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 if (!req->ki_list.next) {
228 struct kioctx *ctx = req->ki_ctx;
231 spin_lock_irqsave(&ctx->ctx_lock, flags);
232 list_add(&req->ki_list, &ctx->active_reqs);
233 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
236 req->ki_cancel = cancel;
238 EXPORT_SYMBOL(kiocb_set_cancel_fn);
240 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
241 struct io_event *res)
243 kiocb_cancel_fn *cancel;
246 cancel = xchg(&kiocb->ki_cancel, KIOCB_CANCELLED);
247 if (!cancel || cancel == KIOCB_CANCELLED)
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);
263 static void free_ioctx_rcu(struct rcu_head *head)
265 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
266 kmem_cache_free(kioctx_cachep, ctx);
270 * When this function runs, the kioctx has been removed from the "hash table"
271 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
272 * now it's safe to cancel any that need to be.
274 static void free_ioctx(struct kioctx *ctx)
279 spin_lock_irq(&ctx->ctx_lock);
281 while (!list_empty(&ctx->active_reqs)) {
282 req = list_first_entry(&ctx->active_reqs,
283 struct kiocb, ki_list);
285 list_del_init(&req->ki_list);
286 kiocb_cancel(ctx, req, &res);
289 spin_unlock_irq(&ctx->ctx_lock);
291 wait_event(ctx->wait, !atomic_read(&ctx->reqs_active));
295 spin_lock(&aio_nr_lock);
296 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
297 aio_nr -= ctx->max_reqs;
298 spin_unlock(&aio_nr_lock);
300 pr_debug("freeing %p\n", ctx);
303 * Here the call_rcu() is between the wait_event() for reqs_active to
304 * hit 0, and freeing the ioctx.
306 * aio_complete() decrements reqs_active, but it has to touch the ioctx
307 * after to issue a wakeup so we use rcu.
309 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
312 static void put_ioctx(struct kioctx *ctx)
314 if (unlikely(atomic_dec_and_test(&ctx->users)))
319 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
321 static struct kioctx *ioctx_alloc(unsigned nr_events)
323 struct mm_struct *mm = current->mm;
327 /* Prevent overflows */
328 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
329 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
330 pr_debug("ENOMEM: nr_events too high\n");
331 return ERR_PTR(-EINVAL);
334 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
335 return ERR_PTR(-EAGAIN);
337 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
339 return ERR_PTR(-ENOMEM);
341 ctx->max_reqs = nr_events;
343 atomic_set(&ctx->users, 2);
344 atomic_set(&ctx->dead, 0);
345 spin_lock_init(&ctx->ctx_lock);
346 spin_lock_init(&ctx->completion_lock);
347 mutex_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|__GFP_ZERO);
493 atomic_set(&req->ki_users, 2);
500 * struct kiocb's are allocated in batches to reduce the number of
501 * times the ctx lock is acquired and released.
503 #define KIOCB_BATCH_SIZE 32L
505 struct list_head head;
506 long count; /* number of requests left to allocate */
509 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
511 INIT_LIST_HEAD(&batch->head);
512 batch->count = total;
515 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
517 struct kiocb *req, *n;
519 if (list_empty(&batch->head))
522 spin_lock_irq(&ctx->ctx_lock);
523 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
524 list_del(&req->ki_batch);
525 kmem_cache_free(kiocb_cachep, req);
526 atomic_dec(&ctx->reqs_active);
528 spin_unlock_irq(&ctx->ctx_lock);
532 * Allocate a batch of kiocbs. This avoids taking and dropping the
533 * context lock a lot during setup.
