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
66 /* This needs improving */
67 unsigned long user_id;
68 struct hlist_node list;
70 wait_queue_head_t wait;
75 struct list_head active_reqs; /* used for cancellation */
78 * This is what userspace passed to io_setup(), it's not used for
79 * anything but counting against the global max_reqs quota.
81 * The real limit is nr_events - 1, which will be larger (see
86 /* Size of ringbuffer, in units of struct io_event */
89 unsigned long mmap_base;
90 unsigned long mmap_size;
92 struct page **ring_pages;
96 struct mutex ring_lock;
97 } ____cacheline_aligned;
101 spinlock_t completion_lock;
102 } ____cacheline_aligned;
104 struct page *internal_pages[AIO_RING_PAGES];
106 struct rcu_head rcu_head;
107 struct work_struct rcu_work;
110 /*------ sysctl variables----*/
111 static DEFINE_SPINLOCK(aio_nr_lock);
112 unsigned long aio_nr; /* current system wide number of aio requests */
113 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
114 /*----end sysctl variables---*/
116 static struct kmem_cache *kiocb_cachep;
117 static struct kmem_cache *kioctx_cachep;
120 * Creates the slab caches used by the aio routines, panic on
121 * failure as this is done early during the boot sequence.
123 static int __init aio_setup(void)
125 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
126 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
128 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
132 __initcall(aio_setup);
134 static void aio_free_ring(struct kioctx *ctx)
138 for (i = 0; i < ctx->nr_pages; i++)
139 put_page(ctx->ring_pages[i]);
142 vm_munmap(ctx->mmap_base, ctx->mmap_size);
144 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
145 kfree(ctx->ring_pages);
148 static int aio_setup_ring(struct kioctx *ctx)
150 struct aio_ring *ring;
151 unsigned nr_events = ctx->max_reqs;
152 struct mm_struct *mm = current->mm;
153 unsigned long size, populate;
156 /* Compensate for the ring buffer's head/tail overlap entry */
157 nr_events += 2; /* 1 is required, 2 for good luck */
159 size = sizeof(struct aio_ring);
160 size += sizeof(struct io_event) * nr_events;
161 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
166 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
169 ctx->ring_pages = ctx->internal_pages;
170 if (nr_pages > AIO_RING_PAGES) {
171 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
173 if (!ctx->ring_pages)
177 ctx->mmap_size = nr_pages * PAGE_SIZE;
178 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
179 down_write(&mm->mmap_sem);
180 ctx->mmap_base = do_mmap_pgoff(NULL, 0, ctx->mmap_size,
181 PROT_READ|PROT_WRITE,
182 MAP_ANONYMOUS|MAP_PRIVATE, 0, &populate);
183 if (IS_ERR((void *)ctx->mmap_base)) {
184 up_write(&mm->mmap_sem);
190 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
191 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
192 1, 0, ctx->ring_pages, NULL);
193 up_write(&mm->mmap_sem);
195 if (unlikely(ctx->nr_pages != nr_pages)) {
200 mm_populate(ctx->mmap_base, populate);
202 ctx->user_id = ctx->mmap_base;
203 ctx->nr_events = nr_events; /* trusted copy */
205 ring = kmap_atomic(ctx->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(ctx->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 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
225 struct kioctx *ctx = req->ki_ctx;
228 spin_lock_irqsave(&ctx->ctx_lock, flags);
230 if (!req->ki_list.next)
231 list_add(&req->ki_list, &ctx->active_reqs);
233 req->ki_cancel = cancel;
235 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
237 EXPORT_SYMBOL(kiocb_set_cancel_fn);
239 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
240 struct io_event *res)
242 kiocb_cancel_fn *old, *cancel;
246 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
247 * actually has a cancel function, hence the cmpxchg()
250 cancel = ACCESS_ONCE(kiocb->ki_cancel);
252 if (!cancel || cancel == KIOCB_CANCELLED)
256 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
257 } while (cancel != old);
259 atomic_inc(&kiocb->ki_users);
260 spin_unlock_irq(&ctx->ctx_lock);
262 memset(res, 0, sizeof(*res));
263 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
264 res->data = kiocb->ki_user_data;
265 ret = cancel(kiocb, res);
267 spin_lock_irq(&ctx->ctx_lock);
272 static void free_ioctx_rcu(struct rcu_head *head)
274 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
275 kmem_cache_free(kioctx_cachep, ctx);
279 * When this function runs, the kioctx has been removed from the "hash table"
280 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
281 * now it's safe to cancel any that need to be.
