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 */
89 /* sys_io_setup currently limits this to an unsigned int */
92 struct aio_ring_info ring_info;
94 spinlock_t completion_lock;
96 struct rcu_head rcu_head;
97 struct work_struct rcu_work;
100 /*------ sysctl variables----*/
101 static DEFINE_SPINLOCK(aio_nr_lock);
102 unsigned long aio_nr; /* current system wide number of aio requests */
103 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
104 /*----end sysctl variables---*/
106 static struct kmem_cache *kiocb_cachep;
107 static struct kmem_cache *kioctx_cachep;
110 * Creates the slab caches used by the aio routines, panic on
111 * failure as this is done early during the boot sequence.
113 static int __init aio_setup(void)
115 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
116 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
118 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
122 __initcall(aio_setup);
124 static void aio_free_ring(struct kioctx *ctx)
126 struct aio_ring_info *info = &ctx->ring_info;
129 for (i=0; i<info->nr_pages; i++)
130 put_page(info->ring_pages[i]);
132 if (info->mmap_size) {
133 vm_munmap(info->mmap_base, info->mmap_size);
136 if (info->ring_pages && info->ring_pages != info->internal_pages)
137 kfree(info->ring_pages);
138 info->ring_pages = NULL;
142 static int aio_setup_ring(struct kioctx *ctx)
144 struct aio_ring *ring;
145 struct aio_ring_info *info = &ctx->ring_info;
146 unsigned nr_events = ctx->max_reqs;
147 struct mm_struct *mm = current->mm;
148 unsigned long size, populate;
151 /* Compensate for the ring buffer's head/tail overlap entry */
152 nr_events += 2; /* 1 is required, 2 for good luck */
154 size = sizeof(struct aio_ring);
155 size += sizeof(struct io_event) * nr_events;
156 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
161 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
164 info->ring_pages = info->internal_pages;
165 if (nr_pages > AIO_RING_PAGES) {
166 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
167 if (!info->ring_pages)
171 info->mmap_size = nr_pages * PAGE_SIZE;
172 pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
173 down_write(&mm->mmap_sem);
174 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
175 PROT_READ|PROT_WRITE,
176 MAP_ANONYMOUS|MAP_PRIVATE, 0,
178 if (IS_ERR((void *)info->mmap_base)) {
179 up_write(&mm->mmap_sem);
185 pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
186 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
187 1, 0, info->ring_pages, NULL);
188 up_write(&mm->mmap_sem);
190 if (unlikely(info->nr_pages != nr_pages)) {
195 mm_populate(info->mmap_base, populate);
197 ctx->user_id = info->mmap_base;
199 info->nr = nr_events; /* trusted copy */
201 ring = kmap_atomic(info->ring_pages[0]);
202 ring->nr = nr_events; /* user copy */
203 ring->id = ctx->user_id;
204 ring->head = ring->tail = 0;
205 ring->magic = AIO_RING_MAGIC;
206 ring->compat_features = AIO_RING_COMPAT_FEATURES;
207 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
208 ring->header_length = sizeof(struct aio_ring);
210 flush_dcache_page(info->ring_pages[0]);
215 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
216 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
217 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
219 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
221 struct kioctx *ctx = req->ki_ctx;
224 spin_lock_irqsave(&ctx->ctx_lock, flags);
226 if (!req->ki_list.next)
227 list_add(&req->ki_list, &ctx->active_reqs);
229 req->ki_cancel = cancel;
231 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
233 EXPORT_SYMBOL(kiocb_set_cancel_fn);
235 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
236 struct io_event *res)
238 kiocb_cancel_fn *old, *cancel;
242 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
243 * actually has a cancel function, hence the cmpxchg()
246 cancel = ACCESS_ONCE(kiocb->ki_cancel);
248 if (!cancel || cancel == KIOCB_CANCELLED)
252 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
253 } while (cancel != old);
255 atomic_inc(&kiocb->ki_users);
256 spin_unlock_irq(&ctx->ctx_lock);
258 memset(res, 0, sizeof(*res));
259 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
260 res->data = kiocb->ki_user_data;
261 ret = cancel(kiocb, res);
263 spin_lock_irq(&ctx->ctx_lock);
268 static void free_ioctx_rcu(struct rcu_head *head)
270 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
271 kmem_cache_free(kioctx_cachep, ctx);
275 * When this function runs, the kioctx has been removed from the "hash table"
276 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
277 * now it's safe to cancel any that need to be.
