
93 * The real limit is ring->nr - 1, which will be larger (see
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)
287 struct aio_ring_info *info = &ctx->ring_info;
288 struct aio_ring *ring;
291 unsigned head, avail;
293 spin_lock_irq(&ctx->ctx_lock);
295 while (!list_empty(&ctx->active_reqs)) {
296 req = list_first_entry(&ctx->active_reqs,
297 struct kiocb, ki_list);
299 list_del_init(&req->ki_list);
300 kiocb_cancel(ctx, req, &res);
303 spin_unlock_irq(&ctx->ctx_lock);
305 ring = kmap_atomic(info->ring_pages[0]);
309 while (atomic_read(&ctx->reqs_active) > 0) {
310 wait_event(ctx->wait, head != info->tail);
312 avail = (head <= info->tail ? info->tail : info->nr) - head;
314 atomic_sub(avail, &ctx->reqs_active);
319 WARN_ON(atomic_read(&ctx->reqs_active) < 0);
323 spin_lock(&aio_nr_lock);
324 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
325 aio_nr -= ctx->max_reqs;
326 spin_unlock(&aio_nr_lock);
328 pr_debug("freeing %p\n", ctx);
331 * Here the call_rcu() is between the wait_event() for reqs_active to
332 * hit 0, and freeing the ioctx.
334 * aio_complete() decrements reqs_active, but it has to touch the ioctx
335 * after to issue a wakeup so we use rcu.
337 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
340 static void put_ioctx(struct kioctx *ctx)
342 if (unlikely(atomic_dec_and_test(&ctx->users)))
347 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
349 static struct kioctx *ioctx_alloc(unsigned nr_events)
351 struct mm_struct *mm = current->mm;
355 /* Prevent overflows */
356 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
357 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
358 pr_debug("ENOMEM: nr_events too high\n");
359 return ERR_PTR(-EINVAL);
362 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
363 return ERR_PTR(-EAGAIN);
365 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
367 return ERR_PTR(-ENOMEM);
369 ctx->max_reqs = nr_events;
371 atomic_set(&ctx->users, 2);
372 atomic_set(&ctx->dead, 0);
373 spin_lock_init(&ctx->ctx_lock);
374 spin_lock_init(&ctx->completion_lock);
375 mutex_init(&ctx->ring_info.ring_lock);
376 init_waitqueue_head(&ctx->wait);
378 INIT_LIST_HEAD(&ctx->active_reqs);
380 if (aio_setup_ring(ctx) < 0)
383 /* limit the number of system wide aios */
384 spin_lock(&aio_nr_lock);
385 if (aio_nr + nr_events > aio_max_nr ||
386 aio_nr + nr_events < aio_nr) {
387 spin_unlock(&aio_nr_lock);
390 aio_nr += ctx->max_reqs;
391 spin_unlock(&aio_nr_lock);
393 /* now link into global list. */
394 spin_lock(&mm->ioctx_lock);
395 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
396 spin_unlock(&mm->ioctx_lock);
398 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
399 ctx, ctx->user_id, mm, ctx->ring_info.nr);
406 kmem_cache_free(kioctx_cachep, ctx);
407 pr_debug("error allocating ioctx %d\n", err);
411 static void kill_ioctx_work(struct work_struct *work)
413 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
415 wake_up_all(&ctx->wait);
419 static void kill_ioctx_rcu(struct rcu_head *head)
421 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
423 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
424 schedule_work(&ctx->rcu_work);
428 * Cancels all outstanding aio requests on an aio context. Used
429 * when the processes owning a context have all exited to encourage
430 * the rapid destruction of the kioctx.
432 static void kill_ioctx(struct kioctx *ctx)
434 if (!atomic_xchg(&ctx->dead, 1)) {
435 hlist_del_rcu(&ctx->list);
436 /* Between hlist_del_rcu() and dropping the initial ref */
440 * We can't punt to workqueue here because put_ioctx() ->
441 * free_ioctx() will unmap the ringbuffer, and that has to be
442 * done in the original process's context. kill_ioctx_rcu/work()
443 * exist for exit_aio(), as in that path free_ioctx() won't do
446 kill_ioctx_work(&ctx->rcu_work);
450 /* wait_on_sync_kiocb:
451 * Waits on the given sync kiocb to complete.
