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/module.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 #include <linux/poll.h>
44 #include <linux/anon_inodes.h>
48 #define dprintk printk
50 #define dprintk(x...) do { ; } while (0)
53 /*------ sysctl variables----*/
54 static DEFINE_SPINLOCK(aio_nr_lock);
55 unsigned long aio_nr; /* current system wide number of aio requests */
56 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
57 /*----end sysctl variables---*/
59 static struct kmem_cache *kiocb_cachep;
60 static struct kmem_cache *kioctx_cachep;
62 static struct workqueue_struct *aio_wq;
64 /* Used for rare fput completion. */
65 static void aio_fput_routine(struct work_struct *);
66 static DECLARE_WORK(fput_work, aio_fput_routine);
68 static DEFINE_SPINLOCK(fput_lock);
69 static LIST_HEAD(fput_head);
71 static void aio_kick_handler(struct work_struct *);
72 static void aio_queue_work(struct kioctx *);
75 * Creates the slab caches used by the aio routines, panic on
76 * failure as this is done early during the boot sequence.
78 static int __init aio_setup(void)
80 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
81 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
83 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
86 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
90 __initcall(aio_setup);
92 static void aio_free_ring(struct kioctx *ctx)
94 struct aio_ring_info *info = &ctx->ring_info;
97 for (i=0; i<info->nr_pages; i++)
98 put_page(info->ring_pages[i]);
100 if (info->mmap_size) {
101 down_write(&ctx->mm->mmap_sem);
102 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
103 up_write(&ctx->mm->mmap_sem);
106 if (info->ring_pages && info->ring_pages != info->internal_pages)
107 kfree(info->ring_pages);
108 info->ring_pages = NULL;
112 static int aio_setup_ring(struct kioctx *ctx)
114 struct aio_ring *ring;
115 struct aio_ring_info *info = &ctx->ring_info;
116 unsigned nr_events = ctx->max_reqs;
120 /* Compensate for the ring buffer's head/tail overlap entry */
121 nr_events += 2; /* 1 is required, 2 for good luck */
123 size = sizeof(struct aio_ring);
124 size += sizeof(struct io_event) * nr_events;
125 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
130 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
133 info->ring_pages = info->internal_pages;
134 if (nr_pages > AIO_RING_PAGES) {
135 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
136 if (!info->ring_pages)
140 info->mmap_size = nr_pages * PAGE_SIZE;
141 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
142 down_write(&ctx->mm->mmap_sem);
143 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
144 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
146 if (IS_ERR((void *)info->mmap_base)) {
147 up_write(&ctx->mm->mmap_sem);
153 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
154 info->nr_pages = get_user_pages(current, ctx->mm,
155 info->mmap_base, nr_pages,
156 1, 0, info->ring_pages, NULL);
157 up_write(&ctx->mm->mmap_sem);
159 if (unlikely(info->nr_pages != nr_pages)) {
164 ctx->user_id = info->mmap_base;
166 info->nr = nr_events; /* trusted copy */
168 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
169 ring->nr = nr_events; /* user copy */
170 ring->id = ctx->user_id;
171 ring->head = ring->tail = 0;
172 ring->magic = AIO_RING_MAGIC;
173 ring->compat_features = AIO_RING_COMPAT_FEATURES;
174 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
175 ring->header_length = sizeof(struct aio_ring);
176 kunmap_atomic(ring, KM_USER0);
182 /* aio_ring_event: returns a pointer to the event at the given index from
183 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
185 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
186 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
187 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
189 #define aio_ring_event(info, nr, km) ({ \
190 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
191 struct io_event *__event; \
192 __event = kmap_atomic( \
193 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
194 __event += pos % AIO_EVENTS_PER_PAGE; \
198 #define put_aio_ring_event(event, km) do { \
199 struct io_event *__event = (event); \
201 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
204 static void ctx_rcu_free(struct rcu_head *head)
206 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
207 unsigned nr_events = ctx->max_reqs;
209 kmem_cache_free(kioctx_cachep, ctx);
212 spin_lock(&aio_nr_lock);
213 BUG_ON(aio_nr - nr_events > aio_nr);
215 spin_unlock(&aio_nr_lock);
220 * Called when the last user of an aio context has gone away,
221 * and the struct needs to be freed.
