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 kmem_cache_free(kioctx_cachep, ctx);
206 * Called when the last user of an aio context has gone away,
207 * and the struct needs to be freed.
209 static void __put_ioctx(struct kioctx *ctx)
211 unsigned nr_events = ctx->max_reqs;
212 BUG_ON(ctx->reqs_active);
214 cancel_delayed_work(&ctx->wq);
215 cancel_work_sync(&ctx->wq.work);
220 spin_lock(&aio_nr_lock);
221 BUG_ON(aio_nr - nr_events > aio_nr);
223 spin_unlock(&aio_nr_lock);
225 pr_debug("__put_ioctx: freeing %p\n", ctx);
226 call_rcu(&ctx->rcu_head, ctx_rcu_free);
229 static inline int try_get_ioctx(struct kioctx *kioctx)
231 return atomic_inc_not_zero(&kioctx->users);
234 static inline void put_ioctx(struct kioctx *kioctx)
236 BUG_ON(atomic_read(&kioctx->users) <= 0);
237 if (unlikely(atomic_dec_and_test(&kioctx->users)))
242 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
244 static struct kioctx *ioctx_alloc(unsigned nr_events)
246 struct mm_struct *mm;
250 /* Prevent overflows */
251 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
252 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
253 pr_debug("ENOMEM: nr_events too high\n");
254 return ERR_PTR(-EINVAL);
257 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
258 return ERR_PTR(-EAGAIN);
260 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
262 return ERR_PTR(-ENOMEM);
264 ctx->max_reqs = nr_events;
265 mm = ctx->mm = current->mm;
266 atomic_inc(&mm->mm_count);
268 atomic_set(&ctx->users, 2);
269 spin_lock_init(&ctx->ctx_lock);
270 spin_lock_init(&ctx->ring_info.ring_lock);
271 init_waitqueue_head(&ctx->wait);
273 INIT_LIST_HEAD(&ctx->active_reqs);
274 INIT_LIST_HEAD(&ctx->run_list);
275 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
277 if (aio_setup_ring(ctx) < 0)
280 /* limit the number of system wide aios */
281 spin_lock_bh(&aio_nr_lock);
282 if (aio_nr + nr_events > aio_max_nr ||
283 aio_nr + nr_events < aio_nr) {
284 spin_unlock_bh(&aio_nr_lock);
287 aio_nr += ctx->max_reqs;
288 spin_unlock_bh(&aio_nr_lock);
290 /* now link into global list. */
291 spin_lock(&mm->ioctx_lock);
292 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
293 spin_unlock(&mm->ioctx_lock);
295 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
296 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
304 kmem_cache_free(kioctx_cachep, ctx);
305 dprintk("aio: error allocating ioctx %d\n", err);
310 * Cancels all outstanding aio requests on an aio context. Used
311 * when the processes owning a context have all exited to encourage
312 * the rapid destruction of the kioctx.
314 static void aio_cancel_all(struct kioctx *ctx)
316 int (*cancel)(struct kiocb *, struct io_event *);
318 spin_lock_irq(&ctx->ctx_lock);
320 while (!list_empty(&ctx->active_reqs)) {
321 struct list_head *pos = ctx->active_reqs.next;
322 struct kiocb *iocb = list_kiocb(pos);
323 list_del_init(&iocb->ki_list);
324 cancel = iocb->ki_cancel;
325 kiocbSetCancelled(iocb);
328 spin_unlock_irq(&ctx->ctx_lock);
330 spin_lock_irq(&ctx->ctx_lock);
333 spin_unlock_irq(&ctx->ctx_lock);
336 static void wait_for_all_aios(struct kioctx *ctx)
338 struct task_struct *tsk = current;
339 DECLARE_WAITQUEUE(wait, tsk);
341 spin_lock_irq(&ctx->ctx_lock);
342 if (!ctx->reqs_active)
345 add_wait_queue(&ctx->wait, &wait);
346 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
347 while (ctx->reqs_active) {
348 spin_unlock_irq(&ctx->ctx_lock);
350 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
351 spin_lock_irq(&ctx->ctx_lock);
353 __set_task_state(tsk, TASK_RUNNING);
354 remove_wait_queue(&ctx->wait, &wait);
357 spin_unlock_irq(&ctx->ctx_lock);
360 /* wait_on_sync_kiocb:
361 * Waits on the given sync kiocb to complete.
363 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
365 while (iocb->ki_users) {
366 set_current_state(TASK_UNINTERRUPTIBLE);
371 __set_current_state(TASK_RUNNING);
372 return iocb->ki_user_data;
374 EXPORT_SYMBOL(wait_on_sync_kiocb);
376 /* exit_aio: called when the last user of mm goes away. At this point,
377 * there is no way for any new requests to be submited or any of the
378 * io_* syscalls to be called on the context. However, there may be
379 * outstanding requests which hold references to the context; as they
380 * go away, they will call put_ioctx and release any pinned memory
381 * associated with the request (held via struct page * references).
