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 struct rcu_head rcu_head;
101 struct work_struct rcu_work;
104 /*------ sysctl variables----*/
105 static DEFINE_SPINLOCK(aio_nr_lock);
106 unsigned long aio_nr; /* current system wide number of aio requests */
107 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
108 /*----end sysctl variables---*/
110 static struct kmem_cache *kiocb_cachep;
111 static struct kmem_cache *kioctx_cachep;
114 * Creates the slab caches used by the aio routines, panic on
115 * failure as this is done early during the boot sequence.
117 static int __init aio_setup(void)
119 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
120 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
122 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
126 __initcall(aio_setup);
128 static void aio_free_ring(struct kioctx *ctx)
130 struct aio_ring_info *info = &ctx->ring_info;
133 for (i=0; i<info->nr_pages; i++)
134 put_page(info->ring_pages[i]);
136 if (info->mmap_size) {
137 vm_munmap(info->mmap_base, info->mmap_size);
140 if (info->ring_pages && info->ring_pages != info->internal_pages)
141 kfree(info->ring_pages);
142 info->ring_pages = NULL;
146 static int aio_setup_ring(struct kioctx *ctx)
148 struct aio_ring *ring;
149 struct aio_ring_info *info = &ctx->ring_info;
150 unsigned nr_events = ctx->max_reqs;
151 struct mm_struct *mm = current->mm;
152 unsigned long size, populate;
155 /* Compensate for the ring buffer's head/tail overlap entry */
156 nr_events += 2; /* 1 is required, 2 for good luck */
158 size = sizeof(struct aio_ring);
159 size += sizeof(struct io_event) * nr_events;
160 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
165 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
168 info->ring_pages = info->internal_pages;
169 if (nr_pages > AIO_RING_PAGES) {
170 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
171 if (!info->ring_pages)
175 info->mmap_size = nr_pages * PAGE_SIZE;
176 pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
177 down_write(&mm->mmap_sem);
178 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
179 PROT_READ|PROT_WRITE,
180 MAP_ANONYMOUS|MAP_PRIVATE, 0,
182 if (IS_ERR((void *)info->mmap_base)) {
183 up_write(&mm->mmap_sem);
189 pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
190 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
191 1, 0, info->ring_pages, NULL);
192 up_write(&mm->mmap_sem);
194 if (unlikely(info->nr_pages != nr_pages)) {
199 mm_populate(info->mmap_base, populate);
201 ctx->user_id = info->mmap_base;
203 info->nr = nr_events; /* trusted copy */
205 ring = kmap_atomic(info->ring_pages[0]);
206 ring->nr = nr_events; /* user copy */
207 ring->id = ctx->user_id;
208 ring->head = ring->tail = 0;
209 ring->magic = AIO_RING_MAGIC;
210 ring->compat_features = AIO_RING_COMPAT_FEATURES;
211 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
212 ring->header_length = sizeof(struct aio_ring);
214 flush_dcache_page(info->ring_pages[0]);
219 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
220 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
221 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
223 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
224 struct io_event *res)
226 int (*cancel)(struct kiocb *, struct io_event *);
229 cancel = kiocb->ki_cancel;
230 kiocbSetCancelled(kiocb);
232 atomic_inc(&kiocb->ki_users);
233 spin_unlock_irq(&ctx->ctx_lock);
235 memset(res, 0, sizeof(*res));
236 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
237 res->data = kiocb->ki_user_data;
238 ret = cancel(kiocb, res);
240 spin_lock_irq(&ctx->ctx_lock);
246 static void free_ioctx_rcu(struct rcu_head *head)
248 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
249 kmem_cache_free(kioctx_cachep, ctx);
253 * When this function runs, the kioctx has been removed from the "hash table"
254 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
255 * now it's safe to cancel any that need to be.
257 static void free_ioctx(struct kioctx *ctx)
262 spin_lock_irq(&ctx->ctx_lock);
264 while (!list_empty(&ctx->active_reqs)) {
265 req = list_first_entry(&ctx->active_reqs,
266 struct kiocb, ki_list);
268 list_del_init(&req->ki_list);
269 kiocb_cancel(ctx, req, &res);
272 spin_unlock_irq(&ctx->ctx_lock);
274 wait_event(ctx->wait, !atomic_read(&ctx->reqs_active));
278 spin_lock(&aio_nr_lock);
279 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
280 aio_nr -= ctx->max_reqs;
281 spin_unlock(&aio_nr_lock);
283 pr_debug("freeing %p\n", ctx);
286 * Here the call_rcu() is between the wait_event() for reqs_active to
287 * hit 0, and freeing the ioctx.