453 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
455 while (atomic_read(&iocb->ki_users)) {
456 set_current_state(TASK_UNINTERRUPTIBLE);
457 if (!atomic_read(&iocb->ki_users))
461 __set_current_state(TASK_RUNNING);
462 return iocb->ki_user_data;
464 EXPORT_SYMBOL(wait_on_sync_kiocb);
467 * exit_aio: called when the last user of mm goes away. At this point, there is
468 * no way for any new requests to be submited or any of the io_* syscalls to be
469 * called on the context.
471 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
474 void exit_aio(struct mm_struct *mm)
477 struct hlist_node *n;
479 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
480 if (1 != atomic_read(&ctx->users))
482 "exit_aio:ioctx still alive: %d %d %d\n",
483 atomic_read(&ctx->users),
484 atomic_read(&ctx->dead),
485 atomic_read(&ctx->reqs_active));
487 * We don't need to bother with munmap() here -
488 * exit_mmap(mm) is coming and it'll unmap everything.
489 * Since aio_free_ring() uses non-zero ->mmap_size
490 * as indicator that it needs to unmap the area,
491 * just set it to 0; aio_free_ring() is the only
492 * place that uses ->mmap_size, so it's safe.
494 ctx->ring_info.mmap_size = 0;
496 if (!atomic_xchg(&ctx->dead, 1)) {
497 hlist_del_rcu(&ctx->list);
498 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
504 * Allocate a slot for an aio request. Increments the ki_users count
505 * of the kioctx so that the kioctx stays around until all requests are
506 * complete. Returns NULL if no requests are free.
508 * Returns with kiocb->ki_users set to 2. The io submit code path holds
509 * an extra reference while submitting the i/o.
510 * This prevents races between the aio code path referencing the
511 * req (after submitting it) and aio_complete() freeing the req.
513 static struct kiocb *__aio_get_req(struct kioctx *ctx)
515 struct kiocb *req = NULL;
517 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
521 atomic_set(&req->ki_users, 2);
528 * struct kiocb's are allocated in batches to reduce the number of
529 * times the ctx lock is acquired and released.
531 #define KIOCB_BATCH_SIZE 32L
533 struct list_head head;
534 long count; /* number of requests left to allocate */
537 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
539 INIT_LIST_HEAD(&batch->head);
540 batch->count = total;
543 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
545 struct kiocb *req, *n;
547 if (list_empty(&batch->head))
550 spin_lock_irq(&ctx->ctx_lock);
551 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
552 list_del(&req->ki_batch);
553 kmem_cache_free(kiocb_cachep, req);
554 atomic_dec(&ctx->reqs_active);
556 spin_unlock_irq(&ctx->ctx_lock);
560 * Allocate a batch of kiocbs. This avoids taking and dropping the
561 * context lock a lot during setup.
563 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
565 unsigned short allocated, to_alloc;
567 struct kiocb *req, *n;
569 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
570 for (allocated = 0; allocated < to_alloc; allocated++) {
571 req = __aio_get_req(ctx);
573 /* allocation failed, go with what we've got */
575 list_add(&req->ki_batch, &batch->head);
581 spin_lock_irq(&ctx->ctx_lock);
583 avail = ctx->ring_info.nr - atomic_read(&ctx->reqs_active) - 1;
585 if (avail < allocated) {
586 /* Trim back the number of requests. */
587 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
588 list_del(&req->ki_batch);
589 kmem_cache_free(kiocb_cachep, req);
590 if (--allocated <= avail)
595 batch->count -= allocated;
596 atomic_add(allocated, &ctx->reqs_active);
598 spin_unlock_irq(&ctx->ctx_lock);
604 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
605 struct kiocb_batch *batch)
609 if (list_empty(&batch->head))
610 if (kiocb_batch_refill(ctx, batch) == 0)
612 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
613 list_del(&req->ki_batch);
617 static void kiocb_free(struct kiocb *req)
621 if (req->ki_eventfd != NULL)
622 eventfd_ctx_put(req->ki_eventfd);
625 if (req->ki_iovec != &req->ki_inline_vec)
626 kfree(req->ki_iovec);
627 kmem_cache_free(kiocb_cachep, req);
630 void aio_put_req(struct kiocb *req)
632 if (atomic_dec_and_test(&req->ki_users))
635 EXPORT_SYMBOL(aio_put_req);
637 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
639 struct mm_struct *mm = current->mm;
640 struct kioctx *ctx, *ret = NULL;
644 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
645 if (ctx->user_id == ctx_id) {
646 atomic_inc(&ctx->users);
657 * Called when the io request on the given iocb is complete.