383 void exit_aio(struct mm_struct *mm)
387 while (!hlist_empty(&mm->ioctx_list)) {
388 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
389 hlist_del_rcu(&ctx->list);
393 wait_for_all_aios(ctx);
395 * Ensure we don't leave the ctx on the aio_wq
397 cancel_work_sync(&ctx->wq.work);
399 if (1 != atomic_read(&ctx->users))
401 "exit_aio:ioctx still alive: %d %d %d\n",
402 atomic_read(&ctx->users), ctx->dead,
409 * Allocate a slot for an aio request. Increments the users count
410 * of the kioctx so that the kioctx stays around until all requests are
411 * complete. Returns NULL if no requests are free.
413 * Returns with kiocb->users set to 2. The io submit code path holds
414 * an extra reference while submitting the i/o.
415 * This prevents races between the aio code path referencing the
416 * req (after submitting it) and aio_complete() freeing the req.
418 static struct kiocb *__aio_get_req(struct kioctx *ctx)
420 struct kiocb *req = NULL;
422 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
430 req->ki_cancel = NULL;
431 req->ki_retry = NULL;
434 req->ki_iovec = NULL;
435 INIT_LIST_HEAD(&req->ki_run_list);
436 req->ki_eventfd = NULL;
442 * struct kiocb's are allocated in batches to reduce the number of
443 * times the ctx lock is acquired and released.
445 #define KIOCB_BATCH_SIZE 32L
447 struct list_head head;
448 long count; /* number of requests left to allocate */
451 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
453 INIT_LIST_HEAD(&batch->head);
454 batch->count = total;
457 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
459 struct kiocb *req, *n;
461 if (list_empty(&batch->head))
464 spin_lock_irq(&ctx->ctx_lock);
465 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
466 list_del(&req->ki_batch);
467 list_del(&req->ki_list);
468 kmem_cache_free(kiocb_cachep, req);
471 if (unlikely(!ctx->reqs_active && ctx->dead))
472 wake_up_all(&ctx->wait);
473 spin_unlock_irq(&ctx->ctx_lock);
477 * Allocate a batch of kiocbs. This avoids taking and dropping the
478 * context lock a lot during setup.
480 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
482 unsigned short allocated, to_alloc;
484 bool called_fput = false;
485 struct kiocb *req, *n;
486 struct aio_ring *ring;
488 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
489 for (allocated = 0; allocated < to_alloc; allocated++) {
490 req = __aio_get_req(ctx);
492 /* allocation failed, go with what we've got */
494 list_add(&req->ki_batch, &batch->head);
501 spin_lock_irq(&ctx->ctx_lock);
502 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
504 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
506 if (avail == 0 && !called_fput) {
508 * Handle a potential starvation case. It is possible that
509 * we hold the last reference on a struct file, causing us
510 * to delay the final fput to non-irq context. In this case,
511 * ctx->reqs_active is artificially high. Calling the fput
512 * routine here may free up a slot in the event completion
513 * ring, allowing this allocation to succeed.
516 spin_unlock_irq(&ctx->ctx_lock);
517 aio_fput_routine(NULL);
522 if (avail < allocated) {
523 /* Trim back the number of requests. */
524 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
525 list_del(&req->ki_batch);
526 kmem_cache_free(kiocb_cachep, req);
527 if (--allocated <= avail)
532 batch->count -= allocated;
533 list_for_each_entry(req, &batch->head, ki_batch) {
534 list_add(&req->ki_list, &ctx->active_reqs);
539 spin_unlock_irq(&ctx->ctx_lock);
545 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
546 struct kiocb_batch *batch)
550 if (list_empty(&batch->head))
551 if (kiocb_batch_refill(ctx, batch) == 0)
553 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
554 list_del(&req->ki_batch);
558 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
560 assert_spin_locked(&ctx->ctx_lock);
562 if (req->ki_eventfd != NULL)
563 eventfd_ctx_put(req->ki_eventfd);
566 if (req->ki_iovec != &req->ki_inline_vec)
567 kfree(req->ki_iovec);
568 kmem_cache_free(kiocb_cachep, req);
571 if (unlikely(!ctx->reqs_active && ctx->dead))
572 wake_up_all(&ctx->wait);
575 static void aio_fput_routine(struct work_struct *data)
577 spin_lock_irq(&fput_lock);
578 while (likely(!list_empty(&fput_head))) {
579 struct kiocb *req = list_kiocb(fput_head.next);
580 struct kioctx *ctx = req->ki_ctx;
582 list_del(&req->ki_list);
583 spin_unlock_irq(&fput_lock);
585 /* Complete the fput(s) */
586 if (req->ki_filp != NULL)
589 /* Link the iocb into the context's free list */
591 spin_lock_irq(&ctx->ctx_lock);
592 really_put_req(ctx, req);
594 * at that point ctx might've been killed, but actual
597 spin_unlock_irq(&ctx->ctx_lock);
600 spin_lock_irq(&fput_lock);
602 spin_unlock_irq(&fput_lock);
606 * Returns true if this put was the last user of the request.