223 static void __put_ioctx(struct kioctx *ctx)
225 BUG_ON(ctx->reqs_active);
227 cancel_delayed_work(&ctx->wq);
228 cancel_work_sync(&ctx->wq.work);
232 pr_debug("__put_ioctx: freeing %p\n", ctx);
233 call_rcu(&ctx->rcu_head, ctx_rcu_free);
236 static inline void get_ioctx(struct kioctx *kioctx)
238 BUG_ON(atomic_read(&kioctx->users) <= 0);
239 atomic_inc(&kioctx->users);
242 static inline int try_get_ioctx(struct kioctx *kioctx)
244 return atomic_inc_not_zero(&kioctx->users);
247 static inline void put_ioctx(struct kioctx *kioctx)
249 BUG_ON(atomic_read(&kioctx->users) <= 0);
250 if (unlikely(atomic_dec_and_test(&kioctx->users)))
255 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
257 static struct kioctx *ioctx_alloc(unsigned nr_events)
259 struct mm_struct *mm;
263 /* Prevent overflows */
264 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
265 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
266 pr_debug("ENOMEM: nr_events too high\n");
267 return ERR_PTR(-EINVAL);
270 if ((unsigned long)nr_events > aio_max_nr)
271 return ERR_PTR(-EAGAIN);
273 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
275 return ERR_PTR(-ENOMEM);
277 ctx->max_reqs = nr_events;
278 mm = ctx->mm = current->mm;
279 atomic_inc(&mm->mm_count);
281 atomic_set(&ctx->users, 1);
282 spin_lock_init(&ctx->ctx_lock);
283 spin_lock_init(&ctx->ring_info.ring_lock);
284 init_waitqueue_head(&ctx->wait);
286 INIT_LIST_HEAD(&ctx->active_reqs);
287 INIT_LIST_HEAD(&ctx->run_list);
288 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
290 if (aio_setup_ring(ctx) < 0)
293 /* limit the number of system wide aios */
295 spin_lock_bh(&aio_nr_lock);
296 if (aio_nr + nr_events > aio_max_nr ||
297 aio_nr + nr_events < aio_nr)
300 aio_nr += ctx->max_reqs;
301 spin_unlock_bh(&aio_nr_lock);
302 if (ctx->max_reqs || did_sync)
305 /* wait for rcu callbacks to have completed before giving up */
308 ctx->max_reqs = nr_events;
311 if (ctx->max_reqs == 0)
314 /* now link into global list. */
315 spin_lock(&mm->ioctx_lock);
316 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
317 spin_unlock(&mm->ioctx_lock);
319 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
320 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
325 return ERR_PTR(-EAGAIN);
329 kmem_cache_free(kioctx_cachep, ctx);
330 ctx = ERR_PTR(-ENOMEM);
332 dprintk("aio: error allocating ioctx %p\n", ctx);
337 * Cancels all outstanding aio requests on an aio context. Used
338 * when the processes owning a context have all exited to encourage
339 * the rapid destruction of the kioctx.
341 static void aio_cancel_all(struct kioctx *ctx)
343 int (*cancel)(struct kiocb *, struct io_event *);
345 spin_lock_irq(&ctx->ctx_lock);
347 while (!list_empty(&ctx->active_reqs)) {
348 struct list_head *pos = ctx->active_reqs.next;
349 struct kiocb *iocb = list_kiocb(pos);
350 list_del_init(&iocb->ki_list);
351 cancel = iocb->ki_cancel;
352 kiocbSetCancelled(iocb);
355 spin_unlock_irq(&ctx->ctx_lock);
357 spin_lock_irq(&ctx->ctx_lock);
360 spin_unlock_irq(&ctx->ctx_lock);
363 static void wait_for_all_aios(struct kioctx *ctx)
365 struct task_struct *tsk = current;
366 DECLARE_WAITQUEUE(wait, tsk);
368 spin_lock_irq(&ctx->ctx_lock);
369 if (!ctx->reqs_active)
372 add_wait_queue(&ctx->wait, &wait);
373 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
374 while (ctx->reqs_active) {
375 spin_unlock_irq(&ctx->ctx_lock);
377 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
378 spin_lock_irq(&ctx->ctx_lock);
380 __set_task_state(tsk, TASK_RUNNING);
381 remove_wait_queue(&ctx->wait, &wait);
384 spin_unlock_irq(&ctx->ctx_lock);
387 /* wait_on_sync_kiocb:
388 * Waits on the given sync kiocb to complete.
390 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
392 while (iocb->ki_users) {
393 set_current_state(TASK_UNINTERRUPTIBLE);
398 __set_current_state(TASK_RUNNING);
399 return iocb->ki_user_data;
401 EXPORT_SYMBOL(wait_on_sync_kiocb);
403 /* exit_aio: called when the last user of mm goes away. At this point,
404 * there is no way for any new requests to be submited or any of the
405 * io_* syscalls to be called on the context. However, there may be
406 * outstanding requests which hold references to the context; as they
407 * go away, they will call put_ioctx and release any pinned memory
408 * associated with the request (held via struct page * references).
410 void exit_aio(struct mm_struct *mm)
414 while (!hlist_empty(&mm->ioctx_list)) {
415 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
416 hlist_del_rcu(&ctx->list);
420 wait_for_all_aios(ctx);
422 * Ensure we don't leave the ctx on the aio_wq
424 cancel_work_sync(&ctx->wq.work);
426 if (1 != atomic_read(&ctx->users))
428 "exit_aio:ioctx still alive: %d %d %d\n",
429 atomic_read(&ctx->users), ctx->dead,
436 * Allocate a slot for an aio request. Increments the users count
437 * of the kioctx so that the kioctx stays around until all requests are
438 * complete. Returns NULL if no requests are free.
440 * Returns with kiocb->users set to 2. The io submit code path holds
441 * an extra reference while submitting the i/o.
442 * This prevents races between the aio code path referencing the
443 * req (after submitting it) and aio_complete() freeing the req.
445 static struct kiocb *__aio_get_req(struct kioctx *ctx)
447 struct kiocb *req = NULL;
449 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
457 req->ki_cancel = NULL;
458 req->ki_retry = NULL;
461 req->ki_iovec = NULL;
462 INIT_LIST_HEAD(&req->ki_run_list);
463 req->ki_eventfd = NULL;
469 * struct kiocb's are allocated in batches to reduce the number of
470 * times the ctx lock is acquired and released.
472 #define KIOCB_BATCH_SIZE 32L
474 struct list_head head;
475 long count; /* number of requests left to allocate */
478 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
480 INIT_LIST_HEAD(&batch->head);
481 batch->count = total;
484 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
486 struct kiocb *req, *n;
488 if (list_empty(&batch->head))
491 spin_lock_irq(&ctx->ctx_lock);
492 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
493 list_del(&req->ki_batch);
494 list_del(&req->ki_list);
495 kmem_cache_free(kiocb_cachep, req);
498 spin_unlock_irq(&ctx->ctx_lock);
502 * Allocate a batch of kiocbs. This avoids taking and dropping the
503 * context lock a lot during setup.
