2 * An async IO implementation for Linux
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/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/anon_inodes.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
64 struct io_event io_events[0];
65 }; /* 128 bytes + ring size */
67 #define AIO_RING_PAGES 8
70 unsigned reqs_available;
74 struct percpu_ref users;
77 /* This needs improving */
78 unsigned long user_id;
79 struct hlist_node list;
81 struct __percpu kioctx_cpu *cpu;
84 * For percpu reqs_available, number of slots we move to/from global
89 * This is what userspace passed to io_setup(), it's not used for
90 * anything but counting against the global max_reqs quota.
92 * The real limit is nr_events - 1, which will be larger (see
97 /* Size of ringbuffer, in units of struct io_event */
100 unsigned long mmap_base;
101 unsigned long mmap_size;
103 struct page **ring_pages;
106 struct rcu_head rcu_head;
107 struct work_struct free_work;
111 * This counts the number of available slots in the ringbuffer,
112 * so we avoid overflowing it: it's decremented (if positive)
113 * when allocating a kiocb and incremented when the resulting
114 * io_event is pulled off the ringbuffer.
116 * We batch accesses to it with a percpu version.
118 atomic_t reqs_available;
119 } ____cacheline_aligned_in_smp;
123 struct list_head active_reqs; /* used for cancellation */
124 } ____cacheline_aligned_in_smp;
127 struct mutex ring_lock;
128 wait_queue_head_t wait;
129 } ____cacheline_aligned_in_smp;
133 spinlock_t completion_lock;
134 } ____cacheline_aligned_in_smp;
136 struct page *internal_pages[AIO_RING_PAGES];
137 struct file *aio_ring_file;
140 /*------ sysctl variables----*/
141 static DEFINE_SPINLOCK(aio_nr_lock);
142 unsigned long aio_nr; /* current system wide number of aio requests */
143 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
144 /*----end sysctl variables---*/
146 static struct kmem_cache *kiocb_cachep;
147 static struct kmem_cache *kioctx_cachep;
150 * Creates the slab caches used by the aio routines, panic on
151 * failure as this is done early during the boot sequence.
153 static int __init aio_setup(void)
155 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
156 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
158 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
162 __initcall(aio_setup);
164 static void aio_free_ring(struct kioctx *ctx)
167 struct file *aio_ring_file = ctx->aio_ring_file;
169 for (i = 0; i < ctx->nr_pages; i++) {
170 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
171 page_count(ctx->ring_pages[i]));
172 put_page(ctx->ring_pages[i]);
175 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
176 kfree(ctx->ring_pages);
179 truncate_setsize(aio_ring_file->f_inode, 0);
180 pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n",
181 current->pid, aio_ring_file->f_inode->i_nlink,
182 aio_ring_file->f_path.dentry->d_count,
183 d_unhashed(aio_ring_file->f_path.dentry),
184 atomic_read(&aio_ring_file->f_inode->i_count));
186 ctx->aio_ring_file = NULL;
190 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
192 vma->vm_ops = &generic_file_vm_ops;
196 static const struct file_operations aio_ring_fops = {
197 .mmap = aio_ring_mmap,
200 static int aio_set_page_dirty(struct page *page)
205 #if IS_ENABLED(CONFIG_MIGRATION)
206 static int aio_migratepage(struct address_space *mapping, struct page *new,
207 struct page *old, enum migrate_mode mode)
209 struct kioctx *ctx = mapping->private_data;
211 unsigned idx = old->index;
214 /* Writeback must be complete */
215 BUG_ON(PageWriteback(old));
218 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
219 if (rc != MIGRATEPAGE_SUCCESS) {
226 spin_lock_irqsave(&ctx->completion_lock, flags);
227 migrate_page_copy(new, old);
228 ctx->ring_pages[idx] = new;
229 spin_unlock_irqrestore(&ctx->completion_lock, flags);
235 static const struct address_space_operations aio_ctx_aops = {
236 .set_page_dirty = aio_set_page_dirty,