608 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
610 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
611 req, atomic_long_read(&req->ki_filp->f_count));
613 assert_spin_locked(&ctx->ctx_lock);
616 BUG_ON(req->ki_users < 0);
617 if (likely(req->ki_users))
619 list_del(&req->ki_list); /* remove from active_reqs */
620 req->ki_cancel = NULL;
621 req->ki_retry = NULL;
624 * Try to optimize the aio and eventfd file* puts, by avoiding to
625 * schedule work in case it is not final fput() time. In normal cases,
626 * we would not be holding the last reference to the file*, so
627 * this function will be executed w/out any aio kthread wakeup.
629 if (unlikely(!fput_atomic(req->ki_filp))) {
630 spin_lock(&fput_lock);
631 list_add(&req->ki_list, &fput_head);
632 spin_unlock(&fput_lock);
633 schedule_work(&fput_work);
636 really_put_req(ctx, req);
642 * Returns true if this put was the last user of the kiocb,
643 * false if the request is still in use.
645 int aio_put_req(struct kiocb *req)
647 struct kioctx *ctx = req->ki_ctx;
649 spin_lock_irq(&ctx->ctx_lock);
650 ret = __aio_put_req(ctx, req);
651 spin_unlock_irq(&ctx->ctx_lock);
654 EXPORT_SYMBOL(aio_put_req);
656 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
658 struct mm_struct *mm = current->mm;
659 struct kioctx *ctx, *ret = NULL;
660 struct hlist_node *n;
664 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
666 * RCU protects us against accessing freed memory but
667 * we have to be careful not to get a reference when the
668 * reference count already dropped to 0 (ctx->dead test
669 * is unreliable because of races).
671 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
682 * Queue up a kiocb to be retried. Assumes that the kiocb
683 * has already been marked as kicked, and places it on
684 * the retry run list for the corresponding ioctx, if it
685 * isn't already queued. Returns 1 if it actually queued
686 * the kiocb (to tell the caller to activate the work
687 * queue to process it), or 0, if it found that it was
690 static inline int __queue_kicked_iocb(struct kiocb *iocb)
692 struct kioctx *ctx = iocb->ki_ctx;
694 assert_spin_locked(&ctx->ctx_lock);
696 if (list_empty(&iocb->ki_run_list)) {
697 list_add_tail(&iocb->ki_run_list,
705 * This is the core aio execution routine. It is
706 * invoked both for initial i/o submission and
707 * subsequent retries via the aio_kick_handler.
708 * Expects to be invoked with iocb->ki_ctx->lock
709 * already held. The lock is released and reacquired
710 * as needed during processing.
712 * Calls the iocb retry method (already setup for the
713 * iocb on initial submission) for operation specific
714 * handling, but takes care of most of common retry
715 * execution details for a given iocb. The retry method
716 * needs to be non-blocking as far as possible, to avoid
717 * holding up other iocbs waiting to be serviced by the
718 * retry kernel thread.
720 * The trickier parts in this code have to do with
721 * ensuring that only one retry instance is in progress
722 * for a given iocb at any time. Providing that guarantee
723 * simplifies the coding of individual aio operations as
724 * it avoids various potential races.
726 static ssize_t aio_run_iocb(struct kiocb *iocb)
728 struct kioctx *ctx = iocb->ki_ctx;
729 ssize_t (*retry)(struct kiocb *);
732 if (!(retry = iocb->ki_retry)) {
733 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
738 * We don't want the next retry iteration for this
739 * operation to start until this one has returned and
740 * updated the iocb state. However, wait_queue functions
741 * can trigger a kick_iocb from interrupt context in the
742 * meantime, indicating that data is available for the next
743 * iteration. We want to remember that and enable the
744 * next retry iteration _after_ we are through with
747 * So, in order to be able to register a "kick", but
748 * prevent it from being queued now, we clear the kick
749 * flag, but make the kick code *think* that the iocb is
750 * still on the run list until we are actually done.
751 * When we are done with this iteration, we check if
752 * the iocb was kicked in the meantime and if so, queue
756 kiocbClearKicked(iocb);
759 * This is so that aio_complete knows it doesn't need to
760 * pull the iocb off the run list (We can't just call
761 * INIT_LIST_HEAD because we don't want a kick_iocb to
762 * queue this on the run list yet)
764 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
765 spin_unlock_irq(&ctx->ctx_lock);
767 /* Quit retrying if the i/o has been cancelled */
768 if (kiocbIsCancelled(iocb)) {
770 aio_complete(iocb, ret, 0);
771 /* must not access the iocb after this */
776 * Now we are all set to call the retry method in async
781 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
783 * There's no easy way to restart the syscall since other AIO's
784 * may be already running. Just fail this IO with EINTR.
