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 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/export.h>
17 #include <linux/syscalls.h>
18 #include <linux/backing-dev.h>
19 #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>
43 #define dprintk printk
45 #define dprintk(x...) do { ; } while (0)
48 /*------ sysctl variables----*/
49 static DEFINE_SPINLOCK(aio_nr_lock);
50 unsigned long aio_nr; /* current system wide number of aio requests */
51 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
52 /*----end sysctl variables---*/
54 static struct kmem_cache *kiocb_cachep;
55 static struct kmem_cache *kioctx_cachep;
58 * Creates the slab caches used by the aio routines, panic on
59 * failure as this is done early during the boot sequence.
61 static int __init aio_setup(void)
63 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
64 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
66 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
70 __initcall(aio_setup);
72 static void aio_free_ring(struct kioctx *ctx)
74 struct aio_ring_info *info = &ctx->ring_info;
77 for (i=0; i<info->nr_pages; i++)
78 put_page(info->ring_pages[i]);
80 if (info->mmap_size) {
81 vm_munmap(info->mmap_base, info->mmap_size);
84 if (info->ring_pages && info->ring_pages != info->internal_pages)
85 kfree(info->ring_pages);
86 info->ring_pages = NULL;
90 static int aio_setup_ring(struct kioctx *ctx)
92 struct aio_ring *ring;
93 struct aio_ring_info *info = &ctx->ring_info;
94 unsigned nr_events = ctx->max_reqs;
95 struct mm_struct *mm = current->mm;
96 unsigned long size, populate;
99 /* Compensate for the ring buffer's head/tail overlap entry */
100 nr_events += 2; /* 1 is required, 2 for good luck */
102 size = sizeof(struct aio_ring);
103 size += sizeof(struct io_event) * nr_events;
104 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
109 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
112 info->ring_pages = info->internal_pages;
113 if (nr_pages > AIO_RING_PAGES) {
114 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
115 if (!info->ring_pages)
119 info->mmap_size = nr_pages * PAGE_SIZE;
120 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
121 down_write(&mm->mmap_sem);
122 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
123 PROT_READ|PROT_WRITE,
124 MAP_ANONYMOUS|MAP_PRIVATE, 0,
126 if (IS_ERR((void *)info->mmap_base)) {
127 up_write(&mm->mmap_sem);
133 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
134 info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages,
135 1, 0, info->ring_pages, NULL);
136 up_write(&mm->mmap_sem);
138 if (unlikely(info->nr_pages != nr_pages)) {
143 mm_populate(info->mmap_base, populate);
145 ctx->user_id = info->mmap_base;
147 info->nr = nr_events; /* trusted copy */
149 ring = kmap_atomic(info->ring_pages[0]);
150 ring->nr = nr_events; /* user copy */
151 ring->id = ctx->user_id;
152 ring->head = ring->tail = 0;
153 ring->magic = AIO_RING_MAGIC;
154 ring->compat_features = AIO_RING_COMPAT_FEATURES;
155 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
156 ring->header_length = sizeof(struct aio_ring);
163 /* aio_ring_event: returns a pointer to the event at the given index from
164 * kmap_atomic(). Release the pointer with put_aio_ring_event();
166 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
167 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
168 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
170 #define aio_ring_event(info, nr) ({ \
171 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
172 struct io_event *__event; \
173 __event = kmap_atomic( \
174 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
175 __event += pos % AIO_EVENTS_PER_PAGE; \
179 #define put_aio_ring_event(event) do { \
180 struct io_event *__event = (event); \
182 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
185 static void ctx_rcu_free(struct rcu_head *head)
187 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
188 kmem_cache_free(kioctx_cachep, ctx);
192 * Called when the last user of an aio context has gone away,
193 * and the struct needs to be freed.
