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 #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]);
95 if (info->mmap_size) {
96 down_write(&ctx->mm->mmap_sem);
97 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
98 up_write(&ctx->mm->mmap_sem);
101 if (info->ring_pages && info->ring_pages != info->internal_pages)
102 kfree(info->ring_pages);
103 info->ring_pages = NULL;
107 static int aio_setup_ring(struct kioctx *ctx)
109 struct aio_ring *ring;
110 struct aio_ring_info *info = &ctx->ring_info;
111 unsigned nr_events = ctx->max_reqs;
115 /* Compensate for the ring buffer's head/tail overlap entry */
116 nr_events += 2; /* 1 is required, 2 for good luck */
118 size = sizeof(struct aio_ring);
119 size += sizeof(struct io_event) * nr_events;
120 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
125 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
128 info->ring_pages = info->internal_pages;
129 if (nr_pages > AIO_RING_PAGES) {
130 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
131 if (!info->ring_pages)
135 info->mmap_size = nr_pages * PAGE_SIZE;
136 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
137 down_write(&ctx->mm->mmap_sem);
138 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
139 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
141 if (IS_ERR((void *)info->mmap_base)) {
142 up_write(&ctx->mm->mmap_sem);
148 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
149 info->nr_pages = get_user_pages(current, ctx->mm,
150 info->mmap_base, nr_pages,
151 1, 0, info->ring_pages, NULL);
152 up_write(&ctx->mm->mmap_sem);
154 if (unlikely(info->nr_pages != nr_pages)) {
159 ctx->user_id = info->mmap_base;
161 info->nr = nr_events; /* trusted copy */
163 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
164 ring->nr = nr_events; /* user copy */
165 ring->id = ctx->user_id;
166 ring->head = ring->tail = 0;
167 ring->magic = AIO_RING_MAGIC;
168 ring->compat_features = AIO_RING_COMPAT_FEATURES;
169 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
170 ring->header_length = sizeof(struct aio_ring);
171 kunmap_atomic(ring, KM_USER0);
177 /* aio_ring_event: returns a pointer to the event at the given index from
178 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
180 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
181 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
182 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
184 #define aio_ring_event(info, nr, km) ({ \
185 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
186 struct io_event *__event; \
187 __event = kmap_atomic( \
188 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
189 __event += pos % AIO_EVENTS_PER_PAGE; \
193 #define put_aio_ring_event(event, km) do { \
194 struct io_event *__event = (event); \
196 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
199 static void ctx_rcu_free(struct rcu_head *head)
201 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
202 unsigned nr_events = ctx->max_reqs;
204 kmem_cache_free(kioctx_cachep, ctx);
207 spin_lock(&aio_nr_lock);
208 BUG_ON(aio_nr - nr_events > aio_nr);
210 spin_unlock(&aio_nr_lock);
215 * Called when the last user of an aio context has gone away,
216 * and the struct needs to be freed.
218 static void __put_ioctx(struct kioctx *ctx)
220 BUG_ON(ctx->reqs_active);
222 cancel_delayed_work(&ctx->wq);
223 cancel_work_sync(&ctx->wq.work);
227 pr_debug("__put_ioctx: freeing %p\n", ctx);
228 call_rcu(&ctx->rcu_head, ctx_rcu_free);
231 static inline void get_ioctx(struct kioctx *kioctx)
233 BUG_ON(atomic_read(&kioctx->users) <= 0);
234 atomic_inc(&kioctx->users);
237 static inline int try_get_ioctx(struct kioctx *kioctx)
239 return atomic_inc_not_zero(&kioctx->users);
242 static inline void put_ioctx(struct kioctx *kioctx)
244 BUG_ON(atomic_read(&kioctx->users) <= 0);
245 if (unlikely(atomic_dec_and_test(&kioctx->users)))
250 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
252 static struct kioctx *ioctx_alloc(unsigned nr_events)
254 struct mm_struct *mm;
258 /* Prevent overflows */
259 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
260 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
261 pr_debug("ENOMEM: nr_events too high\n");
262 return ERR_PTR(-EINVAL);
265 if ((unsigned long)nr_events > aio_max_nr)
266 return ERR_PTR(-EAGAIN);
268 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
270 return ERR_PTR(-ENOMEM);
272 ctx->max_reqs = nr_events;
273 mm = ctx->mm = current->mm;
274 atomic_inc(&mm->mm_count);
276 atomic_set(&ctx->users, 1);
277 spin_lock_init(&ctx->ctx_lock);
278 spin_lock_init(&ctx->ring_info.ring_lock);
279 init_waitqueue_head(&ctx->wait);
281 INIT_LIST_HEAD(&ctx->active_reqs);
282 INIT_LIST_HEAD(&ctx->run_list);
283 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
285 if (aio_setup_ring(ctx) < 0)
288 /* limit the number of system wide aios */
290 spin_lock_bh(&aio_nr_lock);
291 if (aio_nr + nr_events > aio_max_nr ||
292 aio_nr + nr_events < aio_nr)
295 aio_nr += ctx->max_reqs;
296 spin_unlock_bh(&aio_nr_lock);
297 if (ctx->max_reqs || did_sync)
300 /* wait for rcu callbacks to have completed before giving up */
303 ctx->max_reqs = nr_events;
306 if (ctx->max_reqs == 0)
309 /* now link into global list. */
310 spin_lock(&mm->ioctx_lock);
311 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
312 spin_unlock(&mm->ioctx_lock);
314 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
315 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
320 return ERR_PTR(-EAGAIN);
324 kmem_cache_free(kioctx_cachep, ctx);
325 ctx = ERR_PTR(-ENOMEM);
327 dprintk("aio: error allocating ioctx %p\n", ctx);
332 * Cancels all outstanding aio requests on an aio context. Used
333 * when the processes owning a context have all exited to encourage
334 * the rapid destruction of the kioctx.