283 static void free_ioctx(struct kioctx *ctx)
285 struct aio_ring *ring;
288 unsigned head, avail;
290 spin_lock_irq(&ctx->ctx_lock);
292 while (!list_empty(&ctx->active_reqs)) {
293 req = list_first_entry(&ctx->active_reqs,
294 struct kiocb, ki_list);
296 list_del_init(&req->ki_list);
297 kiocb_cancel(ctx, req, &res);
300 spin_unlock_irq(&ctx->ctx_lock);
302 ring = kmap_atomic(ctx->ring_pages[0]);
306 while (atomic_read(&ctx->reqs_active) > 0) {
307 wait_event(ctx->wait, head != ctx->tail);
309 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
311 atomic_sub(avail, &ctx->reqs_active);
313 head %= ctx->nr_events;
316 WARN_ON(atomic_read(&ctx->reqs_active) < 0);
320 spin_lock(&aio_nr_lock);
321 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
322 aio_nr -= ctx->max_reqs;
323 spin_unlock(&aio_nr_lock);
325 pr_debug("freeing %p\n", ctx);
328 * Here the call_rcu() is between the wait_event() for reqs_active to
329 * hit 0, and freeing the ioctx.
331 * aio_complete() decrements reqs_active, but it has to touch the ioctx
332 * after to issue a wakeup so we use rcu.
334 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
337 static void put_ioctx(struct kioctx *ctx)
339 if (unlikely(atomic_dec_and_test(&ctx->users)))
344 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
346 static struct kioctx *ioctx_alloc(unsigned nr_events)
348 struct mm_struct *mm = current->mm;
352 /* Prevent overflows */
353 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
354 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
355 pr_debug("ENOMEM: nr_events too high\n");
356 return ERR_PTR(-EINVAL);
359 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
360 return ERR_PTR(-EAGAIN);
362 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
364 return ERR_PTR(-ENOMEM);
366 ctx->max_reqs = nr_events;
368 atomic_set(&ctx->users, 2);
369 atomic_set(&ctx->dead, 0);
370 spin_lock_init(&ctx->ctx_lock);
371 spin_lock_init(&ctx->completion_lock);
372 mutex_init(&ctx->ring_lock);
373 init_waitqueue_head(&ctx->wait);
375 INIT_LIST_HEAD(&ctx->active_reqs);
377 if (aio_setup_ring(ctx) < 0)
380 /* limit the number of system wide aios */
381 spin_lock(&aio_nr_lock);
382 if (aio_nr + nr_events > aio_max_nr ||
383 aio_nr + nr_events < aio_nr) {
384 spin_unlock(&aio_nr_lock);
387 aio_nr += ctx->max_reqs;
388 spin_unlock(&aio_nr_lock);
390 /* now link into global list. */
391 spin_lock(&mm->ioctx_lock);
392 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
393 spin_unlock(&mm->ioctx_lock);
395 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
396 ctx, ctx->user_id, mm, ctx->nr_events);
403 kmem_cache_free(kioctx_cachep, ctx);
404 pr_debug("error allocating ioctx %d\n", err);
408 static void kill_ioctx_work(struct work_struct *work)
410 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
412 wake_up_all(&ctx->wait);
416 static void kill_ioctx_rcu(struct rcu_head *head)
418 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
420 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
421 schedule_work(&ctx->rcu_work);
425 * Cancels all outstanding aio requests on an aio context. Used
426 * when the processes owning a context have all exited to encourage
427 * the rapid destruction of the kioctx.
