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>
38 #include <linux/anon_inodes.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
42 #include <asm/kmap_types.h>
43 #include <asm/uaccess.h>
47 #define AIO_RING_MAGIC 0xa10a10a1
48 #define AIO_RING_COMPAT_FEATURES 1
49 #define AIO_RING_INCOMPAT_FEATURES 0
51 unsigned id; /* kernel internal index number */
52 unsigned nr; /* number of io_events */
57 unsigned compat_features;
58 unsigned incompat_features;
59 unsigned header_length; /* size of aio_ring */
62 struct io_event io_events[0];
63 }; /* 128 bytes + ring size */
65 #define AIO_RING_PAGES 8
71 /* This needs improving */
72 unsigned long user_id;
73 struct hlist_node list;
76 * This is what userspace passed to io_setup(), it's not used for
77 * anything but counting against the global max_reqs quota.
79 * The real limit is nr_events - 1, which will be larger (see
84 /* Size of ringbuffer, in units of struct io_event */
87 unsigned long mmap_base;
88 unsigned long mmap_size;
90 struct page **ring_pages;
93 struct rcu_head rcu_head;
94 struct work_struct rcu_work;
98 * This counts the number of available slots in the ringbuffer,
99 * so we avoid overflowing it: it's decremented (if positive)
100 * when allocating a kiocb and incremented when the resulting
101 * io_event is pulled off the ringbuffer.
103 atomic_t reqs_available;
104 } ____cacheline_aligned_in_smp;
108 struct list_head active_reqs; /* used for cancellation */
109 } ____cacheline_aligned_in_smp;
112 struct mutex ring_lock;
113 wait_queue_head_t wait;
114 } ____cacheline_aligned_in_smp;
118 spinlock_t completion_lock;
119 } ____cacheline_aligned_in_smp;
121 struct page *internal_pages[AIO_RING_PAGES];
122 struct file *aio_ring_file;
125 /*------ sysctl variables----*/
126 static DEFINE_SPINLOCK(aio_nr_lock);
127 unsigned long aio_nr; /* current system wide number of aio requests */
128 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
129 /*----end sysctl variables---*/
131 static struct kmem_cache *kiocb_cachep;
132 static struct kmem_cache *kioctx_cachep;
135 * Creates the slab caches used by the aio routines, panic on
136 * failure as this is done early during the boot sequence.
138 static int __init aio_setup(void)
140 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
141 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
143 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
147 __initcall(aio_setup);
149 static void aio_free_ring(struct kioctx *ctx)
152 struct file *aio_ring_file = ctx->aio_ring_file;
154 for (i = 0; i < ctx->nr_pages; i++) {
155 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
156 page_count(ctx->ring_pages[i]));
157 put_page(ctx->ring_pages[i]);
160 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
161 kfree(ctx->ring_pages);
164 truncate_setsize(aio_ring_file->f_inode, 0);
165 pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n",
166 current->pid, aio_ring_file->f_inode->i_nlink,
167 aio_ring_file->f_path.dentry->d_count,
168 d_unhashed(aio_ring_file->f_path.dentry),
169 atomic_read(&aio_ring_file->f_inode->i_count));
171 ctx->aio_ring_file = NULL;
175 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
177 vma->vm_ops = &generic_file_vm_ops;
181 static const struct file_operations aio_ring_fops = {
182 .mmap = aio_ring_mmap,
185 static int aio_set_page_dirty(struct page *page)
190 #if IS_ENABLED(CONFIG_MIGRATION)
191 static int aio_migratepage(struct address_space *mapping, struct page *new,
192 struct page *old, enum migrate_mode mode)
194 struct kioctx *ctx = mapping->private_data;
196 unsigned idx = old->index;
199 /* Writeback must be complete */
200 BUG_ON(PageWriteback(old));
203 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
204 if (rc != MIGRATEPAGE_SUCCESS) {
211 spin_lock_irqsave(&ctx->completion_lock, flags);
212 migrate_page_copy(new, old);
213 ctx->ring_pages[idx] = new;
214 spin_unlock_irqrestore(&ctx->completion_lock, flags);
220 static const struct address_space_operations aio_ctx_aops = {