336 static void aio_cancel_all(struct kioctx *ctx)
338 int (*cancel)(struct kiocb *, struct io_event *);
340 spin_lock_irq(&ctx->ctx_lock);
342 while (!list_empty(&ctx->active_reqs)) {
343 struct list_head *pos = ctx->active_reqs.next;
344 struct kiocb *iocb = list_kiocb(pos);
345 list_del_init(&iocb->ki_list);
346 cancel = iocb->ki_cancel;
347 kiocbSetCancelled(iocb);
350 spin_unlock_irq(&ctx->ctx_lock);
352 spin_lock_irq(&ctx->ctx_lock);
355 spin_unlock_irq(&ctx->ctx_lock);
358 static void wait_for_all_aios(struct kioctx *ctx)
360 struct task_struct *tsk = current;
361 DECLARE_WAITQUEUE(wait, tsk);
363 spin_lock_irq(&ctx->ctx_lock);
364 if (!ctx->reqs_active)
367 add_wait_queue(&ctx->wait, &wait);
368 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
369 while (ctx->reqs_active) {
370 spin_unlock_irq(&ctx->ctx_lock);
372 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
373 spin_lock_irq(&ctx->ctx_lock);
375 __set_task_state(tsk, TASK_RUNNING);
376 remove_wait_queue(&ctx->wait, &wait);
379 spin_unlock_irq(&ctx->ctx_lock);
382 /* wait_on_sync_kiocb:
383 * Waits on the given sync kiocb to complete.
385 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
387 while (iocb->ki_users) {
388 set_current_state(TASK_UNINTERRUPTIBLE);
393 __set_current_state(TASK_RUNNING);
394 return iocb->ki_user_data;
396 EXPORT_SYMBOL(wait_on_sync_kiocb);
398 /* exit_aio: called when the last user of mm goes away. At this point,
399 * there is no way for any new requests to be submited or any of the
400 * io_* syscalls to be called on the context. However, there may be
401 * outstanding requests which hold references to the context; as they
402 * go away, they will call put_ioctx and release any pinned memory
403 * associated with the request (held via struct page * references).
405 void exit_aio(struct mm_struct *mm)
409 while (!hlist_empty(&mm->ioctx_list)) {
410 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
411 hlist_del_rcu(&ctx->list);
415 wait_for_all_aios(ctx);
417 * Ensure we don't leave the ctx on the aio_wq
419 cancel_work_sync(&ctx->wq.work);
421 if (1 != atomic_read(&ctx->users))
423 "exit_aio:ioctx still alive: %d %d %d\n",
424 atomic_read(&ctx->users), ctx->dead,
431 * Allocate a slot for an aio request. Increments the users count
432 * of the kioctx so that the kioctx stays around until all requests are
433 * complete. Returns NULL if no requests are free.
435 * Returns with kiocb->users set to 2. The io submit code path holds
436 * an extra reference while submitting the i/o.
437 * This prevents races between the aio code path referencing the
438 * req (after submitting it) and aio_complete() freeing the req.
440 static struct kiocb *__aio_get_req(struct kioctx *ctx)
442 struct kiocb *req = NULL;
444 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
452 req->ki_cancel = NULL;
453 req->ki_retry = NULL;
456 req->ki_iovec = NULL;
457 INIT_LIST_HEAD(&req->ki_run_list);
458 req->ki_eventfd = NULL;
464 * struct kiocb's are allocated in batches to reduce the number of
465 * times the ctx lock is acquired and released.
467 #define KIOCB_BATCH_SIZE 32L
469 struct list_head head;
470 long count; /* number of requests left to allocate */
473 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
475 INIT_LIST_HEAD(&batch->head);
476 batch->count = total;
479 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
481 struct kiocb *req, *n;
483 if (list_empty(&batch->head))
486 spin_lock_irq(&ctx->ctx_lock);
487 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
488 list_del(&req->ki_batch);
489 list_del(&req->ki_list);
490 kmem_cache_free(kiocb_cachep, req);
493 if (unlikely(!ctx->reqs_active && ctx->dead))
494 wake_up_all(&ctx->wait);
495 spin_unlock_irq(&ctx->ctx_lock);
499 * Allocate a batch of kiocbs. This avoids taking and dropping the
500 * context lock a lot during setup.