289 * aio_complete() decrements reqs_active, but it has to touch the ioctx
290 * after to issue a wakeup so we use rcu.
292 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
295 static void put_ioctx(struct kioctx *ctx)
297 if (unlikely(atomic_dec_and_test(&ctx->users)))
302 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
304 static struct kioctx *ioctx_alloc(unsigned nr_events)
306 struct mm_struct *mm = current->mm;
310 /* Prevent overflows */
311 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
312 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
313 pr_debug("ENOMEM: nr_events too high\n");
314 return ERR_PTR(-EINVAL);
317 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
318 return ERR_PTR(-EAGAIN);
320 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
322 return ERR_PTR(-ENOMEM);
324 ctx->max_reqs = nr_events;
326 atomic_set(&ctx->users, 2);
327 atomic_set(&ctx->dead, 0);
328 spin_lock_init(&ctx->ctx_lock);
329 mutex_init(&ctx->ring_info.ring_lock);
330 init_waitqueue_head(&ctx->wait);
332 INIT_LIST_HEAD(&ctx->active_reqs);
334 if (aio_setup_ring(ctx) < 0)
337 /* limit the number of system wide aios */
338 spin_lock(&aio_nr_lock);
339 if (aio_nr + nr_events > aio_max_nr ||
340 aio_nr + nr_events < aio_nr) {
341 spin_unlock(&aio_nr_lock);
344 aio_nr += ctx->max_reqs;
345 spin_unlock(&aio_nr_lock);
347 /* now link into global list. */
348 spin_lock(&mm->ioctx_lock);
349 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
350 spin_unlock(&mm->ioctx_lock);
352 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
353 ctx, ctx->user_id, mm, ctx->ring_info.nr);
360 kmem_cache_free(kioctx_cachep, ctx);
361 pr_debug("error allocating ioctx %d\n", err);
365 static void kill_ioctx_work(struct work_struct *work)
367 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
369 wake_up_all(&ctx->wait);
373 static void kill_ioctx_rcu(struct rcu_head *head)
375 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
377 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
378 schedule_work(&ctx->rcu_work);
382 * Cancels all outstanding aio requests on an aio context. Used
383 * when the processes owning a context have all exited to encourage
384 * the rapid destruction of the kioctx.
386 static void kill_ioctx(struct kioctx *ctx)
388 if (!atomic_xchg(&ctx->dead, 1)) {
389 hlist_del_rcu(&ctx->list);
390 /* Between hlist_del_rcu() and dropping the initial ref */
394 * We can't punt to workqueue here because put_ioctx() ->
395 * free_ioctx() will unmap the ringbuffer, and that has to be
396 * done in the original process's context. kill_ioctx_rcu/work()
397 * exist for exit_aio(), as in that path free_ioctx() won't do
400 kill_ioctx_work(&ctx->rcu_work);
404 /* wait_on_sync_kiocb:
405 * Waits on the given sync kiocb to complete.
407 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
409 while (atomic_read(&iocb->ki_users)) {
410 set_current_state(TASK_UNINTERRUPTIBLE);
411 if (!atomic_read(&iocb->ki_users))
415 __set_current_state(TASK_RUNNING);
416 return iocb->ki_user_data;
418 EXPORT_SYMBOL(wait_on_sync_kiocb);
421 * exit_aio: called when the last user of mm goes away. At this point, there is
422 * no way for any new requests to be submited or any of the io_* syscalls to be
423 * called on the context.
425 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
428 void exit_aio(struct mm_struct *mm)
431 struct hlist_node *n;
433 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
434 if (1 != atomic_read(&ctx->users))
436 "exit_aio:ioctx still alive: %d %d %d\n",
437 atomic_read(&ctx->users),
438 atomic_read(&ctx->dead),
439 atomic_read(&ctx->reqs_active));
441 * We don't need to bother with munmap() here -
442 * exit_mmap(mm) is coming and it'll unmap everything.