505 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
507 unsigned short allocated, to_alloc;
509 bool called_fput = false;
510 struct kiocb *req, *n;
511 struct aio_ring *ring;
513 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
514 for (allocated = 0; allocated < to_alloc; allocated++) {
515 req = __aio_get_req(ctx);
517 /* allocation failed, go with what we've got */
519 list_add(&req->ki_batch, &batch->head);
526 spin_lock_irq(&ctx->ctx_lock);
527 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
529 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
531 if (avail == 0 && !called_fput) {
533 * Handle a potential starvation case. It is possible that
534 * we hold the last reference on a struct file, causing us
535 * to delay the final fput to non-irq context. In this case,
536 * ctx->reqs_active is artificially high. Calling the fput
537 * routine here may free up a slot in the event completion
538 * ring, allowing this allocation to succeed.
541 spin_unlock_irq(&ctx->ctx_lock);
542 aio_fput_routine(NULL);
547 if (avail < allocated) {
548 /* Trim back the number of requests. */
549 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
550 list_del(&req->ki_batch);
551 kmem_cache_free(kiocb_cachep, req);
552 if (--allocated <= avail)
557 batch->count -= allocated;
558 list_for_each_entry(req, &batch->head, ki_batch) {
559 list_add(&req->ki_list, &ctx->active_reqs);
564 spin_unlock_irq(&ctx->ctx_lock);
570 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
571 struct kiocb_batch *batch)
575 if (list_empty(&batch->head))
576 if (kiocb_batch_refill(ctx, batch) == 0)
578 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
579 list_del(&req->ki_batch);
583 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
585 assert_spin_locked(&ctx->ctx_lock);
587 if (req->ki_eventfd != NULL)
588 eventfd_ctx_put(req->ki_eventfd);
591 if (req->ki_iovec != &req->ki_inline_vec)
592 kfree(req->ki_iovec);
593 kmem_cache_free(kiocb_cachep, req);
596 if (unlikely(!ctx->reqs_active && ctx->dead))
597 wake_up_all(&ctx->wait);
600 static void aio_fput_routine(struct work_struct *data)
602 spin_lock_irq(&fput_lock);
603 while (likely(!list_empty(&fput_head))) {
604 struct kiocb *req = list_kiocb(fput_head.next);
605 struct kioctx *ctx = req->ki_ctx;
607 list_del(&req->ki_list);
608 spin_unlock_irq(&fput_lock);
610 /* Complete the fput(s) */
611 if (req->ki_filp != NULL)
614 /* Link the iocb into the context's free list */
615 spin_lock_irq(&ctx->ctx_lock);
616 really_put_req(ctx, req);
617 spin_unlock_irq(&ctx->ctx_lock);
620 spin_lock_irq(&fput_lock);
622 spin_unlock_irq(&fput_lock);
626 * Returns true if this put was the last user of the request.
628 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
630 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
631 req, atomic_long_read(&req->ki_filp->f_count));
633 assert_spin_locked(&ctx->ctx_lock);
636 BUG_ON(req->ki_users < 0);
637 if (likely(req->ki_users))
639 list_del(&req->ki_list); /* remove from active_reqs */
640 req->ki_cancel = NULL;
641 req->ki_retry = NULL;
644 * Try to optimize the aio and eventfd file* puts, by avoiding to
645 * schedule work in case it is not final fput() time. In normal cases,
646 * we would not be holding the last reference to the file*, so
647 * this function will be executed w/out any aio kthread wakeup.
649 if (unlikely(!fput_atomic(req->ki_filp))) {
651 spin_lock(&fput_lock);
652 list_add(&req->ki_list, &fput_head);
653 spin_unlock(&fput_lock);
654 schedule_work(&fput_work);
657 really_put_req(ctx, req);
663 * Returns true if this put was the last user of the kiocb,
664 * false if the request is still in use.
666 int aio_put_req(struct kiocb *req)
668 struct kioctx *ctx = req->ki_ctx;
670 spin_lock_irq(&ctx->ctx_lock);
671 ret = __aio_put_req(ctx, req);
672 spin_unlock_irq(&ctx->ctx_lock);
675 EXPORT_SYMBOL(aio_put_req);
677 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
679 struct mm_struct *mm = current->mm;
680 struct kioctx *ctx, *ret = NULL;
681 struct hlist_node *n;
685 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
687 * RCU protects us against accessing freed memory but
688 * we have to be careful not to get a reference when the
689 * reference count already dropped to 0 (ctx->dead test
690 * is unreliable because of races).
692 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
703 * Queue up a kiocb to be retried. Assumes that the kiocb
704 * has already been marked as kicked, and places it on
705 * the retry run list for the corresponding ioctx, if it
706 * isn't already queued. Returns 1 if it actually queued
707 * the kiocb (to tell the caller to activate the work
708 * queue to process it), or 0, if it found that it was
711 static inline int __queue_kicked_iocb(struct kiocb *iocb)
713 struct kioctx *ctx = iocb->ki_ctx;
715 assert_spin_locked(&ctx->ctx_lock);
717 if (list_empty(&iocb->ki_run_list)) {
718 list_add_tail(&iocb->ki_run_list,
726 * This is the core aio execution routine. It is
727 * invoked both for initial i/o submission and
728 * subsequent retries via the aio_kick_handler.
729 * Expects to be invoked with iocb->ki_ctx->lock
730 * already held. The lock is released and reacquired
731 * as needed during processing.
733 * Calls the iocb retry method (already setup for the
734 * iocb on initial submission) for operation specific
735 * handling, but takes care of most of common retry
736 * execution details for a given iocb. The retry method
737 * needs to be non-blocking as far as possible, to avoid
738 * holding up other iocbs waiting to be serviced by the
739 * retry kernel thread.