221 .set_page_dirty = aio_set_page_dirty,
222 #if IS_ENABLED(CONFIG_MIGRATION)
223 .migratepage = aio_migratepage,
227 static int aio_setup_ring(struct kioctx *ctx)
229 struct aio_ring *ring;
230 unsigned nr_events = ctx->max_reqs;
231 struct mm_struct *mm = current->mm;
232 unsigned long size, populate;
237 /* Compensate for the ring buffer's head/tail overlap entry */
238 nr_events += 2; /* 1 is required, 2 for good luck */
240 size = sizeof(struct aio_ring);
241 size += sizeof(struct io_event) * nr_events;
243 nr_pages = PFN_UP(size);
247 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
249 ctx->aio_ring_file = NULL;
253 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
254 file->f_inode->i_mapping->private_data = ctx;
255 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
257 for (i = 0; i < nr_pages; i++) {
259 page = find_or_create_page(file->f_inode->i_mapping,
260 i, GFP_HIGHUSER | __GFP_ZERO);
263 pr_debug("pid(%d) page[%d]->count=%d\n",
264 current->pid, i, page_count(page));
265 SetPageUptodate(page);
269 ctx->aio_ring_file = file;
270 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
271 / sizeof(struct io_event);
273 ctx->ring_pages = ctx->internal_pages;
274 if (nr_pages > AIO_RING_PAGES) {
275 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
277 if (!ctx->ring_pages)
281 ctx->mmap_size = nr_pages * PAGE_SIZE;
282 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
284 down_write(&mm->mmap_sem);
285 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
286 PROT_READ | PROT_WRITE,
287 MAP_SHARED | MAP_POPULATE, 0, &populate);
288 if (IS_ERR((void *)ctx->mmap_base)) {
289 up_write(&mm->mmap_sem);
294 up_write(&mm->mmap_sem);
296 mm_populate(ctx->mmap_base, populate);
298 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
299 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
300 1, 0, ctx->ring_pages, NULL);
301 for (i = 0; i < ctx->nr_pages; i++)
302 put_page(ctx->ring_pages[i]);
304 if (unlikely(ctx->nr_pages != nr_pages)) {
309 ctx->user_id = ctx->mmap_base;
310 ctx->nr_events = nr_events; /* trusted copy */
312 ring = kmap_atomic(ctx->ring_pages[0]);
313 ring->nr = nr_events; /* user copy */
314 ring->id = ctx->user_id;
315 ring->head = ring->tail = 0;
316 ring->magic = AIO_RING_MAGIC;
317 ring->compat_features = AIO_RING_COMPAT_FEATURES;
318 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
319 ring->header_length = sizeof(struct aio_ring);
321 flush_dcache_page(ctx->ring_pages[0]);
326 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
327 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
328 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
330 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
332 struct kioctx *ctx = req->ki_ctx;
335 spin_lock_irqsave(&ctx->ctx_lock, flags);
337 if (!req->ki_list.next)
338 list_add(&req->ki_list, &ctx->active_reqs);
340 req->ki_cancel = cancel;
342 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
344 EXPORT_SYMBOL(kiocb_set_cancel_fn);
346 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
347 struct io_event *res)
349 kiocb_cancel_fn *old, *cancel;
353 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
354 * actually has a cancel function, hence the cmpxchg()
357 cancel = ACCESS_ONCE(kiocb->ki_cancel);
359 if (!cancel || cancel == KIOCB_CANCELLED)
363 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
364 } while (cancel != old);
366 atomic_inc(&kiocb->ki_users);
367 spin_unlock_irq(&ctx->ctx_lock);
369 memset(res, 0, sizeof(*res));
370 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
371 res->data = kiocb->ki_user_data;
372 ret = cancel(kiocb, res);
374 spin_lock_irq(&ctx->ctx_lock);
379 static void free_ioctx_rcu(struct rcu_head *head)
381 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
382 kmem_cache_free(kioctx_cachep, ctx);
386 * When this function runs, the kioctx has been removed from the "hash table"
387 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
388 * now it's safe to cancel any that need to be.