443 * Since aio_free_ring() uses non-zero ->mmap_size
444 * as indicator that it needs to unmap the area,
445 * just set it to 0; aio_free_ring() is the only
446 * place that uses ->mmap_size, so it's safe.
448 ctx->ring_info.mmap_size = 0;
450 if (!atomic_xchg(&ctx->dead, 1)) {
451 hlist_del_rcu(&ctx->list);
452 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
458 * Allocate a slot for an aio request. Increments the ki_users count
459 * of the kioctx so that the kioctx stays around until all requests are
460 * complete. Returns NULL if no requests are free.
462 * Returns with kiocb->ki_users set to 2. The io submit code path holds
463 * an extra reference while submitting the i/o.
464 * This prevents races between the aio code path referencing the
465 * req (after submitting it) and aio_complete() freeing the req.
467 static struct kiocb *__aio_get_req(struct kioctx *ctx)
469 struct kiocb *req = NULL;
471 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
476 atomic_set(&req->ki_users, 2);
479 req->ki_cancel = NULL;
480 req->ki_retry = NULL;
483 req->ki_iovec = NULL;
484 req->ki_eventfd = NULL;
490 * struct kiocb's are allocated in batches to reduce the number of
491 * times the ctx lock is acquired and released.
493 #define KIOCB_BATCH_SIZE 32L
495 struct list_head head;
496 long count; /* number of requests left to allocate */
499 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
501 INIT_LIST_HEAD(&batch->head);
502 batch->count = total;
505 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
507 struct kiocb *req, *n;
509 if (list_empty(&batch->head))
512 spin_lock_irq(&ctx->ctx_lock);
513 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
514 list_del(&req->ki_batch);
515 list_del(&req->ki_list);
516 kmem_cache_free(kiocb_cachep, req);
517 atomic_dec(&ctx->reqs_active);
519 spin_unlock_irq(&ctx->ctx_lock);
523 * Allocate a batch of kiocbs. This avoids taking and dropping the
524 * context lock a lot during setup.
526 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
528 unsigned short allocated, to_alloc;
530 struct kiocb *req, *n;
531 struct aio_ring *ring;
533 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
534 for (allocated = 0; allocated < to_alloc; allocated++) {
535 req = __aio_get_req(ctx);
537 /* allocation failed, go with what we've got */
539 list_add(&req->ki_batch, &batch->head);
545 spin_lock_irq(&ctx->ctx_lock);
546 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
548 avail = aio_ring_avail(&ctx->ring_info, ring) -
549 atomic_read(&ctx->reqs_active);
551 if (avail < allocated) {
552 /* Trim back the number of requests. */
553 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
554 list_del(&req->ki_batch);
555 kmem_cache_free(kiocb_cachep, req);
556 if (--allocated <= avail)
561 batch->count -= allocated;
562 list_for_each_entry(req, &batch->head, ki_batch) {
563 list_add(&req->ki_list, &ctx->active_reqs);
564 atomic_inc(&ctx->reqs_active);
568 spin_unlock_irq(&ctx->ctx_lock);
574 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
575 struct kiocb_batch *batch)
579 if (list_empty(&batch->head))
580 if (kiocb_batch_refill(ctx, batch) == 0)
582 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
583 list_del(&req->ki_batch);
587 static void kiocb_free(struct kiocb *req)
591 if (req->ki_eventfd != NULL)
592 eventfd_ctx_put(req->ki_eventfd);
595 if (req->ki_iovec != &req->ki_inline_vec)
596 kfree(req->ki_iovec);
597 kmem_cache_free(kiocb_cachep, req);
600 void aio_put_req(struct kiocb *req)
602 if (atomic_dec_and_test(&req->ki_users))
605 EXPORT_SYMBOL(aio_put_req);
607 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
609 struct mm_struct *mm = current->mm;
610 struct kioctx *ctx, *ret = NULL;
614 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
615 if (ctx->user_id == ctx_id) {
616 atomic_inc(&ctx->users);
627 * Called when the io request on the given iocb is complete.