659 void aio_complete(struct kiocb *iocb, long res, long res2)
661 struct kioctx *ctx = iocb->ki_ctx;
662 struct aio_ring_info *info;
663 struct aio_ring *ring;
664 struct io_event *ev_page, *event;
669 * Special case handling for sync iocbs:
670 * - events go directly into the iocb for fast handling
671 * - the sync task with the iocb in its stack holds the single iocb
672 * ref, no other paths have a way to get another ref
673 * - the sync task helpfully left a reference to itself in the iocb
675 if (is_sync_kiocb(iocb)) {
676 BUG_ON(atomic_read(&iocb->ki_users) != 1);
677 iocb->ki_user_data = res;
678 atomic_set(&iocb->ki_users, 0);
679 wake_up_process(iocb->ki_obj.tsk);
683 info = &ctx->ring_info;
686 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
687 * need to issue a wakeup after decrementing reqs_active.
691 if (iocb->ki_list.next) {
694 spin_lock_irqsave(&ctx->ctx_lock, flags);
695 list_del(&iocb->ki_list);
696 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
700 * cancelled requests don't get events, userland was given one
701 * when the event got cancelled.
703 if (unlikely(xchg(&iocb->ki_cancel,
704 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
705 atomic_dec(&ctx->reqs_active);
706 /* Still need the wake_up in case free_ioctx is waiting */
711 * Add a completion event to the ring buffer. Must be done holding
712 * ctx->ctx_lock to prevent other code from messing with the tail
713 * pointer since we might be called from irq context.
715 spin_lock_irqsave(&ctx->completion_lock, flags);
718 pos = tail + AIO_EVENTS_OFFSET;
720 if (++tail >= info->nr)
723 ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
724 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
726 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
727 event->data = iocb->ki_user_data;
731 kunmap_atomic(ev_page);
732 flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
734 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
735 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
738 /* after flagging the request as done, we
739 * must never even look at it again
741 smp_wmb(); /* make event visible before updating tail */
745 ring = kmap_atomic(info->ring_pages[0]);
748 flush_dcache_page(info->ring_pages[0]);
750 spin_unlock_irqrestore(&ctx->completion_lock, flags);
752 pr_debug("added to ring %p at [%u]\n", iocb, tail);
755 * Check if the user asked us to deliver the result through an
756 * eventfd. The eventfd_signal() function is safe to be called
759 if (iocb->ki_eventfd != NULL)
760 eventfd_signal(iocb->ki_eventfd, 1);
763 /* everything turned out well, dispose of the aiocb. */
767 * We have to order our ring_info tail store above and test
768 * of the wait list below outside the wait lock. This is
769 * like in wake_up_bit() where clearing a bit has to be
770 * ordered with the unlocked test.
774 if (waitqueue_active(&ctx->wait))
779 EXPORT_SYMBOL(aio_complete);
782 * Pull an event off of the ioctx's event ring. Returns the number of
785 static long aio_read_events_ring(struct kioctx *ctx,
786 struct io_event __user *event, long nr)
788 struct aio_ring_info *info = &ctx->ring_info;
789 struct aio_ring *ring;
794 mutex_lock(&info->ring_lock);
796 ring = kmap_atomic(info->ring_pages[0]);
800 pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
802 if (head == info->tail)
810 avail = (head <= info->tail ? info->tail : info->nr) - head;
811 if (head == info->tail)
814 avail = min(avail, nr - ret);
815 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
816 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
818 pos = head + AIO_EVENTS_OFFSET;
819 page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
820 pos %= AIO_EVENTS_PER_PAGE;
823 copy_ret = copy_to_user(event + ret, ev + pos,
824 sizeof(*ev) * avail);
827 if (unlikely(copy_ret)) {
837 ring = kmap_atomic(info->ring_pages[0]);
840 flush_dcache_page(info->ring_pages[0]);
842 pr_debug("%li h%u t%u\n", ret, head, info->tail);
844 atomic_sub(ret, &ctx->reqs_active);
846 mutex_unlock(&info->ring_lock);
851 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
852 struct io_event __user *event, long *i)
854 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
859 if (unlikely(atomic_read(&ctx->dead)))
865 return ret < 0 || *i >= min_nr;
868 static long read_events(struct kioctx *ctx, long min_nr, long nr,
869 struct io_event __user *event,
870 struct timespec __user *timeout)
872 ktime_t until = { .tv64 = KTIME_MAX };
878 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
881 until = timespec_to_ktime(ts);
885 * Note that aio_read_events() is being called as the conditional - i.e.