237 #if IS_ENABLED(CONFIG_MIGRATION)
238 .migratepage = aio_migratepage,
242 static int aio_setup_ring(struct kioctx *ctx)
244 struct aio_ring *ring;
245 unsigned nr_events = ctx->max_reqs;
246 struct mm_struct *mm = current->mm;
247 unsigned long size, populate;
252 /* Compensate for the ring buffer's head/tail overlap entry */
253 nr_events += 2; /* 1 is required, 2 for good luck */
255 size = sizeof(struct aio_ring);
256 size += sizeof(struct io_event) * nr_events;
258 nr_pages = PFN_UP(size);
262 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
264 ctx->aio_ring_file = NULL;
268 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
269 file->f_inode->i_mapping->private_data = ctx;
270 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
272 for (i = 0; i < nr_pages; i++) {
274 page = find_or_create_page(file->f_inode->i_mapping,
275 i, GFP_HIGHUSER | __GFP_ZERO);
278 pr_debug("pid(%d) page[%d]->count=%d\n",
279 current->pid, i, page_count(page));
280 SetPageUptodate(page);
284 ctx->aio_ring_file = file;
285 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
286 / sizeof(struct io_event);
288 ctx->ring_pages = ctx->internal_pages;
289 if (nr_pages > AIO_RING_PAGES) {
290 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
292 if (!ctx->ring_pages)
296 ctx->mmap_size = nr_pages * PAGE_SIZE;
297 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
299 down_write(&mm->mmap_sem);
300 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
301 PROT_READ | PROT_WRITE,
302 MAP_SHARED | MAP_POPULATE, 0, &populate);
303 if (IS_ERR((void *)ctx->mmap_base)) {
304 up_write(&mm->mmap_sem);
309 up_write(&mm->mmap_sem);
311 mm_populate(ctx->mmap_base, populate);
313 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
314 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
315 1, 0, ctx->ring_pages, NULL);
316 for (i = 0; i < ctx->nr_pages; i++)
317 put_page(ctx->ring_pages[i]);
319 if (unlikely(ctx->nr_pages != nr_pages)) {
324 ctx->user_id = ctx->mmap_base;
325 ctx->nr_events = nr_events; /* trusted copy */
327 ring = kmap_atomic(ctx->ring_pages[0]);
328 ring->nr = nr_events; /* user copy */
329 ring->id = ctx->user_id;
330 ring->head = ring->tail = 0;
331 ring->magic = AIO_RING_MAGIC;
332 ring->compat_features = AIO_RING_COMPAT_FEATURES;
333 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
334 ring->header_length = sizeof(struct aio_ring);
336 flush_dcache_page(ctx->ring_pages[0]);
341 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
342 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
343 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
345 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
347 struct kioctx *ctx = req->ki_ctx;
350 spin_lock_irqsave(&ctx->ctx_lock, flags);
352 if (!req->ki_list.next)
353 list_add(&req->ki_list, &ctx->active_reqs);
355 req->ki_cancel = cancel;
357 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
359 EXPORT_SYMBOL(kiocb_set_cancel_fn);
361 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
362 struct io_event *res)
364 kiocb_cancel_fn *old, *cancel;
368 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
369 * actually has a cancel function, hence the cmpxchg()
372 cancel = ACCESS_ONCE(kiocb->ki_cancel);
374 if (!cancel || cancel == KIOCB_CANCELLED)
378 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
379 } while (cancel != old);
381 atomic_inc(&kiocb->ki_users);
382 spin_unlock_irq(&ctx->ctx_lock);
384 memset(res, 0, sizeof(*res));
385 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
386 res->data = kiocb->ki_user_data;
387 ret = cancel(kiocb, res);
389 spin_lock_irq(&ctx->ctx_lock);
394 static void free_ioctx_rcu(struct rcu_head *head)
396 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
398 free_percpu(ctx->cpu);
399 kmem_cache_free(kioctx_cachep, ctx);
403 * When this function runs, the kioctx has been removed from the "hash table"
404 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
405 * now it's safe to cancel any that need to be.
