2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
11 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/export.h>
17 #include <linux/syscalls.h>
18 #include <linux/backing-dev.h>
19 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/compat.h>
39 #include <asm/kmap_types.h>
40 #include <asm/uaccess.h>
43 #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 BUG_ON(ctx->mm != current->mm);
102 vm_munmap(info->mmap_base, info->mmap_size);
105 if (info->ring_pages && info->ring_pages != info->internal_pages)
106 kfree(info->ring_pages);
107 info->ring_pages = NULL;
111 static int aio_setup_ring(struct kioctx *ctx)
113 struct aio_ring *ring;
114 struct aio_ring_info *info = &ctx->ring_info;
115 unsigned nr_events = ctx->max_reqs;
119 /* Compensate for the ring buffer's head/tail overlap entry */
120 nr_events += 2; /* 1 is required, 2 for good luck */
122 size = sizeof(struct aio_ring);
123 size += sizeof(struct io_event) * nr_events;
124 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
129 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
132 info->ring_pages = info->internal_pages;
133 if (nr_pages > AIO_RING_PAGES) {
134 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
135 if (!info->ring_pages)
139 info->mmap_size = nr_pages * PAGE_SIZE;
140 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
141 down_write(&ctx->mm->mmap_sem);
142 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
143 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
145 if (IS_ERR((void *)info->mmap_base)) {
146 up_write(&ctx->mm->mmap_sem);
152 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
153 info->nr_pages = get_user_pages(current, ctx->mm,
154 info->mmap_base, nr_pages,
155 1, 0, info->ring_pages, NULL);
156 up_write(&ctx->mm->mmap_sem);
158 if (unlikely(info->nr_pages != nr_pages)) {
163 ctx->user_id = info->mmap_base;
165 info->nr = nr_events; /* trusted copy */
167 ring = kmap_atomic(info->ring_pages[0]);
168 ring->nr = nr_events; /* user copy */
169 ring->id = ctx->user_id;
170 ring->head = ring->tail = 0;
171 ring->magic = AIO_RING_MAGIC;
172 ring->compat_features = AIO_RING_COMPAT_FEATURES;
173 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
174 ring->header_length = sizeof(struct aio_ring);
181 /* aio_ring_event: returns a pointer to the event at the given index from
182 * kmap_atomic(). Release the pointer with put_aio_ring_event();
184 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
185 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
186 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
188 #define aio_ring_event(info, nr) ({ \
189 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
190 struct io_event *__event; \
191 __event = kmap_atomic( \
192 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
193 __event += pos % AIO_EVENTS_PER_PAGE; \
197 #define put_aio_ring_event(event) do { \
198 struct io_event *__event = (event); \
200 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
203 static void ctx_rcu_free(struct rcu_head *head)
205 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
206 kmem_cache_free(kioctx_cachep, ctx);
210 * Called when the last user of an aio context has gone away,
211 * and the struct needs to be freed.