502 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
504 unsigned short allocated, to_alloc;
506 bool called_fput = false;
507 struct kiocb *req, *n;
508 struct aio_ring *ring;
510 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
511 for (allocated = 0; allocated < to_alloc; allocated++) {
512 req = __aio_get_req(ctx);
514 /* allocation failed, go with what we've got */
516 list_add(&req->ki_batch, &batch->head);
523 spin_lock_irq(&ctx->ctx_lock);
524 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
526 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
528 if (avail == 0 && !called_fput) {
530 * Handle a potential starvation case. It is possible that
531 * we hold the last reference on a struct file, causing us
532 * to delay the final fput to non-irq context. In this case,
533 * ctx->reqs_active is artificially high. Calling the fput
534 * routine here may free up a slot in the event completion
535 * ring, allowing this allocation to succeed.
538 spin_unlock_irq(&ctx->ctx_lock);
539 aio_fput_routine(NULL);
544 if (avail < allocated) {
545 /* Trim back the number of requests. */
546 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
547 list_del(&req->ki_batch);
548 kmem_cache_free(kiocb_cachep, req);
549 if (--allocated <= avail)
554 batch->count -= allocated;
555 list_for_each_entry(req, &batch->head, ki_batch) {
556 list_add(&req->ki_list, &ctx->active_reqs);
561 spin_unlock_irq(&ctx->ctx_lock);
567 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
568 struct kiocb_batch *batch)
572 if (list_empty(&batch->head))
573 if (kiocb_batch_refill(ctx, batch) == 0)
575 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
576 list_del(&req->ki_batch);
580 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
582 assert_spin_locked(&ctx->ctx_lock);
584 if (req->ki_eventfd != NULL)
585 eventfd_ctx_put(req->ki_eventfd);
588 if (req->ki_iovec != &req->ki_inline_vec)
589 kfree(req->ki_iovec);
590 kmem_cache_free(kiocb_cachep, req);
593 if (unlikely(!ctx->reqs_active && ctx->dead))
594 wake_up_all(&ctx->wait);
597 static void aio_fput_routine(struct work_struct *data)
599 spin_lock_irq(&fput_lock);
600 while (likely(!list_empty(&fput_head))) {
601 struct kiocb *req = list_kiocb(fput_head.next);
602 struct kioctx *ctx = req->ki_ctx;
604 list_del(&req->ki_list);
605 spin_unlock_irq(&fput_lock);
607 /* Complete the fput(s) */
608 if (req->ki_filp != NULL)
611 /* Link the iocb into the context's free list */
612 spin_lock_irq(&ctx->ctx_lock);
613 really_put_req(ctx, req);
614 spin_unlock_irq(&ctx->ctx_lock);
617 spin_lock_irq(&fput_lock);
619 spin_unlock_irq(&fput_lock);
623 * Returns true if this put was the last user of the request.
625 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
627 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
628 req, atomic_long_read(&req->ki_filp->f_count));
630 assert_spin_locked(&ctx->ctx_lock);
633 BUG_ON(req->ki_users < 0);
634 if (likely(req->ki_users))
636 list_del(&req->ki_list); /* remove from active_reqs */
637 req->ki_cancel = NULL;
638 req->ki_retry = NULL;
641 * Try to optimize the aio and eventfd file* puts, by avoiding to
642 * schedule work in case it is not final fput() time. In normal cases,
643 * we would not be holding the last reference to the file*, so
644 * this function will be executed w/out any aio kthread wakeup.
646 if (unlikely(!fput_atomic(req->ki_filp))) {
648 spin_lock(&fput_lock);
649 list_add(&req->ki_list, &fput_head);
650 spin_unlock(&fput_lock);
651 schedule_work(&fput_work);
654 really_put_req(ctx, req);
660 * Returns true if this put was the last user of the kiocb,
661 * false if the request is still in use.
663 int aio_put_req(struct kiocb *req)
665 struct kioctx *ctx = req->ki_ctx;
667 spin_lock_irq(&ctx->ctx_lock);
668 ret = __aio_put_req(ctx, req);
669 spin_unlock_irq(&ctx->ctx_lock);
672 EXPORT_SYMBOL(aio_put_req);
674 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
676 struct mm_struct *mm = current->mm;
677 struct kioctx *ctx, *ret = NULL;
678 struct hlist_node *n;
682 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
684 * RCU protects us against accessing freed memory but
685 * we have to be careful not to get a reference when the
686 * reference count already dropped to 0 (ctx->dead test
687 * is unreliable because of races).
689 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
700 * Queue up a kiocb to be retried. Assumes that the kiocb
701 * has already been marked as kicked, and places it on
702 * the retry run list for the corresponding ioctx, if it
703 * isn't already queued. Returns 1 if it actually queued
704 * the kiocb (to tell the caller to activate the work
705 * queue to process it), or 0, if it found that it was
708 static inline int __queue_kicked_iocb(struct kiocb *iocb)
710 struct kioctx *ctx = iocb->ki_ctx;
712 assert_spin_locked(&ctx->ctx_lock);
714 if (list_empty(&iocb->ki_run_list)) {
715 list_add_tail(&iocb->ki_run_list,
723 * This is the core aio execution routine. It is
724 * invoked both for initial i/o submission and
725 * subsequent retries via the aio_kick_handler.
726 * Expects to be invoked with iocb->ki_ctx->lock
727 * already held. The lock is released and reacquired
728 * as needed during processing.
730 * Calls the iocb retry method (already setup for the
731 * iocb on initial submission) for operation specific
732 * handling, but takes care of most of common retry
733 * execution details for a given iocb. The retry method
734 * needs to be non-blocking as far as possible, to avoid
735 * holding up other iocbs waiting to be serviced by the
736 * retry kernel thread.