786 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
787 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
789 aio_complete(iocb, ret, 0);
792 spin_lock_irq(&ctx->ctx_lock);
794 if (-EIOCBRETRY == ret) {
796 * OK, now that we are done with this iteration
797 * and know that there is more left to go,
798 * this is where we let go so that a subsequent
799 * "kick" can start the next iteration
802 /* will make __queue_kicked_iocb succeed from here on */
803 INIT_LIST_HEAD(&iocb->ki_run_list);
804 /* we must queue the next iteration ourselves, if it
805 * has already been kicked */
806 if (kiocbIsKicked(iocb)) {
807 __queue_kicked_iocb(iocb);
810 * __queue_kicked_iocb will always return 1 here, because
811 * iocb->ki_run_list is empty at this point so it should
812 * be safe to unconditionally queue the context into the
823 * Process all pending retries queued on the ioctx
825 * Assumes it is operating within the aio issuer's mm
828 static int __aio_run_iocbs(struct kioctx *ctx)
831 struct list_head run_list;
833 assert_spin_locked(&ctx->ctx_lock);
835 list_replace_init(&ctx->run_list, &run_list);
836 while (!list_empty(&run_list)) {
837 iocb = list_entry(run_list.next, struct kiocb,
839 list_del(&iocb->ki_run_list);
841 * Hold an extra reference while retrying i/o.
843 iocb->ki_users++; /* grab extra reference */
845 __aio_put_req(ctx, iocb);
847 if (!list_empty(&ctx->run_list))
852 static void aio_queue_work(struct kioctx * ctx)
854 unsigned long timeout;
856 * if someone is waiting, get the work started right
857 * away, otherwise, use a longer delay
860 if (waitqueue_active(&ctx->wait))
864 queue_delayed_work(aio_wq, &ctx->wq, timeout);
869 * Process all pending retries queued on the ioctx
870 * run list, and keep running them until the list
872 * Assumes it is operating within the aio issuer's mm context.
874 static inline void aio_run_all_iocbs(struct kioctx *ctx)
876 spin_lock_irq(&ctx->ctx_lock);
877 while (__aio_run_iocbs(ctx))
879 spin_unlock_irq(&ctx->ctx_lock);
884 * Work queue handler triggered to process pending
885 * retries on an ioctx. Takes on the aio issuer's
886 * mm context before running the iocbs, so that
887 * copy_xxx_user operates on the issuer's address
889 * Run on aiod's context.
891 static void aio_kick_handler(struct work_struct *work)
893 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
894 mm_segment_t oldfs = get_fs();
895 struct mm_struct *mm;
900 spin_lock_irq(&ctx->ctx_lock);
901 requeue =__aio_run_iocbs(ctx);
903 spin_unlock_irq(&ctx->ctx_lock);
907 * we're in a worker thread already, don't use queue_delayed_work,
910 queue_delayed_work(aio_wq, &ctx->wq, 0);
915 * Called by kick_iocb to queue the kiocb for retry
916 * and if required activate the aio work queue to process
919 static void try_queue_kicked_iocb(struct kiocb *iocb)
921 struct kioctx *ctx = iocb->ki_ctx;
925 spin_lock_irqsave(&ctx->ctx_lock, flags);
926 /* set this inside the lock so that we can't race with aio_run_iocb()
927 * testing it and putting the iocb on the run list under the lock */
928 if (!kiocbTryKick(iocb))
929 run = __queue_kicked_iocb(iocb);
930 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
937 * Called typically from a wait queue callback context
938 * to trigger a retry of the iocb.
939 * The retry is usually executed by aio workqueue
940 * threads (See aio_kick_handler).
942 void kick_iocb(struct kiocb *iocb)
944 /* sync iocbs are easy: they can only ever be executing from a
946 if (is_sync_kiocb(iocb)) {
947 kiocbSetKicked(iocb);
948 wake_up_process(iocb->ki_obj.tsk);
952 try_queue_kicked_iocb(iocb);
954 EXPORT_SYMBOL(kick_iocb);
957 * Called when the io request on the given iocb is complete.
958 * Returns true if this is the last user of the request. The
959 * only other user of the request can be the cancellation code.
961 int aio_complete(struct kiocb *iocb, long res, long res2)
963 struct kioctx *ctx = iocb->ki_ctx;
964 struct aio_ring_info *info;
965 struct aio_ring *ring;
966 struct io_event *event;
972 * Special case handling for sync iocbs:
973 * - events go directly into the iocb for fast handling
974 * - the sync task with the iocb in its stack holds the single iocb
975 * ref, no other paths have a way to get another ref
976 * - the sync task helpfully left a reference to itself in the iocb
978 if (is_sync_kiocb(iocb)) {
979 BUG_ON(iocb->ki_users != 1);
980 iocb->ki_user_data = res;
982 wake_up_process(iocb->ki_obj.tsk);
986 info = &ctx->ring_info;
988 /* add a completion event to the ring buffer.