213 static void __put_ioctx(struct kioctx *ctx)
215 unsigned nr_events = ctx->max_reqs;
216 BUG_ON(ctx->reqs_active);
218 cancel_delayed_work_sync(&ctx->wq);
223 spin_lock(&aio_nr_lock);
224 BUG_ON(aio_nr - nr_events > aio_nr);
226 spin_unlock(&aio_nr_lock);
228 pr_debug("__put_ioctx: freeing %p\n", ctx);
229 call_rcu(&ctx->rcu_head, ctx_rcu_free);
232 static inline int try_get_ioctx(struct kioctx *kioctx)
234 return atomic_inc_not_zero(&kioctx->users);
237 static inline void put_ioctx(struct kioctx *kioctx)
239 BUG_ON(atomic_read(&kioctx->users) <= 0);
240 if (unlikely(atomic_dec_and_test(&kioctx->users)))
245 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
247 static struct kioctx *ioctx_alloc(unsigned nr_events)
249 struct mm_struct *mm;
253 /* Prevent overflows */
254 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
255 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
256 pr_debug("ENOMEM: nr_events too high\n");
257 return ERR_PTR(-EINVAL);
260 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
261 return ERR_PTR(-EAGAIN);
263 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
265 return ERR_PTR(-ENOMEM);
267 ctx->max_reqs = nr_events;
268 mm = ctx->mm = current->mm;
269 atomic_inc(&mm->mm_count);
271 atomic_set(&ctx->users, 2);
272 spin_lock_init(&ctx->ctx_lock);
273 spin_lock_init(&ctx->ring_info.ring_lock);
274 init_waitqueue_head(&ctx->wait);
276 INIT_LIST_HEAD(&ctx->active_reqs);
277 INIT_LIST_HEAD(&ctx->run_list);
278 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
280 if (aio_setup_ring(ctx) < 0)
283 /* limit the number of system wide aios */
284 spin_lock(&aio_nr_lock);
285 if (aio_nr + nr_events > aio_max_nr ||
286 aio_nr + nr_events < aio_nr) {
287 spin_unlock(&aio_nr_lock);
290 aio_nr += ctx->max_reqs;
291 spin_unlock(&aio_nr_lock);
293 /* now link into global list. */
294 spin_lock(&mm->ioctx_lock);
295 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
296 spin_unlock(&mm->ioctx_lock);
298 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
299 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
307 kmem_cache_free(kioctx_cachep, ctx);
308 dprintk("aio: error allocating ioctx %d\n", err);
313 * Cancels all outstanding aio requests on an aio context. Used
314 * when the processes owning a context have all exited to encourage
315 * the rapid destruction of the kioctx.
317 static void kill_ctx(struct kioctx *ctx)
319 int (*cancel)(struct kiocb *, struct io_event *);
320 struct task_struct *tsk = current;
321 DECLARE_WAITQUEUE(wait, tsk);
324 spin_lock_irq(&ctx->ctx_lock);
326 while (!list_empty(&ctx->active_reqs)) {
327 struct list_head *pos = ctx->active_reqs.next;
328 struct kiocb *iocb = list_kiocb(pos);
329 list_del_init(&iocb->ki_list);
330 cancel = iocb->ki_cancel;
331 kiocbSetCancelled(iocb);
334 spin_unlock_irq(&ctx->ctx_lock);
336 spin_lock_irq(&ctx->ctx_lock);
340 if (!ctx->reqs_active)
343 add_wait_queue(&ctx->wait, &wait);
344 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
345 while (ctx->reqs_active) {
346 spin_unlock_irq(&ctx->ctx_lock);
348 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
349 spin_lock_irq(&ctx->ctx_lock);
351 __set_task_state(tsk, TASK_RUNNING);
352 remove_wait_queue(&ctx->wait, &wait);
355 spin_unlock_irq(&ctx->ctx_lock);
358 /* wait_on_sync_kiocb:
359 * Waits on the given sync kiocb to complete.
361 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
363 while (iocb->ki_users) {
364 set_current_state(TASK_UNINTERRUPTIBLE);
369 __set_current_state(TASK_RUNNING);
370 return iocb->ki_user_data;
372 EXPORT_SYMBOL(wait_on_sync_kiocb);
374 /* exit_aio: called when the last user of mm goes away. At this point,
375 * there is no way for any new requests to be submited or any of the
376 * io_* syscalls to be called on the context. However, there may be
377 * outstanding requests which hold references to the context; as they
378 * go away, they will call put_ioctx and release any pinned memory
379 * associated with the request (held via struct page * references).
381 void exit_aio(struct mm_struct *mm)
385 while (!hlist_empty(&mm->ioctx_list)) {
386 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
387 hlist_del_rcu(&ctx->list);
391 if (1 != atomic_read(&ctx->users))
393 "exit_aio:ioctx still alive: %d %d %d\n",
394 atomic_read(&ctx->users), ctx->dead,
397 * We don't need to bother with munmap() here -
398 * exit_mmap(mm) is coming and it'll unmap everything.