407 static void free_ioctx(struct work_struct *work)
409 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
410 struct aio_ring *ring;
413 unsigned cpu, head, avail;
415 spin_lock_irq(&ctx->ctx_lock);
417 while (!list_empty(&ctx->active_reqs)) {
418 req = list_first_entry(&ctx->active_reqs,
419 struct kiocb, ki_list);
421 list_del_init(&req->ki_list);
422 kiocb_cancel(ctx, req, &res);
425 spin_unlock_irq(&ctx->ctx_lock);
427 for_each_possible_cpu(cpu) {
428 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
430 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
431 kcpu->reqs_available = 0;
434 ring = kmap_atomic(ctx->ring_pages[0]);
438 while (atomic_read(&ctx->reqs_available) < ctx->nr_events - 1) {
439 wait_event(ctx->wait,
440 (head != ctx->tail) ||
441 (atomic_read(&ctx->reqs_available) >=
442 ctx->nr_events - 1));
444 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
446 atomic_add(avail, &ctx->reqs_available);
448 head %= ctx->nr_events;
451 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
455 pr_debug("freeing %p\n", ctx);
458 * Here the call_rcu() is between the wait_event() for reqs_active to
459 * hit 0, and freeing the ioctx.
461 * aio_complete() decrements reqs_active, but it has to touch the ioctx
462 * after to issue a wakeup so we use rcu.
464 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
467 static void free_ioctx_ref(struct percpu_ref *ref)
469 struct kioctx *ctx = container_of(ref, struct kioctx, users);
471 INIT_WORK(&ctx->free_work, free_ioctx);
472 schedule_work(&ctx->free_work);
476 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
478 static struct kioctx *ioctx_alloc(unsigned nr_events)
480 struct mm_struct *mm = current->mm;
485 * We keep track of the number of available ringbuffer slots, to prevent
486 * overflow (reqs_available), and we also use percpu counters for this.
488 * So since up to half the slots might be on other cpu's percpu counters
489 * and unavailable, double nr_events so userspace sees what they
490 * expected: additionally, we move req_batch slots to/from percpu
491 * counters at a time, so make sure that isn't 0:
493 nr_events = max(nr_events, num_possible_cpus() * 4);
496 /* Prevent overflows */
497 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
498 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
499 pr_debug("ENOMEM: nr_events too high\n");
500 return ERR_PTR(-EINVAL);
503 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
504 return ERR_PTR(-EAGAIN);
506 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
508 return ERR_PTR(-ENOMEM);
510 ctx->max_reqs = nr_events;
512 if (percpu_ref_init(&ctx->users, free_ioctx_ref))
515 spin_lock_init(&ctx->ctx_lock);
516 spin_lock_init(&ctx->completion_lock);
517 mutex_init(&ctx->ring_lock);
518 init_waitqueue_head(&ctx->wait);
520 INIT_LIST_HEAD(&ctx->active_reqs);
522 ctx->cpu = alloc_percpu(struct kioctx_cpu);
526 if (aio_setup_ring(ctx) < 0)
529 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
530 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
531 BUG_ON(!ctx->req_batch);
533 /* limit the number of system wide aios */
534 spin_lock(&aio_nr_lock);
535 if (aio_nr + nr_events > aio_max_nr ||
536 aio_nr + nr_events < aio_nr) {
537 spin_unlock(&aio_nr_lock);
540 aio_nr += ctx->max_reqs;
541 spin_unlock(&aio_nr_lock);
543 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
545 /* now link into global list. */
546 spin_lock(&mm->ioctx_lock);
547 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
548 spin_unlock(&mm->ioctx_lock);
550 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
551 ctx, ctx->user_id, mm, ctx->nr_events);
558 free_percpu(ctx->cpu);
560 free_percpu(ctx->users.pcpu_count);
562 if (ctx->aio_ring_file)
563 fput(ctx->aio_ring_file);
564 kmem_cache_free(kioctx_cachep, ctx);
565 pr_debug("error allocating ioctx %d\n", err);
570 * Cancels all outstanding aio requests on an aio context. Used
571 * when the processes owning a context have all exited to encourage
572 * the rapid destruction of the kioctx.