535 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
537 unsigned short allocated, to_alloc;
539 struct kiocb *req, *n;
540 struct aio_ring *ring;
542 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
543 for (allocated = 0; allocated < to_alloc; allocated++) {
544 req = __aio_get_req(ctx);
546 /* allocation failed, go with what we've got */
548 list_add(&req->ki_batch, &batch->head);
554 spin_lock_irq(&ctx->ctx_lock);
555 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
557 avail = aio_ring_avail(&ctx->ring_info, ring) - atomic_read(&ctx->reqs_active);
559 if (avail < allocated) {
560 /* Trim back the number of requests. */
561 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
562 list_del(&req->ki_batch);
563 kmem_cache_free(kiocb_cachep, req);
564 if (--allocated <= avail)
569 batch->count -= allocated;
570 atomic_add(allocated, &ctx->reqs_active);
573 spin_unlock_irq(&ctx->ctx_lock);
579 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
580 struct kiocb_batch *batch)
584 if (list_empty(&batch->head))
585 if (kiocb_batch_refill(ctx, batch) == 0)
587 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
588 list_del(&req->ki_batch);
592 static void kiocb_free(struct kiocb *req)
596 if (req->ki_eventfd != NULL)
597 eventfd_ctx_put(req->ki_eventfd);
600 if (req->ki_iovec != &req->ki_inline_vec)
601 kfree(req->ki_iovec);
602 kmem_cache_free(kiocb_cachep, req);
605 void aio_put_req(struct kiocb *req)
607 if (atomic_dec_and_test(&req->ki_users))
610 EXPORT_SYMBOL(aio_put_req);
612 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
614 struct mm_struct *mm = current->mm;
615 struct kioctx *ctx, *ret = NULL;
619 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
620 if (ctx->user_id == ctx_id){
621 atomic_inc(&ctx->users);
632 * Called when the io request on the given iocb is complete.
634 void aio_complete(struct kiocb *iocb, long res, long res2)
636 struct kioctx *ctx = iocb->ki_ctx;
637 struct aio_ring_info *info;
638 struct aio_ring *ring;
639 struct io_event *ev_page, *event;
644 * Special case handling for sync iocbs:
645 * - events go directly into the iocb for fast handling
646 * - the sync task with the iocb in its stack holds the single iocb
647 * ref, no other paths have a way to get another ref
648 * - the sync task helpfully left a reference to itself in the iocb
650 if (is_sync_kiocb(iocb)) {
651 BUG_ON(atomic_read(&iocb->ki_users) != 1);
652 iocb->ki_user_data = res;
653 atomic_set(&iocb->ki_users, 0);
654 wake_up_process(iocb->ki_obj.tsk);
658 info = &ctx->ring_info;
661 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
662 * need to issue a wakeup after decrementing reqs_active.
666 if (iocb->ki_list.next) {
669 spin_lock_irqsave(&ctx->ctx_lock, flags);
670 list_del(&iocb->ki_list);
671 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
675 * cancelled requests don't get events, userland was given one
676 * when the event got cancelled.
678 if (unlikely(xchg(&iocb->ki_cancel,
679 KIOCB_CANCELLED) == KIOCB_CANCELLED))
683 * Add a completion event to the ring buffer. Must be done holding
684 * ctx->ctx_lock to prevent other code from messing with the tail
685 * pointer since we might be called from irq context.
687 spin_lock_irqsave(&ctx->completion_lock, flags);
690 pos = tail + AIO_EVENTS_OFFSET;
692 if (++tail >= info->nr)
695 ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
696 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
698 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
699 event->data = iocb->ki_user_data;
703 kunmap_atomic(ev_page);
704 flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
706 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
707 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
710 /* after flagging the request as done, we
711 * must never even look at it again
713 smp_wmb(); /* make event visible before updating tail */
717 ring = kmap_atomic(info->ring_pages[0]);
720 flush_dcache_page(info->ring_pages[0]);
722 spin_unlock_irqrestore(&ctx->completion_lock, flags);
724 pr_debug("added to ring %p at [%u]\n", iocb, tail);
727 * Check if the user asked us to deliver the result through an
728 * eventfd. The eventfd_signal() function is safe to be called
731 if (iocb->ki_eventfd != NULL)
732 eventfd_signal(iocb->ki_eventfd, 1);
735 /* everything turned out well, dispose of the aiocb. */
737 atomic_dec(&ctx->reqs_active);
740 * We have to order our ring_info tail store above and test
741 * of the wait list below outside the wait lock. This is
742 * like in wake_up_bit() where clearing a bit has to be
743 * ordered with the unlocked test.