399 * Since aio_free_ring() uses non-zero ->mmap_size
400 * as indicator that it needs to unmap the area,
401 * just set it to 0; aio_free_ring() is the only
402 * place that uses ->mmap_size, so it's safe.
403 * That way we get all munmap done to current->mm -
404 * all other callers have ctx->mm == current->mm.
406 ctx->ring_info.mmap_size = 0;
412 * Allocate a slot for an aio request. Increments the users count
413 * of the kioctx so that the kioctx stays around until all requests are
414 * complete. Returns NULL if no requests are free.
416 * Returns with kiocb->users set to 2. The io submit code path holds
417 * an extra reference while submitting the i/o.
418 * This prevents races between the aio code path referencing the
419 * req (after submitting it) and aio_complete() freeing the req.
421 static struct kiocb *__aio_get_req(struct kioctx *ctx)
423 struct kiocb *req = NULL;
425 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
433 req->ki_cancel = NULL;
434 req->ki_retry = NULL;
437 req->ki_iovec = NULL;
438 INIT_LIST_HEAD(&req->ki_run_list);
439 req->ki_eventfd = NULL;
445 * struct kiocb's are allocated in batches to reduce the number of
446 * times the ctx lock is acquired and released.
448 #define KIOCB_BATCH_SIZE 32L
450 struct list_head head;
451 long count; /* number of requests left to allocate */
454 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
456 INIT_LIST_HEAD(&batch->head);
457 batch->count = total;
460 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
462 struct kiocb *req, *n;
464 if (list_empty(&batch->head))
467 spin_lock_irq(&ctx->ctx_lock);
468 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
469 list_del(&req->ki_batch);
470 list_del(&req->ki_list);
471 kmem_cache_free(kiocb_cachep, req);
474 if (unlikely(!ctx->reqs_active && ctx->dead))
475 wake_up_all(&ctx->wait);
476 spin_unlock_irq(&ctx->ctx_lock);
480 * Allocate a batch of kiocbs. This avoids taking and dropping the
481 * context lock a lot during setup.
483 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
485 unsigned short allocated, to_alloc;
487 bool called_fput = false;
488 struct kiocb *req, *n;
489 struct aio_ring *ring;
491 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
492 for (allocated = 0; allocated < to_alloc; allocated++) {
493 req = __aio_get_req(ctx);
495 /* allocation failed, go with what we've got */
497 list_add(&req->ki_batch, &batch->head);
504 spin_lock_irq(&ctx->ctx_lock);
505 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
507 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
509 if (avail == 0 && !called_fput) {
511 * Handle a potential starvation case. It is possible that
512 * we hold the last reference on a struct file, causing us
513 * to delay the final fput to non-irq context. In this case,
514 * ctx->reqs_active is artificially high. Calling the fput
515 * routine here may free up a slot in the event completion
516 * ring, allowing this allocation to succeed.