741 * The trickier parts in this code have to do with
742 * ensuring that only one retry instance is in progress
743 * for a given iocb at any time. Providing that guarantee
744 * simplifies the coding of individual aio operations as
745 * it avoids various potential races.
747 static ssize_t aio_run_iocb(struct kiocb *iocb)
749 struct kioctx *ctx = iocb->ki_ctx;
750 ssize_t (*retry)(struct kiocb *);
753 if (!(retry = iocb->ki_retry)) {
754 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
759 * We don't want the next retry iteration for this
760 * operation to start until this one has returned and
761 * updated the iocb state. However, wait_queue functions
762 * can trigger a kick_iocb from interrupt context in the
763 * meantime, indicating that data is available for the next
764 * iteration. We want to remember that and enable the
765 * next retry iteration _after_ we are through with
768 * So, in order to be able to register a "kick", but
769 * prevent it from being queued now, we clear the kick
770 * flag, but make the kick code *think* that the iocb is
771 * still on the run list until we are actually done.
772 * When we are done with this iteration, we check if
773 * the iocb was kicked in the meantime and if so, queue
777 kiocbClearKicked(iocb);
780 * This is so that aio_complete knows it doesn't need to
781 * pull the iocb off the run list (We can't just call
782 * INIT_LIST_HEAD because we don't want a kick_iocb to
783 * queue this on the run list yet)
785 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
786 spin_unlock_irq(&ctx->ctx_lock);
788 /* Quit retrying if the i/o has been cancelled */
789 if (kiocbIsCancelled(iocb)) {
791 aio_complete(iocb, ret, 0);
792 /* must not access the iocb after this */
797 * Now we are all set to call the retry method in async
802 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
804 * There's no easy way to restart the syscall since other AIO's
805 * may be already running. Just fail this IO with EINTR.
807 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
808 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
810 aio_complete(iocb, ret, 0);
813 spin_lock_irq(&ctx->ctx_lock);
815 if (-EIOCBRETRY == ret) {
817 * OK, now that we are done with this iteration
818 * and know that there is more left to go,
819 * this is where we let go so that a subsequent
820 * "kick" can start the next iteration
823 /* will make __queue_kicked_iocb succeed from here on */
824 INIT_LIST_HEAD(&iocb->ki_run_list);
825 /* we must queue the next iteration ourselves, if it
826 * has already been kicked */
827 if (kiocbIsKicked(iocb)) {
828 __queue_kicked_iocb(iocb);
831 * __queue_kicked_iocb will always return 1 here, because
832 * iocb->ki_run_list is empty at this point so it should
833 * be safe to unconditionally queue the context into the
844 * Process all pending retries queued on the ioctx
846 * Assumes it is operating within the aio issuer's mm
849 static int __aio_run_iocbs(struct kioctx *ctx)
852 struct list_head run_list;
854 assert_spin_locked(&ctx->ctx_lock);
856 list_replace_init(&ctx->run_list, &run_list);
857 while (!list_empty(&run_list)) {
858 iocb = list_entry(run_list.next, struct kiocb,
860 list_del(&iocb->ki_run_list);
862 * Hold an extra reference while retrying i/o.
864 iocb->ki_users++; /* grab extra reference */
866 __aio_put_req(ctx, iocb);
868 if (!list_empty(&ctx->run_list))
873 static void aio_queue_work(struct kioctx * ctx)
875 unsigned long timeout;
877 * if someone is waiting, get the work started right
878 * away, otherwise, use a longer delay
881 if (waitqueue_active(&ctx->wait))
885 queue_delayed_work(aio_wq, &ctx->wq, timeout);
890 * Process all pending retries queued on the ioctx
891 * run list, and keep running them until the list
893 * Assumes it is operating within the aio issuer's mm context.
895 static inline void aio_run_all_iocbs(struct kioctx *ctx)
897 spin_lock_irq(&ctx->ctx_lock);
898 while (__aio_run_iocbs(ctx))
900 spin_unlock_irq(&ctx->ctx_lock);
905 * Work queue handler triggered to process pending
906 * retries on an ioctx. Takes on the aio issuer's
907 * mm context before running the iocbs, so that
908 * copy_xxx_user operates on the issuer's address
910 * Run on aiod's context.
912 static void aio_kick_handler(struct work_struct *work)
914 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
915 mm_segment_t oldfs = get_fs();
916 struct mm_struct *mm;
921 spin_lock_irq(&ctx->ctx_lock);
922 requeue =__aio_run_iocbs(ctx);
924 spin_unlock_irq(&ctx->ctx_lock);
928 * we're in a worker thread already, don't use queue_delayed_work,
931 queue_delayed_work(aio_wq, &ctx->wq, 0);
936 * Called by kick_iocb to queue the kiocb for retry
937 * and if required activate the aio work queue to process
940 static void try_queue_kicked_iocb(struct kiocb *iocb)
942 struct kioctx *ctx = iocb->ki_ctx;
946 spin_lock_irqsave(&ctx->ctx_lock, flags);
947 /* set this inside the lock so that we can't race with aio_run_iocb()
948 * testing it and putting the iocb on the run list under the lock */
949 if (!kiocbTryKick(iocb))
950 run = __queue_kicked_iocb(iocb);
951 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
958 * Called typically from a wait queue callback context
959 * to trigger a retry of the iocb.
960 * The retry is usually executed by aio workqueue
961 * threads (See aio_kick_handler).
963 void kick_iocb(struct kiocb *iocb)
965 /* sync iocbs are easy: they can only ever be executing from a
967 if (is_sync_kiocb(iocb)) {
968 kiocbSetKicked(iocb);
969 wake_up_process(iocb->ki_obj.tsk);
973 try_queue_kicked_iocb(iocb);
975 EXPORT_SYMBOL(kick_iocb);
978 * Called when the io request on the given iocb is complete.