390 static void free_ioctx(struct kioctx *ctx)
392 struct aio_ring *ring;
395 unsigned head, avail;
397 spin_lock_irq(&ctx->ctx_lock);
399 while (!list_empty(&ctx->active_reqs)) {
400 req = list_first_entry(&ctx->active_reqs,
401 struct kiocb, ki_list);
403 list_del_init(&req->ki_list);
404 kiocb_cancel(ctx, req, &res);
407 spin_unlock_irq(&ctx->ctx_lock);
409 ring = kmap_atomic(ctx->ring_pages[0]);
413 while (atomic_read(&ctx->reqs_available) < ctx->nr_events - 1) {
414 wait_event(ctx->wait,
415 (head != ctx->tail) ||
416 (atomic_read(&ctx->reqs_available) >=
417 ctx->nr_events - 1));
419 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
421 atomic_add(avail, &ctx->reqs_available);
423 head %= ctx->nr_events;
426 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
430 pr_debug("freeing %p\n", ctx);
433 * Here the call_rcu() is between the wait_event() for reqs_active to
434 * hit 0, and freeing the ioctx.
436 * aio_complete() decrements reqs_active, but it has to touch the ioctx
437 * after to issue a wakeup so we use rcu.
439 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
442 static void put_ioctx(struct kioctx *ctx)
444 if (unlikely(atomic_dec_and_test(&ctx->users)))
449 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
451 static struct kioctx *ioctx_alloc(unsigned nr_events)
453 struct mm_struct *mm = current->mm;
457 /* Prevent overflows */
458 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
459 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
460 pr_debug("ENOMEM: nr_events too high\n");
461 return ERR_PTR(-EINVAL);
464 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
465 return ERR_PTR(-EAGAIN);
467 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
469 return ERR_PTR(-ENOMEM);
471 ctx->max_reqs = nr_events;
473 atomic_set(&ctx->users, 2);
474 atomic_set(&ctx->dead, 0);
475 spin_lock_init(&ctx->ctx_lock);
476 spin_lock_init(&ctx->completion_lock);
477 mutex_init(&ctx->ring_lock);
478 init_waitqueue_head(&ctx->wait);
480 INIT_LIST_HEAD(&ctx->active_reqs);
482 if (aio_setup_ring(ctx) < 0)
485 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
487 /* limit the number of system wide aios */
488 spin_lock(&aio_nr_lock);
489 if (aio_nr + nr_events > aio_max_nr ||
490 aio_nr + nr_events < aio_nr) {
491 spin_unlock(&aio_nr_lock);
494 aio_nr += ctx->max_reqs;
495 spin_unlock(&aio_nr_lock);
497 /* now link into global list. */
498 spin_lock(&mm->ioctx_lock);
499 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
500 spin_unlock(&mm->ioctx_lock);
502 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
503 ctx, ctx->user_id, mm, ctx->nr_events);
510 if (ctx->aio_ring_file)
511 fput(ctx->aio_ring_file);
512 kmem_cache_free(kioctx_cachep, ctx);
513 pr_debug("error allocating ioctx %d\n", err);
517 static void kill_ioctx_work(struct work_struct *work)
519 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
521 wake_up_all(&ctx->wait);
525 static void kill_ioctx_rcu(struct rcu_head *head)
527 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
529 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
530 schedule_work(&ctx->rcu_work);
534 * Cancels all outstanding aio requests on an aio context. Used
535 * when the processes owning a context have all exited to encourage
536 * the rapid destruction of the kioctx.