279 static void free_ioctx(struct kioctx *ctx)
281 struct aio_ring_info *info = &ctx->ring_info;
282 struct aio_ring *ring;
285 unsigned head, avail;
287 spin_lock_irq(&ctx->ctx_lock);
289 while (!list_empty(&ctx->active_reqs)) {
290 req = list_first_entry(&ctx->active_reqs,
291 struct kiocb, ki_list);
293 list_del_init(&req->ki_list);
294 kiocb_cancel(ctx, req, &res);
297 spin_unlock_irq(&ctx->ctx_lock);
299 ring = kmap_atomic(info->ring_pages[0]);
303 while (atomic_read(&ctx->reqs_active) > 0) {
304 wait_event(ctx->wait, head != info->tail);
306 avail = (head < info->tail ? info->tail : info->nr) - head;
308 atomic_sub(avail, &ctx->reqs_active);
313 WARN_ON(atomic_read(&ctx->reqs_active) < 0);
317 spin_lock(&aio_nr_lock);
318 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
319 aio_nr -= ctx->max_reqs;
320 spin_unlock(&aio_nr_lock);
322 pr_debug("freeing %p\n", ctx);
325 * Here the call_rcu() is between the wait_event() for reqs_active to
326 * hit 0, and freeing the ioctx.
328 * aio_complete() decrements reqs_active, but it has to touch the ioctx
329 * after to issue a wakeup so we use rcu.
331 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
334 static void put_ioctx(struct kioctx *ctx)
336 if (unlikely(atomic_dec_and_test(&ctx->users)))
341 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
343 static struct kioctx *ioctx_alloc(unsigned nr_events)
345 struct mm_struct *mm = current->mm;
349 /* Prevent overflows */
350 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
351 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
352 pr_debug("ENOMEM: nr_events too high\n");
353 return ERR_PTR(-EINVAL);
356 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
357 return ERR_PTR(-EAGAIN);
359 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
361 return ERR_PTR(-ENOMEM);
363 ctx->max_reqs = nr_events;
365 atomic_set(&ctx->users, 2);
366 atomic_set(&ctx->dead, 0);
367 spin_lock_init(&ctx->ctx_lock);
368 spin_lock_init(&ctx->completion_lock);
369 mutex_init(&ctx->ring_info.ring_lock);
370 init_waitqueue_head(&ctx->wait);
372 INIT_LIST_HEAD(&ctx->active_reqs);
374 if (aio_setup_ring(ctx) < 0)
377 /* limit the number of system wide aios */
378 spin_lock(&aio_nr_lock);
379 if (aio_nr + nr_events > aio_max_nr ||
380 aio_nr + nr_events < aio_nr) {
381 spin_unlock(&aio_nr_lock);
384 aio_nr += ctx->max_reqs;
385 spin_unlock(&aio_nr_lock);
387 /* now link into global list. */
388 spin_lock(&mm->ioctx_lock);
389 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
390 spin_unlock(&mm->ioctx_lock);
392 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
393 ctx, ctx->user_id, mm, ctx->ring_info.nr);
400 kmem_cache_free(kioctx_cachep, ctx);
401 pr_debug("error allocating ioctx %d\n", err);
405 static void kill_ioctx_work(struct work_struct *work)
407 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
409 wake_up_all(&ctx->wait);
413 static void kill_ioctx_rcu(struct rcu_head *head)
415 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
417 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
418 schedule_work(&ctx->rcu_work);
422 * Cancels all outstanding aio requests on an aio context. Used
423 * when the processes owning a context have all exited to encourage
424 * the rapid destruction of the kioctx.
426 static void kill_ioctx(struct kioctx *ctx)
428 if (!atomic_xchg(&ctx->dead, 1)) {
429 hlist_del_rcu(&ctx->list);
430 /* Between hlist_del_rcu() and dropping the initial ref */
434 * We can't punt to workqueue here because put_ioctx() ->
435 * free_ioctx() will unmap the ringbuffer, and that has to be
436 * done in the original process's context. kill_ioctx_rcu/work()
437 * exist for exit_aio(), as in that path free_ioctx() won't do
440 kill_ioctx_work(&ctx->rcu_work);
444 /* wait_on_sync_kiocb:
445 * Waits on the given sync kiocb to complete.