429 static void kill_ioctx(struct kioctx *ctx)
431 if (!atomic_xchg(&ctx->dead, 1)) {
432 hlist_del_rcu(&ctx->list);
433 /* Between hlist_del_rcu() and dropping the initial ref */
437 * We can't punt to workqueue here because put_ioctx() ->
438 * free_ioctx() will unmap the ringbuffer, and that has to be
439 * done in the original process's context. kill_ioctx_rcu/work()
440 * exist for exit_aio(), as in that path free_ioctx() won't do
443 kill_ioctx_work(&ctx->rcu_work);
447 /* wait_on_sync_kiocb:
448 * Waits on the given sync kiocb to complete.
450 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
452 while (atomic_read(&iocb->ki_users)) {
453 set_current_state(TASK_UNINTERRUPTIBLE);
454 if (!atomic_read(&iocb->ki_users))
458 __set_current_state(TASK_RUNNING);
459 return iocb->ki_user_data;
461 EXPORT_SYMBOL(wait_on_sync_kiocb);
464 * exit_aio: called when the last user of mm goes away. At this point, there is
465 * no way for any new requests to be submited or any of the io_* syscalls to be
466 * called on the context.
468 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
471 void exit_aio(struct mm_struct *mm)
474 struct hlist_node *n;
476 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
477 if (1 != atomic_read(&ctx->users))
479 "exit_aio:ioctx still alive: %d %d %d\n",
480 atomic_read(&ctx->users),
481 atomic_read(&ctx->dead),
482 atomic_read(&ctx->reqs_active));
484 * We don't need to bother with munmap() here -
485 * exit_mmap(mm) is coming and it'll unmap everything.
486 * Since aio_free_ring() uses non-zero ->mmap_size
487 * as indicator that it needs to unmap the area,
488 * just set it to 0; aio_free_ring() is the only
489 * place that uses ->mmap_size, so it's safe.
493 if (!atomic_xchg(&ctx->dead, 1)) {
494 hlist_del_rcu(&ctx->list);
495 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
501 * Allocate a slot for an aio request. Increments the ki_users count
502 * of the kioctx so that the kioctx stays around until all requests are
503 * complete. Returns NULL if no requests are free.
505 * Returns with kiocb->ki_users set to 2. The io submit code path holds
506 * an extra reference while submitting the i/o.
507 * This prevents races between the aio code path referencing the
508 * req (after submitting it) and aio_complete() freeing the req.
510 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
514 if (atomic_read(&ctx->reqs_active) >= ctx->nr_events)
517 if (atomic_inc_return(&ctx->reqs_active) > ctx->nr_events - 1)
520 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
524 atomic_set(&req->ki_users, 2);
529 atomic_dec(&ctx->reqs_active);
533 static void kiocb_free(struct kiocb *req)
537 if (req->ki_eventfd != NULL)
538 eventfd_ctx_put(req->ki_eventfd);
541 if (req->ki_iovec != &req->ki_inline_vec)
542 kfree(req->ki_iovec);
543 kmem_cache_free(kiocb_cachep, req);
546 void aio_put_req(struct kiocb *req)
548 if (atomic_dec_and_test(&req->ki_users))
551 EXPORT_SYMBOL(aio_put_req);
553 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
555 struct mm_struct *mm = current->mm;
556 struct kioctx *ctx, *ret = NULL;
560 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
561 if (ctx->user_id == ctx_id) {
562 atomic_inc(&ctx->users);
573 * Called when the io request on the given iocb is complete.
575 void aio_complete(struct kiocb *iocb, long res, long res2)
577 struct kioctx *ctx = iocb->ki_ctx;
578 struct aio_ring *ring;
579 struct io_event *ev_page, *event;
584 * Special case handling for sync iocbs:
585 * - events go directly into the iocb for fast handling
586 * - the sync task with the iocb in its stack holds the single iocb
587 * ref, no other paths have a way to get another ref
588 * - the sync task helpfully left a reference to itself in the iocb
590 if (is_sync_kiocb(iocb)) {
591 BUG_ON(atomic_read(&iocb->ki_users) != 1);
592 iocb->ki_user_data = res;
593 atomic_set(&iocb->ki_users, 0);
594 wake_up_process(iocb->ki_obj.tsk);
599 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
600 * need to issue a wakeup after decrementing reqs_active.