195 static void __put_ioctx(struct kioctx *ctx)
197 unsigned nr_events = ctx->max_reqs;
198 BUG_ON(ctx->reqs_active);
202 spin_lock(&aio_nr_lock);
203 BUG_ON(aio_nr - nr_events > aio_nr);
205 spin_unlock(&aio_nr_lock);
207 pr_debug("__put_ioctx: freeing %p\n", ctx);
208 call_rcu(&ctx->rcu_head, ctx_rcu_free);
211 static inline int try_get_ioctx(struct kioctx *kioctx)
213 return atomic_inc_not_zero(&kioctx->users);
216 static inline void put_ioctx(struct kioctx *kioctx)
218 BUG_ON(atomic_read(&kioctx->users) <= 0);
219 if (unlikely(atomic_dec_and_test(&kioctx->users)))
224 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
226 static struct kioctx *ioctx_alloc(unsigned nr_events)
228 struct mm_struct *mm = current->mm;
232 /* Prevent overflows */
233 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
234 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
235 pr_debug("ENOMEM: nr_events too high\n");
236 return ERR_PTR(-EINVAL);
239 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
240 return ERR_PTR(-EAGAIN);
242 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
244 return ERR_PTR(-ENOMEM);
246 ctx->max_reqs = nr_events;
248 atomic_set(&ctx->users, 2);
249 spin_lock_init(&ctx->ctx_lock);
250 spin_lock_init(&ctx->ring_info.ring_lock);
251 init_waitqueue_head(&ctx->wait);
253 INIT_LIST_HEAD(&ctx->active_reqs);
255 if (aio_setup_ring(ctx) < 0)
258 /* limit the number of system wide aios */
259 spin_lock(&aio_nr_lock);
260 if (aio_nr + nr_events > aio_max_nr ||
261 aio_nr + nr_events < aio_nr) {
262 spin_unlock(&aio_nr_lock);
265 aio_nr += ctx->max_reqs;
266 spin_unlock(&aio_nr_lock);
268 /* now link into global list. */
269 spin_lock(&mm->ioctx_lock);
270 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
271 spin_unlock(&mm->ioctx_lock);
273 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
274 ctx, ctx->user_id, mm, ctx->ring_info.nr);
281 kmem_cache_free(kioctx_cachep, ctx);
282 dprintk("aio: error allocating ioctx %d\n", err);
287 * Cancels all outstanding aio requests on an aio context. Used
288 * when the processes owning a context have all exited to encourage
289 * the rapid destruction of the kioctx.
291 static void kill_ctx(struct kioctx *ctx)
293 int (*cancel)(struct kiocb *, struct io_event *);
294 struct task_struct *tsk = current;
295 DECLARE_WAITQUEUE(wait, tsk);
298 spin_lock_irq(&ctx->ctx_lock);
300 while (!list_empty(&ctx->active_reqs)) {
301 struct list_head *pos = ctx->active_reqs.next;
302 struct kiocb *iocb = list_kiocb(pos);
303 list_del_init(&iocb->ki_list);
304 cancel = iocb->ki_cancel;
305 kiocbSetCancelled(iocb);
308 spin_unlock_irq(&ctx->ctx_lock);
310 spin_lock_irq(&ctx->ctx_lock);
314 if (!ctx->reqs_active)
317 add_wait_queue(&ctx->wait, &wait);
318 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
319 while (ctx->reqs_active) {
320 spin_unlock_irq(&ctx->ctx_lock);
322 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
323 spin_lock_irq(&ctx->ctx_lock);
325 __set_task_state(tsk, TASK_RUNNING);
326 remove_wait_queue(&ctx->wait, &wait);
329 spin_unlock_irq(&ctx->ctx_lock);
332 /* wait_on_sync_kiocb:
333 * Waits on the given sync kiocb to complete.
335 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
337 while (iocb->ki_users) {
338 set_current_state(TASK_UNINTERRUPTIBLE);
343 __set_current_state(TASK_RUNNING);
344 return iocb->ki_user_data;
346 EXPORT_SYMBOL(wait_on_sync_kiocb);
348 /* exit_aio: called when the last user of mm goes away. At this point,
349 * there is no way for any new requests to be submited or any of the
350 * io_* syscalls to be called on the context. However, there may be
351 * outstanding requests which hold references to the context; as they
352 * go away, they will call put_ioctx and release any pinned memory
353 * associated with the request (held via struct page * references).