447 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
449 while (atomic_read(&iocb->ki_users)) {
450 set_current_state(TASK_UNINTERRUPTIBLE);
451 if (!atomic_read(&iocb->ki_users))
455 __set_current_state(TASK_RUNNING);
456 return iocb->ki_user_data;
458 EXPORT_SYMBOL(wait_on_sync_kiocb);
461 * exit_aio: called when the last user of mm goes away. At this point, there is
462 * no way for any new requests to be submited or any of the io_* syscalls to be
463 * called on the context.
465 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
468 void exit_aio(struct mm_struct *mm)
471 struct hlist_node *n;
473 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
474 if (1 != atomic_read(&ctx->users))
476 "exit_aio:ioctx still alive: %d %d %d\n",
477 atomic_read(&ctx->users),
478 atomic_read(&ctx->dead),
479 atomic_read(&ctx->reqs_active));
481 * We don't need to bother with munmap() here -
482 * exit_mmap(mm) is coming and it'll unmap everything.
483 * Since aio_free_ring() uses non-zero ->mmap_size
484 * as indicator that it needs to unmap the area,
485 * just set it to 0; aio_free_ring() is the only
486 * place that uses ->mmap_size, so it's safe.
488 ctx->ring_info.mmap_size = 0;
490 if (!atomic_xchg(&ctx->dead, 1)) {
491 hlist_del_rcu(&ctx->list);
492 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
498 * Allocate a slot for an aio request. Increments the ki_users count
499 * of the kioctx so that the kioctx stays around until all requests are
500 * complete. Returns NULL if no requests are free.
502 * Returns with kiocb->ki_users set to 2. The io submit code path holds
503 * an extra reference while submitting the i/o.
504 * This prevents races between the aio code path referencing the
505 * req (after submitting it) and aio_complete() freeing the req.
507 static struct kiocb *__aio_get_req(struct kioctx *ctx)
509 struct kiocb *req = NULL;
511 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
515 atomic_set(&req->ki_users, 2);
522 * struct kiocb's are allocated in batches to reduce the number of
523 * times the ctx lock is acquired and released.
525 #define KIOCB_BATCH_SIZE 32L
527 struct list_head head;
528 long count; /* number of requests left to allocate */
531 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
533 INIT_LIST_HEAD(&batch->head);
534 batch->count = total;
537 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
539 struct kiocb *req, *n;
541 if (list_empty(&batch->head))
544 spin_lock_irq(&ctx->ctx_lock);
545 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
546 list_del(&req->ki_batch);
547 kmem_cache_free(kiocb_cachep, req);
548 atomic_dec(&ctx->reqs_active);
550 spin_unlock_irq(&ctx->ctx_lock);
554 * Allocate a batch of kiocbs. This avoids taking and dropping the
555 * context lock a lot during setup.
557 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
559 unsigned short allocated, to_alloc;
561 struct kiocb *req, *n;
563 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
564 for (allocated = 0; allocated < to_alloc; allocated++) {
565 req = __aio_get_req(ctx);
567 /* allocation failed, go with what we've got */
569 list_add(&req->ki_batch, &batch->head);
575 spin_lock_irq(&ctx->ctx_lock);
577 avail = ctx->ring_info.nr - atomic_read(&ctx->reqs_active);
579 if (avail < allocated) {
580 /* Trim back the number of requests. */
581 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
582 list_del(&req->ki_batch);
583 kmem_cache_free(kiocb_cachep, req);
584 if (--allocated <= avail)
589 batch->count -= allocated;
590 atomic_add(allocated, &ctx->reqs_active);
592 spin_unlock_irq(&ctx->ctx_lock);
598 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
599 struct kiocb_batch *batch)
603 if (list_empty(&batch->head))
604 if (kiocb_batch_refill(ctx, batch) == 0)
606 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
607 list_del(&req->ki_batch);
611 static void kiocb_free(struct kiocb *req)
615 if (req->ki_eventfd != NULL)
616 eventfd_ctx_put(req->ki_eventfd);
619 if (req->ki_iovec != &req->ki_inline_vec)
620 kfree(req->ki_iovec);
621 kmem_cache_free(kiocb_cachep, req);
624 void aio_put_req(struct kiocb *req)
626 if (atomic_dec_and_test(&req->ki_users))
629 EXPORT_SYMBOL(aio_put_req);
631 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
633 struct mm_struct *mm = current->mm;
634 struct kioctx *ctx, *ret = NULL;
638 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
639 if (ctx->user_id == ctx_id){
640 atomic_inc(&ctx->users);
651 * Called when the io request on the given iocb is complete.