629 void aio_complete(struct kiocb *iocb, long res, long res2)
631 struct kioctx *ctx = iocb->ki_ctx;
632 struct aio_ring_info *info;
633 struct aio_ring *ring;
634 struct io_event *ev_page, *event;
639 * Special case handling for sync iocbs:
640 * - events go directly into the iocb for fast handling
641 * - the sync task with the iocb in its stack holds the single iocb
642 * ref, no other paths have a way to get another ref
643 * - the sync task helpfully left a reference to itself in the iocb
645 if (is_sync_kiocb(iocb)) {
646 BUG_ON(atomic_read(&iocb->ki_users) != 1);
647 iocb->ki_user_data = res;
648 atomic_set(&iocb->ki_users, 0);
649 wake_up_process(iocb->ki_obj.tsk);
653 info = &ctx->ring_info;
656 * Add a completion event to the ring buffer. Must be done holding
657 * ctx->ctx_lock to prevent other code from messing with the tail
658 * pointer since we might be called from irq context.
660 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
661 * need to issue a wakeup after decrementing reqs_active.
664 spin_lock_irqsave(&ctx->ctx_lock, flags);
666 list_del(&iocb->ki_list); /* remove from active_reqs */
669 * cancelled requests don't get events, userland was given one
670 * when the event got cancelled.
672 if (kiocbIsCancelled(iocb))
676 pos = tail + AIO_EVENTS_OFFSET;
678 if (++tail >= info->nr)
681 ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
682 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
684 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
685 event->data = iocb->ki_user_data;
689 kunmap_atomic(ev_page);
690 flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
692 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
693 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
696 /* after flagging the request as done, we
697 * must never even look at it again
699 smp_wmb(); /* make event visible before updating tail */
703 ring = kmap_atomic(info->ring_pages[0]);
706 flush_dcache_page(info->ring_pages[0]);
708 pr_debug("added to ring %p at [%u]\n", iocb, tail);
711 * Check if the user asked us to deliver the result through an
712 * eventfd. The eventfd_signal() function is safe to be called
715 if (iocb->ki_eventfd != NULL)
716 eventfd_signal(iocb->ki_eventfd, 1);
719 /* everything turned out well, dispose of the aiocb. */
721 atomic_dec(&ctx->reqs_active);
724 * We have to order our ring_info tail store above and test
725 * of the wait list below outside the wait lock. This is
726 * like in wake_up_bit() where clearing a bit has to be
727 * ordered with the unlocked test.
731 if (waitqueue_active(&ctx->wait))
734 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
737 EXPORT_SYMBOL(aio_complete);
740 * Pull an event off of the ioctx's event ring. Returns the number of
743 static long aio_read_events_ring(struct kioctx *ctx,
744 struct io_event __user *event, long nr)
746 struct aio_ring_info *info = &ctx->ring_info;
747 struct aio_ring *ring;
752 mutex_lock(&info->ring_lock);
754 ring = kmap_atomic(info->ring_pages[0]);
758 pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
760 if (head == info->tail)
768 avail = (head <= info->tail ? info->tail : info->nr) - head;
769 if (head == info->tail)
772 avail = min(avail, nr - ret);
773 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
774 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
776 pos = head + AIO_EVENTS_OFFSET;
777 page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
778 pos %= AIO_EVENTS_PER_PAGE;
781 copy_ret = copy_to_user(event + ret, ev + pos,
782 sizeof(*ev) * avail);
785 if (unlikely(copy_ret)) {
795 ring = kmap_atomic(info->ring_pages[0]);
798 flush_dcache_page(info->ring_pages[0]);
800 pr_debug("%li h%u t%u\n", ret, head, info->tail);
802 mutex_unlock(&info->ring_lock);
807 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
808 struct io_event __user *event, long *i)
810 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
815 if (unlikely(atomic_read(&ctx->dead)))
821 return ret < 0 || *i >= min_nr;
824 static long read_events(struct kioctx *ctx, long min_nr, long nr,
825 struct io_event __user *event,
826 struct timespec __user *timeout)
828 ktime_t until = { .tv64 = KTIME_MAX };
834 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
837 until = timespec_to_ktime(ts);
841 * Note that aio_read_events() is being called as the conditional - i.e.
842 * we're calling it after prepare_to_wait() has set task state to
843 * TASK_INTERRUPTIBLE.
845 * But aio_read_events() can block, and if it blocks it's going to flip
846 * the task state back to TASK_RUNNING.
848 * This should be ok, provided it doesn't flip the state back to
849 * TASK_RUNNING and return 0 too much - that causes us to spin. That
850 * will only happen if the mutex_lock() call blocks, and we then find
851 * the ringbuffer empty. So in practice we should be ok, but it's
852 * something to be aware of when touching this code.