574 static void kill_ioctx(struct kioctx *ctx)
576 if (!atomic_xchg(&ctx->dead, 1)) {
577 hlist_del_rcu(&ctx->list);
578 /* percpu_ref_kill() will do the necessary call_rcu() */
579 wake_up_all(&ctx->wait);
582 * It'd be more correct to do this in free_ioctx(), after all
583 * the outstanding kiocbs have finished - but by then io_destroy
584 * has already returned, so io_setup() could potentially return
585 * -EAGAIN with no ioctxs actually in use (as far as userspace
588 spin_lock(&aio_nr_lock);
589 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
590 aio_nr -= ctx->max_reqs;
591 spin_unlock(&aio_nr_lock);
594 vm_munmap(ctx->mmap_base, ctx->mmap_size);
596 percpu_ref_kill(&ctx->users);
600 /* wait_on_sync_kiocb:
601 * Waits on the given sync kiocb to complete.
603 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
605 while (atomic_read(&iocb->ki_users)) {
606 set_current_state(TASK_UNINTERRUPTIBLE);
607 if (!atomic_read(&iocb->ki_users))
611 __set_current_state(TASK_RUNNING);
612 return iocb->ki_user_data;
614 EXPORT_SYMBOL(wait_on_sync_kiocb);
617 * exit_aio: called when the last user of mm goes away. At this point, there is
618 * no way for any new requests to be submited or any of the io_* syscalls to be
619 * called on the context.
621 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
624 void exit_aio(struct mm_struct *mm)
627 struct hlist_node *n;
629 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
631 * We don't need to bother with munmap() here -
632 * exit_mmap(mm) is coming and it'll unmap everything.
633 * Since aio_free_ring() uses non-zero ->mmap_size
634 * as indicator that it needs to unmap the area,
635 * just set it to 0; aio_free_ring() is the only
636 * place that uses ->mmap_size, so it's safe.
644 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
646 struct kioctx_cpu *kcpu;
649 kcpu = this_cpu_ptr(ctx->cpu);
651 kcpu->reqs_available += nr;
652 while (kcpu->reqs_available >= ctx->req_batch * 2) {
653 kcpu->reqs_available -= ctx->req_batch;
654 atomic_add(ctx->req_batch, &ctx->reqs_available);
660 static bool get_reqs_available(struct kioctx *ctx)
662 struct kioctx_cpu *kcpu;
666 kcpu = this_cpu_ptr(ctx->cpu);
668 if (!kcpu->reqs_available) {
669 int old, avail = atomic_read(&ctx->reqs_available);
672 if (avail < ctx->req_batch)
676 avail = atomic_cmpxchg(&ctx->reqs_available,
677 avail, avail - ctx->req_batch);
678 } while (avail != old);
680 kcpu->reqs_available += ctx->req_batch;
684 kcpu->reqs_available--;
691 * Allocate a slot for an aio request. Increments the ki_users count
692 * of the kioctx so that the kioctx stays around until all requests are
693 * complete. Returns NULL if no requests are free.
695 * Returns with kiocb->ki_users set to 2. The io submit code path holds
696 * an extra reference while submitting the i/o.
697 * This prevents races between the aio code path referencing the
698 * req (after submitting it) and aio_complete() freeing the req.
700 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
704 if (!get_reqs_available(ctx))
707 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
711 atomic_set(&req->ki_users, 2);
715 put_reqs_available(ctx, 1);
719 static void kiocb_free(struct kiocb *req)
723 if (req->ki_eventfd != NULL)
724 eventfd_ctx_put(req->ki_eventfd);
727 if (req->ki_iovec != &req->ki_inline_vec)
728 kfree(req->ki_iovec);
729 kmem_cache_free(kiocb_cachep, req);
732 void aio_put_req(struct kiocb *req)
734 if (atomic_dec_and_test(&req->ki_users))
737 EXPORT_SYMBOL(aio_put_req);
739 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
741 struct mm_struct *mm = current->mm;
742 struct kioctx *ctx, *ret = NULL;
746 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
747 if (ctx->user_id == ctx_id) {
748 percpu_ref_get(&ctx->users);
759 * Called when the io request on the given iocb is complete.