355 void exit_aio(struct mm_struct *mm)
359 while (!hlist_empty(&mm->ioctx_list)) {
360 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
361 hlist_del_rcu(&ctx->list);
365 if (1 != atomic_read(&ctx->users))
367 "exit_aio:ioctx still alive: %d %d %d\n",
368 atomic_read(&ctx->users), ctx->dead,
371 * We don't need to bother with munmap() here -
372 * exit_mmap(mm) is coming and it'll unmap everything.
373 * Since aio_free_ring() uses non-zero ->mmap_size
374 * as indicator that it needs to unmap the area,
375 * just set it to 0; aio_free_ring() is the only
376 * place that uses ->mmap_size, so it's safe.
378 ctx->ring_info.mmap_size = 0;
384 * Allocate a slot for an aio request. Increments the users count
385 * of the kioctx so that the kioctx stays around until all requests are
386 * complete. Returns NULL if no requests are free.
388 * Returns with kiocb->users set to 2. The io submit code path holds
389 * an extra reference while submitting the i/o.
390 * This prevents races between the aio code path referencing the
391 * req (after submitting it) and aio_complete() freeing the req.
393 static struct kiocb *__aio_get_req(struct kioctx *ctx)
395 struct kiocb *req = NULL;
397 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
405 req->ki_cancel = NULL;
406 req->ki_retry = NULL;
409 req->ki_iovec = NULL;
410 req->ki_eventfd = NULL;
416 * struct kiocb's are allocated in batches to reduce the number of
417 * times the ctx lock is acquired and released.
419 #define KIOCB_BATCH_SIZE 32L
421 struct list_head head;
422 long count; /* number of requests left to allocate */
425 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
427 INIT_LIST_HEAD(&batch->head);
428 batch->count = total;
431 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
433 struct kiocb *req, *n;
435 if (list_empty(&batch->head))
438 spin_lock_irq(&ctx->ctx_lock);
439 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
440 list_del(&req->ki_batch);
441 list_del(&req->ki_list);
442 kmem_cache_free(kiocb_cachep, req);
445 if (unlikely(!ctx->reqs_active && ctx->dead))
446 wake_up_all(&ctx->wait);
447 spin_unlock_irq(&ctx->ctx_lock);
451 * Allocate a batch of kiocbs. This avoids taking and dropping the
452 * context lock a lot during setup.
454 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
456 unsigned short allocated, to_alloc;
458 struct kiocb *req, *n;
459 struct aio_ring *ring;
461 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
462 for (allocated = 0; allocated < to_alloc; allocated++) {
463 req = __aio_get_req(ctx);
465 /* allocation failed, go with what we've got */
467 list_add(&req->ki_batch, &batch->head);
473 spin_lock_irq(&ctx->ctx_lock);
474 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
476 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
478 if (avail < allocated) {
479 /* Trim back the number of requests. */
480 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
481 list_del(&req->ki_batch);
482 kmem_cache_free(kiocb_cachep, req);
483 if (--allocated <= avail)
488 batch->count -= allocated;
489 list_for_each_entry(req, &batch->head, ki_batch) {
490 list_add(&req->ki_list, &ctx->active_reqs);
495 spin_unlock_irq(&ctx->ctx_lock);
501 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
502 struct kiocb_batch *batch)
506 if (list_empty(&batch->head))
507 if (kiocb_batch_refill(ctx, batch) == 0)
509 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
510 list_del(&req->ki_batch);
514 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
516 assert_spin_locked(&ctx->ctx_lock);
518 if (req->ki_eventfd != NULL)
519 eventfd_ctx_put(req->ki_eventfd);
522 if (req->ki_iovec != &req->ki_inline_vec)
523 kfree(req->ki_iovec);
524 kmem_cache_free(kiocb_cachep, req);
527 if (unlikely(!ctx->reqs_active && ctx->dead))
528 wake_up_all(&ctx->wait);
532 * Returns true if this put was the last user of the request.