604 if (iocb->ki_list.next) {
607 spin_lock_irqsave(&ctx->ctx_lock, flags);
608 list_del(&iocb->ki_list);
609 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
613 * cancelled requests don't get events, userland was given one
614 * when the event got cancelled.
616 if (unlikely(xchg(&iocb->ki_cancel,
617 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
618 atomic_dec(&ctx->reqs_active);
619 /* Still need the wake_up in case free_ioctx is waiting */
624 * Add a completion event to the ring buffer. Must be done holding
625 * ctx->ctx_lock to prevent other code from messing with the tail
626 * pointer since we might be called from irq context.
628 spin_lock_irqsave(&ctx->completion_lock, flags);
631 pos = tail + AIO_EVENTS_OFFSET;
633 if (++tail >= ctx->nr_events)
636 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
637 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
639 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
640 event->data = iocb->ki_user_data;
644 kunmap_atomic(ev_page);
645 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
647 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
648 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
651 /* after flagging the request as done, we
652 * must never even look at it again
654 smp_wmb(); /* make event visible before updating tail */
658 ring = kmap_atomic(ctx->ring_pages[0]);
661 flush_dcache_page(ctx->ring_pages[0]);
663 spin_unlock_irqrestore(&ctx->completion_lock, flags);
665 pr_debug("added to ring %p at [%u]\n", iocb, tail);
668 * Check if the user asked us to deliver the result through an
669 * eventfd. The eventfd_signal() function is safe to be called
672 if (iocb->ki_eventfd != NULL)
673 eventfd_signal(iocb->ki_eventfd, 1);
676 /* everything turned out well, dispose of the aiocb. */
680 * We have to order our ring_info tail store above and test
681 * of the wait list below outside the wait lock. This is
682 * like in wake_up_bit() where clearing a bit has to be
683 * ordered with the unlocked test.
687 if (waitqueue_active(&ctx->wait))
692 EXPORT_SYMBOL(aio_complete);
695 * Pull an event off of the ioctx's event ring. Returns the number of
698 static long aio_read_events_ring(struct kioctx *ctx,
699 struct io_event __user *event, long nr)
701 struct aio_ring *ring;
706 mutex_lock(&ctx->ring_lock);
708 ring = kmap_atomic(ctx->ring_pages[0]);
712 pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events);
714 if (head == ctx->tail)
722 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
723 if (head == ctx->tail)
726 avail = min(avail, nr - ret);
727 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
728 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
730 pos = head + AIO_EVENTS_OFFSET;
731 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
732 pos %= AIO_EVENTS_PER_PAGE;
735 copy_ret = copy_to_user(event + ret, ev + pos,
736 sizeof(*ev) * avail);
739 if (unlikely(copy_ret)) {
746 head %= ctx->nr_events;
749 ring = kmap_atomic(ctx->ring_pages[0]);
752 flush_dcache_page(ctx->ring_pages[0]);
754 pr_debug("%li h%u t%u\n", ret, head, ctx->tail);
756 atomic_sub(ret, &ctx->reqs_active);
758 mutex_unlock(&ctx->ring_lock);
763 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
764 struct io_event __user *event, long *i)
766 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
771 if (unlikely(atomic_read(&ctx->dead)))
777 return ret < 0 || *i >= min_nr;
780 static long read_events(struct kioctx *ctx, long min_nr, long nr,
781 struct io_event __user *event,
782 struct timespec __user *timeout)
784 ktime_t until = { .tv64 = KTIME_MAX };
790 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
793 until = timespec_to_ktime(ts);
797 * Note that aio_read_events() is being called as the conditional - i.e.
798 * we're calling it after prepare_to_wait() has set task state to
799 * TASK_INTERRUPTIBLE.
801 * But aio_read_events() can block, and if it blocks it's going to flip
802 * the task state back to TASK_RUNNING.
804 * This should be ok, provided it doesn't flip the state back to
805 * TASK_RUNNING and return 0 too much - that causes us to spin. That
806 * will only happen if the mutex_lock() call blocks, and we then find
807 * the ringbuffer empty. So in practice we should be ok, but it's
808 * something to be aware of when touching this code.