979 * Returns true if this is the last user of the request. The
980 * only other user of the request can be the cancellation code.
982 int aio_complete(struct kiocb *iocb, long res, long res2)
984 struct kioctx *ctx = iocb->ki_ctx;
985 struct aio_ring_info *info;
986 struct aio_ring *ring;
987 struct io_event *event;
993 * Special case handling for sync iocbs:
994 * - events go directly into the iocb for fast handling
995 * - the sync task with the iocb in its stack holds the single iocb
996 * ref, no other paths have a way to get another ref
997 * - the sync task helpfully left a reference to itself in the iocb
999 if (is_sync_kiocb(iocb)) {
1000 BUG_ON(iocb->ki_users != 1);
1001 iocb->ki_user_data = res;
1003 wake_up_process(iocb->ki_obj.tsk);
1007 info = &ctx->ring_info;
1009 /* add a completion event to the ring buffer.
1010 * must be done holding ctx->ctx_lock to prevent
1011 * other code from messing with the tail
1012 * pointer since we might be called from irq
1015 spin_lock_irqsave(&ctx->ctx_lock, flags);
1017 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
1018 list_del_init(&iocb->ki_run_list);
1021 * cancelled requests don't get events, userland was given one
1022 * when the event got cancelled.
1024 if (kiocbIsCancelled(iocb))
1027 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
1030 event = aio_ring_event(info, tail, KM_IRQ0);
1031 if (++tail >= info->nr)
1034 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1035 event->data = iocb->ki_user_data;
1039 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1040 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1043 /* after flagging the request as done, we
1044 * must never even look at it again
1046 smp_wmb(); /* make event visible before updating tail */
1051 put_aio_ring_event(event, KM_IRQ0);
1052 kunmap_atomic(ring, KM_IRQ1);
1054 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1057 * Check if the user asked us to deliver the result through an
1058 * eventfd. The eventfd_signal() function is safe to be called
1061 if (iocb->ki_eventfd != NULL)
1062 eventfd_signal(iocb->ki_eventfd, 1);
1065 /* everything turned out well, dispose of the aiocb. */
1066 ret = __aio_put_req(ctx, iocb);
1069 * We have to order our ring_info tail store above and test
1070 * of the wait list below outside the wait lock. This is
1071 * like in wake_up_bit() where clearing a bit has to be
1072 * ordered with the unlocked test.
1076 if (waitqueue_active(&ctx->wait))
1077 wake_up(&ctx->wait);
1080 if (ctx->file && waitqueue_active(&ctx->poll_wait))
1081 wake_up(&ctx->poll_wait);
1084 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1087 EXPORT_SYMBOL(aio_complete);
1090 * Pull an event off of the ioctx's event ring. Returns the number of
1091 * events fetched (0 or 1 ;-)
1092 * If ent parameter is 0, just returns the number of events that would
1094 * FIXME: make this use cmpxchg.
1095 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1097 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1099 struct aio_ring_info *info = &ioctx->ring_info;
1100 struct aio_ring *ring;
1104 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1105 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1106 (unsigned long)ring->head, (unsigned long)ring->tail,
1107 (unsigned long)ring->nr);
1109 if (ring->head == ring->tail)
1112 spin_lock(&info->ring_lock);
1114 head = ring->head % info->nr;
1115 if (head != ring->tail) {
1116 if (ent) { /* event requested */
1117 struct io_event *evp =
1118 aio_ring_event(info, head, KM_USER1);
1120 head = (head + 1) % info->nr;
1121 /* finish reading the event before updatng the head */
1125 put_aio_ring_event(evp, KM_USER1);
1126 } else /* only need to know availability */
1129 spin_unlock(&info->ring_lock);
1132 kunmap_atomic(ring, KM_USER0);
1133 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1134 (unsigned long)ring->head, (unsigned long)ring->tail);
1138 struct aio_timeout {
1139 struct timer_list timer;
1141 struct task_struct *p;
1144 static void timeout_func(unsigned long data)
1146 struct aio_timeout *to = (struct aio_timeout *)data;
1149 wake_up_process(to->p);
1152 static inline void init_timeout(struct aio_timeout *to)
1154 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1159 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1160 const struct timespec *ts)
1162 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1163 if (time_after(to->timer.expires, jiffies))
1164 add_timer(&to->timer);
1169 static inline void clear_timeout(struct aio_timeout *to)
1171 del_singleshot_timer_sync(&to->timer);
1174 static int read_events(struct kioctx *ctx,
1175 long min_nr, long nr,
1176 struct io_event __user *event,
1177 struct timespec __user *timeout)
1179 long start_jiffies = jiffies;
1180 struct task_struct *tsk = current;
1181 DECLARE_WAITQUEUE(wait, tsk);
1184 struct io_event ent;
1185 struct aio_timeout to;
1188 /* needed to zero any padding within an entry (there shouldn't be
1189 * any, but C is fun!