534 static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
536 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
537 req, atomic_long_read(&req->ki_filp->f_count));
539 assert_spin_locked(&ctx->ctx_lock);
542 BUG_ON(req->ki_users < 0);
543 if (likely(req->ki_users))
545 list_del(&req->ki_list); /* remove from active_reqs */
546 req->ki_cancel = NULL;
547 req->ki_retry = NULL;
551 really_put_req(ctx, req);
555 * Returns true if this put was the last user of the kiocb,
556 * false if the request is still in use.
558 void aio_put_req(struct kiocb *req)
560 struct kioctx *ctx = req->ki_ctx;
561 spin_lock_irq(&ctx->ctx_lock);
562 __aio_put_req(ctx, req);
563 spin_unlock_irq(&ctx->ctx_lock);
565 EXPORT_SYMBOL(aio_put_req);
567 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
569 struct mm_struct *mm = current->mm;
570 struct kioctx *ctx, *ret = NULL;
574 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
576 * RCU protects us against accessing freed memory but
577 * we have to be careful not to get a reference when the
578 * reference count already dropped to 0 (ctx->dead test
579 * is unreliable because of races).
581 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
592 * Called when the io request on the given iocb is complete.
594 void aio_complete(struct kiocb *iocb, long res, long res2)
596 struct kioctx *ctx = iocb->ki_ctx;
597 struct aio_ring_info *info;
598 struct aio_ring *ring;
599 struct io_event *event;
604 * Special case handling for sync iocbs:
605 * - events go directly into the iocb for fast handling
606 * - the sync task with the iocb in its stack holds the single iocb
607 * ref, no other paths have a way to get another ref
608 * - the sync task helpfully left a reference to itself in the iocb
610 if (is_sync_kiocb(iocb)) {
611 BUG_ON(iocb->ki_users != 1);
612 iocb->ki_user_data = res;
614 wake_up_process(iocb->ki_obj.tsk);
618 info = &ctx->ring_info;
620 /* add a completion event to the ring buffer.
621 * must be done holding ctx->ctx_lock to prevent
622 * other code from messing with the tail
623 * pointer since we might be called from irq
626 spin_lock_irqsave(&ctx->ctx_lock, flags);
629 * cancelled requests don't get events, userland was given one
630 * when the event got cancelled.
632 if (kiocbIsCancelled(iocb))
635 ring = kmap_atomic(info->ring_pages[0]);
638 event = aio_ring_event(info, tail);
639 if (++tail >= info->nr)
642 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
643 event->data = iocb->ki_user_data;
647 dprintk("aio_complete: %p[%lu]: %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 */
659 put_aio_ring_event(event);
662 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
665 * Check if the user asked us to deliver the result through an
666 * eventfd. The eventfd_signal() function is safe to be called
669 if (iocb->ki_eventfd != NULL)
670 eventfd_signal(iocb->ki_eventfd, 1);
673 /* everything turned out well, dispose of the aiocb. */
674 __aio_put_req(ctx, iocb);
677 * We have to order our ring_info tail store above and test
678 * of the wait list below outside the wait lock. This is
679 * like in wake_up_bit() where clearing a bit has to be
680 * ordered with the unlocked test.
684 if (waitqueue_active(&ctx->wait))
687 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
689 EXPORT_SYMBOL(aio_complete);
692 * Pull an event off of the ioctx's event ring. Returns the number of
693 * events fetched (0 or 1 ;-)
694 * FIXME: make this use cmpxchg.
695 * TODO: make the ringbuffer user mmap()able (requires FIXME).