810 wait_event_interruptible_hrtimeout(ctx->wait,
811 aio_read_events(ctx, min_nr, nr, event, &ret), until);
813 if (!ret && signal_pending(current))
820 * Create an aio_context capable of receiving at least nr_events.
821 * ctxp must not point to an aio_context that already exists, and
822 * must be initialized to 0 prior to the call. On successful
823 * creation of the aio_context, *ctxp is filled in with the resulting
824 * handle. May fail with -EINVAL if *ctxp is not initialized,
825 * if the specified nr_events exceeds internal limits. May fail
826 * with -EAGAIN if the specified nr_events exceeds the user's limit
827 * of available events. May fail with -ENOMEM if insufficient kernel
828 * resources are available. May fail with -EFAULT if an invalid
829 * pointer is passed for ctxp. Will fail with -ENOSYS if not
832 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
834 struct kioctx *ioctx = NULL;
838 ret = get_user(ctx, ctxp);
843 if (unlikely(ctx || nr_events == 0)) {
844 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
849 ioctx = ioctx_alloc(nr_events);
850 ret = PTR_ERR(ioctx);
851 if (!IS_ERR(ioctx)) {
852 ret = put_user(ioctx->user_id, ctxp);
863 * Destroy the aio_context specified. May cancel any outstanding
864 * AIOs and block on completion. Will fail with -ENOSYS if not
865 * implemented. May fail with -EINVAL if the context pointed to
868 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
870 struct kioctx *ioctx = lookup_ioctx(ctx);
871 if (likely(NULL != ioctx)) {
876 pr_debug("EINVAL: io_destroy: invalid context id\n");
880 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
882 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
886 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
887 ssize_t this = min((ssize_t)iov->iov_len, ret);
888 iov->iov_base += this;
889 iov->iov_len -= this;
890 iocb->ki_left -= this;
892 if (iov->iov_len == 0) {
898 /* the caller should not have done more io than what fit in
899 * the remaining iovecs */
900 BUG_ON(ret > 0 && iocb->ki_left == 0);
903 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
905 struct file *file = iocb->ki_filp;
906 struct address_space *mapping = file->f_mapping;
907 struct inode *inode = mapping->host;
908 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
909 unsigned long, loff_t);
911 unsigned short opcode;
913 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
914 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
915 rw_op = file->f_op->aio_read;
916 opcode = IOCB_CMD_PREADV;
918 rw_op = file->f_op->aio_write;
919 opcode = IOCB_CMD_PWRITEV;
922 /* This matches the pread()/pwrite() logic */
923 if (iocb->ki_pos < 0)
926 if (opcode == IOCB_CMD_PWRITEV)
927 file_start_write(file);
929 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
930 iocb->ki_nr_segs - iocb->ki_cur_seg,
933 aio_advance_iovec(iocb, ret);
935 /* retry all partial writes. retry partial reads as long as its a
937 } while (ret > 0 && iocb->ki_left > 0 &&
938 (opcode == IOCB_CMD_PWRITEV ||
939 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
940 if (opcode == IOCB_CMD_PWRITEV)
941 file_end_write(file);
943 /* This means we must have transferred all that we could */
944 /* No need to retry anymore */
945 if ((ret == 0) || (iocb->ki_left == 0))
946 ret = iocb->ki_nbytes - iocb->ki_left;
948 /* If we managed to write some out we return that, rather than
949 * the eventual error. */
950 if (opcode == IOCB_CMD_PWRITEV
951 && ret < 0 && ret != -EIOCBQUEUED
952 && iocb->ki_nbytes - iocb->ki_left)
953 ret = iocb->ki_nbytes - iocb->ki_left;
958 static ssize_t aio_fdsync(struct kiocb *iocb)
960 struct file *file = iocb->ki_filp;
961 ssize_t ret = -EINVAL;
963 if (file->f_op->aio_fsync)
964 ret = file->f_op->aio_fsync(iocb, 1);
968 static ssize_t aio_fsync(struct kiocb *iocb)
970 struct file *file = iocb->ki_filp;
971 ssize_t ret = -EINVAL;
973 if (file->f_op->aio_fsync)
974 ret = file->f_op->aio_fsync(iocb, 0);
978 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
984 ret = compat_rw_copy_check_uvector(type,
985 (struct compat_iovec __user *)kiocb->ki_buf,
986 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
990 ret = rw_copy_check_uvector(type,
991 (struct iovec __user *)kiocb->ki_buf,
992 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
997 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1001 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1002 kiocb->ki_cur_seg = 0;
1003 /* ki_nbytes/left now reflect bytes instead of segs */
1004 kiocb->ki_nbytes = ret;
1005 kiocb->ki_left = ret;
1012 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1016 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1020 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1021 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1022 kiocb->ki_iovec->iov_len = bytes;
1023 kiocb->ki_nr_segs = 1;
1024 kiocb->ki_cur_seg = 0;
1030 * Performs the initial checks and aio retry method
1031 * setup for the kiocb at the time of io submission.