1191 memset(&ent, 0, sizeof(ent));
1194 while (likely(i < nr)) {
1195 ret = aio_read_evt(ctx, &ent);
1196 if (unlikely(ret <= 0))
1199 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1200 ent.data, ent.obj, ent.res, ent.res2);
1202 /* Could we split the check in two? */
1204 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1205 dprintk("aio: lost an event due to EFAULT.\n");
1210 /* Good, event copied to userland, update counts. */
1222 /* racey check, but it gets redone */
1223 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1225 aio_run_all_iocbs(ctx);
1233 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1236 set_timeout(start_jiffies, &to, &ts);
1239 while (likely(i < nr)) {
1240 add_wait_queue_exclusive(&ctx->wait, &wait);
1242 set_task_state(tsk, TASK_INTERRUPTIBLE);
1243 ret = aio_read_evt(ctx, &ent);
1248 if (unlikely(ctx->dead)) {
1252 if (to.timed_out) /* Only check after read evt */
1254 /* Try to only show up in io wait if there are ops
1256 if (ctx->reqs_active)
1260 if (signal_pending(tsk)) {
1264 /*ret = aio_read_evt(ctx, &ent);*/
1267 set_task_state(tsk, TASK_RUNNING);
1268 remove_wait_queue(&ctx->wait, &wait);
1270 if (unlikely(ret <= 0))
1274 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1275 dprintk("aio: lost an event due to EFAULT.\n");
1279 /* Good, event copied to userland, update counts. */
1287 destroy_timer_on_stack(&to.timer);
1291 /* Take an ioctx and remove it from the list of ioctx's. Protects
1292 * against races with itself via ->dead.
1294 static void io_destroy(struct kioctx *ioctx)
1296 struct mm_struct *mm = current->mm;
1299 /* delete the entry from the list is someone else hasn't already */
1300 spin_lock(&mm->ioctx_lock);
1301 was_dead = ioctx->dead;
1303 hlist_del_rcu(&ioctx->list);
1304 spin_unlock(&mm->ioctx_lock);
1306 dprintk("aio_release(%p)\n", ioctx);
1307 if (likely(!was_dead))
1308 put_ioctx(ioctx); /* twice for the list */
1310 aio_cancel_all(ioctx);
1311 wait_for_all_aios(ioctx);
1313 /* forget the poll file, but it's up to the user to close it */
1316 ioctx->file->private_data = 0;
1322 * Wake up any waiters. The setting of ctx->dead must be seen
1323 * by other CPUs at this point. Right now, we rely on the
1324 * locking done by the above calls to ensure this consistency.
1326 wake_up_all(&ioctx->wait);
1327 put_ioctx(ioctx); /* once for the lookup */
1332 static int aio_queue_fd_close(struct inode *inode, struct file *file)
1334 struct kioctx *ioctx = file->private_data;
1336 file->private_data = 0;
1337 spin_lock_irq(&ioctx->ctx_lock);
1339 spin_unlock_irq(&ioctx->ctx_lock);
1345 static unsigned int aio_queue_fd_poll(struct file *file, poll_table *wait)
1346 { unsigned int pollflags = 0;
1347 struct kioctx *ioctx = file->private_data;
1351 spin_lock_irq(&ioctx->ctx_lock);
1352 /* Insert inside our poll wait queue */
1353 poll_wait(file, &ioctx->poll_wait, wait);
1355 /* Check our condition */
1356 if (aio_read_evt(ioctx, 0))
1357 pollflags = POLLIN | POLLRDNORM;
1358 spin_unlock_irq(&ioctx->ctx_lock);
1364 static const struct file_operations aioq_fops = {
1365 .release = aio_queue_fd_close,
1366 .poll = aio_queue_fd_poll
1370 * Create a file descriptor that can be used to poll the event queue.
1371 * Based on the excellent epoll code.
1374 static int make_aio_fd(struct kioctx *ioctx)
1379 fd = anon_inode_getfd("[aioq]", &aioq_fops, ioctx, 0);
1383 /* associate the file with the IO context */
1387 file->private_data = ioctx;
1389 init_waitqueue_head(&ioctx->poll_wait);
1395 * Create an aio_context capable of receiving at least nr_events.
1396 * ctxp must not point to an aio_context that already exists, and
1397 * must be initialized to 0 prior to the call. On successful
1398 * creation of the aio_context, *ctxp is filled in with the resulting
1399 * handle. May fail with -EINVAL if *ctxp is not initialized,
1400 * if the specified nr_events exceeds internal limits. May fail
1401 * with -EAGAIN if the specified nr_events exceeds the user's limit
1402 * of available events. May fail with -ENOMEM if insufficient kernel
1403 * resources are available. May fail with -EFAULT if an invalid
1404 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1407 * To request a selectable fd, the user context has to be initialized
1408 * to 1, instead of 0, and the return value is the fd.
1409 * This keeps the system call compatible, since a non-zero value
1410 * was not allowed so far.