747 if (waitqueue_active(&ctx->wait))
752 EXPORT_SYMBOL(aio_complete);
755 * Pull an event off of the ioctx's event ring. Returns the number of
758 static int aio_read_events_ring(struct kioctx *ctx,
759 struct io_event __user *event, long nr)
761 struct aio_ring_info *info = &ctx->ring_info;
762 struct aio_ring *ring;
764 int ret = 0, copy_ret;
766 if (!mutex_trylock(&info->ring_lock)) {
767 __set_current_state(TASK_RUNNING);
768 mutex_lock(&info->ring_lock);
771 ring = kmap_atomic(info->ring_pages[0]);
775 pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
777 if (head == info->tail)
780 __set_current_state(TASK_RUNNING);
783 unsigned i = (head < info->tail ? info->tail : info->nr) - head;
787 if (head == info->tail)
790 i = min_t(int, i, nr - ret);
791 i = min_t(int, i, AIO_EVENTS_PER_PAGE -
792 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
794 pos = head + AIO_EVENTS_OFFSET;
795 page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
796 pos %= AIO_EVENTS_PER_PAGE;
799 copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * i);
802 if (unlikely(copy_ret)) {
812 ring = kmap_atomic(info->ring_pages[0]);
815 flush_dcache_page(info->ring_pages[0]);
817 pr_debug("%d h%u t%u\n", ret, head, info->tail);
819 mutex_unlock(&info->ring_lock);
824 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
825 struct io_event __user *event, long *i)
827 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
832 if (unlikely(atomic_read(&ctx->dead)))
838 return ret < 0 || *i >= min_nr;
841 static long read_events(struct kioctx *ctx, long min_nr, long nr,
842 struct io_event __user *event,
843 struct timespec __user *timeout)
845 ktime_t until = { .tv64 = KTIME_MAX };
851 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
854 until = timespec_to_ktime(ts);
857 wait_event_interruptible_hrtimeout(ctx->wait,
858 aio_read_events(ctx, min_nr, nr, event, &ret), until);
860 if (!ret && signal_pending(current))
867 * Create an aio_context capable of receiving at least nr_events.
868 * ctxp must not point to an aio_context that already exists, and
869 * must be initialized to 0 prior to the call. On successful
870 * creation of the aio_context, *ctxp is filled in with the resulting
871 * handle. May fail with -EINVAL if *ctxp is not initialized,
872 * if the specified nr_events exceeds internal limits. May fail
873 * with -EAGAIN if the specified nr_events exceeds the user's limit
874 * of available events. May fail with -ENOMEM if insufficient kernel
875 * resources are available. May fail with -EFAULT if an invalid
876 * pointer is passed for ctxp. Will fail with -ENOSYS if not
879 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
881 struct kioctx *ioctx = NULL;
885 ret = get_user(ctx, ctxp);
890 if (unlikely(ctx || nr_events == 0)) {
891 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
896 ioctx = ioctx_alloc(nr_events);
897 ret = PTR_ERR(ioctx);
898 if (!IS_ERR(ioctx)) {
899 ret = put_user(ioctx->user_id, ctxp);
910 * Destroy the aio_context specified. May cancel any outstanding
911 * AIOs and block on completion. Will fail with -ENOSYS if not
912 * implemented. May fail with -EINVAL if the context pointed to
915 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
917 struct kioctx *ioctx = lookup_ioctx(ctx);
918 if (likely(NULL != ioctx)) {
923 pr_debug("EINVAL: io_destroy: invalid context id\n");
927 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
929 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
933 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
934 ssize_t this = min((ssize_t)iov->iov_len, ret);
935 iov->iov_base += this;
936 iov->iov_len -= this;
937 iocb->ki_left -= this;
939 if (iov->iov_len == 0) {
945 /* the caller should not have done more io than what fit in
946 * the remaining iovecs */
947 BUG_ON(ret > 0 && iocb->ki_left == 0);
950 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
952 struct file *file = iocb->ki_filp;
953 struct address_space *mapping = file->f_mapping;
954 struct inode *inode = mapping->host;
955 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
956 unsigned long, loff_t);
958 unsigned short opcode;
960 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
961 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
962 rw_op = file->f_op->aio_read;
963 opcode = IOCB_CMD_PREADV;
965 rw_op = file->f_op->aio_write;
966 opcode = IOCB_CMD_PWRITEV;
969 /* This matches the pread()/pwrite() logic */
970 if (iocb->ki_pos < 0)
974 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
975 iocb->ki_nr_segs - iocb->ki_cur_seg,
978 aio_advance_iovec(iocb, ret);
980 /* retry all partial writes. retry partial reads as long as its a
982 } while (ret > 0 && iocb->ki_left > 0 &&
983 (opcode == IOCB_CMD_PWRITEV ||
984 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
986 /* This means we must have transferred all that we could */