538 static void kill_ioctx(struct kioctx *ctx)
540 if (!atomic_xchg(&ctx->dead, 1)) {
541 hlist_del_rcu(&ctx->list);
544 * It'd be more correct to do this in free_ioctx(), after all
545 * the outstanding kiocbs have finished - but by then io_destroy
546 * has already returned, so io_setup() could potentially return
547 * -EAGAIN with no ioctxs actually in use (as far as userspace
550 spin_lock(&aio_nr_lock);
551 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
552 aio_nr -= ctx->max_reqs;
553 spin_unlock(&aio_nr_lock);
556 vm_munmap(ctx->mmap_base, ctx->mmap_size);
558 /* Between hlist_del_rcu() and dropping the initial ref */
559 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
563 /* wait_on_sync_kiocb:
564 * Waits on the given sync kiocb to complete.
566 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
568 while (atomic_read(&iocb->ki_users)) {
569 set_current_state(TASK_UNINTERRUPTIBLE);
570 if (!atomic_read(&iocb->ki_users))
574 __set_current_state(TASK_RUNNING);
575 return iocb->ki_user_data;
577 EXPORT_SYMBOL(wait_on_sync_kiocb);
580 * exit_aio: called when the last user of mm goes away. At this point, there is
581 * no way for any new requests to be submited or any of the io_* syscalls to be
582 * called on the context.
584 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
587 void exit_aio(struct mm_struct *mm)
590 struct hlist_node *n;
592 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
593 if (1 != atomic_read(&ctx->users))
595 "exit_aio:ioctx still alive: %d %d %d\n",
596 atomic_read(&ctx->users),
597 atomic_read(&ctx->dead),
598 atomic_read(&ctx->reqs_available));
600 * We don't need to bother with munmap() here -
601 * exit_mmap(mm) is coming and it'll unmap everything.
602 * Since aio_free_ring() uses non-zero ->mmap_size
603 * as indicator that it needs to unmap the area,
604 * just set it to 0; aio_free_ring() is the only
605 * place that uses ->mmap_size, so it's safe.
614 * Allocate a slot for an aio request. Increments the ki_users count
615 * of the kioctx so that the kioctx stays around until all requests are
616 * complete. Returns NULL if no requests are free.
618 * Returns with kiocb->ki_users set to 2. The io submit code path holds
619 * an extra reference while submitting the i/o.
620 * This prevents races between the aio code path referencing the
621 * req (after submitting it) and aio_complete() freeing the req.
623 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
627 if (atomic_dec_if_positive(&ctx->reqs_available) <= 0)
630 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
634 atomic_set(&req->ki_users, 2);
639 atomic_inc(&ctx->reqs_available);
643 static void kiocb_free(struct kiocb *req)
647 if (req->ki_eventfd != NULL)
648 eventfd_ctx_put(req->ki_eventfd);
651 if (req->ki_iovec != &req->ki_inline_vec)
652 kfree(req->ki_iovec);
653 kmem_cache_free(kiocb_cachep, req);
656 void aio_put_req(struct kiocb *req)
658 if (atomic_dec_and_test(&req->ki_users))
661 EXPORT_SYMBOL(aio_put_req);
663 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
665 struct mm_struct *mm = current->mm;
666 struct kioctx *ctx, *ret = NULL;
670 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
671 if (ctx->user_id == ctx_id) {
672 atomic_inc(&ctx->users);
683 * Called when the io request on the given iocb is complete.
685 void aio_complete(struct kiocb *iocb, long res, long res2)
687 struct kioctx *ctx = iocb->ki_ctx;
688 struct aio_ring *ring;
689 struct io_event *ev_page, *event;
694 * Special case handling for sync iocbs:
695 * - events go directly into the iocb for fast handling
696 * - the sync task with the iocb in its stack holds the single iocb
697 * ref, no other paths have a way to get another ref
698 * - the sync task helpfully left a reference to itself in the iocb
700 if (is_sync_kiocb(iocb)) {
701 BUG_ON(atomic_read(&iocb->ki_users) != 1);
702 iocb->ki_user_data = res;
703 atomic_set(&iocb->ki_users, 0);
704 wake_up_process(iocb->ki_obj.tsk);
709 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
710 * need to issue a wakeup after incrementing reqs_available.
