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;
57 static struct workqueue_struct *aio_wq;
59 /* Used for rare fput completion. */
60 static void aio_fput_routine(struct work_struct *);
61 static DECLARE_WORK(fput_work, aio_fput_routine);
63 static DEFINE_SPINLOCK(fput_lock);
64 static LIST_HEAD(fput_head);
66 static void aio_kick_handler(struct work_struct *);
67 static void aio_queue_work(struct kioctx *);
70 * Creates the slab caches used by the aio routines, panic on
71 * failure as this is done early during the boot sequence.
73 static int __init aio_setup(void)
75 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
78 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
85 __initcall(aio_setup);
87 static void aio_free_ring(struct kioctx *ctx)
89 struct aio_ring_info *info = &ctx->ring_info;
92 for (i=0; i<info->nr_pages; i++)
93 put_page(info->ring_pages[i]);
96 vm_munmap(ctx->mm, info->mmap_base, info->mmap_size);
98 if (info->ring_pages && info->ring_pages != info->internal_pages)
99 kfree(info->ring_pages);
100 info->ring_pages = NULL;
104 static int aio_setup_ring(struct kioctx *ctx)
106 struct aio_ring *ring;
107 struct aio_ring_info *info = &ctx->ring_info;
108 unsigned nr_events = ctx->max_reqs;
112 /* Compensate for the ring buffer's head/tail overlap entry */
113 nr_events += 2; /* 1 is required, 2 for good luck */
115 size = sizeof(struct aio_ring);
116 size += sizeof(struct io_event) * nr_events;
117 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
122 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
125 info->ring_pages = info->internal_pages;
126 if (nr_pages > AIO_RING_PAGES) {
127 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
128 if (!info->ring_pages)
132 info->mmap_size = nr_pages * PAGE_SIZE;
133 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
134 down_write(&ctx->mm->mmap_sem);
135 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
136 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
138 if (IS_ERR((void *)info->mmap_base)) {
139 up_write(&ctx->mm->mmap_sem);
145 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
146 info->nr_pages = get_user_pages(current, ctx->mm,
147 info->mmap_base, nr_pages,
148 1, 0, info->ring_pages, NULL);
149 up_write(&ctx->mm->mmap_sem);
151 if (unlikely(info->nr_pages != nr_pages)) {
156 ctx->user_id = info->mmap_base;
158 info->nr = nr_events; /* trusted copy */
160 ring = kmap_atomic(info->ring_pages[0]);
161 ring->nr = nr_events; /* user copy */
162 ring->id = ctx->user_id;
163 ring->head = ring->tail = 0;
164 ring->magic = AIO_RING_MAGIC;
165 ring->compat_features = AIO_RING_COMPAT_FEATURES;
166 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
167 ring->header_length = sizeof(struct aio_ring);
174 /* aio_ring_event: returns a pointer to the event at the given index from
175 * kmap_atomic(). Release the pointer with put_aio_ring_event();
177 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
178 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
179 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
181 #define aio_ring_event(info, nr) ({ \
182 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
183 struct io_event *__event; \
184 __event = kmap_atomic( \
185 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
186 __event += pos % AIO_EVENTS_PER_PAGE; \
190 #define put_aio_ring_event(event) do { \
191 struct io_event *__event = (event); \
193 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
196 static void ctx_rcu_free(struct rcu_head *head)
198 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
199 kmem_cache_free(kioctx_cachep, ctx);
203 * Called when the last user of an aio context has gone away,
204 * and the struct needs to be freed.
206 static void __put_ioctx(struct kioctx *ctx)
208 unsigned nr_events = ctx->max_reqs;
209 BUG_ON(ctx->reqs_active);
211 cancel_delayed_work_sync(&ctx->wq);
216 spin_lock(&aio_nr_lock);
217 BUG_ON(aio_nr - nr_events > aio_nr);
219 spin_unlock(&aio_nr_lock);
221 pr_debug("__put_ioctx: freeing %p\n", ctx);
222 call_rcu(&ctx->rcu_head, ctx_rcu_free);
225 static inline int try_get_ioctx(struct kioctx *kioctx)
227 return atomic_inc_not_zero(&kioctx->users);
230 static inline void put_ioctx(struct kioctx *kioctx)
232 BUG_ON(atomic_read(&kioctx->users) <= 0);
233 if (unlikely(atomic_dec_and_test(&kioctx->users)))
238 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
240 static struct kioctx *ioctx_alloc(unsigned nr_events)
242 struct mm_struct *mm;
246 /* Prevent overflows */
247 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
248 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
249 pr_debug("ENOMEM: nr_events too high\n");
250 return ERR_PTR(-EINVAL);
253 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
254 return ERR_PTR(-EAGAIN);
256 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
258 return ERR_PTR(-ENOMEM);
260 ctx->max_reqs = nr_events;
261 mm = ctx->mm = current->mm;
262 atomic_inc(&mm->mm_count);
264 atomic_set(&ctx->users, 2);
265 spin_lock_init(&ctx->ctx_lock);
266 spin_lock_init(&ctx->ring_info.ring_lock);
267 init_waitqueue_head(&ctx->wait);
269 INIT_LIST_HEAD(&ctx->active_reqs);
270 INIT_LIST_HEAD(&ctx->run_list);
271 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
273 if (aio_setup_ring(ctx) < 0)
276 /* limit the number of system wide aios */
277 spin_lock(&aio_nr_lock);
278 if (aio_nr + nr_events > aio_max_nr ||
279 aio_nr + nr_events < aio_nr) {
280 spin_unlock(&aio_nr_lock);
283 aio_nr += ctx->max_reqs;
284 spin_unlock(&aio_nr_lock);
286 /* now link into global list. */
287 spin_lock(&mm->ioctx_lock);
288 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
289 spin_unlock(&mm->ioctx_lock);
291 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
292 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
300 kmem_cache_free(kioctx_cachep, ctx);
301 dprintk("aio: error allocating ioctx %d\n", err);
306 * Cancels all outstanding aio requests on an aio context. Used
307 * when the processes owning a context have all exited to encourage
308 * the rapid destruction of the kioctx.