653 void aio_complete(struct kiocb *iocb, long res, long res2)
655 struct kioctx *ctx = iocb->ki_ctx;
656 struct aio_ring_info *info;
657 struct aio_ring *ring;
658 struct io_event *ev_page, *event;
663 * Special case handling for sync iocbs:
664 * - events go directly into the iocb for fast handling
665 * - the sync task with the iocb in its stack holds the single iocb
666 * ref, no other paths have a way to get another ref
667 * - the sync task helpfully left a reference to itself in the iocb
669 if (is_sync_kiocb(iocb)) {
670 BUG_ON(atomic_read(&iocb->ki_users) != 1);
671 iocb->ki_user_data = res;
672 atomic_set(&iocb->ki_users, 0);
673 wake_up_process(iocb->ki_obj.tsk);
677 info = &ctx->ring_info;
680 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
681 * need to issue a wakeup after decrementing reqs_active.
685 if (iocb->ki_list.next) {
688 spin_lock_irqsave(&ctx->ctx_lock, flags);
689 list_del(&iocb->ki_list);
690 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
694 * cancelled requests don't get events, userland was given one
695 * when the event got cancelled.
697 if (unlikely(xchg(&iocb->ki_cancel,
698 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
699 atomic_dec(&ctx->reqs_active);
700 /* Still need the wake_up in case free_ioctx is waiting */
705 * Add a completion event to the ring buffer. Must be done holding
706 * ctx->ctx_lock to prevent other code from messing with the tail
707 * pointer since we might be called from irq context.
709 spin_lock_irqsave(&ctx->completion_lock, flags);
712 pos = tail + AIO_EVENTS_OFFSET;
714 if (++tail >= info->nr)
717 ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
718 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
720 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
721 event->data = iocb->ki_user_data;
725 kunmap_atomic(ev_page);
726 flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
728 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
729 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
732 /* after flagging the request as done, we
733 * must never even look at it again
735 smp_wmb(); /* make event visible before updating tail */
739 ring = kmap_atomic(info->ring_pages[0]);
742 flush_dcache_page(info->ring_pages[0]);
744 spin_unlock_irqrestore(&ctx->completion_lock, flags);
746 pr_debug("added to ring %p at [%u]\n", iocb, tail);
749 * Check if the user asked us to deliver the result through an
750 * eventfd. The eventfd_signal() function is safe to be called
753 if (iocb->ki_eventfd != NULL)
754 eventfd_signal(iocb->ki_eventfd, 1);
757 /* everything turned out well, dispose of the aiocb. */
761 * We have to order our ring_info tail store above and test
762 * of the wait list below outside the wait lock. This is
763 * like in wake_up_bit() where clearing a bit has to be
764 * ordered with the unlocked test.
768 if (waitqueue_active(&ctx->wait))
773 EXPORT_SYMBOL(aio_complete);
776 * Pull an event off of the ioctx's event ring. Returns the number of
779 static int aio_read_events_ring(struct kioctx *ctx,
780 struct io_event __user *event, long nr)
782 struct aio_ring_info *info = &ctx->ring_info;
783 struct aio_ring *ring;
785 int ret = 0, copy_ret;
787 if (!mutex_trylock(&info->ring_lock)) {
788 __set_current_state(TASK_RUNNING);
789 mutex_lock(&info->ring_lock);
792 ring = kmap_atomic(info->ring_pages[0]);
796 pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
798 if (head == info->tail)
801 __set_current_state(TASK_RUNNING);
804 unsigned i = (head < info->tail ? info->tail : info->nr) - head;
808 if (head == info->tail)
811 i = min_t(int, i, nr - ret);
812 i = min_t(int, i, AIO_EVENTS_PER_PAGE -
813 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
815 pos = head + AIO_EVENTS_OFFSET;
816 page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
817 pos %= AIO_EVENTS_PER_PAGE;
820 copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * i);
823 if (unlikely(copy_ret)) {
833 ring = kmap_atomic(info->ring_pages[0]);
836 flush_dcache_page(info->ring_pages[0]);
838 pr_debug("%d h%u t%u\n", ret, head, info->tail);
840 atomic_sub(ret, &ctx->reqs_active);
842 mutex_unlock(&info->ring_lock);
847 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
848 struct io_event __user *event, long *i)
850 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
855 if (unlikely(atomic_read(&ctx->dead)))
861 return ret < 0 || *i >= min_nr;
864 static long read_events(struct kioctx *ctx, long min_nr, long nr,
865 struct io_event __user *event,
866 struct timespec __user *timeout)
868 ktime_t until = { .tv64 = KTIME_MAX };
874 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
877 until = timespec_to_ktime(ts);
880 wait_event_interruptible_hrtimeout(ctx->wait,
881 aio_read_events(ctx, min_nr, nr, event, &ret), until);
883 if (!ret && signal_pending(current))
890 * Create an aio_context capable of receiving at least nr_events.