1412 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1414 struct kioctx *ioctx = NULL;
1419 ret = get_user(ctx, ctxp);
1430 if (unlikely(ctx || nr_events == 0)) {
1431 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1436 ioctx = ioctx_alloc(nr_events);
1437 ret = PTR_ERR(ioctx);
1438 if (!IS_ERR(ioctx)) {
1439 ret = put_user(ioctx->user_id, ctxp);
1441 if (make_fd && ret >= 0)
1442 ret = make_aio_fd(ioctx);
1447 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1456 * Destroy the aio_context specified. May cancel any outstanding
1457 * AIOs and block on completion. Will fail with -ENOSYS if not
1458 * implemented. May fail with -EINVAL if the context pointed to
1461 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1463 struct kioctx *ioctx = lookup_ioctx(ctx);
1464 if (likely(NULL != ioctx)) {
1468 pr_debug("EINVAL: io_destroy: invalid context id\n");
1472 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1474 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1478 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1479 ssize_t this = min((ssize_t)iov->iov_len, ret);
1480 iov->iov_base += this;
1481 iov->iov_len -= this;
1482 iocb->ki_left -= this;
1484 if (iov->iov_len == 0) {
1490 /* the caller should not have done more io than what fit in
1491 * the remaining iovecs */
1492 BUG_ON(ret > 0 && iocb->ki_left == 0);
1495 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1497 struct file *file = iocb->ki_filp;
1498 struct address_space *mapping = file->f_mapping;
1499 struct inode *inode = mapping->host;
1500 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1501 unsigned long, loff_t);
1503 unsigned short opcode;
1505 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1506 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1507 rw_op = file->f_op->aio_read;
1508 opcode = IOCB_CMD_PREADV;
1510 rw_op = file->f_op->aio_write;
1511 opcode = IOCB_CMD_PWRITEV;
1514 /* This matches the pread()/pwrite() logic */
1515 if (iocb->ki_pos < 0)
1519 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1520 iocb->ki_nr_segs - iocb->ki_cur_seg,
1523 aio_advance_iovec(iocb, ret);
1525 /* retry all partial writes. retry partial reads as long as its a
1527 } while (ret > 0 && iocb->ki_left > 0 &&
1528 (opcode == IOCB_CMD_PWRITEV ||
1529 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1531 /* This means we must have transferred all that we could */
1532 /* No need to retry anymore */
1533 if ((ret == 0) || (iocb->ki_left == 0))
1534 ret = iocb->ki_nbytes - iocb->ki_left;
1536 /* If we managed to write some out we return that, rather than
1537 * the eventual error. */
1538 if (opcode == IOCB_CMD_PWRITEV
1539 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1540 && iocb->ki_nbytes - iocb->ki_left)
1541 ret = iocb->ki_nbytes - iocb->ki_left;
1546 static ssize_t aio_fdsync(struct kiocb *iocb)
1548 struct file *file = iocb->ki_filp;
1549 ssize_t ret = -EINVAL;
1551 if (file->f_op->aio_fsync)
1552 ret = file->f_op->aio_fsync(iocb, 1);
1556 static ssize_t aio_fsync(struct kiocb *iocb)
1558 struct file *file = iocb->ki_filp;
1559 ssize_t ret = -EINVAL;
1561 if (file->f_op->aio_fsync)
1562 ret = file->f_op->aio_fsync(iocb, 0);
1566 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1570 #ifdef CONFIG_COMPAT
1572 ret = compat_rw_copy_check_uvector(type,
1573 (struct compat_iovec __user *)kiocb->ki_buf,
1574 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1575 &kiocb->ki_iovec, 1);
1578 ret = rw_copy_check_uvector(type,
1579 (struct iovec __user *)kiocb->ki_buf,
1580 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1581 &kiocb->ki_iovec, 1);
1585 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1586 kiocb->ki_cur_seg = 0;
1587 /* ki_nbytes/left now reflect bytes instead of segs */
1588 kiocb->ki_nbytes = ret;
1589 kiocb->ki_left = ret;
1596 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1598 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1599 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1600 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1601 kiocb->ki_nr_segs = 1;
1602 kiocb->ki_cur_seg = 0;
1608 * Performs the initial checks and aio retry method
1609 * setup for the kiocb at the time of io submission.
1611 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1613 struct file *file = kiocb->ki_filp;
1616 switch (kiocb->ki_opcode) {
1617 case IOCB_CMD_PREAD:
1619 if (unlikely(!(file->f_mode & FMODE_READ)))
1622 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1625 ret = security_file_permission(file, MAY_READ);
1628 ret = aio_setup_single_vector(kiocb);
1632 if (file->f_op->aio_read)
1633 kiocb->ki_retry = aio_rw_vect_retry;
1635 case IOCB_CMD_PWRITE:
1637 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1640 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1643 ret = security_file_permission(file, MAY_WRITE);
1646 ret = aio_setup_single_vector(kiocb);
1650 if (file->f_op->aio_write)
1651 kiocb->ki_retry = aio_rw_vect_retry;
1653 case IOCB_CMD_PREADV:
1655 if (unlikely(!(file->f_mode & FMODE_READ)))
1657 ret = security_file_permission(file, MAY_READ);
1660 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1664 if (file->f_op->aio_read)
1665 kiocb->ki_retry = aio_rw_vect_retry;
1667 case IOCB_CMD_PWRITEV:
1669 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1671 ret = security_file_permission(file, MAY_WRITE);
1674 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1678 if (file->f_op->aio_write)
1679 kiocb->ki_retry = aio_rw_vect_retry;
1681 case IOCB_CMD_FDSYNC:
1683 if (file->f_op->aio_fsync)
1684 kiocb->ki_retry = aio_fdsync;
1686 case IOCB_CMD_FSYNC:
1688 if (file->f_op->aio_fsync)
1689 kiocb->ki_retry = aio_fsync;
1692 dprintk("EINVAL: io_submit: no operation provided\n");
1696 if (!kiocb->ki_retry)
1702 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1703 struct iocb *iocb, struct kiocb_batch *batch,
1710 /* enforce forwards compatibility on users */
1711 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1712 pr_debug("EINVAL: io_submit: reserve field set\n");
1716 /* prevent overflows */
1718 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1719 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1720 ((ssize_t)iocb->aio_nbytes < 0)
1722 pr_debug("EINVAL: io_submit: overflow check\n");
1726 file = fget(iocb->aio_fildes);
1727 if (unlikely(!file))
1730 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1731 if (unlikely(!req)) {
1735 req->ki_filp = file;
1736 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1738 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1739 * instance of the file* now. The file descriptor must be
1740 * an eventfd() fd, and will be signaled for each completed
1741 * event using the eventfd_signal() function.