738 * The trickier parts in this code have to do with
739 * ensuring that only one retry instance is in progress
740 * for a given iocb at any time. Providing that guarantee
741 * simplifies the coding of individual aio operations as
742 * it avoids various potential races.
744 static ssize_t aio_run_iocb(struct kiocb *iocb)
746 struct kioctx *ctx = iocb->ki_ctx;
747 ssize_t (*retry)(struct kiocb *);
750 if (!(retry = iocb->ki_retry)) {
751 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
756 * We don't want the next retry iteration for this
757 * operation to start until this one has returned and
758 * updated the iocb state. However, wait_queue functions
759 * can trigger a kick_iocb from interrupt context in the
760 * meantime, indicating that data is available for the next
761 * iteration. We want to remember that and enable the
762 * next retry iteration _after_ we are through with
765 * So, in order to be able to register a "kick", but
766 * prevent it from being queued now, we clear the kick
767 * flag, but make the kick code *think* that the iocb is
768 * still on the run list until we are actually done.
769 * When we are done with this iteration, we check if
770 * the iocb was kicked in the meantime and if so, queue
774 kiocbClearKicked(iocb);
777 * This is so that aio_complete knows it doesn't need to
778 * pull the iocb off the run list (We can't just call
779 * INIT_LIST_HEAD because we don't want a kick_iocb to
780 * queue this on the run list yet)
782 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
783 spin_unlock_irq(&ctx->ctx_lock);
785 /* Quit retrying if the i/o has been cancelled */
786 if (kiocbIsCancelled(iocb)) {
788 aio_complete(iocb, ret, 0);
789 /* must not access the iocb after this */
794 * Now we are all set to call the retry method in async
799 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
801 * There's no easy way to restart the syscall since other AIO's
802 * may be already running. Just fail this IO with EINTR.
804 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
805 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
807 aio_complete(iocb, ret, 0);
810 spin_lock_irq(&ctx->ctx_lock);
812 if (-EIOCBRETRY == ret) {
814 * OK, now that we are done with this iteration
815 * and know that there is more left to go,
816 * this is where we let go so that a subsequent
817 * "kick" can start the next iteration
820 /* will make __queue_kicked_iocb succeed from here on */
821 INIT_LIST_HEAD(&iocb->ki_run_list);
822 /* we must queue the next iteration ourselves, if it
823 * has already been kicked */
824 if (kiocbIsKicked(iocb)) {
825 __queue_kicked_iocb(iocb);
828 * __queue_kicked_iocb will always return 1 here, because
829 * iocb->ki_run_list is empty at this point so it should
830 * be safe to unconditionally queue the context into the
841 * Process all pending retries queued on the ioctx
843 * Assumes it is operating within the aio issuer's mm
846 static int __aio_run_iocbs(struct kioctx *ctx)
849 struct list_head run_list;
851 assert_spin_locked(&ctx->ctx_lock);
853 list_replace_init(&ctx->run_list, &run_list);
854 while (!list_empty(&run_list)) {
855 iocb = list_entry(run_list.next, struct kiocb,
857 list_del(&iocb->ki_run_list);
859 * Hold an extra reference while retrying i/o.
861 iocb->ki_users++; /* grab extra reference */
863 __aio_put_req(ctx, iocb);
865 if (!list_empty(&ctx->run_list))
870 static void aio_queue_work(struct kioctx * ctx)
872 unsigned long timeout;
874 * if someone is waiting, get the work started right
875 * away, otherwise, use a longer delay
878 if (waitqueue_active(&ctx->wait))
882 queue_delayed_work(aio_wq, &ctx->wq, timeout);
887 * Process all pending retries queued on the ioctx
888 * run list, and keep running them until the list
890 * Assumes it is operating within the aio issuer's mm context.
892 static inline void aio_run_all_iocbs(struct kioctx *ctx)
894 spin_lock_irq(&ctx->ctx_lock);
895 while (__aio_run_iocbs(ctx))
897 spin_unlock_irq(&ctx->ctx_lock);
902 * Work queue handler triggered to process pending
903 * retries on an ioctx. Takes on the aio issuer's
904 * mm context before running the iocbs, so that
905 * copy_xxx_user operates on the issuer's address
907 * Run on aiod's context.
909 static void aio_kick_handler(struct work_struct *work)
911 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
912 mm_segment_t oldfs = get_fs();
913 struct mm_struct *mm;
918 spin_lock_irq(&ctx->ctx_lock);
919 requeue =__aio_run_iocbs(ctx);
921 spin_unlock_irq(&ctx->ctx_lock);
925 * we're in a worker thread already, don't use queue_delayed_work,
928 queue_delayed_work(aio_wq, &ctx->wq, 0);
933 * Called by kick_iocb to queue the kiocb for retry
934 * and if required activate the aio work queue to process
937 static void try_queue_kicked_iocb(struct kiocb *iocb)
939 struct kioctx *ctx = iocb->ki_ctx;
943 spin_lock_irqsave(&ctx->ctx_lock, flags);
944 /* set this inside the lock so that we can't race with aio_run_iocb()
945 * testing it and putting the iocb on the run list under the lock */
946 if (!kiocbTryKick(iocb))
947 run = __queue_kicked_iocb(iocb);
948 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
955 * Called typically from a wait queue callback context
956 * to trigger a retry of the iocb.