989 * must be done holding ctx->ctx_lock to prevent
990 * other code from messing with the tail
991 * pointer since we might be called from irq
994 spin_lock_irqsave(&ctx->ctx_lock, flags);
996 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
997 list_del_init(&iocb->ki_run_list);
1000 * cancelled requests don't get events, userland was given one
1001 * when the event got cancelled.
1003 if (kiocbIsCancelled(iocb))
1006 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
1009 event = aio_ring_event(info, tail, KM_IRQ0);
1010 if (++tail >= info->nr)
1013 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1014 event->data = iocb->ki_user_data;
1018 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1019 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1022 /* after flagging the request as done, we
1023 * must never even look at it again
1025 smp_wmb(); /* make event visible before updating tail */
1030 put_aio_ring_event(event, KM_IRQ0);
1031 kunmap_atomic(ring, KM_IRQ1);
1033 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1036 * Check if the user asked us to deliver the result through an
1037 * eventfd. The eventfd_signal() function is safe to be called
1040 if (iocb->ki_eventfd != NULL)
1041 eventfd_signal(iocb->ki_eventfd, 1);
1044 /* everything turned out well, dispose of the aiocb. */
1045 ret = __aio_put_req(ctx, iocb);
1048 * We have to order our ring_info tail store above and test
1049 * of the wait list below outside the wait lock. This is
1050 * like in wake_up_bit() where clearing a bit has to be
1051 * ordered with the unlocked test.
1055 if (waitqueue_active(&ctx->wait))
1056 wake_up(&ctx->wait);
1058 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1061 EXPORT_SYMBOL(aio_complete);
1064 * Pull an event off of the ioctx's event ring. Returns the number of
1065 * events fetched (0 or 1 ;-)
1066 * FIXME: make this use cmpxchg.
1067 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1069 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1071 struct aio_ring_info *info = &ioctx->ring_info;
1072 struct aio_ring *ring;
1076 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1077 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1078 (unsigned long)ring->head, (unsigned long)ring->tail,
1079 (unsigned long)ring->nr);
1081 if (ring->head == ring->tail)
1084 spin_lock(&info->ring_lock);
1086 head = ring->head % info->nr;
1087 if (head != ring->tail) {
1088 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1090 head = (head + 1) % info->nr;
1091 smp_mb(); /* finish reading the event before updatng the head */
1094 put_aio_ring_event(evp, KM_USER1);
1096 spin_unlock(&info->ring_lock);
1099 kunmap_atomic(ring, KM_USER0);
1100 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1101 (unsigned long)ring->head, (unsigned long)ring->tail);
1105 struct aio_timeout {
1106 struct timer_list timer;
1108 struct task_struct *p;
1111 static void timeout_func(unsigned long data)
1113 struct aio_timeout *to = (struct aio_timeout *)data;
1116 wake_up_process(to->p);
1119 static inline void init_timeout(struct aio_timeout *to)
1121 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1126 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1127 const struct timespec *ts)
1129 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1130 if (time_after(to->timer.expires, jiffies))
1131 add_timer(&to->timer);
1136 static inline void clear_timeout(struct aio_timeout *to)
1138 del_singleshot_timer_sync(&to->timer);
1141 static int read_events(struct kioctx *ctx,
1142 long min_nr, long nr,
1143 struct io_event __user *event,
1144 struct timespec __user *timeout)
1146 long start_jiffies = jiffies;
1147 struct task_struct *tsk = current;
1148 DECLARE_WAITQUEUE(wait, tsk);
1151 struct io_event ent;
1152 struct aio_timeout to;
1155 /* needed to zero any padding within an entry (there shouldn't be
1156 * any, but C is fun!
1158 memset(&ent, 0, sizeof(ent));
1161 while (likely(i < nr)) {
1162 ret = aio_read_evt(ctx, &ent);
1163 if (unlikely(ret <= 0))
1166 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1167 ent.data, ent.obj, ent.res, ent.res2);
1169 /* Could we split the check in two? */
1171 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1172 dprintk("aio: lost an event due to EFAULT.\n");
1177 /* Good, event copied to userland, update counts. */
1189 /* racey check, but it gets redone */
1190 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1192 aio_run_all_iocbs(ctx);
1200 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1203 set_timeout(start_jiffies, &to, &ts);
1206 while (likely(i < nr)) {
1207 add_wait_queue_exclusive(&ctx->wait, &wait);
1209 set_task_state(tsk, TASK_INTERRUPTIBLE);
1210 ret = aio_read_evt(ctx, &ent);
1215 if (unlikely(ctx->dead)) {
1219 if (to.timed_out) /* Only check after read evt */
1221 /* Try to only show up in io wait if there are ops
1223 if (ctx->reqs_active)
1227 if (signal_pending(tsk)) {
1231 /*ret = aio_read_evt(ctx, &ent);*/
1234 set_task_state(tsk, TASK_RUNNING);
1235 remove_wait_queue(&ctx->wait, &wait);
1237 if (unlikely(ret <= 0))
1241 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1242 dprintk("aio: lost an event due to EFAULT.\n");
1246 /* Good, event copied to userland, update counts. */
1254 destroy_timer_on_stack(&to.timer);
1258 /* Take an ioctx and remove it from the list of ioctx's. Protects
1259 * against races with itself via ->dead.