519 spin_unlock_irq(&ctx->ctx_lock);
520 aio_fput_routine(NULL);
525 if (avail < allocated) {
526 /* Trim back the number of requests. */
527 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
528 list_del(&req->ki_batch);
529 kmem_cache_free(kiocb_cachep, req);
530 if (--allocated <= avail)
535 batch->count -= allocated;
536 list_for_each_entry(req, &batch->head, ki_batch) {
537 list_add(&req->ki_list, &ctx->active_reqs);
542 spin_unlock_irq(&ctx->ctx_lock);
548 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
549 struct kiocb_batch *batch)
553 if (list_empty(&batch->head))
554 if (kiocb_batch_refill(ctx, batch) == 0)
556 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
557 list_del(&req->ki_batch);
561 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
563 assert_spin_locked(&ctx->ctx_lock);
565 if (req->ki_eventfd != NULL)
566 eventfd_ctx_put(req->ki_eventfd);
569 if (req->ki_iovec != &req->ki_inline_vec)
570 kfree(req->ki_iovec);
571 kmem_cache_free(kiocb_cachep, req);
574 if (unlikely(!ctx->reqs_active && ctx->dead))
575 wake_up_all(&ctx->wait);
578 static void aio_fput_routine(struct work_struct *data)
580 spin_lock_irq(&fput_lock);
581 while (likely(!list_empty(&fput_head))) {
582 struct kiocb *req = list_kiocb(fput_head.next);
583 struct kioctx *ctx = req->ki_ctx;
585 list_del(&req->ki_list);
586 spin_unlock_irq(&fput_lock);
588 /* Complete the fput(s) */
589 if (req->ki_filp != NULL)
592 /* Link the iocb into the context's free list */
594 spin_lock_irq(&ctx->ctx_lock);
595 really_put_req(ctx, req);
597 * at that point ctx might've been killed, but actual
600 spin_unlock_irq(&ctx->ctx_lock);
603 spin_lock_irq(&fput_lock);
605 spin_unlock_irq(&fput_lock);
609 * Returns true if this put was the last user of the request.
611 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
613 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
614 req, atomic_long_read(&req->ki_filp->f_count));
616 assert_spin_locked(&ctx->ctx_lock);
619 BUG_ON(req->ki_users < 0);
620 if (likely(req->ki_users))
622 list_del(&req->ki_list); /* remove from active_reqs */
623 req->ki_cancel = NULL;
624 req->ki_retry = NULL;
627 * Try to optimize the aio and eventfd file* puts, by avoiding to
628 * schedule work in case it is not final fput() time. In normal cases,
629 * we would not be holding the last reference to the file*, so
630 * this function will be executed w/out any aio kthread wakeup.
632 if (unlikely(!fput_atomic(req->ki_filp))) {
633 spin_lock(&fput_lock);
634 list_add(&req->ki_list, &fput_head);
635 spin_unlock(&fput_lock);
636 schedule_work(&fput_work);
639 really_put_req(ctx, req);
645 * Returns true if this put was the last user of the kiocb,
646 * false if the request is still in use.
648 int aio_put_req(struct kiocb *req)
650 struct kioctx *ctx = req->ki_ctx;
652 spin_lock_irq(&ctx->ctx_lock);
653 ret = __aio_put_req(ctx, req);
654 spin_unlock_irq(&ctx->ctx_lock);
657 EXPORT_SYMBOL(aio_put_req);
659 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
661 struct mm_struct *mm = current->mm;
662 struct kioctx *ctx, *ret = NULL;
663 struct hlist_node *n;
667 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
669 * RCU protects us against accessing freed memory but
670 * we have to be careful not to get a reference when the
671 * reference count already dropped to 0 (ctx->dead test
672 * is unreliable because of races).
674 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
685 * Queue up a kiocb to be retried. Assumes that the kiocb
686 * has already been marked as kicked, and places it on
687 * the retry run list for the corresponding ioctx, if it
688 * isn't already queued. Returns 1 if it actually queued
689 * the kiocb (to tell the caller to activate the work
690 * queue to process it), or 0, if it found that it was
693 static inline int __queue_kicked_iocb(struct kiocb *iocb)
695 struct kioctx *ctx = iocb->ki_ctx;
697 assert_spin_locked(&ctx->ctx_lock);
699 if (list_empty(&iocb->ki_run_list)) {
700 list_add_tail(&iocb->ki_run_list,
708 * This is the core aio execution routine. It is
709 * invoked both for initial i/o submission and
710 * subsequent retries via the aio_kick_handler.
711 * Expects to be invoked with iocb->ki_ctx->lock
712 * already held. The lock is released and reacquired
713 * as needed during processing.
715 * Calls the iocb retry method (already setup for the
716 * iocb on initial submission) for operation specific
717 * handling, but takes care of most of common retry
718 * execution details for a given iocb. The retry method
719 * needs to be non-blocking as far as possible, to avoid
720 * holding up other iocbs waiting to be serviced by the
721 * retry kernel thread.