714 if (iocb->ki_list.next) {
717 spin_lock_irqsave(&ctx->ctx_lock, flags);
718 list_del(&iocb->ki_list);
719 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
723 * cancelled requests don't get events, userland was given one
724 * when the event got cancelled.
726 if (unlikely(xchg(&iocb->ki_cancel,
727 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
728 atomic_inc(&ctx->reqs_available);
729 /* Still need the wake_up in case free_ioctx is waiting */
734 * Add a completion event to the ring buffer. Must be done holding
735 * ctx->completion_lock to prevent other code from messing with the tail
736 * pointer since we might be called from irq context.
738 spin_lock_irqsave(&ctx->completion_lock, flags);
741 pos = tail + AIO_EVENTS_OFFSET;
743 if (++tail >= ctx->nr_events)
746 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
747 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
749 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
750 event->data = iocb->ki_user_data;
754 kunmap_atomic(ev_page);
755 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
757 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
758 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
761 /* after flagging the request as done, we
762 * must never even look at it again
764 smp_wmb(); /* make event visible before updating tail */
768 ring = kmap_atomic(ctx->ring_pages[0]);
771 flush_dcache_page(ctx->ring_pages[0]);
773 spin_unlock_irqrestore(&ctx->completion_lock, flags);
775 pr_debug("added to ring %p at [%u]\n", iocb, tail);
778 * Check if the user asked us to deliver the result through an
779 * eventfd. The eventfd_signal() function is safe to be called
782 if (iocb->ki_eventfd != NULL)
783 eventfd_signal(iocb->ki_eventfd, 1);
786 /* everything turned out well, dispose of the aiocb. */
790 * We have to order our ring_info tail store above and test
791 * of the wait list below outside the wait lock. This is
792 * like in wake_up_bit() where clearing a bit has to be
793 * ordered with the unlocked test.
797 if (waitqueue_active(&ctx->wait))
802 EXPORT_SYMBOL(aio_complete);
805 * Pull an event off of the ioctx's event ring. Returns the number of
808 static long aio_read_events_ring(struct kioctx *ctx,
809 struct io_event __user *event, long nr)
811 struct aio_ring *ring;
816 mutex_lock(&ctx->ring_lock);
818 ring = kmap_atomic(ctx->ring_pages[0]);
822 pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events);
824 if (head == ctx->tail)
832 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
833 if (head == ctx->tail)
836 avail = min(avail, nr - ret);
837 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
838 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
840 pos = head + AIO_EVENTS_OFFSET;
841 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
842 pos %= AIO_EVENTS_PER_PAGE;
845 copy_ret = copy_to_user(event + ret, ev + pos,
846 sizeof(*ev) * avail);
849 if (unlikely(copy_ret)) {
856 head %= ctx->nr_events;
859 ring = kmap_atomic(ctx->ring_pages[0]);
862 flush_dcache_page(ctx->ring_pages[0]);
864 pr_debug("%li h%u t%u\n", ret, head, ctx->tail);
866 atomic_add(ret, &ctx->reqs_available);
868 mutex_unlock(&ctx->ring_lock);
873 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
874 struct io_event __user *event, long *i)
876 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
881 if (unlikely(atomic_read(&ctx->dead)))
887 return ret < 0 || *i >= min_nr;
890 static long read_events(struct kioctx *ctx, long min_nr, long nr,
891 struct io_event __user *event,
892 struct timespec __user *timeout)
894 ktime_t until = { .tv64 = KTIME_MAX };
900 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
903 until = timespec_to_ktime(ts);
907 * Note that aio_read_events() is being called as the conditional - i.e.