697 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
699 struct aio_ring_info *info = &ioctx->ring_info;
700 struct aio_ring *ring;
704 ring = kmap_atomic(info->ring_pages[0]);
705 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
706 (unsigned long)ring->head, (unsigned long)ring->tail,
707 (unsigned long)ring->nr);
709 if (ring->head == ring->tail)
712 spin_lock(&info->ring_lock);
714 head = ring->head % info->nr;
715 if (head != ring->tail) {
716 struct io_event *evp = aio_ring_event(info, head);
718 head = (head + 1) % info->nr;
719 smp_mb(); /* finish reading the event before updatng the head */
722 put_aio_ring_event(evp);
724 spin_unlock(&info->ring_lock);
728 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
729 (unsigned long)ring->head, (unsigned long)ring->tail);
734 struct timer_list timer;
736 struct task_struct *p;
739 static void timeout_func(unsigned long data)
741 struct aio_timeout *to = (struct aio_timeout *)data;
744 wake_up_process(to->p);
747 static inline void init_timeout(struct aio_timeout *to)
749 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
754 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
755 const struct timespec *ts)
757 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
758 if (time_after(to->timer.expires, jiffies))
759 add_timer(&to->timer);
764 static inline void clear_timeout(struct aio_timeout *to)
766 del_singleshot_timer_sync(&to->timer);
769 static int read_events(struct kioctx *ctx,
770 long min_nr, long nr,
771 struct io_event __user *event,
772 struct timespec __user *timeout)
774 long start_jiffies = jiffies;
775 struct task_struct *tsk = current;
776 DECLARE_WAITQUEUE(wait, tsk);
780 struct aio_timeout to;
782 /* needed to zero any padding within an entry (there shouldn't be
785 memset(&ent, 0, sizeof(ent));
787 while (likely(i < nr)) {
788 ret = aio_read_evt(ctx, &ent);
789 if (unlikely(ret <= 0))
792 dprintk("read event: %Lx %Lx %Lx %Lx\n",
793 ent.data, ent.obj, ent.res, ent.res2);
795 /* Could we split the check in two? */
797 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
798 dprintk("aio: lost an event due to EFAULT.\n");
803 /* Good, event copied to userland, update counts. */
819 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
822 set_timeout(start_jiffies, &to, &ts);
825 while (likely(i < nr)) {
826 add_wait_queue_exclusive(&ctx->wait, &wait);
828 set_task_state(tsk, TASK_INTERRUPTIBLE);
829 ret = aio_read_evt(ctx, &ent);
834 if (unlikely(ctx->dead)) {
838 if (to.timed_out) /* Only check after read evt */
840 /* Try to only show up in io wait if there are ops
842 if (ctx->reqs_active)
846 if (signal_pending(tsk)) {
850 /*ret = aio_read_evt(ctx, &ent);*/
853 set_task_state(tsk, TASK_RUNNING);
854 remove_wait_queue(&ctx->wait, &wait);
856 if (unlikely(ret <= 0))
860 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
861 dprintk("aio: lost an event due to EFAULT.\n");
865 /* Good, event copied to userland, update counts. */
873 destroy_timer_on_stack(&to.timer);
877 /* Take an ioctx and remove it from the list of ioctx's. Protects
878 * against races with itself via ->dead.
880 static void io_destroy(struct kioctx *ioctx)
882 struct mm_struct *mm = current->mm;
885 /* delete the entry from the list is someone else hasn't already */
886 spin_lock(&mm->ioctx_lock);
887 was_dead = ioctx->dead;
889 hlist_del_rcu(&ioctx->list);
890 spin_unlock(&mm->ioctx_lock);
892 dprintk("aio_release(%p)\n", ioctx);
893 if (likely(!was_dead))
894 put_ioctx(ioctx); /* twice for the list */
899 * Wake up any waiters. The setting of ctx->dead must be seen
900 * by other CPUs at this point. Right now, we rely on the
901 * locking done by the above calls to ensure this consistency.
903 wake_up_all(&ioctx->wait);
907 * Create an aio_context capable of receiving at least nr_events.