987 /* No need to retry anymore */
988 if ((ret == 0) || (iocb->ki_left == 0))
989 ret = iocb->ki_nbytes - iocb->ki_left;
991 /* If we managed to write some out we return that, rather than
992 * the eventual error. */
993 if (opcode == IOCB_CMD_PWRITEV
994 && ret < 0 && ret != -EIOCBQUEUED
995 && iocb->ki_nbytes - iocb->ki_left)
996 ret = iocb->ki_nbytes - iocb->ki_left;
1001 static ssize_t aio_fdsync(struct kiocb *iocb)
1003 struct file *file = iocb->ki_filp;
1004 ssize_t ret = -EINVAL;
1006 if (file->f_op->aio_fsync)
1007 ret = file->f_op->aio_fsync(iocb, 1);
1011 static ssize_t aio_fsync(struct kiocb *iocb)
1013 struct file *file = iocb->ki_filp;
1014 ssize_t ret = -EINVAL;
1016 if (file->f_op->aio_fsync)
1017 ret = file->f_op->aio_fsync(iocb, 0);
1021 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1025 #ifdef CONFIG_COMPAT
1027 ret = compat_rw_copy_check_uvector(type,
1028 (struct compat_iovec __user *)kiocb->ki_buf,
1029 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1033 ret = rw_copy_check_uvector(type,
1034 (struct iovec __user *)kiocb->ki_buf,
1035 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1040 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1044 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1045 kiocb->ki_cur_seg = 0;
1046 /* ki_nbytes/left now reflect bytes instead of segs */
1047 kiocb->ki_nbytes = ret;
1048 kiocb->ki_left = ret;
1055 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1059 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1063 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1064 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1065 kiocb->ki_iovec->iov_len = bytes;
1066 kiocb->ki_nr_segs = 1;
1067 kiocb->ki_cur_seg = 0;
1073 * Performs the initial checks and aio retry method
1074 * setup for the kiocb at the time of io submission.
1076 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1078 struct file *file = kiocb->ki_filp;
1081 switch (kiocb->ki_opcode) {
1082 case IOCB_CMD_PREAD:
1084 if (unlikely(!(file->f_mode & FMODE_READ)))
1087 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1090 ret = aio_setup_single_vector(READ, file, kiocb);
1094 if (file->f_op->aio_read)
1095 kiocb->ki_retry = aio_rw_vect_retry;
1097 case IOCB_CMD_PWRITE:
1099 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1102 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1105 ret = aio_setup_single_vector(WRITE, file, kiocb);
1109 if (file->f_op->aio_write)
1110 kiocb->ki_retry = aio_rw_vect_retry;
1112 case IOCB_CMD_PREADV:
1114 if (unlikely(!(file->f_mode & FMODE_READ)))
1116 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1120 if (file->f_op->aio_read)
1121 kiocb->ki_retry = aio_rw_vect_retry;
1123 case IOCB_CMD_PWRITEV:
1125 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1127 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1131 if (file->f_op->aio_write)
1132 kiocb->ki_retry = aio_rw_vect_retry;
1134 case IOCB_CMD_FDSYNC:
1136 if (file->f_op->aio_fsync)
1137 kiocb->ki_retry = aio_fdsync;
1139 case IOCB_CMD_FSYNC:
1141 if (file->f_op->aio_fsync)
1142 kiocb->ki_retry = aio_fsync;
1145 pr_debug("EINVAL: no operation provided\n");
1149 if (!kiocb->ki_retry)
1155 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1156 struct iocb *iocb, struct kiocb_batch *batch,
1162 /* enforce forwards compatibility on users */
1163 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1164 pr_debug("EINVAL: reserve field set\n");
1168 /* prevent overflows */
1170 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1171 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1172 ((ssize_t)iocb->aio_nbytes < 0)
1174 pr_debug("EINVAL: io_submit: overflow check\n");
1178 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1182 req->ki_filp = fget(iocb->aio_fildes);
1183 if (unlikely(!req->ki_filp)) {
1188 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1190 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1191 * instance of the file* now. The file descriptor must be
1192 * an eventfd() fd, and will be signaled for each completed
1193 * event using the eventfd_signal() function.
1195 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1196 if (IS_ERR(req->ki_eventfd)) {
1197 ret = PTR_ERR(req->ki_eventfd);
1198 req->ki_eventfd = NULL;
1203 ret = put_user(req->ki_key, &user_iocb->aio_key);
1204 if (unlikely(ret)) {
1205 pr_debug("EFAULT: aio_key\n");
1209 req->ki_obj.user = user_iocb;
1210 req->ki_user_data = iocb->aio_data;
1211 req->ki_pos = iocb->aio_offset;
1213 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1214 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1215 req->ki_opcode = iocb->aio_lio_opcode;
1217 ret = aio_setup_iocb(req, compat);
1221 ret = req->ki_retry(req);
1222 if (ret != -EIOCBQUEUED) {
1224 * There's no easy way to restart the syscall since other AIO's
1225 * may be already running. Just fail this IO with EINTR.