310 static void kill_ctx(struct kioctx *ctx)
312 int (*cancel)(struct kiocb *, struct io_event *);
313 struct task_struct *tsk = current;
314 DECLARE_WAITQUEUE(wait, tsk);
317 spin_lock_irq(&ctx->ctx_lock);
319 while (!list_empty(&ctx->active_reqs)) {
320 struct list_head *pos = ctx->active_reqs.next;
321 struct kiocb *iocb = list_kiocb(pos);
322 list_del_init(&iocb->ki_list);
323 cancel = iocb->ki_cancel;
324 kiocbSetCancelled(iocb);
327 spin_unlock_irq(&ctx->ctx_lock);
329 spin_lock_irq(&ctx->ctx_lock);
333 if (!ctx->reqs_active)
336 add_wait_queue(&ctx->wait, &wait);
337 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
338 while (ctx->reqs_active) {
339 spin_unlock_irq(&ctx->ctx_lock);
341 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
342 spin_lock_irq(&ctx->ctx_lock);
344 __set_task_state(tsk, TASK_RUNNING);
345 remove_wait_queue(&ctx->wait, &wait);
348 spin_unlock_irq(&ctx->ctx_lock);
351 /* wait_on_sync_kiocb:
352 * Waits on the given sync kiocb to complete.
354 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
356 while (iocb->ki_users) {
357 set_current_state(TASK_UNINTERRUPTIBLE);
362 __set_current_state(TASK_RUNNING);
363 return iocb->ki_user_data;
365 EXPORT_SYMBOL(wait_on_sync_kiocb);
367 /* exit_aio: called when the last user of mm goes away. At this point,
368 * there is no way for any new requests to be submited or any of the
369 * io_* syscalls to be called on the context. However, there may be
370 * outstanding requests which hold references to the context; as they
371 * go away, they will call put_ioctx and release any pinned memory
372 * associated with the request (held via struct page * references).
374 void exit_aio(struct mm_struct *mm)
378 while (!hlist_empty(&mm->ioctx_list)) {
379 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
380 hlist_del_rcu(&ctx->list);
384 if (1 != atomic_read(&ctx->users))
386 "exit_aio:ioctx still alive: %d %d %d\n",
387 atomic_read(&ctx->users), ctx->dead,
394 * Allocate a slot for an aio request. Increments the users count
395 * of the kioctx so that the kioctx stays around until all requests are
396 * complete. Returns NULL if no requests are free.
398 * Returns with kiocb->users set to 2. The io submit code path holds
399 * an extra reference while submitting the i/o.
400 * This prevents races between the aio code path referencing the
401 * req (after submitting it) and aio_complete() freeing the req.
403 static struct kiocb *__aio_get_req(struct kioctx *ctx)
405 struct kiocb *req = NULL;
407 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
415 req->ki_cancel = NULL;
416 req->ki_retry = NULL;
419 req->ki_iovec = NULL;
420 INIT_LIST_HEAD(&req->ki_run_list);
421 req->ki_eventfd = NULL;
427 * struct kiocb's are allocated in batches to reduce the number of
428 * times the ctx lock is acquired and released.
430 #define KIOCB_BATCH_SIZE 32L
432 struct list_head head;
433 long count; /* number of requests left to allocate */
436 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
438 INIT_LIST_HEAD(&batch->head);
439 batch->count = total;
442 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
444 struct kiocb *req, *n;
446 if (list_empty(&batch->head))
449 spin_lock_irq(&ctx->ctx_lock);
450 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
451 list_del(&req->ki_batch);
452 list_del(&req->ki_list);
453 kmem_cache_free(kiocb_cachep, req);
456 if (unlikely(!ctx->reqs_active && ctx->dead))
457 wake_up_all(&ctx->wait);
458 spin_unlock_irq(&ctx->ctx_lock);
462 * Allocate a batch of kiocbs. This avoids taking and dropping the
463 * context lock a lot during setup.
465 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
467 unsigned short allocated, to_alloc;
469 bool called_fput = false;
470 struct kiocb *req, *n;
471 struct aio_ring *ring;
473 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
474 for (allocated = 0; allocated < to_alloc; allocated++) {
475 req = __aio_get_req(ctx);
477 /* allocation failed, go with what we've got */
479 list_add(&req->ki_batch, &batch->head);
486 spin_lock_irq(&ctx->ctx_lock);
487 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
489 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
491 if (avail == 0 && !called_fput) {
493 * Handle a potential starvation case. It is possible that
494 * we hold the last reference on a struct file, causing us
495 * to delay the final fput to non-irq context. In this case,
496 * ctx->reqs_active is artificially high. Calling the fput
497 * routine here may free up a slot in the event completion
498 * ring, allowing this allocation to succeed.