854 wait_event_interruptible_hrtimeout(ctx->wait,
855 aio_read_events(ctx, min_nr, nr, event, &ret), until);
857 if (!ret && signal_pending(current))
864 * Create an aio_context capable of receiving at least nr_events.
865 * ctxp must not point to an aio_context that already exists, and
866 * must be initialized to 0 prior to the call. On successful
867 * creation of the aio_context, *ctxp is filled in with the resulting
868 * handle. May fail with -EINVAL if *ctxp is not initialized,
869 * if the specified nr_events exceeds internal limits. May fail
870 * with -EAGAIN if the specified nr_events exceeds the user's limit
871 * of available events. May fail with -ENOMEM if insufficient kernel
872 * resources are available. May fail with -EFAULT if an invalid
873 * pointer is passed for ctxp. Will fail with -ENOSYS if not
876 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
878 struct kioctx *ioctx = NULL;
882 ret = get_user(ctx, ctxp);
887 if (unlikely(ctx || nr_events == 0)) {
888 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
893 ioctx = ioctx_alloc(nr_events);
894 ret = PTR_ERR(ioctx);
895 if (!IS_ERR(ioctx)) {
896 ret = put_user(ioctx->user_id, ctxp);
907 * Destroy the aio_context specified. May cancel any outstanding
908 * AIOs and block on completion. Will fail with -ENOSYS if not
909 * implemented. May fail with -EINVAL if the context pointed to
912 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
914 struct kioctx *ioctx = lookup_ioctx(ctx);
915 if (likely(NULL != ioctx)) {
920 pr_debug("EINVAL: io_destroy: invalid context id\n");
924 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
926 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
930 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
931 ssize_t this = min((ssize_t)iov->iov_len, ret);
932 iov->iov_base += this;
933 iov->iov_len -= this;
934 iocb->ki_left -= this;
936 if (iov->iov_len == 0) {
942 /* the caller should not have done more io than what fit in
943 * the remaining iovecs */
944 BUG_ON(ret > 0 && iocb->ki_left == 0);
947 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
949 struct file *file = iocb->ki_filp;
950 struct address_space *mapping = file->f_mapping;
951 struct inode *inode = mapping->host;
952 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
953 unsigned long, loff_t);
955 unsigned short opcode;
957 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
958 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
959 rw_op = file->f_op->aio_read;
960 opcode = IOCB_CMD_PREADV;
962 rw_op = file->f_op->aio_write;
963 opcode = IOCB_CMD_PWRITEV;
966 /* This matches the pread()/pwrite() logic */
967 if (iocb->ki_pos < 0)
970 if (opcode == IOCB_CMD_PWRITEV)
971 file_start_write(file);
973 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
974 iocb->ki_nr_segs - iocb->ki_cur_seg,
977 aio_advance_iovec(iocb, ret);
979 /* retry all partial writes. retry partial reads as long as its a
981 } while (ret > 0 && iocb->ki_left > 0 &&
982 (opcode == IOCB_CMD_PWRITEV ||
983 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
984 if (opcode == IOCB_CMD_PWRITEV)
985 file_end_write(file);
987 /* This means we must have transferred all that we could */
988 /* No need to retry anymore */
989 if ((ret == 0) || (iocb->ki_left == 0))
990 ret = iocb->ki_nbytes - iocb->ki_left;
992 /* If we managed to write some out we return that, rather than
993 * the eventual error. */
994 if (opcode == IOCB_CMD_PWRITEV
995 && ret < 0 && ret != -EIOCBQUEUED
996 && iocb->ki_nbytes - iocb->ki_left)
997 ret = iocb->ki_nbytes - iocb->ki_left;
1002 static ssize_t aio_fdsync(struct kiocb *iocb)
1004 struct file *file = iocb->ki_filp;
1005 ssize_t ret = -EINVAL;
1007 if (file->f_op->aio_fsync)
1008 ret = file->f_op->aio_fsync(iocb, 1);
1012 static ssize_t aio_fsync(struct kiocb *iocb)
1014 struct file *file = iocb->ki_filp;
1015 ssize_t ret = -EINVAL;
1017 if (file->f_op->aio_fsync)
1018 ret = file->f_op->aio_fsync(iocb, 0);
1022 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1026 #ifdef CONFIG_COMPAT
1028 ret = compat_rw_copy_check_uvector(type,
1029 (struct compat_iovec __user *)kiocb->ki_buf,
1030 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1034 ret = rw_copy_check_uvector(type,
1035 (struct iovec __user *)kiocb->ki_buf,
1036 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1041 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1045 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1046 kiocb->ki_cur_seg = 0;
1047 /* ki_nbytes/left now reflect bytes instead of segs */
1048 kiocb->ki_nbytes = ret;
1049 kiocb->ki_left = ret;
1056 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1060 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1064 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1065 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1066 kiocb->ki_iovec->iov_len = bytes;
1067 kiocb->ki_nr_segs = 1;
1068 kiocb->ki_cur_seg = 0;
1074 * Performs the initial checks and aio retry method
1075 * setup for the kiocb at the time of io submission.