1261 static void io_destroy(struct kioctx *ioctx)
1263 struct mm_struct *mm = current->mm;
1266 /* delete the entry from the list is someone else hasn't already */
1267 spin_lock(&mm->ioctx_lock);
1268 was_dead = ioctx->dead;
1270 hlist_del_rcu(&ioctx->list);
1271 spin_unlock(&mm->ioctx_lock);
1273 dprintk("aio_release(%p)\n", ioctx);
1274 if (likely(!was_dead))
1275 put_ioctx(ioctx); /* twice for the list */
1277 aio_cancel_all(ioctx);
1278 wait_for_all_aios(ioctx);
1281 * Wake up any waiters. The setting of ctx->dead must be seen
1282 * by other CPUs at this point. Right now, we rely on the
1283 * locking done by the above calls to ensure this consistency.
1285 wake_up_all(&ioctx->wait);
1286 put_ioctx(ioctx); /* once for the lookup */
1290 * Create an aio_context capable of receiving at least nr_events.
1291 * ctxp must not point to an aio_context that already exists, and
1292 * must be initialized to 0 prior to the call. On successful
1293 * creation of the aio_context, *ctxp is filled in with the resulting
1294 * handle. May fail with -EINVAL if *ctxp is not initialized,
1295 * if the specified nr_events exceeds internal limits. May fail
1296 * with -EAGAIN if the specified nr_events exceeds the user's limit
1297 * of available events. May fail with -ENOMEM if insufficient kernel
1298 * resources are available. May fail with -EFAULT if an invalid
1299 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1302 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1304 struct kioctx *ioctx = NULL;
1308 ret = get_user(ctx, ctxp);
1313 if (unlikely(ctx || nr_events == 0)) {
1314 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1319 ioctx = ioctx_alloc(nr_events);
1320 ret = PTR_ERR(ioctx);
1321 if (!IS_ERR(ioctx)) {
1322 ret = put_user(ioctx->user_id, ctxp);
1335 * Destroy the aio_context specified. May cancel any outstanding
1336 * AIOs and block on completion. Will fail with -ENOSYS if not
1337 * implemented. May fail with -EINVAL if the context pointed to
1340 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1342 struct kioctx *ioctx = lookup_ioctx(ctx);
1343 if (likely(NULL != ioctx)) {
1347 pr_debug("EINVAL: io_destroy: invalid context id\n");
1351 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1353 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1357 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1358 ssize_t this = min((ssize_t)iov->iov_len, ret);
1359 iov->iov_base += this;
1360 iov->iov_len -= this;
1361 iocb->ki_left -= this;
1363 if (iov->iov_len == 0) {
1369 /* the caller should not have done more io than what fit in
1370 * the remaining iovecs */
1371 BUG_ON(ret > 0 && iocb->ki_left == 0);
1374 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1376 struct file *file = iocb->ki_filp;
1377 struct address_space *mapping = file->f_mapping;
1378 struct inode *inode = mapping->host;
1379 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1380 unsigned long, loff_t);
1382 unsigned short opcode;
1384 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1385 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1386 rw_op = file->f_op->aio_read;
1387 opcode = IOCB_CMD_PREADV;
1389 rw_op = file->f_op->aio_write;
1390 opcode = IOCB_CMD_PWRITEV;
1393 /* This matches the pread()/pwrite() logic */
1394 if (iocb->ki_pos < 0)
1398 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1399 iocb->ki_nr_segs - iocb->ki_cur_seg,
1402 aio_advance_iovec(iocb, ret);
1404 /* retry all partial writes. retry partial reads as long as its a
1406 } while (ret > 0 && iocb->ki_left > 0 &&
1407 (opcode == IOCB_CMD_PWRITEV ||
1408 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1410 /* This means we must have transferred all that we could */
1411 /* No need to retry anymore */
1412 if ((ret == 0) || (iocb->ki_left == 0))
1413 ret = iocb->ki_nbytes - iocb->ki_left;
1415 /* If we managed to write some out we return that, rather than
1416 * the eventual error. */
1417 if (opcode == IOCB_CMD_PWRITEV
1418 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1419 && iocb->ki_nbytes - iocb->ki_left)
1420 ret = iocb->ki_nbytes - iocb->ki_left;
1425 static ssize_t aio_fdsync(struct kiocb *iocb)
1427 struct file *file = iocb->ki_filp;
1428 ssize_t ret = -EINVAL;
1430 if (file->f_op->aio_fsync)
1431 ret = file->f_op->aio_fsync(iocb, 1);
1435 static ssize_t aio_fsync(struct kiocb *iocb)
1437 struct file *file = iocb->ki_filp;
1438 ssize_t ret = -EINVAL;
1440 if (file->f_op->aio_fsync)
1441 ret = file->f_op->aio_fsync(iocb, 0);
1445 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1449 #ifdef CONFIG_COMPAT
1451 ret = compat_rw_copy_check_uvector(type,
1452 (struct compat_iovec __user *)kiocb->ki_buf,
1453 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1454 &kiocb->ki_iovec, 1);
1457 ret = rw_copy_check_uvector(type,
1458 (struct iovec __user *)kiocb->ki_buf,
1459 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1460 &kiocb->ki_iovec, 1);
1464 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1465 kiocb->ki_cur_seg = 0;
1466 /* ki_nbytes/left now reflect bytes instead of segs */
1467 kiocb->ki_nbytes = ret;
1468 kiocb->ki_left = ret;
1475 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1477 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1478 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1479 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1480 kiocb->ki_nr_segs = 1;
1481 kiocb->ki_cur_seg = 0;
1487 * Performs the initial checks and aio retry method
1488 * setup for the kiocb at the time of io submission.