891 * ctxp must not point to an aio_context that already exists, and
892 * must be initialized to 0 prior to the call. On successful
893 * creation of the aio_context, *ctxp is filled in with the resulting
894 * handle. May fail with -EINVAL if *ctxp is not initialized,
895 * if the specified nr_events exceeds internal limits. May fail
896 * with -EAGAIN if the specified nr_events exceeds the user's limit
897 * of available events. May fail with -ENOMEM if insufficient kernel
898 * resources are available. May fail with -EFAULT if an invalid
899 * pointer is passed for ctxp. Will fail with -ENOSYS if not
902 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
904 struct kioctx *ioctx = NULL;
908 ret = get_user(ctx, ctxp);
913 if (unlikely(ctx || nr_events == 0)) {
914 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
919 ioctx = ioctx_alloc(nr_events);
920 ret = PTR_ERR(ioctx);
921 if (!IS_ERR(ioctx)) {
922 ret = put_user(ioctx->user_id, ctxp);
933 * Destroy the aio_context specified. May cancel any outstanding
934 * AIOs and block on completion. Will fail with -ENOSYS if not
935 * implemented. May fail with -EINVAL if the context pointed to
938 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
940 struct kioctx *ioctx = lookup_ioctx(ctx);
941 if (likely(NULL != ioctx)) {
946 pr_debug("EINVAL: io_destroy: invalid context id\n");
950 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
952 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
956 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
957 ssize_t this = min((ssize_t)iov->iov_len, ret);
958 iov->iov_base += this;
959 iov->iov_len -= this;
960 iocb->ki_left -= this;
962 if (iov->iov_len == 0) {
968 /* the caller should not have done more io than what fit in
969 * the remaining iovecs */
970 BUG_ON(ret > 0 && iocb->ki_left == 0);
973 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
975 struct file *file = iocb->ki_filp;
976 struct address_space *mapping = file->f_mapping;
977 struct inode *inode = mapping->host;
978 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
979 unsigned long, loff_t);
981 unsigned short opcode;
983 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
984 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
985 rw_op = file->f_op->aio_read;
986 opcode = IOCB_CMD_PREADV;
988 rw_op = file->f_op->aio_write;
989 opcode = IOCB_CMD_PWRITEV;
992 /* This matches the pread()/pwrite() logic */
993 if (iocb->ki_pos < 0)
997 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
998 iocb->ki_nr_segs - iocb->ki_cur_seg,
1001 aio_advance_iovec(iocb, ret);
1003 /* retry all partial writes. retry partial reads as long as its a
1005 } while (ret > 0 && iocb->ki_left > 0 &&
1006 (opcode == IOCB_CMD_PWRITEV ||
1007 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1009 /* This means we must have transferred all that we could */
1010 /* No need to retry anymore */
1011 if ((ret == 0) || (iocb->ki_left == 0))
1012 ret = iocb->ki_nbytes - iocb->ki_left;
1014 /* If we managed to write some out we return that, rather than
1015 * the eventual error. */
1016 if (opcode == IOCB_CMD_PWRITEV
1017 && ret < 0 && ret != -EIOCBQUEUED
1018 && iocb->ki_nbytes - iocb->ki_left)
1019 ret = iocb->ki_nbytes - iocb->ki_left;
1024 static ssize_t aio_fdsync(struct kiocb *iocb)
1026 struct file *file = iocb->ki_filp;
1027 ssize_t ret = -EINVAL;
1029 if (file->f_op->aio_fsync)
1030 ret = file->f_op->aio_fsync(iocb, 1);
1034 static ssize_t aio_fsync(struct kiocb *iocb)
1036 struct file *file = iocb->ki_filp;
1037 ssize_t ret = -EINVAL;
1039 if (file->f_op->aio_fsync)
1040 ret = file->f_op->aio_fsync(iocb, 0);
1044 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1048 #ifdef CONFIG_COMPAT
1050 ret = compat_rw_copy_check_uvector(type,
1051 (struct compat_iovec __user *)kiocb->ki_buf,
1052 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1056 ret = rw_copy_check_uvector(type,
1057 (struct iovec __user *)kiocb->ki_buf,
1058 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1063 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1067 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1068 kiocb->ki_cur_seg = 0;
1069 /* ki_nbytes/left now reflect bytes instead of segs */
1070 kiocb->ki_nbytes = ret;
1071 kiocb->ki_left = ret;
1078 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1082 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1086 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1087 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1088 kiocb->ki_iovec->iov_len = bytes;
1089 kiocb->ki_nr_segs = 1;
1090 kiocb->ki_cur_seg = 0;
1096 * Performs the initial checks and aio retry method
1097 * setup for the kiocb at the time of io submission.