1077 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1079 struct file *file = kiocb->ki_filp;
1082 switch (kiocb->ki_opcode) {
1083 case IOCB_CMD_PREAD:
1085 if (unlikely(!(file->f_mode & FMODE_READ)))
1088 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1091 ret = aio_setup_single_vector(READ, file, kiocb);
1095 if (file->f_op->aio_read)
1096 kiocb->ki_retry = aio_rw_vect_retry;
1098 case IOCB_CMD_PWRITE:
1100 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1103 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1106 ret = aio_setup_single_vector(WRITE, file, kiocb);
1110 if (file->f_op->aio_write)
1111 kiocb->ki_retry = aio_rw_vect_retry;
1113 case IOCB_CMD_PREADV:
1115 if (unlikely(!(file->f_mode & FMODE_READ)))
1117 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1121 if (file->f_op->aio_read)
1122 kiocb->ki_retry = aio_rw_vect_retry;
1124 case IOCB_CMD_PWRITEV:
1126 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1128 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1132 if (file->f_op->aio_write)
1133 kiocb->ki_retry = aio_rw_vect_retry;
1135 case IOCB_CMD_FDSYNC:
1137 if (file->f_op->aio_fsync)
1138 kiocb->ki_retry = aio_fdsync;
1140 case IOCB_CMD_FSYNC:
1142 if (file->f_op->aio_fsync)
1143 kiocb->ki_retry = aio_fsync;
1146 pr_debug("EINVAL: no operation provided\n");
1150 if (!kiocb->ki_retry)
1156 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1157 struct iocb *iocb, struct kiocb_batch *batch,
1163 /* enforce forwards compatibility on users */
1164 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1165 pr_debug("EINVAL: reserve field set\n");
1169 /* prevent overflows */
1171 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1172 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1173 ((ssize_t)iocb->aio_nbytes < 0)
1175 pr_debug("EINVAL: io_submit: overflow check\n");
1179 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1183 req->ki_filp = fget(iocb->aio_fildes);
1184 if (unlikely(!req->ki_filp)) {
1189 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1191 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1192 * instance of the file* now. The file descriptor must be
1193 * an eventfd() fd, and will be signaled for each completed
1194 * event using the eventfd_signal() function.
1196 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1197 if (IS_ERR(req->ki_eventfd)) {
1198 ret = PTR_ERR(req->ki_eventfd);
1199 req->ki_eventfd = NULL;
1204 ret = put_user(req->ki_key, &user_iocb->aio_key);
1205 if (unlikely(ret)) {
1206 pr_debug("EFAULT: aio_key\n");
1210 req->ki_obj.user = user_iocb;
1211 req->ki_user_data = iocb->aio_data;
1212 req->ki_pos = iocb->aio_offset;
1214 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1215 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1216 req->ki_opcode = iocb->aio_lio_opcode;
1218 ret = aio_setup_iocb(req, compat);
1223 if (unlikely(kiocbIsCancelled(req)))
1226 ret = req->ki_retry(req);
1228 if (ret != -EIOCBQUEUED) {
1230 * There's no easy way to restart the syscall since other AIO's
1231 * may be already running. Just fail this IO with EINTR.