1033 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1035 struct file *file = kiocb->ki_filp;
1038 switch (kiocb->ki_opcode) {
1039 case IOCB_CMD_PREAD:
1041 if (unlikely(!(file->f_mode & FMODE_READ)))
1044 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1047 ret = aio_setup_single_vector(READ, file, kiocb);
1051 if (file->f_op->aio_read)
1052 kiocb->ki_retry = aio_rw_vect_retry;
1054 case IOCB_CMD_PWRITE:
1056 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1059 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1062 ret = aio_setup_single_vector(WRITE, file, kiocb);
1066 if (file->f_op->aio_write)
1067 kiocb->ki_retry = aio_rw_vect_retry;
1069 case IOCB_CMD_PREADV:
1071 if (unlikely(!(file->f_mode & FMODE_READ)))
1073 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1077 if (file->f_op->aio_read)
1078 kiocb->ki_retry = aio_rw_vect_retry;
1080 case IOCB_CMD_PWRITEV:
1082 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1084 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1088 if (file->f_op->aio_write)
1089 kiocb->ki_retry = aio_rw_vect_retry;
1091 case IOCB_CMD_FDSYNC:
1093 if (file->f_op->aio_fsync)
1094 kiocb->ki_retry = aio_fdsync;
1096 case IOCB_CMD_FSYNC:
1098 if (file->f_op->aio_fsync)
1099 kiocb->ki_retry = aio_fsync;
1102 pr_debug("EINVAL: no operation provided\n");
1106 if (!kiocb->ki_retry)
1112 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1113 struct iocb *iocb, bool compat)
1118 /* enforce forwards compatibility on users */
1119 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1120 pr_debug("EINVAL: reserve field set\n");
1124 /* prevent overflows */
1126 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1127 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1128 ((ssize_t)iocb->aio_nbytes < 0)
1130 pr_debug("EINVAL: io_submit: overflow check\n");
1134 req = aio_get_req(ctx); /* returns with 2 references to req */
1138 req->ki_filp = fget(iocb->aio_fildes);
1139 if (unlikely(!req->ki_filp)) {
1144 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1146 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1147 * instance of the file* now. The file descriptor must be
1148 * an eventfd() fd, and will be signaled for each completed
1149 * event using the eventfd_signal() function.
1151 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1152 if (IS_ERR(req->ki_eventfd)) {
1153 ret = PTR_ERR(req->ki_eventfd);
1154 req->ki_eventfd = NULL;
1159 ret = put_user(req->ki_key, &user_iocb->aio_key);
1160 if (unlikely(ret)) {
1161 pr_debug("EFAULT: aio_key\n");
1165 req->ki_obj.user = user_iocb;
1166 req->ki_user_data = iocb->aio_data;
1167 req->ki_pos = iocb->aio_offset;
1169 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1170 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1171 req->ki_opcode = iocb->aio_lio_opcode;
1173 ret = aio_setup_iocb(req, compat);
1177 ret = req->ki_retry(req);
1178 if (ret != -EIOCBQUEUED) {
1180 * There's no easy way to restart the syscall since other AIO's
1181 * may be already running. Just fail this IO with EINTR.