886 * we're calling it after prepare_to_wait() has set task state to
887 * TASK_INTERRUPTIBLE.
889 * But aio_read_events() can block, and if it blocks it's going to flip
890 * the task state back to TASK_RUNNING.
892 * This should be ok, provided it doesn't flip the state back to
893 * TASK_RUNNING and return 0 too much - that causes us to spin. That
894 * will only happen if the mutex_lock() call blocks, and we then find
895 * the ringbuffer empty. So in practice we should be ok, but it's
896 * something to be aware of when touching this code.
898 wait_event_interruptible_hrtimeout(ctx->wait,
899 aio_read_events(ctx, min_nr, nr, event, &ret), until);
901 if (!ret && signal_pending(current))
908 * Create an aio_context capable of receiving at least nr_events.
909 * ctxp must not point to an aio_context that already exists, and
910 * must be initialized to 0 prior to the call. On successful
911 * creation of the aio_context, *ctxp is filled in with the resulting
912 * handle. May fail with -EINVAL if *ctxp is not initialized,
913 * if the specified nr_events exceeds internal limits. May fail
914 * with -EAGAIN if the specified nr_events exceeds the user's limit
915 * of available events. May fail with -ENOMEM if insufficient kernel
916 * resources are available. May fail with -EFAULT if an invalid
917 * pointer is passed for ctxp. Will fail with -ENOSYS if not
920 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
922 struct kioctx *ioctx = NULL;
926 ret = get_user(ctx, ctxp);
931 if (unlikely(ctx || nr_events == 0)) {
932 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
937 ioctx = ioctx_alloc(nr_events);
938 ret = PTR_ERR(ioctx);
939 if (!IS_ERR(ioctx)) {
940 ret = put_user(ioctx->user_id, ctxp);
951 * Destroy the aio_context specified. May cancel any outstanding
952 * AIOs and block on completion. Will fail with -ENOSYS if not
953 * implemented. May fail with -EINVAL if the context pointed to
956 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
958 struct kioctx *ioctx = lookup_ioctx(ctx);
959 if (likely(NULL != ioctx)) {
964 pr_debug("EINVAL: io_destroy: invalid context id\n");
968 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
970 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
974 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
975 ssize_t this = min((ssize_t)iov->iov_len, ret);
976 iov->iov_base += this;
977 iov->iov_len -= this;
978 iocb->ki_left -= this;
980 if (iov->iov_len == 0) {
986 /* the caller should not have done more io than what fit in
987 * the remaining iovecs */
988 BUG_ON(ret > 0 && iocb->ki_left == 0);
991 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
993 struct file *file = iocb->ki_filp;
994 struct address_space *mapping = file->f_mapping;
995 struct inode *inode = mapping->host;
996 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
997 unsigned long, loff_t);
999 unsigned short opcode;
1001 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1002 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1003 rw_op = file->f_op->aio_read;
1004 opcode = IOCB_CMD_PREADV;
1006 rw_op = file->f_op->aio_write;
1007 opcode = IOCB_CMD_PWRITEV;
1010 /* This matches the pread()/pwrite() logic */
1011 if (iocb->ki_pos < 0)
1014 if (opcode == IOCB_CMD_PWRITEV)
1015 file_start_write(file);
1017 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1018 iocb->ki_nr_segs - iocb->ki_cur_seg,
1021 aio_advance_iovec(iocb, ret);
1023 /* retry all partial writes. retry partial reads as long as its a
1025 } while (ret > 0 && iocb->ki_left > 0 &&
1026 (opcode == IOCB_CMD_PWRITEV ||
1027 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1028 if (opcode == IOCB_CMD_PWRITEV)
1029 file_end_write(file);
1031 /* This means we must have transferred all that we could */
1032 /* No need to retry anymore */
1033 if ((ret == 0) || (iocb->ki_left == 0))
1034 ret = iocb->ki_nbytes - iocb->ki_left;
1036 /* If we managed to write some out we return that, rather than
1037 * the eventual error. */
1038 if (opcode == IOCB_CMD_PWRITEV
1039 && ret < 0 && ret != -EIOCBQUEUED
1040 && iocb->ki_nbytes - iocb->ki_left)
1041 ret = iocb->ki_nbytes - iocb->ki_left;
1046 static ssize_t aio_fdsync(struct kiocb *iocb)
1048 struct file *file = iocb->ki_filp;
1049 ssize_t ret = -EINVAL;
1051 if (file->f_op->aio_fsync)
1052 ret = file->f_op->aio_fsync(iocb, 1);