723 * The trickier parts in this code have to do with
724 * ensuring that only one retry instance is in progress
725 * for a given iocb at any time. Providing that guarantee
726 * simplifies the coding of individual aio operations as
727 * it avoids various potential races.
729 static ssize_t aio_run_iocb(struct kiocb *iocb)
731 struct kioctx *ctx = iocb->ki_ctx;
732 ssize_t (*retry)(struct kiocb *);
735 if (!(retry = iocb->ki_retry)) {
736 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
741 * We don't want the next retry iteration for this
742 * operation to start until this one has returned and
743 * updated the iocb state. However, wait_queue functions
744 * can trigger a kick_iocb from interrupt context in the
745 * meantime, indicating that data is available for the next
746 * iteration. We want to remember that and enable the
747 * next retry iteration _after_ we are through with
750 * So, in order to be able to register a "kick", but
751 * prevent it from being queued now, we clear the kick
752 * flag, but make the kick code *think* that the iocb is
753 * still on the run list until we are actually done.
754 * When we are done with this iteration, we check if
755 * the iocb was kicked in the meantime and if so, queue
759 kiocbClearKicked(iocb);
762 * This is so that aio_complete knows it doesn't need to
763 * pull the iocb off the run list (We can't just call
764 * INIT_LIST_HEAD because we don't want a kick_iocb to
765 * queue this on the run list yet)
767 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
768 spin_unlock_irq(&ctx->ctx_lock);
770 /* Quit retrying if the i/o has been cancelled */
771 if (kiocbIsCancelled(iocb)) {
773 aio_complete(iocb, ret, 0);
774 /* must not access the iocb after this */
779 * Now we are all set to call the retry method in async
784 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
786 * There's no easy way to restart the syscall since other AIO's
787 * may be already running. Just fail this IO with EINTR.
789 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
790 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
792 aio_complete(iocb, ret, 0);
795 spin_lock_irq(&ctx->ctx_lock);
797 if (-EIOCBRETRY == ret) {
799 * OK, now that we are done with this iteration
800 * and know that there is more left to go,
801 * this is where we let go so that a subsequent
802 * "kick" can start the next iteration
805 /* will make __queue_kicked_iocb succeed from here on */
806 INIT_LIST_HEAD(&iocb->ki_run_list);
807 /* we must queue the next iteration ourselves, if it
808 * has already been kicked */
809 if (kiocbIsKicked(iocb)) {
810 __queue_kicked_iocb(iocb);
813 * __queue_kicked_iocb will always return 1 here, because
814 * iocb->ki_run_list is empty at this point so it should
815 * be safe to unconditionally queue the context into the
826 * Process all pending retries queued on the ioctx
828 * Assumes it is operating within the aio issuer's mm
831 static int __aio_run_iocbs(struct kioctx *ctx)
834 struct list_head run_list;
836 assert_spin_locked(&ctx->ctx_lock);
838 list_replace_init(&ctx->run_list, &run_list);
839 while (!list_empty(&run_list)) {
840 iocb = list_entry(run_list.next, struct kiocb,
842 list_del(&iocb->ki_run_list);
844 * Hold an extra reference while retrying i/o.
846 iocb->ki_users++; /* grab extra reference */
848 __aio_put_req(ctx, iocb);
850 if (!list_empty(&ctx->run_list))
855 static void aio_queue_work(struct kioctx * ctx)
857 unsigned long timeout;
859 * if someone is waiting, get the work started right
860 * away, otherwise, use a longer delay
863 if (waitqueue_active(&ctx->wait))
867 queue_delayed_work(aio_wq, &ctx->wq, timeout);
872 * Process all pending retries queued on the ioctx
873 * run list, and keep running them until the list
875 * Assumes it is operating within the aio issuer's mm context.