501 spin_unlock_irq(&ctx->ctx_lock);
502 aio_fput_routine(NULL);
507 if (avail < allocated) {
508 /* Trim back the number of requests. */
509 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
510 list_del(&req->ki_batch);
511 kmem_cache_free(kiocb_cachep, req);
512 if (--allocated <= avail)
517 batch->count -= allocated;
518 list_for_each_entry(req, &batch->head, ki_batch) {
519 list_add(&req->ki_list, &ctx->active_reqs);
524 spin_unlock_irq(&ctx->ctx_lock);
530 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
531 struct kiocb_batch *batch)
535 if (list_empty(&batch->head))
536 if (kiocb_batch_refill(ctx, batch) == 0)
538 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
539 list_del(&req->ki_batch);
543 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
545 assert_spin_locked(&ctx->ctx_lock);
547 if (req->ki_eventfd != NULL)
548 eventfd_ctx_put(req->ki_eventfd);
551 if (req->ki_iovec != &req->ki_inline_vec)
552 kfree(req->ki_iovec);
553 kmem_cache_free(kiocb_cachep, req);
556 if (unlikely(!ctx->reqs_active && ctx->dead))
557 wake_up_all(&ctx->wait);
560 static void aio_fput_routine(struct work_struct *data)
562 spin_lock_irq(&fput_lock);
563 while (likely(!list_empty(&fput_head))) {
564 struct kiocb *req = list_kiocb(fput_head.next);
565 struct kioctx *ctx = req->ki_ctx;
567 list_del(&req->ki_list);
568 spin_unlock_irq(&fput_lock);
570 /* Complete the fput(s) */
571 if (req->ki_filp != NULL)
574 /* Link the iocb into the context's free list */
576 spin_lock_irq(&ctx->ctx_lock);
577 really_put_req(ctx, req);
579 * at that point ctx might've been killed, but actual
582 spin_unlock_irq(&ctx->ctx_lock);
585 spin_lock_irq(&fput_lock);
587 spin_unlock_irq(&fput_lock);
591 * Returns true if this put was the last user of the request.
593 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
595 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
596 req, atomic_long_read(&req->ki_filp->f_count));
598 assert_spin_locked(&ctx->ctx_lock);
601 BUG_ON(req->ki_users < 0);
602 if (likely(req->ki_users))
604 list_del(&req->ki_list); /* remove from active_reqs */
605 req->ki_cancel = NULL;
606 req->ki_retry = NULL;
609 * Try to optimize the aio and eventfd file* puts, by avoiding to
610 * schedule work in case it is not final fput() time. In normal cases,
611 * we would not be holding the last reference to the file*, so
612 * this function will be executed w/out any aio kthread wakeup.
614 if (unlikely(!fput_atomic(req->ki_filp))) {
615 spin_lock(&fput_lock);
616 list_add(&req->ki_list, &fput_head);
617 spin_unlock(&fput_lock);
618 schedule_work(&fput_work);
621 really_put_req(ctx, req);
627 * Returns true if this put was the last user of the kiocb,
628 * false if the request is still in use.
630 int aio_put_req(struct kiocb *req)
632 struct kioctx *ctx = req->ki_ctx;
634 spin_lock_irq(&ctx->ctx_lock);
635 ret = __aio_put_req(ctx, req);
636 spin_unlock_irq(&ctx->ctx_lock);
639 EXPORT_SYMBOL(aio_put_req);
641 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
643 struct mm_struct *mm = current->mm;
644 struct kioctx *ctx, *ret = NULL;
645 struct hlist_node *n;
649 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
651 * RCU protects us against accessing freed memory but
652 * we have to be careful not to get a reference when the
653 * reference count already dropped to 0 (ctx->dead test
654 * is unreliable because of races).
656 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
667 * Queue up a kiocb to be retried. Assumes that the kiocb
668 * has already been marked as kicked, and places it on
669 * the retry run list for the corresponding ioctx, if it
670 * isn't already queued. Returns 1 if it actually queued
671 * the kiocb (to tell the caller to activate the work
672 * queue to process it), or 0, if it found that it was
675 static inline int __queue_kicked_iocb(struct kiocb *iocb)
677 struct kioctx *ctx = iocb->ki_ctx;
679 assert_spin_locked(&ctx->ctx_lock);
681 if (list_empty(&iocb->ki_run_list)) {
682 list_add_tail(&iocb->ki_run_list,
690 * This is the core aio execution routine. It is
691 * invoked both for initial i/o submission and
692 * subsequent retries via the aio_kick_handler.
693 * Expects to be invoked with iocb->ki_ctx->lock
694 * already held. The lock is released and reacquired
695 * as needed during processing.
697 * Calls the iocb retry method (already setup for the
698 * iocb on initial submission) for operation specific
699 * handling, but takes care of most of common retry
700 * execution details for a given iocb. The retry method
701 * needs to be non-blocking as far as possible, to avoid
702 * holding up other iocbs waiting to be serviced by the
703 * retry kernel thread.