908 * ctxp must not point to an aio_context that already exists, and
909 * must be initialized to 0 prior to the call. On successful
910 * creation of the aio_context, *ctxp is filled in with the resulting
911 * handle. May fail with -EINVAL if *ctxp is not initialized,
912 * if the specified nr_events exceeds internal limits. May fail
913 * with -EAGAIN if the specified nr_events exceeds the user's limit
914 * of available events. May fail with -ENOMEM if insufficient kernel
915 * resources are available. May fail with -EFAULT if an invalid
916 * pointer is passed for ctxp. Will fail with -ENOSYS if not
919 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
921 struct kioctx *ioctx = NULL;
925 ret = get_user(ctx, ctxp);
930 if (unlikely(ctx || nr_events == 0)) {
931 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
936 ioctx = ioctx_alloc(nr_events);
937 ret = PTR_ERR(ioctx);
938 if (!IS_ERR(ioctx)) {
939 ret = put_user(ioctx->user_id, ctxp);
950 * Destroy the aio_context specified. May cancel any outstanding
951 * AIOs and block on completion. Will fail with -ENOSYS if not
952 * implemented. May fail with -EINVAL if the context pointed to
955 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
957 struct kioctx *ioctx = lookup_ioctx(ctx);
958 if (likely(NULL != ioctx)) {
963 pr_debug("EINVAL: io_destroy: invalid context id\n");
967 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
969 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
973 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
974 ssize_t this = min((ssize_t)iov->iov_len, ret);
975 iov->iov_base += this;
976 iov->iov_len -= this;
977 iocb->ki_left -= this;
979 if (iov->iov_len == 0) {
985 /* the caller should not have done more io than what fit in
986 * the remaining iovecs */
987 BUG_ON(ret > 0 && iocb->ki_left == 0);
990 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
992 struct file *file = iocb->ki_filp;
993 struct address_space *mapping = file->f_mapping;
994 struct inode *inode = mapping->host;
995 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
996 unsigned long, loff_t);
998 unsigned short opcode;
1000 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1001 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1002 rw_op = file->f_op->aio_read;
1003 opcode = IOCB_CMD_PREADV;
1005 rw_op = file->f_op->aio_write;
1006 opcode = IOCB_CMD_PWRITEV;
1009 /* This matches the pread()/pwrite() logic */
1010 if (iocb->ki_pos < 0)
1014 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1015 iocb->ki_nr_segs - iocb->ki_cur_seg,
1018 aio_advance_iovec(iocb, ret);
1020 /* retry all partial writes. retry partial reads as long as its a
1022 } while (ret > 0 && iocb->ki_left > 0 &&
1023 (opcode == IOCB_CMD_PWRITEV ||
1024 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1026 /* This means we must have transferred all that we could */
1027 /* No need to retry anymore */
1028 if ((ret == 0) || (iocb->ki_left == 0))
1029 ret = iocb->ki_nbytes - iocb->ki_left;
1031 /* If we managed to write some out we return that, rather than
1032 * the eventual error. */
1033 if (opcode == IOCB_CMD_PWRITEV
1034 && ret < 0 && ret != -EIOCBQUEUED
1035 && iocb->ki_nbytes - iocb->ki_left)
1036 ret = iocb->ki_nbytes - iocb->ki_left;
1041 static ssize_t aio_fdsync(struct kiocb *iocb)
1043 struct file *file = iocb->ki_filp;
1044 ssize_t ret = -EINVAL;
1046 if (file->f_op->aio_fsync)
1047 ret = file->f_op->aio_fsync(iocb, 1);
1051 static ssize_t aio_fsync(struct kiocb *iocb)
1053 struct file *file = iocb->ki_filp;
1054 ssize_t ret = -EINVAL;
1056 if (file->f_op->aio_fsync)
1057 ret = file->f_op->aio_fsync(iocb, 0);
1061 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1065 #ifdef CONFIG_COMPAT
1067 ret = compat_rw_copy_check_uvector(type,
1068 (struct compat_iovec __user *)kiocb->ki_buf,
1069 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1073 ret = rw_copy_check_uvector(type,
1074 (struct iovec __user *)kiocb->ki_buf,
1075 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1080 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1084 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1085 kiocb->ki_cur_seg = 0;
1086 /* ki_nbytes/left now reflect bytes instead of segs */
1087 kiocb->ki_nbytes = ret;
1088 kiocb->ki_left = ret;
1095 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1099 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1103 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1104 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1105 kiocb->ki_iovec->iov_len = bytes;
1106 kiocb->ki_nr_segs = 1;
1107 kiocb->ki_cur_seg = 0;
1113 * Performs the initial checks and aio retry method
1114 * setup for the kiocb at the time of io submission.