1056 static ssize_t aio_fsync(struct kiocb *iocb)
1058 struct file *file = iocb->ki_filp;
1059 ssize_t ret = -EINVAL;
1061 if (file->f_op->aio_fsync)
1062 ret = file->f_op->aio_fsync(iocb, 0);
1066 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1070 #ifdef CONFIG_COMPAT
1072 ret = compat_rw_copy_check_uvector(type,
1073 (struct compat_iovec __user *)kiocb->ki_buf,
1074 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1078 ret = rw_copy_check_uvector(type,
1079 (struct iovec __user *)kiocb->ki_buf,
1080 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1085 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1089 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1090 kiocb->ki_cur_seg = 0;
1091 /* ki_nbytes/left now reflect bytes instead of segs */
1092 kiocb->ki_nbytes = ret;
1093 kiocb->ki_left = ret;
1100 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1104 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1108 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1109 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1110 kiocb->ki_iovec->iov_len = bytes;
1111 kiocb->ki_nr_segs = 1;
1112 kiocb->ki_cur_seg = 0;
1118 * Performs the initial checks and aio retry method
1119 * setup for the kiocb at the time of io submission.
1121 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1123 struct file *file = kiocb->ki_filp;
1126 switch (kiocb->ki_opcode) {
1127 case IOCB_CMD_PREAD:
1129 if (unlikely(!(file->f_mode & FMODE_READ)))
1132 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1135 ret = aio_setup_single_vector(READ, file, kiocb);
1139 if (file->f_op->aio_read)
1140 kiocb->ki_retry = aio_rw_vect_retry;
1142 case IOCB_CMD_PWRITE:
1144 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1147 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1150 ret = aio_setup_single_vector(WRITE, file, kiocb);
1154 if (file->f_op->aio_write)
1155 kiocb->ki_retry = aio_rw_vect_retry;
1157 case IOCB_CMD_PREADV:
1159 if (unlikely(!(file->f_mode & FMODE_READ)))
1161 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1165 if (file->f_op->aio_read)
1166 kiocb->ki_retry = aio_rw_vect_retry;
1168 case IOCB_CMD_PWRITEV:
1170 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1172 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1176 if (file->f_op->aio_write)
1177 kiocb->ki_retry = aio_rw_vect_retry;
1179 case IOCB_CMD_FDSYNC:
1181 if (file->f_op->aio_fsync)
1182 kiocb->ki_retry = aio_fdsync;
1184 case IOCB_CMD_FSYNC:
1186 if (file->f_op->aio_fsync)
1187 kiocb->ki_retry = aio_fsync;
1190 pr_debug("EINVAL: no operation provided\n");
1194 if (!kiocb->ki_retry)
1200 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1201 struct iocb *iocb, struct kiocb_batch *batch,
1207 /* enforce forwards compatibility on users */
1208 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1209 pr_debug("EINVAL: reserve field set\n");
1213 /* prevent overflows */
1215 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1216 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1217 ((ssize_t)iocb->aio_nbytes < 0)
1219 pr_debug("EINVAL: io_submit: overflow check\n");
1223 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1227 req->ki_filp = fget(iocb->aio_fildes);
1228 if (unlikely(!req->ki_filp)) {
1233 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1235 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1236 * instance of the file* now. The file descriptor must be
1237 * an eventfd() fd, and will be signaled for each completed
1238 * event using the eventfd_signal() function.
1240 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1241 if (IS_ERR(req->ki_eventfd)) {
1242 ret = PTR_ERR(req->ki_eventfd);
1243 req->ki_eventfd = NULL;
1248 ret = put_user(req->ki_key, &user_iocb->aio_key);
1249 if (unlikely(ret)) {
1250 pr_debug("EFAULT: aio_key\n");
1254 req->ki_obj.user = user_iocb;
1255 req->ki_user_data = iocb->aio_data;
1256 req->ki_pos = iocb->aio_offset;
1258 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1259 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1260 req->ki_opcode = iocb->aio_lio_opcode;
1262 ret = aio_setup_iocb(req, compat);
1266 ret = req->ki_retry(req);
1267 if (ret != -EIOCBQUEUED) {
1269 * There's no easy way to restart the syscall since other AIO's
1270 * may be already running. Just fail this IO with EINTR.