705 * The trickier parts in this code have to do with
706 * ensuring that only one retry instance is in progress
707 * for a given iocb at any time. Providing that guarantee
708 * simplifies the coding of individual aio operations as
709 * it avoids various potential races.
711 static ssize_t aio_run_iocb(struct kiocb *iocb)
713 struct kioctx *ctx = iocb->ki_ctx;
714 ssize_t (*retry)(struct kiocb *);
717 if (!(retry = iocb->ki_retry)) {
718 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
723 * We don't want the next retry iteration for this
724 * operation to start until this one has returned and
725 * updated the iocb state. However, wait_queue functions
726 * can trigger a kick_iocb from interrupt context in the
727 * meantime, indicating that data is available for the next
728 * iteration. We want to remember that and enable the
729 * next retry iteration _after_ we are through with
732 * So, in order to be able to register a "kick", but
733 * prevent it from being queued now, we clear the kick
734 * flag, but make the kick code *think* that the iocb is
735 * still on the run list until we are actually done.
736 * When we are done with this iteration, we check if
737 * the iocb was kicked in the meantime and if so, queue
741 kiocbClearKicked(iocb);
744 * This is so that aio_complete knows it doesn't need to
745 * pull the iocb off the run list (We can't just call
746 * INIT_LIST_HEAD because we don't want a kick_iocb to
747 * queue this on the run list yet)
749 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
750 spin_unlock_irq(&ctx->ctx_lock);
752 /* Quit retrying if the i/o has been cancelled */
753 if (kiocbIsCancelled(iocb)) {
755 aio_complete(iocb, ret, 0);
756 /* must not access the iocb after this */
761 * Now we are all set to call the retry method in async
766 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
768 * There's no easy way to restart the syscall since other AIO's
769 * may be already running. Just fail this IO with EINTR.
771 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
772 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
774 aio_complete(iocb, ret, 0);
777 spin_lock_irq(&ctx->ctx_lock);
779 if (-EIOCBRETRY == ret) {
781 * OK, now that we are done with this iteration
782 * and know that there is more left to go,
783 * this is where we let go so that a subsequent
784 * "kick" can start the next iteration
787 /* will make __queue_kicked_iocb succeed from here on */
788 INIT_LIST_HEAD(&iocb->ki_run_list);
789 /* we must queue the next iteration ourselves, if it
790 * has already been kicked */
791 if (kiocbIsKicked(iocb)) {
792 __queue_kicked_iocb(iocb);
795 * __queue_kicked_iocb will always return 1 here, because
796 * iocb->ki_run_list is empty at this point so it should
797 * be safe to unconditionally queue the context into the
808 * Process all pending retries queued on the ioctx
810 * Assumes it is operating within the aio issuer's mm
813 static int __aio_run_iocbs(struct kioctx *ctx)
816 struct list_head run_list;
818 assert_spin_locked(&ctx->ctx_lock);
820 list_replace_init(&ctx->run_list, &run_list);
821 while (!list_empty(&run_list)) {
822 iocb = list_entry(run_list.next, struct kiocb,
824 list_del(&iocb->ki_run_list);
826 * Hold an extra reference while retrying i/o.
828 iocb->ki_users++; /* grab extra reference */
830 __aio_put_req(ctx, iocb);
832 if (!list_empty(&ctx->run_list))
837 static void aio_queue_work(struct kioctx * ctx)
839 unsigned long timeout;
841 * if someone is waiting, get the work started right
842 * away, otherwise, use a longer delay
845 if (waitqueue_active(&ctx->wait))
849 queue_delayed_work(aio_wq, &ctx->wq, timeout);
854 * Process all pending retries queued on the ioctx
855 * run list, and keep running them until the list
857 * Assumes it is operating within the aio issuer's mm context.
859 static inline void aio_run_all_iocbs(struct kioctx *ctx)
861 spin_lock_irq(&ctx->ctx_lock);
862 while (__aio_run_iocbs(ctx))
864 spin_unlock_irq(&ctx->ctx_lock);
869 * Work queue handler triggered to process pending
870 * retries on an ioctx. Takes on the aio issuer's
871 * mm context before running the iocbs, so that
872 * copy_xxx_user operates on the issuer's address
874 * Run on aiod's context.
876 static void aio_kick_handler(struct work_struct *work)
878 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
879 mm_segment_t oldfs = get_fs();
880 struct mm_struct *mm;
885 spin_lock_irq(&ctx->ctx_lock);
886 requeue =__aio_run_iocbs(ctx);
888 spin_unlock_irq(&ctx->ctx_lock);
892 * we're in a worker thread already; no point using non-zero delay
895 queue_delayed_work(aio_wq, &ctx->wq, 0);
900 * Called by kick_iocb to queue the kiocb for retry
901 * and if required activate the aio work queue to process
904 static void try_queue_kicked_iocb(struct kiocb *iocb)
906 struct kioctx *ctx = iocb->ki_ctx;
910 spin_lock_irqsave(&ctx->ctx_lock, flags);
911 /* set this inside the lock so that we can't race with aio_run_iocb()
912 * testing it and putting the iocb on the run list under the lock */
913 if (!kiocbTryKick(iocb))
914 run = __queue_kicked_iocb(iocb);
915 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
922 * Called typically from a wait queue callback context
923 * to trigger a retry of the iocb.