761 void aio_complete(struct kiocb *iocb, long res, long res2)
763 struct kioctx *ctx = iocb->ki_ctx;
764 struct aio_ring *ring;
765 struct io_event *ev_page, *event;
770 * Special case handling for sync iocbs:
771 * - events go directly into the iocb for fast handling
772 * - the sync task with the iocb in its stack holds the single iocb
773 * ref, no other paths have a way to get another ref
774 * - the sync task helpfully left a reference to itself in the iocb
776 if (is_sync_kiocb(iocb)) {
777 BUG_ON(atomic_read(&iocb->ki_users) != 1);
778 iocb->ki_user_data = res;
779 atomic_set(&iocb->ki_users, 0);
780 wake_up_process(iocb->ki_obj.tsk);
785 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
786 * need to issue a wakeup after incrementing reqs_available.
790 if (iocb->ki_list.next) {
793 spin_lock_irqsave(&ctx->ctx_lock, flags);
794 list_del(&iocb->ki_list);
795 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
799 * cancelled requests don't get events, userland was given one
800 * when the event got cancelled.
802 if (unlikely(xchg(&iocb->ki_cancel,
803 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
805 * Can't use the percpu reqs_available here - could race with
808 atomic_inc(&ctx->reqs_available);
809 /* Still need the wake_up in case free_ioctx is waiting */
814 * Add a completion event to the ring buffer. Must be done holding
815 * ctx->completion_lock to prevent other code from messing with the tail
816 * pointer since we might be called from irq context.
818 spin_lock_irqsave(&ctx->completion_lock, flags);
821 pos = tail + AIO_EVENTS_OFFSET;
823 if (++tail >= ctx->nr_events)
826 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
827 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
829 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
830 event->data = iocb->ki_user_data;
834 kunmap_atomic(ev_page);
835 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
837 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
838 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
841 /* after flagging the request as done, we
842 * must never even look at it again
844 smp_wmb(); /* make event visible before updating tail */
848 ring = kmap_atomic(ctx->ring_pages[0]);
851 flush_dcache_page(ctx->ring_pages[0]);
853 spin_unlock_irqrestore(&ctx->completion_lock, flags);
855 pr_debug("added to ring %p at [%u]\n", iocb, tail);
858 * Check if the user asked us to deliver the result through an
859 * eventfd. The eventfd_signal() function is safe to be called
862 if (iocb->ki_eventfd != NULL)
863 eventfd_signal(iocb->ki_eventfd, 1);
866 /* everything turned out well, dispose of the aiocb. */
870 * We have to order our ring_info tail store above and test
871 * of the wait list below outside the wait lock. This is
872 * like in wake_up_bit() where clearing a bit has to be
873 * ordered with the unlocked test.
877 if (waitqueue_active(&ctx->wait))
882 EXPORT_SYMBOL(aio_complete);
885 * Pull an event off of the ioctx's event ring. Returns the number of
888 static long aio_read_events_ring(struct kioctx *ctx,
889 struct io_event __user *event, long nr)
891 struct aio_ring *ring;
896 mutex_lock(&ctx->ring_lock);
898 ring = kmap_atomic(ctx->ring_pages[0]);
902 pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events);
904 if (head == ctx->tail)
912 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
913 if (head == ctx->tail)
916 avail = min(avail, nr - ret);
917 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
918 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
920 pos = head + AIO_EVENTS_OFFSET;
921 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
922 pos %= AIO_EVENTS_PER_PAGE;
925 copy_ret = copy_to_user(event + ret, ev + pos,
926 sizeof(*ev) * avail);
929 if (unlikely(copy_ret)) {
936 head %= ctx->nr_events;
939 ring = kmap_atomic(ctx->ring_pages[0]);
942 flush_dcache_page(ctx->ring_pages[0]);
944 pr_debug("%li h%u t%u\n", ret, head, ctx->tail);
946 put_reqs_available(ctx, ret);
948 mutex_unlock(&ctx->ring_lock);
953 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
954 struct io_event __user *event, long *i)
956 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
961 if (unlikely(atomic_read(&ctx->dead)))
967 return ret < 0 || *i >= min_nr;
970 static long read_events(struct kioctx *ctx, long min_nr, long nr,
971 struct io_event __user *event,
972 struct timespec __user *timeout)
974 ktime_t until = { .tv64 = KTIME_MAX };
980 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
983 until = timespec_to_ktime(ts);
987 * Note that aio_read_events() is being called as the conditional - i.e.