1490 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1492 struct file *file = kiocb->ki_filp;
1495 switch (kiocb->ki_opcode) {
1496 case IOCB_CMD_PREAD:
1498 if (unlikely(!(file->f_mode & FMODE_READ)))
1501 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1504 ret = security_file_permission(file, MAY_READ);
1507 ret = aio_setup_single_vector(kiocb);
1511 if (file->f_op->aio_read)
1512 kiocb->ki_retry = aio_rw_vect_retry;
1514 case IOCB_CMD_PWRITE:
1516 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1519 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1522 ret = security_file_permission(file, MAY_WRITE);
1525 ret = aio_setup_single_vector(kiocb);
1529 if (file->f_op->aio_write)
1530 kiocb->ki_retry = aio_rw_vect_retry;
1532 case IOCB_CMD_PREADV:
1534 if (unlikely(!(file->f_mode & FMODE_READ)))
1536 ret = security_file_permission(file, MAY_READ);
1539 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1543 if (file->f_op->aio_read)
1544 kiocb->ki_retry = aio_rw_vect_retry;
1546 case IOCB_CMD_PWRITEV:
1548 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1550 ret = security_file_permission(file, MAY_WRITE);
1553 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1557 if (file->f_op->aio_write)
1558 kiocb->ki_retry = aio_rw_vect_retry;
1560 case IOCB_CMD_FDSYNC:
1562 if (file->f_op->aio_fsync)
1563 kiocb->ki_retry = aio_fdsync;
1565 case IOCB_CMD_FSYNC:
1567 if (file->f_op->aio_fsync)
1568 kiocb->ki_retry = aio_fsync;
1571 dprintk("EINVAL: io_submit: no operation provided\n");
1575 if (!kiocb->ki_retry)
1581 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1582 struct iocb *iocb, struct kiocb_batch *batch,
1589 /* enforce forwards compatibility on users */
1590 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1591 pr_debug("EINVAL: io_submit: reserve field set\n");
1595 /* prevent overflows */
1597 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1598 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1599 ((ssize_t)iocb->aio_nbytes < 0)
1601 pr_debug("EINVAL: io_submit: overflow check\n");
1605 file = fget(iocb->aio_fildes);
1606 if (unlikely(!file))
1609 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1610 if (unlikely(!req)) {
1614 req->ki_filp = file;
1615 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1617 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1618 * instance of the file* now. The file descriptor must be
1619 * an eventfd() fd, and will be signaled for each completed
1620 * event using the eventfd_signal() function.