924 * The retry is usually executed by aio workqueue
925 * threads (See aio_kick_handler).
927 void kick_iocb(struct kiocb *iocb)
929 /* sync iocbs are easy: they can only ever be executing from a
931 if (is_sync_kiocb(iocb)) {
932 kiocbSetKicked(iocb);
933 wake_up_process(iocb->ki_obj.tsk);
937 try_queue_kicked_iocb(iocb);
939 EXPORT_SYMBOL(kick_iocb);
942 * Called when the io request on the given iocb is complete.
943 * Returns true if this is the last user of the request. The
944 * only other user of the request can be the cancellation code.
946 int aio_complete(struct kiocb *iocb, long res, long res2)
948 struct kioctx *ctx = iocb->ki_ctx;
949 struct aio_ring_info *info;
950 struct aio_ring *ring;
951 struct io_event *event;
957 * Special case handling for sync iocbs:
958 * - events go directly into the iocb for fast handling
959 * - the sync task with the iocb in its stack holds the single iocb
960 * ref, no other paths have a way to get another ref
961 * - the sync task helpfully left a reference to itself in the iocb
963 if (is_sync_kiocb(iocb)) {
964 BUG_ON(iocb->ki_users != 1);
965 iocb->ki_user_data = res;
967 wake_up_process(iocb->ki_obj.tsk);
971 info = &ctx->ring_info;
973 /* add a completion event to the ring buffer.
974 * must be done holding ctx->ctx_lock to prevent
975 * other code from messing with the tail
976 * pointer since we might be called from irq
979 spin_lock_irqsave(&ctx->ctx_lock, flags);
981 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
982 list_del_init(&iocb->ki_run_list);
985 * cancelled requests don't get events, userland was given one
986 * when the event got cancelled.
988 if (kiocbIsCancelled(iocb))
991 ring = kmap_atomic(info->ring_pages[0]);
994 event = aio_ring_event(info, tail);
995 if (++tail >= info->nr)
998 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
999 event->data = iocb->ki_user_data;
1003 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1004 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1007 /* after flagging the request as done, we
1008 * must never even look at it again
1010 smp_wmb(); /* make event visible before updating tail */
1015 put_aio_ring_event(event);
1016 kunmap_atomic(ring);
1018 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1021 * Check if the user asked us to deliver the result through an
1022 * eventfd. The eventfd_signal() function is safe to be called
1025 if (iocb->ki_eventfd != NULL)
1026 eventfd_signal(iocb->ki_eventfd, 1);
1029 /* everything turned out well, dispose of the aiocb. */
1030 ret = __aio_put_req(ctx, iocb);
1033 * We have to order our ring_info tail store above and test
1034 * of the wait list below outside the wait lock. This is
1035 * like in wake_up_bit() where clearing a bit has to be
1036 * ordered with the unlocked test.
1040 if (waitqueue_active(&ctx->wait))
1041 wake_up(&ctx->wait);
1043 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1046 EXPORT_SYMBOL(aio_complete);
1049 * Pull an event off of the ioctx's event ring. Returns the number of
1050 * events fetched (0 or 1 ;-)
1051 * FIXME: make this use cmpxchg.
1052 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1054 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1056 struct aio_ring_info *info = &ioctx->ring_info;
1057 struct aio_ring *ring;
1061 ring = kmap_atomic(info->ring_pages[0]);
1062 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1063 (unsigned long)ring->head, (unsigned long)ring->tail,
1064 (unsigned long)ring->nr);
1066 if (ring->head == ring->tail)
1069 spin_lock(&info->ring_lock);
1071 head = ring->head % info->nr;
1072 if (head != ring->tail) {
1073 struct io_event *evp = aio_ring_event(info, head);
1075 head = (head + 1) % info->nr;
1076 smp_mb(); /* finish reading the event before updatng the head */
1079 put_aio_ring_event(evp);
1081 spin_unlock(&info->ring_lock);
1084 kunmap_atomic(ring);
1085 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1086 (unsigned long)ring->head, (unsigned long)ring->tail);
1090 struct aio_timeout {
1091 struct timer_list timer;
1093 struct task_struct *p;
1096 static void timeout_func(unsigned long data)
1098 struct aio_timeout *to = (struct aio_timeout *)data;
1101 wake_up_process(to->p);
1104 static inline void init_timeout(struct aio_timeout *to)
1106 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1111 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1112 const struct timespec *ts)
1114 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1115 if (time_after(to->timer.expires, jiffies))
1116 add_timer(&to->timer);
1121 static inline void clear_timeout(struct aio_timeout *to)
1123 del_singleshot_timer_sync(&to->timer);
1126 static int read_events(struct kioctx *ctx,
1127 long min_nr, long nr,
1128 struct io_event __user *event,
1129 struct timespec __user *timeout)
1131 long start_jiffies = jiffies;
1132 struct task_struct *tsk = current;
1133 DECLARE_WAITQUEUE(wait, tsk);
1136 struct io_event ent;
1137 struct aio_timeout to;
1140 /* needed to zero any padding within an entry (there shouldn't be
1141 * any, but C is fun!