957 * The retry is usually executed by aio workqueue
958 * threads (See aio_kick_handler).
960 void kick_iocb(struct kiocb *iocb)
962 /* sync iocbs are easy: they can only ever be executing from a
964 if (is_sync_kiocb(iocb)) {
965 kiocbSetKicked(iocb);
966 wake_up_process(iocb->ki_obj.tsk);
970 try_queue_kicked_iocb(iocb);
972 EXPORT_SYMBOL(kick_iocb);
975 * Called when the io request on the given iocb is complete.
976 * Returns true if this is the last user of the request. The
977 * only other user of the request can be the cancellation code.
979 int aio_complete(struct kiocb *iocb, long res, long res2)
981 struct kioctx *ctx = iocb->ki_ctx;
982 struct aio_ring_info *info;
983 struct aio_ring *ring;
984 struct io_event *event;
990 * Special case handling for sync iocbs:
991 * - events go directly into the iocb for fast handling
992 * - the sync task with the iocb in its stack holds the single iocb
993 * ref, no other paths have a way to get another ref
994 * - the sync task helpfully left a reference to itself in the iocb
996 if (is_sync_kiocb(iocb)) {
997 BUG_ON(iocb->ki_users != 1);
998 iocb->ki_user_data = res;
1000 wake_up_process(iocb->ki_obj.tsk);
1004 info = &ctx->ring_info;
1006 /* add a completion event to the ring buffer.
1007 * must be done holding ctx->ctx_lock to prevent
1008 * other code from messing with the tail
1009 * pointer since we might be called from irq
1012 spin_lock_irqsave(&ctx->ctx_lock, flags);
1014 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
1015 list_del_init(&iocb->ki_run_list);
1018 * cancelled requests don't get events, userland was given one
1019 * when the event got cancelled.
1021 if (kiocbIsCancelled(iocb))
1024 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
1027 event = aio_ring_event(info, tail, KM_IRQ0);
1028 if (++tail >= info->nr)
1031 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1032 event->data = iocb->ki_user_data;
1036 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1037 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1040 /* after flagging the request as done, we
1041 * must never even look at it again
1043 smp_wmb(); /* make event visible before updating tail */
1048 put_aio_ring_event(event, KM_IRQ0);
1049 kunmap_atomic(ring, KM_IRQ1);
1051 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1054 * Check if the user asked us to deliver the result through an
1055 * eventfd. The eventfd_signal() function is safe to be called
1058 if (iocb->ki_eventfd != NULL)
1059 eventfd_signal(iocb->ki_eventfd, 1);
1062 /* everything turned out well, dispose of the aiocb. */
1063 ret = __aio_put_req(ctx, iocb);
1066 * We have to order our ring_info tail store above and test
1067 * of the wait list below outside the wait lock. This is
1068 * like in wake_up_bit() where clearing a bit has to be
1069 * ordered with the unlocked test.
1073 if (waitqueue_active(&ctx->wait))
1074 wake_up(&ctx->wait);
1076 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1079 EXPORT_SYMBOL(aio_complete);
1082 * Pull an event off of the ioctx's event ring. Returns the number of
1083 * events fetched (0 or 1 ;-)
1084 * FIXME: make this use cmpxchg.
1085 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1087 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1089 struct aio_ring_info *info = &ioctx->ring_info;
1090 struct aio_ring *ring;
1094 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1095 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1096 (unsigned long)ring->head, (unsigned long)ring->tail,
1097 (unsigned long)ring->nr);
1099 if (ring->head == ring->tail)
1102 spin_lock(&info->ring_lock);
1104 head = ring->head % info->nr;
1105 if (head != ring->tail) {
1106 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1108 head = (head + 1) % info->nr;
1109 smp_mb(); /* finish reading the event before updatng the head */
1112 put_aio_ring_event(evp, KM_USER1);
1114 spin_unlock(&info->ring_lock);
1117 kunmap_atomic(ring, KM_USER0);
1118 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1119 (unsigned long)ring->head, (unsigned long)ring->tail);
1123 struct aio_timeout {
1124 struct timer_list timer;
1126 struct task_struct *p;
1129 static void timeout_func(unsigned long data)
1131 struct aio_timeout *to = (struct aio_timeout *)data;
1134 wake_up_process(to->p);
1137 static inline void init_timeout(struct aio_timeout *to)
1139 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1144 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1145 const struct timespec *ts)
1147 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1148 if (time_after(to->timer.expires, jiffies))
1149 add_timer(&to->timer);
1154 static inline void clear_timeout(struct aio_timeout *to)
1156 del_singleshot_timer_sync(&to->timer);
1159 static int read_events(struct kioctx *ctx,
1160 long min_nr, long nr,
1161 struct io_event __user *event,
1162 struct timespec __user *timeout)
1164 long start_jiffies = jiffies;
1165 struct task_struct *tsk = current;
1166 DECLARE_WAITQUEUE(wait, tsk);
1169 struct io_event ent;
1170 struct aio_timeout to;
1173 /* needed to zero any padding within an entry (there shouldn't be
1174 * any, but C is fun!