877 static inline void aio_run_all_iocbs(struct kioctx *ctx)
879 spin_lock_irq(&ctx->ctx_lock);
880 while (__aio_run_iocbs(ctx))
882 spin_unlock_irq(&ctx->ctx_lock);
887 * Work queue handler triggered to process pending
888 * retries on an ioctx. Takes on the aio issuer's
889 * mm context before running the iocbs, so that
890 * copy_xxx_user operates on the issuer's address
892 * Run on aiod's context.
894 static void aio_kick_handler(struct work_struct *work)
896 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
897 mm_segment_t oldfs = get_fs();
898 struct mm_struct *mm;
903 spin_lock_irq(&ctx->ctx_lock);
904 requeue =__aio_run_iocbs(ctx);
906 spin_unlock_irq(&ctx->ctx_lock);
910 * we're in a worker thread already; no point using non-zero delay
913 queue_delayed_work(aio_wq, &ctx->wq, 0);
918 * Called by kick_iocb to queue the kiocb for retry
919 * and if required activate the aio work queue to process
922 static void try_queue_kicked_iocb(struct kiocb *iocb)
924 struct kioctx *ctx = iocb->ki_ctx;
928 spin_lock_irqsave(&ctx->ctx_lock, flags);
929 /* set this inside the lock so that we can't race with aio_run_iocb()
930 * testing it and putting the iocb on the run list under the lock */
931 if (!kiocbTryKick(iocb))
932 run = __queue_kicked_iocb(iocb);
933 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
940 * Called typically from a wait queue callback context
941 * to trigger a retry of the iocb.
942 * The retry is usually executed by aio workqueue
943 * threads (See aio_kick_handler).
945 void kick_iocb(struct kiocb *iocb)
947 /* sync iocbs are easy: they can only ever be executing from a
949 if (is_sync_kiocb(iocb)) {
950 kiocbSetKicked(iocb);
951 wake_up_process(iocb->ki_obj.tsk);
955 try_queue_kicked_iocb(iocb);
957 EXPORT_SYMBOL(kick_iocb);
960 * Called when the io request on the given iocb is complete.
961 * Returns true if this is the last user of the request. The
962 * only other user of the request can be the cancellation code.
964 int aio_complete(struct kiocb *iocb, long res, long res2)
966 struct kioctx *ctx = iocb->ki_ctx;
967 struct aio_ring_info *info;
968 struct aio_ring *ring;
969 struct io_event *event;
975 * Special case handling for sync iocbs:
976 * - events go directly into the iocb for fast handling
977 * - the sync task with the iocb in its stack holds the single iocb
978 * ref, no other paths have a way to get another ref
979 * - the sync task helpfully left a reference to itself in the iocb
981 if (is_sync_kiocb(iocb)) {
982 BUG_ON(iocb->ki_users != 1);
983 iocb->ki_user_data = res;
985 wake_up_process(iocb->ki_obj.tsk);
989 info = &ctx->ring_info;
991 /* add a completion event to the ring buffer.
992 * must be done holding ctx->ctx_lock to prevent
993 * other code from messing with the tail
994 * pointer since we might be called from irq
997 spin_lock_irqsave(&ctx->ctx_lock, flags);
999 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
1000 list_del_init(&iocb->ki_run_list);
1003 * cancelled requests don't get events, userland was given one
1004 * when the event got cancelled.
1006 if (kiocbIsCancelled(iocb))
1009 ring = kmap_atomic(info->ring_pages[0]);
1012 event = aio_ring_event(info, tail);
1013 if (++tail >= info->nr)
1016 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1017 event->data = iocb->ki_user_data;
1021 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1022 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1025 /* after flagging the request as done, we
1026 * must never even look at it again
1028 smp_wmb(); /* make event visible before updating tail */
1033 put_aio_ring_event(event);
1034 kunmap_atomic(ring);
1036 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1039 * Check if the user asked us to deliver the result through an
1040 * eventfd. The eventfd_signal() function is safe to be called
1043 if (iocb->ki_eventfd != NULL)
1044 eventfd_signal(iocb->ki_eventfd, 1);
1047 /* everything turned out well, dispose of the aiocb. */
1048 ret = __aio_put_req(ctx, iocb);
1051 * We have to order our ring_info tail store above and test
1052 * of the wait list below outside the wait lock. This is
1053 * like in wake_up_bit() where clearing a bit has to be
1054 * ordered with the unlocked test.