908 * we're calling it after prepare_to_wait() has set task state to
909 * TASK_INTERRUPTIBLE.
911 * But aio_read_events() can block, and if it blocks it's going to flip
912 * the task state back to TASK_RUNNING.
914 * This should be ok, provided it doesn't flip the state back to
915 * TASK_RUNNING and return 0 too much - that causes us to spin. That
916 * will only happen if the mutex_lock() call blocks, and we then find
917 * the ringbuffer empty. So in practice we should be ok, but it's
918 * something to be aware of when touching this code.
920 wait_event_interruptible_hrtimeout(ctx->wait,
921 aio_read_events(ctx, min_nr, nr, event, &ret), until);
923 if (!ret && signal_pending(current))
930 * Create an aio_context capable of receiving at least nr_events.
931 * ctxp must not point to an aio_context that already exists, and
932 * must be initialized to 0 prior to the call. On successful
933 * creation of the aio_context, *ctxp is filled in with the resulting
934 * handle. May fail with -EINVAL if *ctxp is not initialized,
935 * if the specified nr_events exceeds internal limits. May fail
936 * with -EAGAIN if the specified nr_events exceeds the user's limit
937 * of available events. May fail with -ENOMEM if insufficient kernel
938 * resources are available. May fail with -EFAULT if an invalid
939 * pointer is passed for ctxp. Will fail with -ENOSYS if not
942 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
944 struct kioctx *ioctx = NULL;
948 ret = get_user(ctx, ctxp);
953 if (unlikely(ctx || nr_events == 0)) {
954 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
959 ioctx = ioctx_alloc(nr_events);
960 ret = PTR_ERR(ioctx);
961 if (!IS_ERR(ioctx)) {
962 ret = put_user(ioctx->user_id, ctxp);
973 * Destroy the aio_context specified. May cancel any outstanding
974 * AIOs and block on completion. Will fail with -ENOSYS if not
975 * implemented. May fail with -EINVAL if the context pointed to
978 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
980 struct kioctx *ioctx = lookup_ioctx(ctx);
981 if (likely(NULL != ioctx)) {
986 pr_debug("EINVAL: io_destroy: invalid context id\n");
990 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
992 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
996 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
997 ssize_t this = min((ssize_t)iov->iov_len, ret);
998 iov->iov_base += this;
999 iov->iov_len -= this;
1000 iocb->ki_left -= this;
1002 if (iov->iov_len == 0) {
1008 /* the caller should not have done more io than what fit in
1009 * the remaining iovecs */
1010 BUG_ON(ret > 0 && iocb->ki_left == 0);
1013 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1014 unsigned long, loff_t);
1016 static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op)
1018 struct file *file = iocb->ki_filp;
1019 struct address_space *mapping = file->f_mapping;
1020 struct inode *inode = mapping->host;
1023 /* This matches the pread()/pwrite() logic */
1024 if (iocb->ki_pos < 0)
1028 file_start_write(file);
1030 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1031 iocb->ki_nr_segs - iocb->ki_cur_seg,
1034 aio_advance_iovec(iocb, ret);
1036 /* retry all partial writes. retry partial reads as long as its a
1038 } while (ret > 0 && iocb->ki_left > 0 &&
1040 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1042 file_end_write(file);
1044 /* This means we must have transferred all that we could */
1045 /* No need to retry anymore */
1046 if ((ret == 0) || (iocb->ki_left == 0))
1047 ret = iocb->ki_nbytes - iocb->ki_left;
1049 /* If we managed to write some out we return that, rather than
1050 * the eventual error. */
1052 && ret < 0 && ret != -EIOCBQUEUED
1053 && iocb->ki_nbytes - iocb->ki_left)
1054 ret = iocb->ki_nbytes - iocb->ki_left;
1059 static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat)
1063 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1065 #ifdef CONFIG_COMPAT
1067 ret = compat_rw_copy_check_uvector(rw,
1068 (struct compat_iovec __user *)kiocb->ki_buf,
1069 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1073 ret = rw_copy_check_uvector(rw,
1074 (struct iovec __user *)kiocb->ki_buf,
1075 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1080 /* ki_nbytes now reflect bytes instead of segs */
1081 kiocb->ki_nbytes = ret;
1085 static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb)
1087 if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes)))
1090 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1091 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1092 kiocb->ki_iovec->iov_len = kiocb->ki_nbytes;
1093 kiocb->ki_nr_segs = 1;
1099 * Performs the initial checks and aio retry method
1100 * setup for the kiocb at the time of io submission.