988 * we're calling it after prepare_to_wait() has set task state to
989 * TASK_INTERRUPTIBLE.
991 * But aio_read_events() can block, and if it blocks it's going to flip
992 * the task state back to TASK_RUNNING.
994 * This should be ok, provided it doesn't flip the state back to
995 * TASK_RUNNING and return 0 too much - that causes us to spin. That
996 * will only happen if the mutex_lock() call blocks, and we then find
997 * the ringbuffer empty. So in practice we should be ok, but it's
998 * something to be aware of when touching this code.
1000 wait_event_interruptible_hrtimeout(ctx->wait,
1001 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1003 if (!ret && signal_pending(current))
1010 * Create an aio_context capable of receiving at least nr_events.
1011 * ctxp must not point to an aio_context that already exists, and
1012 * must be initialized to 0 prior to the call. On successful
1013 * creation of the aio_context, *ctxp is filled in with the resulting
1014 * handle. May fail with -EINVAL if *ctxp is not initialized,
1015 * if the specified nr_events exceeds internal limits. May fail
1016 * with -EAGAIN if the specified nr_events exceeds the user's limit
1017 * of available events. May fail with -ENOMEM if insufficient kernel
1018 * resources are available. May fail with -EFAULT if an invalid
1019 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1022 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1024 struct kioctx *ioctx = NULL;
1028 ret = get_user(ctx, ctxp);
1033 if (unlikely(ctx || nr_events == 0)) {
1034 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1039 ioctx = ioctx_alloc(nr_events);
1040 ret = PTR_ERR(ioctx);
1041 if (!IS_ERR(ioctx)) {
1042 ret = put_user(ioctx->user_id, ctxp);
1045 percpu_ref_put(&ioctx->users);
1053 * Destroy the aio_context specified. May cancel any outstanding
1054 * AIOs and block on completion. Will fail with -ENOSYS if not
1055 * implemented. May fail with -EINVAL if the context pointed to
1058 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1060 struct kioctx *ioctx = lookup_ioctx(ctx);
1061 if (likely(NULL != ioctx)) {
1063 percpu_ref_put(&ioctx->users);
1066 pr_debug("EINVAL: io_destroy: invalid context id\n");
1070 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1072 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1076 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1077 ssize_t this = min((ssize_t)iov->iov_len, ret);
1078 iov->iov_base += this;
1079 iov->iov_len -= this;
1080 iocb->ki_left -= this;
1082 if (iov->iov_len == 0) {
1088 /* the caller should not have done more io than what fit in
1089 * the remaining iovecs */
1090 BUG_ON(ret > 0 && iocb->ki_left == 0);
1093 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1094 unsigned long, loff_t);
1096 static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op)
1098 struct file *file = iocb->ki_filp;
1099 struct address_space *mapping = file->f_mapping;
1100 struct inode *inode = mapping->host;
1103 /* This matches the pread()/pwrite() logic */
1104 if (iocb->ki_pos < 0)
1108 file_start_write(file);
1110 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1111 iocb->ki_nr_segs - iocb->ki_cur_seg,
1114 aio_advance_iovec(iocb, ret);
1116 /* retry all partial writes. retry partial reads as long as its a
1118 } while (ret > 0 && iocb->ki_left > 0 &&
1120 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1122 file_end_write(file);
1124 /* This means we must have transferred all that we could */
1125 /* No need to retry anymore */
1126 if ((ret == 0) || (iocb->ki_left == 0))
1127 ret = iocb->ki_nbytes - iocb->ki_left;
1129 /* If we managed to write some out we return that, rather than
1130 * the eventual error. */
1132 && ret < 0 && ret != -EIOCBQUEUED
1133 && iocb->ki_nbytes - iocb->ki_left)
1134 ret = iocb->ki_nbytes - iocb->ki_left;
1139 static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat)
1143 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1145 #ifdef CONFIG_COMPAT
1147 ret = compat_rw_copy_check_uvector(rw,
1148 (struct compat_iovec __user *)kiocb->ki_buf,
1149 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1153 ret = rw_copy_check_uvector(rw,
1154 (struct iovec __user *)kiocb->ki_buf,
1155 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1160 /* ki_nbytes now reflect bytes instead of segs */
1161 kiocb->ki_nbytes = ret;
1165 static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb)
1167 if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes)))
1170 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1171 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1172 kiocb->ki_iovec->iov_len = kiocb->ki_nbytes;
1173 kiocb->ki_nr_segs = 1;
1179 * Performs the initial checks and aio retry method
1180 * setup for the kiocb at the time of io submission.