1058 if (waitqueue_active(&ctx->wait))
1059 wake_up(&ctx->wait);
1062 if (ctx->file && waitqueue_active(&ctx->poll_wait))
1063 wake_up(&ctx->poll_wait);
1066 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1069 EXPORT_SYMBOL(aio_complete);
1072 * Pull an event off of the ioctx's event ring. Returns the number of
1073 * events fetched (0 or 1 ;-)
1074 * If ent parameter is 0, just returns the number of events that would
1076 * FIXME: make this use cmpxchg.
1077 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1079 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1081 struct aio_ring_info *info = &ioctx->ring_info;
1082 struct aio_ring *ring;
1086 ring = kmap_atomic(info->ring_pages[0]);
1087 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1088 (unsigned long)ring->head, (unsigned long)ring->tail,
1089 (unsigned long)ring->nr);
1091 if (ring->head == ring->tail)
1094 spin_lock(&info->ring_lock);
1096 head = ring->head % info->nr;
1097 if (head != ring->tail) {
1098 if (ent) { /* event requested */
1099 struct io_event *evp = aio_ring_event(info, head);
1101 head = (head + 1) % info->nr;
1102 /* finish reading the event before updatng the head */
1106 put_aio_ring_event(evp);
1107 } else /* only need to know availability */
1110 spin_unlock(&info->ring_lock);
1113 kunmap_atomic(ring);
1114 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1115 (unsigned long)ring->head, (unsigned long)ring->tail);
1119 struct aio_timeout {
1120 struct timer_list timer;
1122 struct task_struct *p;
1125 static void timeout_func(unsigned long data)
1127 struct aio_timeout *to = (struct aio_timeout *)data;
1130 wake_up_process(to->p);
1133 static inline void init_timeout(struct aio_timeout *to)
1135 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1140 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1141 const struct timespec *ts)
1143 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1144 if (time_after(to->timer.expires, jiffies))
1145 add_timer(&to->timer);
1150 static inline void clear_timeout(struct aio_timeout *to)
1152 del_singleshot_timer_sync(&to->timer);
1155 static int read_events(struct kioctx *ctx,
1156 long min_nr, long nr,
1157 struct io_event __user *event,
1158 struct timespec __user *timeout)
1160 long start_jiffies = jiffies;
1161 struct task_struct *tsk = current;
1162 DECLARE_WAITQUEUE(wait, tsk);
1165 struct io_event ent;
1166 struct aio_timeout to;
1169 /* needed to zero any padding within an entry (there shouldn't be
1170 * any, but C is fun!
1172 memset(&ent, 0, sizeof(ent));
1175 while (likely(i < nr)) {
1176 ret = aio_read_evt(ctx, &ent);
1177 if (unlikely(ret <= 0))
1180 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1181 ent.data, ent.obj, ent.res, ent.res2);
1183 /* Could we split the check in two? */
1185 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1186 dprintk("aio: lost an event due to EFAULT.\n");
1191 /* Good, event copied to userland, update counts. */
1203 /* racey check, but it gets redone */
1204 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1206 aio_run_all_iocbs(ctx);
1214 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1217 set_timeout(start_jiffies, &to, &ts);
1220 while (likely(i < nr)) {
1221 add_wait_queue_exclusive(&ctx->wait, &wait);
1223 set_task_state(tsk, TASK_INTERRUPTIBLE);
1224 ret = aio_read_evt(ctx, &ent);
1229 if (unlikely(ctx->dead)) {
1233 if (to.timed_out) /* Only check after read evt */
1235 /* Try to only show up in io wait if there are ops
1237 if (ctx->reqs_active)
1241 if (signal_pending(tsk)) {
1245 /*ret = aio_read_evt(ctx, &ent);*/
1248 set_task_state(tsk, TASK_RUNNING);
1249 remove_wait_queue(&ctx->wait, &wait);
1251 if (unlikely(ret <= 0))
1255 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1256 dprintk("aio: lost an event due to EFAULT.\n");
1260 /* Good, event copied to userland, update counts. */
1268 destroy_timer_on_stack(&to.timer);
1272 /* Take an ioctx and remove it from the list of ioctx's. Protects
1273 * against races with itself via ->dead.