1233 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1234 ret == -ERESTARTNOHAND ||
1235 ret == -ERESTART_RESTARTBLOCK))
1237 aio_complete(req, ret, 0);
1240 aio_put_req(req); /* drop extra ref to req */
1244 spin_lock_irq(&ctx->ctx_lock);
1245 list_del(&req->ki_list);
1246 spin_unlock_irq(&ctx->ctx_lock);
1248 atomic_dec(&ctx->reqs_active);
1249 aio_put_req(req); /* drop extra ref to req */
1250 aio_put_req(req); /* drop i/o ref to req */
1254 long do_io_submit(aio_context_t ctx_id, long nr,
1255 struct iocb __user *__user *iocbpp, bool compat)
1260 struct blk_plug plug;
1261 struct kiocb_batch batch;
1263 if (unlikely(nr < 0))
1266 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1267 nr = LONG_MAX/sizeof(*iocbpp);
1269 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1272 ctx = lookup_ioctx(ctx_id);
1273 if (unlikely(!ctx)) {
1274 pr_debug("EINVAL: invalid context id\n");
1278 kiocb_batch_init(&batch, nr);
1280 blk_start_plug(&plug);
1283 * AKPM: should this return a partial result if some of the IOs were
1284 * successfully submitted?
1286 for (i=0; i<nr; i++) {
1287 struct iocb __user *user_iocb;
1290 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1295 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1300 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1304 blk_finish_plug(&plug);
1306 kiocb_batch_free(ctx, &batch);
1312 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1313 * the number of iocbs queued. May return -EINVAL if the aio_context
1314 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1315 * *iocbpp[0] is not properly initialized, if the operation specified
1316 * is invalid for the file descriptor in the iocb. May fail with
1317 * -EFAULT if any of the data structures point to invalid data. May
1318 * fail with -EBADF if the file descriptor specified in the first
1319 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1320 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1321 * fail with -ENOSYS if not implemented.
1323 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1324 struct iocb __user * __user *, iocbpp)
1326 return do_io_submit(ctx_id, nr, iocbpp, 0);
1330 * Finds a given iocb for cancellation.
1332 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1335 struct list_head *pos;
1337 assert_spin_locked(&ctx->ctx_lock);
1339 /* TODO: use a hash or array, this sucks. */
1340 list_for_each(pos, &ctx->active_reqs) {
1341 struct kiocb *kiocb = list_kiocb(pos);
1342 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1349 * Attempts to cancel an iocb previously passed to io_submit. If
1350 * the operation is successfully cancelled, the resulting event is
1351 * copied into the memory pointed to by result without being placed
1352 * into the completion queue and 0 is returned. May fail with
1353 * -EFAULT if any of the data structures pointed to are invalid.
1354 * May fail with -EINVAL if aio_context specified by ctx_id is
1355 * invalid. May fail with -EAGAIN if the iocb specified was not
1356 * cancelled. Will fail with -ENOSYS if not implemented.
1358 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1359 struct io_event __user *, result)
1361 struct io_event res;
1363 struct kiocb *kiocb;
1367 ret = get_user(key, &iocb->aio_key);
1371 ctx = lookup_ioctx(ctx_id);
1375 spin_lock_irq(&ctx->ctx_lock);
1377 kiocb = lookup_kiocb(ctx, iocb, key);
1379 ret = kiocb_cancel(ctx, kiocb, &res);
1383 spin_unlock_irq(&ctx->ctx_lock);
1386 /* Cancellation succeeded -- copy the result
1387 * into the user's buffer.
1389 if (copy_to_user(result, &res, sizeof(res)))
1399 * Attempts to read at least min_nr events and up to nr events from
1400 * the completion queue for the aio_context specified by ctx_id. If
1401 * it succeeds, the number of read events is returned. May fail with
1402 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1403 * out of range, if timeout is out of range. May fail with -EFAULT
1404 * if any of the memory specified is invalid. May return 0 or
1405 * < min_nr if the timeout specified by timeout has elapsed
1406 * before sufficient events are available, where timeout == NULL
1407 * specifies an infinite timeout. Note that the timeout pointed to by
1408 * timeout is relative and will be updated if not NULL and the
1409 * operation blocks. Will fail with -ENOSYS if not implemented.
1411 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1414 struct io_event __user *, events,
1415 struct timespec __user *, timeout)
1417 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1420 if (likely(ioctx)) {
1421 if (likely(min_nr <= nr && min_nr >= 0))
1422 ret = read_events(ioctx, min_nr, nr, events, timeout);