1176 memset(&ent, 0, sizeof(ent));
1179 while (likely(i < nr)) {
1180 ret = aio_read_evt(ctx, &ent);
1181 if (unlikely(ret <= 0))
1184 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1185 ent.data, ent.obj, ent.res, ent.res2);
1187 /* Could we split the check in two? */
1189 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1190 dprintk("aio: lost an event due to EFAULT.\n");
1195 /* Good, event copied to userland, update counts. */
1207 /* racey check, but it gets redone */
1208 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1210 aio_run_all_iocbs(ctx);
1218 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1221 set_timeout(start_jiffies, &to, &ts);
1224 while (likely(i < nr)) {
1225 add_wait_queue_exclusive(&ctx->wait, &wait);
1227 set_task_state(tsk, TASK_INTERRUPTIBLE);
1228 ret = aio_read_evt(ctx, &ent);
1233 if (unlikely(ctx->dead)) {
1237 if (to.timed_out) /* Only check after read evt */
1239 /* Try to only show up in io wait if there are ops
1241 if (ctx->reqs_active)
1245 if (signal_pending(tsk)) {
1249 /*ret = aio_read_evt(ctx, &ent);*/
1252 set_task_state(tsk, TASK_RUNNING);
1253 remove_wait_queue(&ctx->wait, &wait);
1255 if (unlikely(ret <= 0))
1259 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1260 dprintk("aio: lost an event due to EFAULT.\n");
1264 /* Good, event copied to userland, update counts. */
1272 destroy_timer_on_stack(&to.timer);
1276 /* Take an ioctx and remove it from the list of ioctx's. Protects
1277 * against races with itself via ->dead.
1279 static void io_destroy(struct kioctx *ioctx)
1281 struct mm_struct *mm = current->mm;
1284 /* delete the entry from the list is someone else hasn't already */
1285 spin_lock(&mm->ioctx_lock);
1286 was_dead = ioctx->dead;
1288 hlist_del_rcu(&ioctx->list);
1289 spin_unlock(&mm->ioctx_lock);
1291 dprintk("aio_release(%p)\n", ioctx);
1292 if (likely(!was_dead))
1293 put_ioctx(ioctx); /* twice for the list */
1295 aio_cancel_all(ioctx);
1296 wait_for_all_aios(ioctx);
1299 * Wake up any waiters. The setting of ctx->dead must be seen
1300 * by other CPUs at this point. Right now, we rely on the
1301 * locking done by the above calls to ensure this consistency.
1303 wake_up_all(&ioctx->wait);
1304 put_ioctx(ioctx); /* once for the lookup */
1308 * Create an aio_context capable of receiving at least nr_events.
1309 * ctxp must not point to an aio_context that already exists, and
1310 * must be initialized to 0 prior to the call. On successful
1311 * creation of the aio_context, *ctxp is filled in with the resulting
1312 * handle. May fail with -EINVAL if *ctxp is not initialized,
1313 * if the specified nr_events exceeds internal limits. May fail
1314 * with -EAGAIN if the specified nr_events exceeds the user's limit
1315 * of available events. May fail with -ENOMEM if insufficient kernel
1316 * resources are available. May fail with -EFAULT if an invalid
1317 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1320 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1322 struct kioctx *ioctx = NULL;
1326 ret = get_user(ctx, ctxp);
1331 if (unlikely(ctx || nr_events == 0)) {
1332 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1337 ioctx = ioctx_alloc(nr_events);
1338 ret = PTR_ERR(ioctx);
1339 if (!IS_ERR(ioctx)) {
1340 ret = put_user(ioctx->user_id, ctxp);
1344 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1353 * Destroy the aio_context specified. May cancel any outstanding
1354 * AIOs and block on completion. Will fail with -ENOSYS if not
1355 * implemented. May fail with -EINVAL if the context pointed to
1358 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1360 struct kioctx *ioctx = lookup_ioctx(ctx);
1361 if (likely(NULL != ioctx)) {
1365 pr_debug("EINVAL: io_destroy: invalid context id\n");
1369 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1371 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1375 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1376 ssize_t this = min((ssize_t)iov->iov_len, ret);
1377 iov->iov_base += this;
1378 iov->iov_len -= this;
1379 iocb->ki_left -= this;
1381 if (iov->iov_len == 0) {
1387 /* the caller should not have done more io than what fit in
1388 * the remaining iovecs */
1389 BUG_ON(ret > 0 && iocb->ki_left == 0);
1392 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1394 struct file *file = iocb->ki_filp;
1395 struct address_space *mapping = file->f_mapping;
1396 struct inode *inode = mapping->host;
1397 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1398 unsigned long, loff_t);
1400 unsigned short opcode;
1402 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1403 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1404 rw_op = file->f_op->aio_read;
1405 opcode = IOCB_CMD_PREADV;
1407 rw_op = file->f_op->aio_write;
1408 opcode = IOCB_CMD_PWRITEV;