1182 static ssize_t aio_run_iocb(struct kiocb *req, bool compat)
1184 struct file *file = req->ki_filp;
1190 switch (req->ki_opcode) {
1191 case IOCB_CMD_PREAD:
1192 case IOCB_CMD_PREADV:
1195 rw_op = file->f_op->aio_read;
1198 case IOCB_CMD_PWRITE:
1199 case IOCB_CMD_PWRITEV:
1202 rw_op = file->f_op->aio_write;
1205 if (unlikely(!(file->f_mode & mode)))
1211 ret = (req->ki_opcode == IOCB_CMD_PREADV ||
1212 req->ki_opcode == IOCB_CMD_PWRITEV)
1213 ? aio_setup_vectored_rw(rw, req, compat)
1214 : aio_setup_single_vector(rw, req);
1218 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1222 req->ki_nbytes = ret;
1225 ret = aio_rw_vect_retry(req, rw, rw_op);
1228 case IOCB_CMD_FDSYNC:
1229 if (!file->f_op->aio_fsync)
1232 ret = file->f_op->aio_fsync(req, 1);
1235 case IOCB_CMD_FSYNC:
1236 if (!file->f_op->aio_fsync)
1239 ret = file->f_op->aio_fsync(req, 0);
1243 pr_debug("EINVAL: no operation provided\n");
1247 if (ret != -EIOCBQUEUED) {
1249 * There's no easy way to restart the syscall since other AIO's
1250 * may be already running. Just fail this IO with EINTR.
1252 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1253 ret == -ERESTARTNOHAND ||
1254 ret == -ERESTART_RESTARTBLOCK))
1256 aio_complete(req, ret, 0);
1262 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1263 struct iocb *iocb, bool compat)
1268 /* enforce forwards compatibility on users */
1269 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1270 pr_debug("EINVAL: reserve field set\n");
1274 /* prevent overflows */
1276 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1277 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1278 ((ssize_t)iocb->aio_nbytes < 0)
1280 pr_debug("EINVAL: io_submit: overflow check\n");
1284 req = aio_get_req(ctx);
1288 req->ki_filp = fget(iocb->aio_fildes);
1289 if (unlikely(!req->ki_filp)) {
1294 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1296 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1297 * instance of the file* now. The file descriptor must be
1298 * an eventfd() fd, and will be signaled for each completed
1299 * event using the eventfd_signal() function.
1301 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1302 if (IS_ERR(req->ki_eventfd)) {
1303 ret = PTR_ERR(req->ki_eventfd);
1304 req->ki_eventfd = NULL;
1309 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1310 if (unlikely(ret)) {
1311 pr_debug("EFAULT: aio_key\n");
1315 req->ki_obj.user = user_iocb;
1316 req->ki_user_data = iocb->aio_data;
1317 req->ki_pos = iocb->aio_offset;
1319 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1320 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1321 req->ki_opcode = iocb->aio_lio_opcode;
1323 ret = aio_run_iocb(req, compat);
1327 aio_put_req(req); /* drop extra ref to req */
1330 put_reqs_available(ctx, 1);
1331 aio_put_req(req); /* drop extra ref to req */
1332 aio_put_req(req); /* drop i/o ref to req */
1336 long do_io_submit(aio_context_t ctx_id, long nr,
1337 struct iocb __user *__user *iocbpp, bool compat)
1342 struct blk_plug plug;
1344 if (unlikely(nr < 0))
1347 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1348 nr = LONG_MAX/sizeof(*iocbpp);
1350 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1353 ctx = lookup_ioctx(ctx_id);
1354 if (unlikely(!ctx)) {
1355 pr_debug("EINVAL: invalid context id\n");
1359 blk_start_plug(&plug);
1362 * AKPM: should this return a partial result if some of the IOs were
1363 * successfully submitted?