1143 memset(&ent, 0, sizeof(ent));
1146 while (likely(i < nr)) {
1147 ret = aio_read_evt(ctx, &ent);
1148 if (unlikely(ret <= 0))
1151 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1152 ent.data, ent.obj, ent.res, ent.res2);
1154 /* Could we split the check in two? */
1156 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1157 dprintk("aio: lost an event due to EFAULT.\n");
1162 /* Good, event copied to userland, update counts. */
1174 /* racey check, but it gets redone */
1175 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1177 aio_run_all_iocbs(ctx);
1185 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1188 set_timeout(start_jiffies, &to, &ts);
1191 while (likely(i < nr)) {
1192 add_wait_queue_exclusive(&ctx->wait, &wait);
1194 set_task_state(tsk, TASK_INTERRUPTIBLE);
1195 ret = aio_read_evt(ctx, &ent);
1200 if (unlikely(ctx->dead)) {
1204 if (to.timed_out) /* Only check after read evt */
1206 /* Try to only show up in io wait if there are ops
1208 if (ctx->reqs_active)
1212 if (signal_pending(tsk)) {
1216 /*ret = aio_read_evt(ctx, &ent);*/
1219 set_task_state(tsk, TASK_RUNNING);
1220 remove_wait_queue(&ctx->wait, &wait);
1222 if (unlikely(ret <= 0))
1226 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1227 dprintk("aio: lost an event due to EFAULT.\n");
1231 /* Good, event copied to userland, update counts. */
1239 destroy_timer_on_stack(&to.timer);
1243 /* Take an ioctx and remove it from the list of ioctx's. Protects
1244 * against races with itself via ->dead.
1246 static void io_destroy(struct kioctx *ioctx)
1248 struct mm_struct *mm = current->mm;
1251 /* delete the entry from the list is someone else hasn't already */
1252 spin_lock(&mm->ioctx_lock);
1253 was_dead = ioctx->dead;
1255 hlist_del_rcu(&ioctx->list);
1256 spin_unlock(&mm->ioctx_lock);
1258 dprintk("aio_release(%p)\n", ioctx);
1259 if (likely(!was_dead))
1260 put_ioctx(ioctx); /* twice for the list */
1265 * Wake up any waiters. The setting of ctx->dead must be seen
1266 * by other CPUs at this point. Right now, we rely on the
1267 * locking done by the above calls to ensure this consistency.
1269 wake_up_all(&ioctx->wait);
1273 * Create an aio_context capable of receiving at least nr_events.
1274 * ctxp must not point to an aio_context that already exists, and
1275 * must be initialized to 0 prior to the call. On successful
1276 * creation of the aio_context, *ctxp is filled in with the resulting
1277 * handle. May fail with -EINVAL if *ctxp is not initialized,
1278 * if the specified nr_events exceeds internal limits. May fail
1279 * with -EAGAIN if the specified nr_events exceeds the user's limit
1280 * of available events. May fail with -ENOMEM if insufficient kernel
1281 * resources are available. May fail with -EFAULT if an invalid
1282 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1285 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1287 struct kioctx *ioctx = NULL;
1291 ret = get_user(ctx, ctxp);
1296 if (unlikely(ctx || nr_events == 0)) {
1297 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1302 ioctx = ioctx_alloc(nr_events);
1303 ret = PTR_ERR(ioctx);
1304 if (!IS_ERR(ioctx)) {
1305 ret = put_user(ioctx->user_id, ctxp);
1316 * Destroy the aio_context specified. May cancel any outstanding
1317 * AIOs and block on completion. Will fail with -ENOSYS if not
1318 * implemented. May fail with -EINVAL if the context pointed to
1321 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1323 struct kioctx *ioctx = lookup_ioctx(ctx);
1324 if (likely(NULL != ioctx)) {
1329 pr_debug("EINVAL: io_destroy: invalid context id\n");
1333 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1335 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1339 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1340 ssize_t this = min((ssize_t)iov->iov_len, ret);
1341 iov->iov_base += this;
1342 iov->iov_len -= this;
1343 iocb->ki_left -= this;
1345 if (iov->iov_len == 0) {
1351 /* the caller should not have done more io than what fit in
1352 * the remaining iovecs */
1353 BUG_ON(ret > 0 && iocb->ki_left == 0);
1356 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1358 struct file *file = iocb->ki_filp;
1359 struct address_space *mapping = file->f_mapping;
1360 struct inode *inode = mapping->host;
1361 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1362 unsigned long, loff_t);
1364 unsigned short opcode;
1366 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1367 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1368 rw_op = file->f_op->aio_read;
1369 opcode = IOCB_CMD_PREADV;
1371 rw_op = file->f_op->aio_write;
1372 opcode = IOCB_CMD_PWRITEV;
1375 /* This matches the pread()/pwrite() logic */
1376 if (iocb->ki_pos < 0)