1411 /* This matches the pread()/pwrite() logic */
1412 if (iocb->ki_pos < 0)
1416 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1417 iocb->ki_nr_segs - iocb->ki_cur_seg,
1420 aio_advance_iovec(iocb, ret);
1422 /* retry all partial writes. retry partial reads as long as its a
1424 } while (ret > 0 && iocb->ki_left > 0 &&
1425 (opcode == IOCB_CMD_PWRITEV ||
1426 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1428 /* This means we must have transferred all that we could */
1429 /* No need to retry anymore */
1430 if ((ret == 0) || (iocb->ki_left == 0))
1431 ret = iocb->ki_nbytes - iocb->ki_left;
1433 /* If we managed to write some out we return that, rather than
1434 * the eventual error. */
1435 if (opcode == IOCB_CMD_PWRITEV
1436 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1437 && iocb->ki_nbytes - iocb->ki_left)
1438 ret = iocb->ki_nbytes - iocb->ki_left;
1443 static ssize_t aio_fdsync(struct kiocb *iocb)
1445 struct file *file = iocb->ki_filp;
1446 ssize_t ret = -EINVAL;
1448 if (file->f_op->aio_fsync)
1449 ret = file->f_op->aio_fsync(iocb, 1);
1453 static ssize_t aio_fsync(struct kiocb *iocb)
1455 struct file *file = iocb->ki_filp;
1456 ssize_t ret = -EINVAL;
1458 if (file->f_op->aio_fsync)
1459 ret = file->f_op->aio_fsync(iocb, 0);
1463 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1467 #ifdef CONFIG_COMPAT
1469 ret = compat_rw_copy_check_uvector(type,
1470 (struct compat_iovec __user *)kiocb->ki_buf,
1471 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1472 &kiocb->ki_iovec, 1);
1475 ret = rw_copy_check_uvector(type,
1476 (struct iovec __user *)kiocb->ki_buf,
1477 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1478 &kiocb->ki_iovec, 1);
1482 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1483 kiocb->ki_cur_seg = 0;
1484 /* ki_nbytes/left now reflect bytes instead of segs */
1485 kiocb->ki_nbytes = ret;
1486 kiocb->ki_left = ret;
1493 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1495 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1496 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1497 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1498 kiocb->ki_nr_segs = 1;
1499 kiocb->ki_cur_seg = 0;
1505 * Performs the initial checks and aio retry method
1506 * setup for the kiocb at the time of io submission.
1508 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1510 struct file *file = kiocb->ki_filp;
1513 switch (kiocb->ki_opcode) {
1514 case IOCB_CMD_PREAD:
1516 if (unlikely(!(file->f_mode & FMODE_READ)))
1519 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1522 ret = security_file_permission(file, MAY_READ);
1525 ret = aio_setup_single_vector(kiocb);
1529 if (file->f_op->aio_read)
1530 kiocb->ki_retry = aio_rw_vect_retry;
1532 case IOCB_CMD_PWRITE:
1534 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1537 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1540 ret = security_file_permission(file, MAY_WRITE);
1543 ret = aio_setup_single_vector(kiocb);
1547 if (file->f_op->aio_write)
1548 kiocb->ki_retry = aio_rw_vect_retry;
1550 case IOCB_CMD_PREADV:
1552 if (unlikely(!(file->f_mode & FMODE_READ)))
1554 ret = security_file_permission(file, MAY_READ);
1557 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1561 if (file->f_op->aio_read)
1562 kiocb->ki_retry = aio_rw_vect_retry;
1564 case IOCB_CMD_PWRITEV:
1566 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1568 ret = security_file_permission(file, MAY_WRITE);
1571 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1575 if (file->f_op->aio_write)
1576 kiocb->ki_retry = aio_rw_vect_retry;
1578 case IOCB_CMD_FDSYNC:
1580 if (file->f_op->aio_fsync)
1581 kiocb->ki_retry = aio_fdsync;
1583 case IOCB_CMD_FSYNC:
1585 if (file->f_op->aio_fsync)
1586 kiocb->ki_retry = aio_fsync;
1589 dprintk("EINVAL: io_submit: no operation provided\n");
1593 if (!kiocb->ki_retry)
1599 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1600 struct iocb *iocb, struct kiocb_batch *batch,
1607 /* enforce forwards compatibility on users */
1608 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1609 pr_debug("EINVAL: io_submit: reserve field set\n");
1613 /* prevent overflows */
1615 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1616 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1617 ((ssize_t)iocb->aio_nbytes < 0)
1619 pr_debug("EINVAL: io_submit: overflow check\n");
1623 file = fget(iocb->aio_fildes);
1624 if (unlikely(!file))
1627 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1628 if (unlikely(!req)) {
1632 req->ki_filp = file;
1633 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1635 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1636 * instance of the file* now. The file descriptor must be
1637 * an eventfd() fd, and will be signaled for each completed
1638 * event using the eventfd_signal() function.