1365 for (i=0; i<nr; i++) {
1366 struct iocb __user *user_iocb;
1369 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1374 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1379 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1383 blk_finish_plug(&plug);
1385 percpu_ref_put(&ctx->users);
1390 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1391 * the number of iocbs queued. May return -EINVAL if the aio_context
1392 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1393 * *iocbpp[0] is not properly initialized, if the operation specified
1394 * is invalid for the file descriptor in the iocb. May fail with
1395 * -EFAULT if any of the data structures point to invalid data. May
1396 * fail with -EBADF if the file descriptor specified in the first
1397 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1398 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1399 * fail with -ENOSYS if not implemented.
1401 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1402 struct iocb __user * __user *, iocbpp)
1404 return do_io_submit(ctx_id, nr, iocbpp, 0);
1408 * Finds a given iocb for cancellation.
1410 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1413 struct list_head *pos;
1415 assert_spin_locked(&ctx->ctx_lock);
1417 if (key != KIOCB_KEY)
1420 /* TODO: use a hash or array, this sucks. */
1421 list_for_each(pos, &ctx->active_reqs) {
1422 struct kiocb *kiocb = list_kiocb(pos);
1423 if (kiocb->ki_obj.user == iocb)
1430 * Attempts to cancel an iocb previously passed to io_submit. If
1431 * the operation is successfully cancelled, the resulting event is
1432 * copied into the memory pointed to by result without being placed
1433 * into the completion queue and 0 is returned. May fail with
1434 * -EFAULT if any of the data structures pointed to are invalid.
1435 * May fail with -EINVAL if aio_context specified by ctx_id is
1436 * invalid. May fail with -EAGAIN if the iocb specified was not
1437 * cancelled. Will fail with -ENOSYS if not implemented.
1439 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1440 struct io_event __user *, result)
1442 struct io_event res;
1444 struct kiocb *kiocb;
1448 ret = get_user(key, &iocb->aio_key);
1452 ctx = lookup_ioctx(ctx_id);
1456 spin_lock_irq(&ctx->ctx_lock);
1458 kiocb = lookup_kiocb(ctx, iocb, key);
1460 ret = kiocb_cancel(ctx, kiocb, &res);
1464 spin_unlock_irq(&ctx->ctx_lock);
1467 /* Cancellation succeeded -- copy the result
1468 * into the user's buffer.
1470 if (copy_to_user(result, &res, sizeof(res)))
1474 percpu_ref_put(&ctx->users);
1480 * Attempts to read at least min_nr events and up to nr events from
1481 * the completion queue for the aio_context specified by ctx_id. If
1482 * it succeeds, the number of read events is returned. May fail with
1483 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1484 * out of range, if timeout is out of range. May fail with -EFAULT
1485 * if any of the memory specified is invalid. May return 0 or
1486 * < min_nr if the timeout specified by timeout has elapsed
1487 * before sufficient events are available, where timeout == NULL
1488 * specifies an infinite timeout. Note that the timeout pointed to by
1489 * timeout is relative. Will fail with -ENOSYS if not implemented.
1491 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1494 struct io_event __user *, events,
1495 struct timespec __user *, timeout)
1497 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1500 if (likely(ioctx)) {
1501 if (likely(min_nr <= nr && min_nr >= 0))
1502 ret = read_events(ioctx, min_nr, nr, events, timeout);
1503 percpu_ref_put(&ioctx->users);