1099 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1101 struct file *file = kiocb->ki_filp;
1104 switch (kiocb->ki_opcode) {
1105 case IOCB_CMD_PREAD:
1107 if (unlikely(!(file->f_mode & FMODE_READ)))
1110 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1113 ret = aio_setup_single_vector(READ, file, kiocb);
1117 if (file->f_op->aio_read)
1118 kiocb->ki_retry = aio_rw_vect_retry;
1120 case IOCB_CMD_PWRITE:
1122 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1125 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1128 ret = aio_setup_single_vector(WRITE, file, kiocb);
1132 if (file->f_op->aio_write)
1133 kiocb->ki_retry = aio_rw_vect_retry;
1135 case IOCB_CMD_PREADV:
1137 if (unlikely(!(file->f_mode & FMODE_READ)))
1139 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1143 if (file->f_op->aio_read)
1144 kiocb->ki_retry = aio_rw_vect_retry;
1146 case IOCB_CMD_PWRITEV:
1148 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1150 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1154 if (file->f_op->aio_write)
1155 kiocb->ki_retry = aio_rw_vect_retry;
1157 case IOCB_CMD_FDSYNC:
1159 if (file->f_op->aio_fsync)
1160 kiocb->ki_retry = aio_fdsync;
1162 case IOCB_CMD_FSYNC:
1164 if (file->f_op->aio_fsync)
1165 kiocb->ki_retry = aio_fsync;
1168 pr_debug("EINVAL: no operation provided\n");
1172 if (!kiocb->ki_retry)
1178 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1179 struct iocb *iocb, struct kiocb_batch *batch,
1185 /* enforce forwards compatibility on users */
1186 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1187 pr_debug("EINVAL: reserve field set\n");
1191 /* prevent overflows */
1193 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1194 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1195 ((ssize_t)iocb->aio_nbytes < 0)
1197 pr_debug("EINVAL: io_submit: overflow check\n");
1201 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1205 req->ki_filp = fget(iocb->aio_fildes);
1206 if (unlikely(!req->ki_filp)) {
1211 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1213 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1214 * instance of the file* now. The file descriptor must be
1215 * an eventfd() fd, and will be signaled for each completed
1216 * event using the eventfd_signal() function.
1218 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1219 if (IS_ERR(req->ki_eventfd)) {
1220 ret = PTR_ERR(req->ki_eventfd);
1221 req->ki_eventfd = NULL;
1226 ret = put_user(req->ki_key, &user_iocb->aio_key);
1227 if (unlikely(ret)) {
1228 pr_debug("EFAULT: aio_key\n");
1232 req->ki_obj.user = user_iocb;
1233 req->ki_user_data = iocb->aio_data;
1234 req->ki_pos = iocb->aio_offset;
1236 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1237 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1238 req->ki_opcode = iocb->aio_lio_opcode;
1240 ret = aio_setup_iocb(req, compat);
1244 ret = req->ki_retry(req);
1245 if (ret != -EIOCBQUEUED) {
1247 * There's no easy way to restart the syscall since other AIO's
1248 * may be already running. Just fail this IO with EINTR.