1183 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1184 ret == -ERESTARTNOHAND ||
1185 ret == -ERESTART_RESTARTBLOCK))
1187 aio_complete(req, ret, 0);
1190 aio_put_req(req); /* drop extra ref to req */
1194 atomic_dec(&ctx->reqs_active);
1195 aio_put_req(req); /* drop extra ref to req */
1196 aio_put_req(req); /* drop i/o ref to req */
1200 long do_io_submit(aio_context_t ctx_id, long nr,
1201 struct iocb __user *__user *iocbpp, bool compat)
1206 struct blk_plug plug;
1208 if (unlikely(nr < 0))
1211 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1212 nr = LONG_MAX/sizeof(*iocbpp);
1214 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1217 ctx = lookup_ioctx(ctx_id);
1218 if (unlikely(!ctx)) {
1219 pr_debug("EINVAL: invalid context id\n");
1223 blk_start_plug(&plug);
1226 * AKPM: should this return a partial result if some of the IOs were
1227 * successfully submitted?
1229 for (i=0; i<nr; i++) {
1230 struct iocb __user *user_iocb;
1233 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1238 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1243 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1247 blk_finish_plug(&plug);
1254 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1255 * the number of iocbs queued. May return -EINVAL if the aio_context
1256 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1257 * *iocbpp[0] is not properly initialized, if the operation specified
1258 * is invalid for the file descriptor in the iocb. May fail with
1259 * -EFAULT if any of the data structures point to invalid data. May
1260 * fail with -EBADF if the file descriptor specified in the first
1261 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1262 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1263 * fail with -ENOSYS if not implemented.
1265 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1266 struct iocb __user * __user *, iocbpp)
1268 return do_io_submit(ctx_id, nr, iocbpp, 0);
1272 * Finds a given iocb for cancellation.
1274 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1277 struct list_head *pos;
1279 assert_spin_locked(&ctx->ctx_lock);
1281 /* TODO: use a hash or array, this sucks. */
1282 list_for_each(pos, &ctx->active_reqs) {
1283 struct kiocb *kiocb = list_kiocb(pos);
1284 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1291 * Attempts to cancel an iocb previously passed to io_submit. If
1292 * the operation is successfully cancelled, the resulting event is
1293 * copied into the memory pointed to by result without being placed
1294 * into the completion queue and 0 is returned. May fail with
1295 * -EFAULT if any of the data structures pointed to are invalid.
1296 * May fail with -EINVAL if aio_context specified by ctx_id is
1297 * invalid. May fail with -EAGAIN if the iocb specified was not
1298 * cancelled. Will fail with -ENOSYS if not implemented.
1300 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1301 struct io_event __user *, result)
1303 struct io_event res;
1305 struct kiocb *kiocb;
1309 ret = get_user(key, &iocb->aio_key);
1313 ctx = lookup_ioctx(ctx_id);
1317 spin_lock_irq(&ctx->ctx_lock);
1319 kiocb = lookup_kiocb(ctx, iocb, key);
1321 ret = kiocb_cancel(ctx, kiocb, &res);
1325 spin_unlock_irq(&ctx->ctx_lock);
1328 /* Cancellation succeeded -- copy the result
1329 * into the user's buffer.
1331 if (copy_to_user(result, &res, sizeof(res)))
1341 * Attempts to read at least min_nr events and up to nr events from
1342 * the completion queue for the aio_context specified by ctx_id. If
1343 * it succeeds, the number of read events is returned. May fail with
1344 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1345 * out of range, if timeout is out of range. May fail with -EFAULT
1346 * if any of the memory specified is invalid. May return 0 or
1347 * < min_nr if the timeout specified by timeout has elapsed
1348 * before sufficient events are available, where timeout == NULL
1349 * specifies an infinite timeout. Note that the timeout pointed to by
1350 * timeout is relative and will be updated if not NULL and the
1351 * operation blocks. Will fail with -ENOSYS if not implemented.
1353 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1356 struct io_event __user *, events,
1357 struct timespec __user *, timeout)
1359 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1362 if (likely(ioctx)) {
1363 if (likely(min_nr <= nr && min_nr >= 0))
1364 ret = read_events(ioctx, min_nr, nr, events, timeout);