1102 static ssize_t aio_run_iocb(struct kiocb *req, bool compat)
1104 struct file *file = req->ki_filp;
1110 switch (req->ki_opcode) {
1111 case IOCB_CMD_PREAD:
1112 case IOCB_CMD_PREADV:
1115 rw_op = file->f_op->aio_read;
1118 case IOCB_CMD_PWRITE:
1119 case IOCB_CMD_PWRITEV:
1122 rw_op = file->f_op->aio_write;
1125 if (unlikely(!(file->f_mode & mode)))
1131 ret = (req->ki_opcode == IOCB_CMD_PREADV ||
1132 req->ki_opcode == IOCB_CMD_PWRITEV)
1133 ? aio_setup_vectored_rw(rw, req, compat)
1134 : aio_setup_single_vector(rw, req);
1138 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1142 req->ki_nbytes = ret;
1145 ret = aio_rw_vect_retry(req, rw, rw_op);
1148 case IOCB_CMD_FDSYNC:
1149 if (!file->f_op->aio_fsync)
1152 ret = file->f_op->aio_fsync(req, 1);
1155 case IOCB_CMD_FSYNC:
1156 if (!file->f_op->aio_fsync)
1159 ret = file->f_op->aio_fsync(req, 0);
1163 pr_debug("EINVAL: no operation provided\n");
1167 if (ret != -EIOCBQUEUED) {
1169 * There's no easy way to restart the syscall since other AIO's
1170 * may be already running. Just fail this IO with EINTR.
1172 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1173 ret == -ERESTARTNOHAND ||
1174 ret == -ERESTART_RESTARTBLOCK))
1176 aio_complete(req, ret, 0);
1182 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1183 struct iocb *iocb, bool compat)
1188 /* enforce forwards compatibility on users */
1189 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1190 pr_debug("EINVAL: reserve field set\n");
1194 /* prevent overflows */
1196 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1197 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1198 ((ssize_t)iocb->aio_nbytes < 0)
1200 pr_debug("EINVAL: io_submit: overflow check\n");
1204 req = aio_get_req(ctx);
1208 req->ki_filp = fget(iocb->aio_fildes);
1209 if (unlikely(!req->ki_filp)) {
1214 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1216 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1217 * instance of the file* now. The file descriptor must be
1218 * an eventfd() fd, and will be signaled for each completed
1219 * event using the eventfd_signal() function.
1221 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1222 if (IS_ERR(req->ki_eventfd)) {
1223 ret = PTR_ERR(req->ki_eventfd);
1224 req->ki_eventfd = NULL;
1229 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1230 if (unlikely(ret)) {
1231 pr_debug("EFAULT: aio_key\n");
1235 req->ki_obj.user = user_iocb;
1236 req->ki_user_data = iocb->aio_data;
1237 req->ki_pos = iocb->aio_offset;
1239 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1240 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1241 req->ki_opcode = iocb->aio_lio_opcode;
1243 ret = aio_run_iocb(req, compat);
1247 aio_put_req(req); /* drop extra ref to req */
1250 atomic_inc(&ctx->reqs_available);
1251 aio_put_req(req); /* drop extra ref to req */
1252 aio_put_req(req); /* drop i/o ref to req */
1256 long do_io_submit(aio_context_t ctx_id, long nr,
1257 struct iocb __user *__user *iocbpp, bool compat)
1262 struct blk_plug plug;
1264 if (unlikely(nr < 0))
1267 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1268 nr = LONG_MAX/sizeof(*iocbpp);
1270 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1273 ctx = lookup_ioctx(ctx_id);
1274 if (unlikely(!ctx)) {
1275 pr_debug("EINVAL: invalid context id\n");
1279 blk_start_plug(&plug);
1282 * AKPM: should this return a partial result if some of the IOs were
1283 * successfully submitted?