1622 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1623 if (IS_ERR(req->ki_eventfd)) {
1624 ret = PTR_ERR(req->ki_eventfd);
1625 req->ki_eventfd = NULL;
1630 ret = put_user(req->ki_key, &user_iocb->aio_key);
1631 if (unlikely(ret)) {
1632 dprintk("EFAULT: aio_key\n");
1636 req->ki_obj.user = user_iocb;
1637 req->ki_user_data = iocb->aio_data;
1638 req->ki_pos = iocb->aio_offset;
1640 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1641 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1642 req->ki_opcode = iocb->aio_lio_opcode;
1644 ret = aio_setup_iocb(req, compat);
1649 spin_lock_irq(&ctx->ctx_lock);
1651 * We could have raced with io_destroy() and are currently holding a
1652 * reference to ctx which should be destroyed. We cannot submit IO
1653 * since ctx gets freed as soon as io_submit() puts its reference. The
1654 * check here is reliable: io_destroy() sets ctx->dead before waiting
1655 * for outstanding IO and the barrier between these two is realized by
1656 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1657 * increment ctx->reqs_active before checking for ctx->dead and the
1658 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1659 * don't see ctx->dead set here, io_destroy() waits for our IO to
1663 spin_unlock_irq(&ctx->ctx_lock);
1668 if (!list_empty(&ctx->run_list)) {
1669 /* drain the run list */
1670 while (__aio_run_iocbs(ctx))
1673 spin_unlock_irq(&ctx->ctx_lock);
1675 aio_put_req(req); /* drop extra ref to req */
1679 aio_put_req(req); /* drop extra ref to req */
1680 aio_put_req(req); /* drop i/o ref to req */
1684 long do_io_submit(aio_context_t ctx_id, long nr,
1685 struct iocb __user *__user *iocbpp, bool compat)
1690 struct blk_plug plug;
1691 struct kiocb_batch batch;
1693 if (unlikely(nr < 0))
1696 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1697 nr = LONG_MAX/sizeof(*iocbpp);
1699 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1702 ctx = lookup_ioctx(ctx_id);
1703 if (unlikely(!ctx)) {
1704 pr_debug("EINVAL: io_submit: invalid context id\n");
1708 kiocb_batch_init(&batch, nr);
1710 blk_start_plug(&plug);
1713 * AKPM: should this return a partial result if some of the IOs were
1714 * successfully submitted?
1716 for (i=0; i<nr; i++) {
1717 struct iocb __user *user_iocb;
1720 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1725 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1730 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1734 blk_finish_plug(&plug);
1736 kiocb_batch_free(ctx, &batch);
1742 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1743 * the number of iocbs queued. May return -EINVAL if the aio_context
1744 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1745 * *iocbpp[0] is not properly initialized, if the operation specified
1746 * is invalid for the file descriptor in the iocb. May fail with
1747 * -EFAULT if any of the data structures point to invalid data. May
1748 * fail with -EBADF if the file descriptor specified in the first
1749 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1750 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1751 * fail with -ENOSYS if not implemented.
1753 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1754 struct iocb __user * __user *, iocbpp)
1756 return do_io_submit(ctx_id, nr, iocbpp, 0);
1760 * Finds a given iocb for cancellation.
1762 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1765 struct list_head *pos;
1767 assert_spin_locked(&ctx->ctx_lock);
1769 /* TODO: use a hash or array, this sucks. */
1770 list_for_each(pos, &ctx->active_reqs) {
1771 struct kiocb *kiocb = list_kiocb(pos);
1772 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1779 * Attempts to cancel an iocb previously passed to io_submit. If
1780 * the operation is successfully cancelled, the resulting event is
1781 * copied into the memory pointed to by result without being placed
1782 * into the completion queue and 0 is returned. May fail with
1783 * -EFAULT if any of the data structures pointed to are invalid.
1784 * May fail with -EINVAL if aio_context specified by ctx_id is
1785 * invalid. May fail with -EAGAIN if the iocb specified was not
1786 * cancelled. Will fail with -ENOSYS if not implemented.
1788 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1789 struct io_event __user *, result)
1791 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1793 struct kiocb *kiocb;
1797 ret = get_user(key, &iocb->aio_key);
1801 ctx = lookup_ioctx(ctx_id);
1805 spin_lock_irq(&ctx->ctx_lock);
1807 kiocb = lookup_kiocb(ctx, iocb, key);
1808 if (kiocb && kiocb->ki_cancel) {
1809 cancel = kiocb->ki_cancel;
1811 kiocbSetCancelled(kiocb);
1814 spin_unlock_irq(&ctx->ctx_lock);
1816 if (NULL != cancel) {
1817 struct io_event tmp;
1818 pr_debug("calling cancel\n");
1819 memset(&tmp, 0, sizeof(tmp));
1820 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1821 tmp.data = kiocb->ki_user_data;
1822 ret = cancel(kiocb, &tmp);
1824 /* Cancellation succeeded -- copy the result
1825 * into the user's buffer.
1827 if (copy_to_user(result, &tmp, sizeof(tmp)))
1839 * Attempts to read at least min_nr events and up to nr events from
1840 * the completion queue for the aio_context specified by ctx_id. If
1841 * it succeeds, the number of read events is returned. May fail with
1842 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1843 * out of range, if timeout is out of range. May fail with -EFAULT
1844 * if any of the memory specified is invalid. May return 0 or
1845 * < min_nr if the timeout specified by timeout has elapsed
1846 * before sufficient events are available, where timeout == NULL
1847 * specifies an infinite timeout. Note that the timeout pointed to by
1848 * timeout is relative and will be updated if not NULL and the
1849 * operation blocks. Will fail with -ENOSYS if not implemented.
1851 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1854 struct io_event __user *, events,
1855 struct timespec __user *, timeout)
1857 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1860 if (likely(ioctx)) {
1861 if (likely(min_nr <= nr && min_nr >= 0))
1862 ret = read_events(ioctx, min_nr, nr, events, timeout);
1866 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);