1116 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1118 struct file *file = kiocb->ki_filp;
1121 switch (kiocb->ki_opcode) {
1122 case IOCB_CMD_PREAD:
1124 if (unlikely(!(file->f_mode & FMODE_READ)))
1127 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1130 ret = aio_setup_single_vector(READ, file, kiocb);
1134 if (file->f_op->aio_read)
1135 kiocb->ki_retry = aio_rw_vect_retry;
1137 case IOCB_CMD_PWRITE:
1139 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1142 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1145 ret = aio_setup_single_vector(WRITE, file, kiocb);
1149 if (file->f_op->aio_write)
1150 kiocb->ki_retry = aio_rw_vect_retry;
1152 case IOCB_CMD_PREADV:
1154 if (unlikely(!(file->f_mode & FMODE_READ)))
1156 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1160 if (file->f_op->aio_read)
1161 kiocb->ki_retry = aio_rw_vect_retry;
1163 case IOCB_CMD_PWRITEV:
1165 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1167 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1171 if (file->f_op->aio_write)
1172 kiocb->ki_retry = aio_rw_vect_retry;
1174 case IOCB_CMD_FDSYNC:
1176 if (file->f_op->aio_fsync)
1177 kiocb->ki_retry = aio_fdsync;
1179 case IOCB_CMD_FSYNC:
1181 if (file->f_op->aio_fsync)
1182 kiocb->ki_retry = aio_fsync;
1185 dprintk("EINVAL: io_submit: no operation provided\n");
1189 if (!kiocb->ki_retry)
1195 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1196 struct iocb *iocb, struct kiocb_batch *batch,
1203 /* enforce forwards compatibility on users */
1204 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1205 pr_debug("EINVAL: io_submit: reserve field set\n");
1209 /* prevent overflows */
1211 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1212 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1213 ((ssize_t)iocb->aio_nbytes < 0)
1215 pr_debug("EINVAL: io_submit: overflow check\n");
1219 file = fget(iocb->aio_fildes);
1220 if (unlikely(!file))
1223 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1224 if (unlikely(!req)) {
1228 req->ki_filp = file;
1229 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1231 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1232 * instance of the file* now. The file descriptor must be
1233 * an eventfd() fd, and will be signaled for each completed
1234 * event using the eventfd_signal() function.
1236 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1237 if (IS_ERR(req->ki_eventfd)) {
1238 ret = PTR_ERR(req->ki_eventfd);
1239 req->ki_eventfd = NULL;
1244 ret = put_user(req->ki_key, &user_iocb->aio_key);
1245 if (unlikely(ret)) {
1246 dprintk("EFAULT: aio_key\n");
1250 req->ki_obj.user = user_iocb;
1251 req->ki_user_data = iocb->aio_data;
1252 req->ki_pos = iocb->aio_offset;
1254 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1255 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1256 req->ki_opcode = iocb->aio_lio_opcode;
1258 ret = aio_setup_iocb(req, compat);
1263 spin_lock_irq(&ctx->ctx_lock);
1265 * We could have raced with io_destroy() and are currently holding a
1266 * reference to ctx which should be destroyed. We cannot submit IO
1267 * since ctx gets freed as soon as io_submit() puts its reference. The
1268 * check here is reliable: io_destroy() sets ctx->dead before waiting
1269 * for outstanding IO and the barrier between these two is realized by
1270 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1271 * increment ctx->reqs_active before checking for ctx->dead and the
1272 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1273 * don't see ctx->dead set here, io_destroy() waits for our IO to
1278 spin_unlock_irq(&ctx->ctx_lock);
1282 if (unlikely(kiocbIsCancelled(req))) {
1285 ret = req->ki_retry(req);
1287 if (ret != -EIOCBQUEUED) {
1289 * There's no easy way to restart the syscall since other AIO's
1290 * may be already running. Just fail this IO with EINTR.