1272 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1273 ret == -ERESTARTNOHAND ||
1274 ret == -ERESTART_RESTARTBLOCK))
1276 aio_complete(req, ret, 0);
1279 aio_put_req(req); /* drop extra ref to req */
1283 atomic_dec(&ctx->reqs_active);
1284 aio_put_req(req); /* drop extra ref to req */
1285 aio_put_req(req); /* drop i/o ref to req */
1289 long do_io_submit(aio_context_t ctx_id, long nr,
1290 struct iocb __user *__user *iocbpp, bool compat)
1295 struct blk_plug plug;
1296 struct kiocb_batch batch;
1298 if (unlikely(nr < 0))
1301 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1302 nr = LONG_MAX/sizeof(*iocbpp);
1304 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1307 ctx = lookup_ioctx(ctx_id);
1308 if (unlikely(!ctx)) {
1309 pr_debug("EINVAL: invalid context id\n");
1313 kiocb_batch_init(&batch, nr);
1315 blk_start_plug(&plug);
1318 * AKPM: should this return a partial result if some of the IOs were
1319 * successfully submitted?
1321 for (i=0; i<nr; i++) {
1322 struct iocb __user *user_iocb;
1325 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1330 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1335 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1339 blk_finish_plug(&plug);
1341 kiocb_batch_free(ctx, &batch);
1347 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1348 * the number of iocbs queued. May return -EINVAL if the aio_context
1349 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1350 * *iocbpp[0] is not properly initialized, if the operation specified
1351 * is invalid for the file descriptor in the iocb. May fail with
1352 * -EFAULT if any of the data structures point to invalid data. May
1353 * fail with -EBADF if the file descriptor specified in the first
1354 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1355 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1356 * fail with -ENOSYS if not implemented.
1358 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1359 struct iocb __user * __user *, iocbpp)
1361 return do_io_submit(ctx_id, nr, iocbpp, 0);
1365 * Finds a given iocb for cancellation.
1367 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1370 struct list_head *pos;
1372 assert_spin_locked(&ctx->ctx_lock);
1374 /* TODO: use a hash or array, this sucks. */
1375 list_for_each(pos, &ctx->active_reqs) {
1376 struct kiocb *kiocb = list_kiocb(pos);
1377 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1384 * Attempts to cancel an iocb previously passed to io_submit. If
1385 * the operation is successfully cancelled, the resulting event is
1386 * copied into the memory pointed to by result without being placed
1387 * into the completion queue and 0 is returned. May fail with
1388 * -EFAULT if any of the data structures pointed to are invalid.
1389 * May fail with -EINVAL if aio_context specified by ctx_id is
1390 * invalid. May fail with -EAGAIN if the iocb specified was not
1391 * cancelled. Will fail with -ENOSYS if not implemented.
1393 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1394 struct io_event __user *, result)
1396 struct io_event res;
1398 struct kiocb *kiocb;
1402 ret = get_user(key, &iocb->aio_key);
1406 ctx = lookup_ioctx(ctx_id);
1410 spin_lock_irq(&ctx->ctx_lock);
1412 kiocb = lookup_kiocb(ctx, iocb, key);
1414 ret = kiocb_cancel(ctx, kiocb, &res);
1418 spin_unlock_irq(&ctx->ctx_lock);
1421 /* Cancellation succeeded -- copy the result
1422 * into the user's buffer.
1424 if (copy_to_user(result, &res, sizeof(res)))
1434 * Attempts to read at least min_nr events and up to nr events from
1435 * the completion queue for the aio_context specified by ctx_id. If
1436 * it succeeds, the number of read events is returned. May fail with
1437 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1438 * out of range, if timeout is out of range. May fail with -EFAULT
1439 * if any of the memory specified is invalid. May return 0 or
1440 * < min_nr if the timeout specified by timeout has elapsed
1441 * before sufficient events are available, where timeout == NULL
1442 * specifies an infinite timeout. Note that the timeout pointed to by
1443 * timeout is relative and will be updated if not NULL and the
1444 * operation blocks. Will fail with -ENOSYS if not implemented.
1446 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1449 struct io_event __user *, events,
1450 struct timespec __user *, timeout)
1452 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1455 if (likely(ioctx)) {
1456 if (likely(min_nr <= nr && min_nr >= 0))
1457 ret = read_events(ioctx, min_nr, nr, events, timeout);