1250 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1251 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
1253 aio_complete(req, ret, 0);
1256 aio_put_req(req); /* drop extra ref to req */
1260 atomic_dec(&ctx->reqs_active);
1261 aio_put_req(req); /* drop extra ref to req */
1262 aio_put_req(req); /* drop i/o ref to req */
1266 long do_io_submit(aio_context_t ctx_id, long nr,
1267 struct iocb __user *__user *iocbpp, bool compat)
1272 struct blk_plug plug;
1273 struct kiocb_batch batch;
1275 if (unlikely(nr < 0))
1278 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1279 nr = LONG_MAX/sizeof(*iocbpp);
1281 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1284 ctx = lookup_ioctx(ctx_id);
1285 if (unlikely(!ctx)) {
1286 pr_debug("EINVAL: invalid context id\n");
1290 kiocb_batch_init(&batch, nr);
1292 blk_start_plug(&plug);
1295 * AKPM: should this return a partial result if some of the IOs were
1296 * successfully submitted?
1298 for (i=0; i<nr; i++) {
1299 struct iocb __user *user_iocb;
1302 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1307 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1312 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1316 blk_finish_plug(&plug);
1318 kiocb_batch_free(ctx, &batch);
1324 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1325 * the number of iocbs queued. May return -EINVAL if the aio_context
1326 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1327 * *iocbpp[0] is not properly initialized, if the operation specified
1328 * is invalid for the file descriptor in the iocb. May fail with
1329 * -EFAULT if any of the data structures point to invalid data. May
1330 * fail with -EBADF if the file descriptor specified in the first
1331 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1332 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1333 * fail with -ENOSYS if not implemented.
1335 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1336 struct iocb __user * __user *, iocbpp)
1338 return do_io_submit(ctx_id, nr, iocbpp, 0);
1342 * Finds a given iocb for cancellation.
1344 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1347 struct list_head *pos;
1349 assert_spin_locked(&ctx->ctx_lock);
1351 /* TODO: use a hash or array, this sucks. */
1352 list_for_each(pos, &ctx->active_reqs) {
1353 struct kiocb *kiocb = list_kiocb(pos);
1354 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1361 * Attempts to cancel an iocb previously passed to io_submit. If
1362 * the operation is successfully cancelled, the resulting event is
1363 * copied into the memory pointed to by result without being placed
1364 * into the completion queue and 0 is returned. May fail with
1365 * -EFAULT if any of the data structures pointed to are invalid.
1366 * May fail with -EINVAL if aio_context specified by ctx_id is
1367 * invalid. May fail with -EAGAIN if the iocb specified was not
1368 * cancelled. Will fail with -ENOSYS if not implemented.
1370 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1371 struct io_event __user *, result)
1373 struct io_event res;
1375 struct kiocb *kiocb;
1379 ret = get_user(key, &iocb->aio_key);
1383 ctx = lookup_ioctx(ctx_id);
1387 spin_lock_irq(&ctx->ctx_lock);
1389 kiocb = lookup_kiocb(ctx, iocb, key);
1391 ret = kiocb_cancel(ctx, kiocb, &res);
1395 spin_unlock_irq(&ctx->ctx_lock);
1398 /* Cancellation succeeded -- copy the result
1399 * into the user's buffer.
1401 if (copy_to_user(result, &res, sizeof(res)))
1411 * Attempts to read at least min_nr events and up to nr events from
1412 * the completion queue for the aio_context specified by ctx_id. If
1413 * it succeeds, the number of read events is returned. May fail with
1414 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1415 * out of range, if timeout is out of range. May fail with -EFAULT
1416 * if any of the memory specified is invalid. May return 0 or
1417 * < min_nr if the timeout specified by timeout has elapsed
1418 * before sufficient events are available, where timeout == NULL
1419 * specifies an infinite timeout. Note that the timeout pointed to by
1420 * timeout is relative and will be updated if not NULL and the
1421 * operation blocks. Will fail with -ENOSYS if not implemented.
1423 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1426 struct io_event __user *, events,
1427 struct timespec __user *, timeout)
1429 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1432 if (likely(ioctx)) {
1433 if (likely(min_nr <= nr && min_nr >= 0))
1434 ret = read_events(ioctx, min_nr, nr, events, timeout);