1743 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1744 if (IS_ERR(req->ki_eventfd)) {
1745 ret = PTR_ERR(req->ki_eventfd);
1746 req->ki_eventfd = NULL;
1751 ret = put_user(req->ki_key, &user_iocb->aio_key);
1752 if (unlikely(ret)) {
1753 dprintk("EFAULT: aio_key\n");
1757 req->ki_obj.user = user_iocb;
1758 req->ki_user_data = iocb->aio_data;
1759 req->ki_pos = iocb->aio_offset;
1761 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1762 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1763 req->ki_opcode = iocb->aio_lio_opcode;
1765 ret = aio_setup_iocb(req, compat);
1770 spin_lock_irq(&ctx->ctx_lock);
1772 * We could have raced with io_destroy() and are currently holding a
1773 * reference to ctx which should be destroyed. We cannot submit IO
1774 * since ctx gets freed as soon as io_submit() puts its reference. The
1775 * check here is reliable: io_destroy() sets ctx->dead before waiting
1776 * for outstanding IO and the barrier between these two is realized by
1777 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1778 * increment ctx->reqs_active before checking for ctx->dead and the
1779 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1780 * don't see ctx->dead set here, io_destroy() waits for our IO to
1784 spin_unlock_irq(&ctx->ctx_lock);
1789 if (!list_empty(&ctx->run_list)) {
1790 /* drain the run list */
1791 while (__aio_run_iocbs(ctx))
1794 spin_unlock_irq(&ctx->ctx_lock);
1796 aio_put_req(req); /* drop extra ref to req */
1800 aio_put_req(req); /* drop extra ref to req */
1801 aio_put_req(req); /* drop i/o ref to req */
1805 long do_io_submit(aio_context_t ctx_id, long nr,
1806 struct iocb __user *__user *iocbpp, bool compat)
1811 struct blk_plug plug;
1812 struct kiocb_batch batch;
1814 if (unlikely(nr < 0))
1817 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1818 nr = LONG_MAX/sizeof(*iocbpp);
1820 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1823 ctx = lookup_ioctx(ctx_id);
1824 if (unlikely(!ctx)) {
1825 pr_debug("EINVAL: io_submit: invalid context id\n");
1829 kiocb_batch_init(&batch, nr);
1831 blk_start_plug(&plug);
1834 * AKPM: should this return a partial result if some of the IOs were
1835 * successfully submitted?
1837 for (i=0; i<nr; i++) {
1838 struct iocb __user *user_iocb;
1841 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1846 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1851 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1855 blk_finish_plug(&plug);
1857 kiocb_batch_free(ctx, &batch);
1863 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1864 * the number of iocbs queued. May return -EINVAL if the aio_context
1865 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1866 * *iocbpp[0] is not properly initialized, if the operation specified
1867 * is invalid for the file descriptor in the iocb. May fail with
1868 * -EFAULT if any of the data structures point to invalid data. May
1869 * fail with -EBADF if the file descriptor specified in the first
1870 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1871 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1872 * fail with -ENOSYS if not implemented.
1874 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1875 struct iocb __user * __user *, iocbpp)
1877 return do_io_submit(ctx_id, nr, iocbpp, 0);
1881 * Finds a given iocb for cancellation.
1883 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1886 struct list_head *pos;
1888 assert_spin_locked(&ctx->ctx_lock);
1890 /* TODO: use a hash or array, this sucks. */
1891 list_for_each(pos, &ctx->active_reqs) {
1892 struct kiocb *kiocb = list_kiocb(pos);
1893 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1900 * Attempts to cancel an iocb previously passed to io_submit. If
1901 * the operation is successfully cancelled, the resulting event is
1902 * copied into the memory pointed to by result without being placed
1903 * into the completion queue and 0 is returned. May fail with
1904 * -EFAULT if any of the data structures pointed to are invalid.
1905 * May fail with -EINVAL if aio_context specified by ctx_id is
1906 * invalid. May fail with -EAGAIN if the iocb specified was not
1907 * cancelled. Will fail with -ENOSYS if not implemented.
1909 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1910 struct io_event __user *, result)
1912 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1914 struct kiocb *kiocb;
1918 ret = get_user(key, &iocb->aio_key);
1922 ctx = lookup_ioctx(ctx_id);
1926 spin_lock_irq(&ctx->ctx_lock);
1928 kiocb = lookup_kiocb(ctx, iocb, key);
1929 if (kiocb && kiocb->ki_cancel) {
1930 cancel = kiocb->ki_cancel;
1932 kiocbSetCancelled(kiocb);
1935 spin_unlock_irq(&ctx->ctx_lock);
1937 if (NULL != cancel) {
1938 struct io_event tmp;
1939 pr_debug("calling cancel\n");
1940 memset(&tmp, 0, sizeof(tmp));
1941 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1942 tmp.data = kiocb->ki_user_data;
1943 ret = cancel(kiocb, &tmp);
1945 /* Cancellation succeeded -- copy the result
1946 * into the user's buffer.
1948 if (copy_to_user(result, &tmp, sizeof(tmp)))
1960 * Attempts to read at least min_nr events and up to nr events from
1961 * the completion queue for the aio_context specified by ctx_id. If
1962 * it succeeds, the number of read events is returned. May fail with
1963 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1964 * out of range, if timeout is out of range. May fail with -EFAULT
1965 * if any of the memory specified is invalid. May return 0 or
1966 * < min_nr if the timeout specified by timeout has elapsed
1967 * before sufficient events are available, where timeout == NULL
1968 * specifies an infinite timeout. Note that the timeout pointed to by
1969 * timeout is relative and will be updated if not NULL and the
1970 * operation blocks. Will fail with -ENOSYS if not implemented.
1972 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1975 struct io_event __user *, events,
1976 struct timespec __user *, timeout)
1978 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1981 if (likely(ioctx)) {
1982 if (likely(min_nr <= nr && min_nr >= 0))
1983 ret = read_events(ioctx, min_nr, nr, events, timeout);
1987 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);