1285 for (i=0; i<nr; i++) {
1286 struct iocb __user *user_iocb;
1289 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1294 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1299 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1303 blk_finish_plug(&plug);
1310 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1311 * the number of iocbs queued. May return -EINVAL if the aio_context
1312 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1313 * *iocbpp[0] is not properly initialized, if the operation specified
1314 * is invalid for the file descriptor in the iocb. May fail with
1315 * -EFAULT if any of the data structures point to invalid data. May
1316 * fail with -EBADF if the file descriptor specified in the first
1317 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1318 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1319 * fail with -ENOSYS if not implemented.
1321 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1322 struct iocb __user * __user *, iocbpp)
1324 return do_io_submit(ctx_id, nr, iocbpp, 0);
1328 * Finds a given iocb for cancellation.
1330 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1333 struct list_head *pos;
1335 assert_spin_locked(&ctx->ctx_lock);
1337 if (key != KIOCB_KEY)
1340 /* TODO: use a hash or array, this sucks. */
1341 list_for_each(pos, &ctx->active_reqs) {
1342 struct kiocb *kiocb = list_kiocb(pos);
1343 if (kiocb->ki_obj.user == iocb)
1350 * Attempts to cancel an iocb previously passed to io_submit. If
1351 * the operation is successfully cancelled, the resulting event is
1352 * copied into the memory pointed to by result without being placed
1353 * into the completion queue and 0 is returned. May fail with
1354 * -EFAULT if any of the data structures pointed to are invalid.
1355 * May fail with -EINVAL if aio_context specified by ctx_id is
1356 * invalid. May fail with -EAGAIN if the iocb specified was not
1357 * cancelled. Will fail with -ENOSYS if not implemented.
1359 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1360 struct io_event __user *, result)
1362 struct io_event res;
1364 struct kiocb *kiocb;
1368 ret = get_user(key, &iocb->aio_key);
1372 ctx = lookup_ioctx(ctx_id);
1376 spin_lock_irq(&ctx->ctx_lock);
1378 kiocb = lookup_kiocb(ctx, iocb, key);
1380 ret = kiocb_cancel(ctx, kiocb, &res);
1384 spin_unlock_irq(&ctx->ctx_lock);
1387 /* Cancellation succeeded -- copy the result
1388 * into the user's buffer.
1390 if (copy_to_user(result, &res, sizeof(res)))
1400 * Attempts to read at least min_nr events and up to nr events from
1401 * the completion queue for the aio_context specified by ctx_id. If
1402 * it succeeds, the number of read events is returned. May fail with
1403 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1404 * out of range, if timeout is out of range. May fail with -EFAULT
1405 * if any of the memory specified is invalid. May return 0 or
1406 * < min_nr if the timeout specified by timeout has elapsed
1407 * before sufficient events are available, where timeout == NULL
1408 * specifies an infinite timeout. Note that the timeout pointed to by
1409 * timeout is relative. Will fail with -ENOSYS if not implemented.
1411 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1414 struct io_event __user *, events,
1415 struct timespec __user *, timeout)
1417 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1420 if (likely(ioctx)) {
1421 if (likely(min_nr <= nr && min_nr >= 0))
1422 ret = read_events(ioctx, min_nr, nr, events, timeout);