1292 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1293 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
1295 aio_complete(req, ret, 0);
1298 aio_put_req(req); /* drop extra ref to req */
1302 aio_put_req(req); /* drop extra ref to req */
1303 aio_put_req(req); /* drop i/o ref to req */
1307 long do_io_submit(aio_context_t ctx_id, long nr,
1308 struct iocb __user *__user *iocbpp, bool compat)
1313 struct blk_plug plug;
1314 struct kiocb_batch batch;
1316 if (unlikely(nr < 0))
1319 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1320 nr = LONG_MAX/sizeof(*iocbpp);
1322 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1325 ctx = lookup_ioctx(ctx_id);
1326 if (unlikely(!ctx)) {
1327 pr_debug("EINVAL: io_submit: invalid context id\n");
1331 kiocb_batch_init(&batch, nr);
1333 blk_start_plug(&plug);
1336 * AKPM: should this return a partial result if some of the IOs were
1337 * successfully submitted?
1339 for (i=0; i<nr; i++) {
1340 struct iocb __user *user_iocb;
1343 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1348 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1353 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1357 blk_finish_plug(&plug);
1359 kiocb_batch_free(ctx, &batch);
1365 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1366 * the number of iocbs queued. May return -EINVAL if the aio_context
1367 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1368 * *iocbpp[0] is not properly initialized, if the operation specified
1369 * is invalid for the file descriptor in the iocb. May fail with
1370 * -EFAULT if any of the data structures point to invalid data. May
1371 * fail with -EBADF if the file descriptor specified in the first
1372 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1373 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1374 * fail with -ENOSYS if not implemented.
1376 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1377 struct iocb __user * __user *, iocbpp)
1379 return do_io_submit(ctx_id, nr, iocbpp, 0);
1383 * Finds a given iocb for cancellation.
1385 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1388 struct list_head *pos;
1390 assert_spin_locked(&ctx->ctx_lock);
1392 /* TODO: use a hash or array, this sucks. */
1393 list_for_each(pos, &ctx->active_reqs) {
1394 struct kiocb *kiocb = list_kiocb(pos);
1395 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1402 * Attempts to cancel an iocb previously passed to io_submit. If
1403 * the operation is successfully cancelled, the resulting event is
1404 * copied into the memory pointed to by result without being placed
1405 * into the completion queue and 0 is returned. May fail with
1406 * -EFAULT if any of the data structures pointed to are invalid.
1407 * May fail with -EINVAL if aio_context specified by ctx_id is
1408 * invalid. May fail with -EAGAIN if the iocb specified was not
1409 * cancelled. Will fail with -ENOSYS if not implemented.
1411 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1412 struct io_event __user *, result)
1414 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1416 struct kiocb *kiocb;
1420 ret = get_user(key, &iocb->aio_key);
1424 ctx = lookup_ioctx(ctx_id);
1428 spin_lock_irq(&ctx->ctx_lock);
1430 kiocb = lookup_kiocb(ctx, iocb, key);
1431 if (kiocb && kiocb->ki_cancel) {
1432 cancel = kiocb->ki_cancel;
1434 kiocbSetCancelled(kiocb);
1437 spin_unlock_irq(&ctx->ctx_lock);
1439 if (NULL != cancel) {
1440 struct io_event tmp;
1441 pr_debug("calling cancel\n");
1442 memset(&tmp, 0, sizeof(tmp));
1443 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1444 tmp.data = kiocb->ki_user_data;
1445 ret = cancel(kiocb, &tmp);
1447 /* Cancellation succeeded -- copy the result
1448 * into the user's buffer.
1450 if (copy_to_user(result, &tmp, sizeof(tmp)))
1462 * Attempts to read at least min_nr events and up to nr events from
1463 * the completion queue for the aio_context specified by ctx_id. If
1464 * it succeeds, the number of read events is returned. May fail with
1465 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1466 * out of range, if timeout is out of range. May fail with -EFAULT
1467 * if any of the memory specified is invalid. May return 0 or
1468 * < min_nr if the timeout specified by timeout has elapsed
1469 * before sufficient events are available, where timeout == NULL
1470 * specifies an infinite timeout. Note that the timeout pointed to by
1471 * timeout is relative and will be updated if not NULL and the
1472 * operation blocks. Will fail with -ENOSYS if not implemented.
1474 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1477 struct io_event __user *, events,
1478 struct timespec __user *, timeout)
1480 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1483 if (likely(ioctx)) {
1484 if (likely(min_nr <= nr && min_nr >= 0))
1485 ret = read_events(ioctx, min_nr, nr, events, timeout);