1640 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1641 if (IS_ERR(req->ki_eventfd)) {
1642 ret = PTR_ERR(req->ki_eventfd);
1643 req->ki_eventfd = NULL;
1648 ret = put_user(req->ki_key, &user_iocb->aio_key);
1649 if (unlikely(ret)) {
1650 dprintk("EFAULT: aio_key\n");
1654 req->ki_obj.user = user_iocb;
1655 req->ki_user_data = iocb->aio_data;
1656 req->ki_pos = iocb->aio_offset;
1658 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1659 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1660 req->ki_opcode = iocb->aio_lio_opcode;
1662 ret = aio_setup_iocb(req, compat);
1667 spin_lock_irq(&ctx->ctx_lock);
1669 * We could have raced with io_destroy() and are currently holding a
1670 * reference to ctx which should be destroyed. We cannot submit IO
1671 * since ctx gets freed as soon as io_submit() puts its reference. The
1672 * check here is reliable: io_destroy() sets ctx->dead before waiting
1673 * for outstanding IO and the barrier between these two is realized by
1674 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1675 * increment ctx->reqs_active before checking for ctx->dead and the
1676 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1677 * don't see ctx->dead set here, io_destroy() waits for our IO to
1681 spin_unlock_irq(&ctx->ctx_lock);
1686 if (!list_empty(&ctx->run_list)) {
1687 /* drain the run list */
1688 while (__aio_run_iocbs(ctx))
1691 spin_unlock_irq(&ctx->ctx_lock);
1693 aio_put_req(req); /* drop extra ref to req */
1697 aio_put_req(req); /* drop extra ref to req */
1698 aio_put_req(req); /* drop i/o ref to req */
1702 long do_io_submit(aio_context_t ctx_id, long nr,
1703 struct iocb __user *__user *iocbpp, bool compat)
1708 struct blk_plug plug;
1709 struct kiocb_batch batch;
1711 if (unlikely(nr < 0))
1714 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1715 nr = LONG_MAX/sizeof(*iocbpp);
1717 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1720 ctx = lookup_ioctx(ctx_id);
1721 if (unlikely(!ctx)) {
1722 pr_debug("EINVAL: io_submit: invalid context id\n");
1726 kiocb_batch_init(&batch, nr);
1728 blk_start_plug(&plug);
1731 * AKPM: should this return a partial result if some of the IOs were
1732 * successfully submitted?
1734 for (i=0; i<nr; i++) {
1735 struct iocb __user *user_iocb;
1738 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1743 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1748 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1752 blk_finish_plug(&plug);
1754 kiocb_batch_free(ctx, &batch);
1760 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1761 * the number of iocbs queued. May return -EINVAL if the aio_context
1762 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1763 * *iocbpp[0] is not properly initialized, if the operation specified
1764 * is invalid for the file descriptor in the iocb. May fail with
1765 * -EFAULT if any of the data structures point to invalid data. May
1766 * fail with -EBADF if the file descriptor specified in the first
1767 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1768 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1769 * fail with -ENOSYS if not implemented.
1771 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1772 struct iocb __user * __user *, iocbpp)
1774 return do_io_submit(ctx_id, nr, iocbpp, 0);
1778 * Finds a given iocb for cancellation.
1780 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1783 struct list_head *pos;
1785 assert_spin_locked(&ctx->ctx_lock);
1787 /* TODO: use a hash or array, this sucks. */
1788 list_for_each(pos, &ctx->active_reqs) {
1789 struct kiocb *kiocb = list_kiocb(pos);
1790 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1797 * Attempts to cancel an iocb previously passed to io_submit. If
1798 * the operation is successfully cancelled, the resulting event is
1799 * copied into the memory pointed to by result without being placed
1800 * into the completion queue and 0 is returned. May fail with
1801 * -EFAULT if any of the data structures pointed to are invalid.
1802 * May fail with -EINVAL if aio_context specified by ctx_id is
1803 * invalid. May fail with -EAGAIN if the iocb specified was not
1804 * cancelled. Will fail with -ENOSYS if not implemented.
1806 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1807 struct io_event __user *, result)
1809 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1811 struct kiocb *kiocb;
1815 ret = get_user(key, &iocb->aio_key);
1819 ctx = lookup_ioctx(ctx_id);
1823 spin_lock_irq(&ctx->ctx_lock);
1825 kiocb = lookup_kiocb(ctx, iocb, key);
1826 if (kiocb && kiocb->ki_cancel) {
1827 cancel = kiocb->ki_cancel;
1829 kiocbSetCancelled(kiocb);
1832 spin_unlock_irq(&ctx->ctx_lock);
1834 if (NULL != cancel) {
1835 struct io_event tmp;
1836 pr_debug("calling cancel\n");
1837 memset(&tmp, 0, sizeof(tmp));
1838 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1839 tmp.data = kiocb->ki_user_data;
1840 ret = cancel(kiocb, &tmp);
1842 /* Cancellation succeeded -- copy the result
1843 * into the user's buffer.
1845 if (copy_to_user(result, &tmp, sizeof(tmp)))
1857 * Attempts to read at least min_nr events and up to nr events from
1858 * the completion queue for the aio_context specified by ctx_id. If
1859 * it succeeds, the number of read events is returned. May fail with
1860 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1861 * out of range, if timeout is out of range. May fail with -EFAULT
1862 * if any of the memory specified is invalid. May return 0 or
1863 * < min_nr if the timeout specified by timeout has elapsed
1864 * before sufficient events are available, where timeout == NULL
1865 * specifies an infinite timeout. Note that the timeout pointed to by
1866 * timeout is relative and will be updated if not NULL and the
1867 * operation blocks. Will fail with -ENOSYS if not implemented.
1869 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1872 struct io_event __user *, events,
1873 struct timespec __user *, timeout)
1875 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1878 if (likely(ioctx)) {
1879 if (likely(min_nr <= nr && min_nr >= 0))
1880 ret = read_events(ioctx, min_nr, nr, events, timeout);
1884 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);