1380 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1381 iocb->ki_nr_segs - iocb->ki_cur_seg,
1384 aio_advance_iovec(iocb, ret);
1386 /* retry all partial writes. retry partial reads as long as its a
1388 } while (ret > 0 && iocb->ki_left > 0 &&
1389 (opcode == IOCB_CMD_PWRITEV ||
1390 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1392 /* This means we must have transferred all that we could */
1393 /* No need to retry anymore */
1394 if ((ret == 0) || (iocb->ki_left == 0))
1395 ret = iocb->ki_nbytes - iocb->ki_left;
1397 /* If we managed to write some out we return that, rather than
1398 * the eventual error. */
1399 if (opcode == IOCB_CMD_PWRITEV
1400 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1401 && iocb->ki_nbytes - iocb->ki_left)
1402 ret = iocb->ki_nbytes - iocb->ki_left;
1407 static ssize_t aio_fdsync(struct kiocb *iocb)
1409 struct file *file = iocb->ki_filp;
1410 ssize_t ret = -EINVAL;
1412 if (file->f_op->aio_fsync)
1413 ret = file->f_op->aio_fsync(iocb, 1);
1417 static ssize_t aio_fsync(struct kiocb *iocb)
1419 struct file *file = iocb->ki_filp;
1420 ssize_t ret = -EINVAL;
1422 if (file->f_op->aio_fsync)
1423 ret = file->f_op->aio_fsync(iocb, 0);
1427 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1431 #ifdef CONFIG_COMPAT
1433 ret = compat_rw_copy_check_uvector(type,
1434 (struct compat_iovec __user *)kiocb->ki_buf,
1435 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1436 &kiocb->ki_iovec, 1);
1439 ret = rw_copy_check_uvector(type,
1440 (struct iovec __user *)kiocb->ki_buf,
1441 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1442 &kiocb->ki_iovec, 1);
1446 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1447 kiocb->ki_cur_seg = 0;
1448 /* ki_nbytes/left now reflect bytes instead of segs */
1449 kiocb->ki_nbytes = ret;
1450 kiocb->ki_left = ret;
1457 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1459 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1460 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1461 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1462 kiocb->ki_nr_segs = 1;
1463 kiocb->ki_cur_seg = 0;
1469 * Performs the initial checks and aio retry method
1470 * setup for the kiocb at the time of io submission.
1472 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1474 struct file *file = kiocb->ki_filp;
1477 switch (kiocb->ki_opcode) {
1478 case IOCB_CMD_PREAD:
1480 if (unlikely(!(file->f_mode & FMODE_READ)))
1483 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1486 ret = security_file_permission(file, MAY_READ);
1489 ret = aio_setup_single_vector(kiocb);
1493 if (file->f_op->aio_read)
1494 kiocb->ki_retry = aio_rw_vect_retry;
1496 case IOCB_CMD_PWRITE:
1498 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1501 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1504 ret = security_file_permission(file, MAY_WRITE);
1507 ret = aio_setup_single_vector(kiocb);
1511 if (file->f_op->aio_write)
1512 kiocb->ki_retry = aio_rw_vect_retry;
1514 case IOCB_CMD_PREADV:
1516 if (unlikely(!(file->f_mode & FMODE_READ)))
1518 ret = security_file_permission(file, MAY_READ);
1521 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1525 if (file->f_op->aio_read)
1526 kiocb->ki_retry = aio_rw_vect_retry;
1528 case IOCB_CMD_PWRITEV:
1530 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1532 ret = security_file_permission(file, MAY_WRITE);
1535 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1539 if (file->f_op->aio_write)
1540 kiocb->ki_retry = aio_rw_vect_retry;
1542 case IOCB_CMD_FDSYNC:
1544 if (file->f_op->aio_fsync)
1545 kiocb->ki_retry = aio_fdsync;
1547 case IOCB_CMD_FSYNC:
1549 if (file->f_op->aio_fsync)
1550 kiocb->ki_retry = aio_fsync;
1553 dprintk("EINVAL: io_submit: no operation provided\n");
1557 if (!kiocb->ki_retry)
1563 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1564 struct iocb *iocb, struct kiocb_batch *batch,
1571 /* enforce forwards compatibility on users */
1572 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1573 pr_debug("EINVAL: io_submit: reserve field set\n");
1577 /* prevent overflows */
1579 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1580 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1581 ((ssize_t)iocb->aio_nbytes < 0)
1583 pr_debug("EINVAL: io_submit: overflow check\n");
1587 file = fget(iocb->aio_fildes);
1588 if (unlikely(!file))
1591 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1592 if (unlikely(!req)) {
1596 req->ki_filp = file;
1597 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1599 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1600 * instance of the file* now. The file descriptor must be
1601 * an eventfd() fd, and will be signaled for each completed
1602 * event using the eventfd_signal() function.