1227 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1228 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
1230 aio_complete(req, ret, 0);
1233 aio_put_req(req); /* drop extra ref to req */
1237 atomic_dec(&ctx->reqs_active);
1238 aio_put_req(req); /* drop extra ref to req */
1239 aio_put_req(req); /* drop i/o ref to req */
1243 long do_io_submit(aio_context_t ctx_id, long nr,
1244 struct iocb __user *__user *iocbpp, bool compat)
1249 struct blk_plug plug;
1250 struct kiocb_batch batch;
1252 if (unlikely(nr < 0))
1255 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1256 nr = LONG_MAX/sizeof(*iocbpp);
1258 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1261 ctx = lookup_ioctx(ctx_id);
1262 if (unlikely(!ctx)) {
1263 pr_debug("EINVAL: invalid context id\n");
1267 kiocb_batch_init(&batch, nr);
1269 blk_start_plug(&plug);
1272 * AKPM: should this return a partial result if some of the IOs were
1273 * successfully submitted?
1275 for (i=0; i<nr; i++) {
1276 struct iocb __user *user_iocb;
1279 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1284 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1289 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1293 blk_finish_plug(&plug);
1295 kiocb_batch_free(ctx, &batch);
1301 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1302 * the number of iocbs queued. May return -EINVAL if the aio_context
1303 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1304 * *iocbpp[0] is not properly initialized, if the operation specified
1305 * is invalid for the file descriptor in the iocb. May fail with
1306 * -EFAULT if any of the data structures point to invalid data. May
1307 * fail with -EBADF if the file descriptor specified in the first
1308 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1309 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1310 * fail with -ENOSYS if not implemented.
1312 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1313 struct iocb __user * __user *, iocbpp)
1315 return do_io_submit(ctx_id, nr, iocbpp, 0);
1319 * Finds a given iocb for cancellation.
1321 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1324 struct list_head *pos;
1326 assert_spin_locked(&ctx->ctx_lock);
1328 /* TODO: use a hash or array, this sucks. */
1329 list_for_each(pos, &ctx->active_reqs) {
1330 struct kiocb *kiocb = list_kiocb(pos);
1331 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1338 * Attempts to cancel an iocb previously passed to io_submit. If
1339 * the operation is successfully cancelled, the resulting event is
1340 * copied into the memory pointed to by result without being placed
1341 * into the completion queue and 0 is returned. May fail with
1342 * -EFAULT if any of the data structures pointed to are invalid.
1343 * May fail with -EINVAL if aio_context specified by ctx_id is
1344 * invalid. May fail with -EAGAIN if the iocb specified was not
1345 * cancelled. Will fail with -ENOSYS if not implemented.
1347 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1348 struct io_event __user *, result)
1350 struct io_event res;
1352 struct kiocb *kiocb;
1356 ret = get_user(key, &iocb->aio_key);
1360 ctx = lookup_ioctx(ctx_id);
1364 spin_lock_irq(&ctx->ctx_lock);
1366 kiocb = lookup_kiocb(ctx, iocb, key);
1368 ret = kiocb_cancel(ctx, kiocb, &res);
1372 spin_unlock_irq(&ctx->ctx_lock);
1375 /* Cancellation succeeded -- copy the result
1376 * into the user's buffer.
1378 if (copy_to_user(result, &res, sizeof(res)))
1388 * Attempts to read at least min_nr events and up to nr events from
1389 * the completion queue for the aio_context specified by ctx_id. If
1390 * it succeeds, the number of read events is returned. May fail with
1391 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1392 * out of range, if timeout is out of range. May fail with -EFAULT
1393 * if any of the memory specified is invalid. May return 0 or
1394 * < min_nr if the timeout specified by timeout has elapsed
1395 * before sufficient events are available, where timeout == NULL
1396 * specifies an infinite timeout. Note that the timeout pointed to by
1397 * timeout is relative and will be updated if not NULL and the
1398 * operation blocks. Will fail with -ENOSYS if not implemented.
1400 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1403 struct io_event __user *, events,
1404 struct timespec __user *, timeout)
1406 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1409 if (likely(ioctx)) {
1410 if (likely(min_nr <= nr && min_nr >= 0))
1411 ret = read_events(ioctx, min_nr, nr, events, timeout);