1604 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1605 if (IS_ERR(req->ki_eventfd)) {
1606 ret = PTR_ERR(req->ki_eventfd);
1607 req->ki_eventfd = NULL;
1612 ret = put_user(req->ki_key, &user_iocb->aio_key);
1613 if (unlikely(ret)) {
1614 dprintk("EFAULT: aio_key\n");
1618 req->ki_obj.user = user_iocb;
1619 req->ki_user_data = iocb->aio_data;
1620 req->ki_pos = iocb->aio_offset;
1622 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1623 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1624 req->ki_opcode = iocb->aio_lio_opcode;
1626 ret = aio_setup_iocb(req, compat);
1631 spin_lock_irq(&ctx->ctx_lock);
1633 * We could have raced with io_destroy() and are currently holding a
1634 * reference to ctx which should be destroyed. We cannot submit IO
1635 * since ctx gets freed as soon as io_submit() puts its reference. The
1636 * check here is reliable: io_destroy() sets ctx->dead before waiting
1637 * for outstanding IO and the barrier between these two is realized by
1638 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1639 * increment ctx->reqs_active before checking for ctx->dead and the
1640 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1641 * don't see ctx->dead set here, io_destroy() waits for our IO to
1645 spin_unlock_irq(&ctx->ctx_lock);
1650 if (!list_empty(&ctx->run_list)) {
1651 /* drain the run list */
1652 while (__aio_run_iocbs(ctx))
1655 spin_unlock_irq(&ctx->ctx_lock);
1657 aio_put_req(req); /* drop extra ref to req */
1661 aio_put_req(req); /* drop extra ref to req */
1662 aio_put_req(req); /* drop i/o ref to req */
1666 long do_io_submit(aio_context_t ctx_id, long nr,
1667 struct iocb __user *__user *iocbpp, bool compat)
1672 struct blk_plug plug;
1673 struct kiocb_batch batch;
1675 if (unlikely(nr < 0))
1678 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1679 nr = LONG_MAX/sizeof(*iocbpp);
1681 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1684 ctx = lookup_ioctx(ctx_id);
1685 if (unlikely(!ctx)) {
1686 pr_debug("EINVAL: io_submit: invalid context id\n");
1690 kiocb_batch_init(&batch, nr);
1692 blk_start_plug(&plug);
1695 * AKPM: should this return a partial result if some of the IOs were
1696 * successfully submitted?
1698 for (i=0; i<nr; i++) {
1699 struct iocb __user *user_iocb;
1702 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1707 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1712 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1716 blk_finish_plug(&plug);
1718 kiocb_batch_free(ctx, &batch);
1724 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1725 * the number of iocbs queued. May return -EINVAL if the aio_context
1726 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1727 * *iocbpp[0] is not properly initialized, if the operation specified
1728 * is invalid for the file descriptor in the iocb. May fail with
1729 * -EFAULT if any of the data structures point to invalid data. May
1730 * fail with -EBADF if the file descriptor specified in the first
1731 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1732 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1733 * fail with -ENOSYS if not implemented.
1735 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1736 struct iocb __user * __user *, iocbpp)
1738 return do_io_submit(ctx_id, nr, iocbpp, 0);
1742 * Finds a given iocb for cancellation.
1744 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1747 struct list_head *pos;
1749 assert_spin_locked(&ctx->ctx_lock);
1751 /* TODO: use a hash or array, this sucks. */
1752 list_for_each(pos, &ctx->active_reqs) {
1753 struct kiocb *kiocb = list_kiocb(pos);
1754 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1761 * Attempts to cancel an iocb previously passed to io_submit. If
1762 * the operation is successfully cancelled, the resulting event is
1763 * copied into the memory pointed to by result without being placed
1764 * into the completion queue and 0 is returned. May fail with
1765 * -EFAULT if any of the data structures pointed to are invalid.
1766 * May fail with -EINVAL if aio_context specified by ctx_id is
1767 * invalid. May fail with -EAGAIN if the iocb specified was not
1768 * cancelled. Will fail with -ENOSYS if not implemented.
1770 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1771 struct io_event __user *, result)
1773 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1775 struct kiocb *kiocb;
1779 ret = get_user(key, &iocb->aio_key);
1783 ctx = lookup_ioctx(ctx_id);
1787 spin_lock_irq(&ctx->ctx_lock);
1789 kiocb = lookup_kiocb(ctx, iocb, key);
1790 if (kiocb && kiocb->ki_cancel) {
1791 cancel = kiocb->ki_cancel;
1793 kiocbSetCancelled(kiocb);
1796 spin_unlock_irq(&ctx->ctx_lock);
1798 if (NULL != cancel) {
1799 struct io_event tmp;
1800 pr_debug("calling cancel\n");
1801 memset(&tmp, 0, sizeof(tmp));
1802 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1803 tmp.data = kiocb->ki_user_data;
1804 ret = cancel(kiocb, &tmp);
1806 /* Cancellation succeeded -- copy the result
1807 * into the user's buffer.
1809 if (copy_to_user(result, &tmp, sizeof(tmp)))
1821 * Attempts to read at least min_nr events and up to nr events from
1822 * the completion queue for the aio_context specified by ctx_id. If
1823 * it succeeds, the number of read events is returned. May fail with
1824 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1825 * out of range, if timeout is out of range. May fail with -EFAULT
1826 * if any of the memory specified is invalid. May return 0 or
1827 * < min_nr if the timeout specified by timeout has elapsed
1828 * before sufficient events are available, where timeout == NULL
1829 * specifies an infinite timeout. Note that the timeout pointed to by
1830 * timeout is relative and will be updated if not NULL and the
1831 * operation blocks. Will fail with -ENOSYS if not implemented.
1833 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1836 struct io_event __user *, events,
1837 struct timespec __user *, timeout)
1839 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1842 if (likely(ioctx)) {
1843 if (likely(min_nr <= nr && min_nr >= 0))
1844 ret = read_events(ioctx, min_nr, nr, events, timeout);
1848 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);