1275 static void io_destroy(struct kioctx *ioctx)
1277 struct mm_struct *mm = current->mm;
1280 /* delete the entry from the list is someone else hasn't already */
1281 spin_lock(&mm->ioctx_lock);
1282 was_dead = ioctx->dead;
1284 hlist_del_rcu(&ioctx->list);
1285 spin_unlock(&mm->ioctx_lock);
1287 dprintk("aio_release(%p)\n", ioctx);
1288 if (likely(!was_dead))
1289 put_ioctx(ioctx); /* twice for the list */
1294 /* forget the poll file, but it's up to the user to close it */
1297 ioctx->file->private_data = 0;
1303 * Wake up any waiters. The setting of ctx->dead must be seen
1304 * by other CPUs at this point. Right now, we rely on the
1305 * locking done by the above calls to ensure this consistency.
1307 wake_up_all(&ioctx->wait);
1312 static int aio_queue_fd_close(struct inode *inode, struct file *file)
1314 struct kioctx *ioctx = file->private_data;
1316 file->private_data = 0;
1317 spin_lock_irq(&ioctx->ctx_lock);
1319 spin_unlock_irq(&ioctx->ctx_lock);
1325 static unsigned int aio_queue_fd_poll(struct file *file, poll_table *wait)
1326 { unsigned int pollflags = 0;
1327 struct kioctx *ioctx = file->private_data;
1331 spin_lock_irq(&ioctx->ctx_lock);
1332 /* Insert inside our poll wait queue */
1333 poll_wait(file, &ioctx->poll_wait, wait);
1335 /* Check our condition */
1336 if (aio_read_evt(ioctx, 0))
1337 pollflags = POLLIN | POLLRDNORM;
1338 spin_unlock_irq(&ioctx->ctx_lock);
1344 static const struct file_operations aioq_fops = {
1345 .release = aio_queue_fd_close,
1346 .poll = aio_queue_fd_poll
1350 * Create a file descriptor that can be used to poll the event queue.
1351 * Based on the excellent epoll code.
1354 static int make_aio_fd(struct kioctx *ioctx)
1359 fd = anon_inode_getfd("[aioq]", &aioq_fops, ioctx, 0);
1363 /* associate the file with the IO context */
1367 file->private_data = ioctx;
1369 init_waitqueue_head(&ioctx->poll_wait);
1375 * Create an aio_context capable of receiving at least nr_events.
1376 * ctxp must not point to an aio_context that already exists, and
1377 * must be initialized to 0 prior to the call. On successful
1378 * creation of the aio_context, *ctxp is filled in with the resulting
1379 * handle. May fail with -EINVAL if *ctxp is not initialized,
1380 * if the specified nr_events exceeds internal limits. May fail
1381 * with -EAGAIN if the specified nr_events exceeds the user's limit
1382 * of available events. May fail with -ENOMEM if insufficient kernel
1383 * resources are available. May fail with -EFAULT if an invalid
1384 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1387 * To request a selectable fd, the user context has to be initialized
1388 * to 1, instead of 0, and the return value is the fd.
1389 * This keeps the system call compatible, since a non-zero value
1390 * was not allowed so far.
1392 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1394 struct kioctx *ioctx = NULL;
1399 ret = get_user(ctx, ctxp);
1410 if (unlikely(ctx || nr_events == 0)) {
1411 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1416 ioctx = ioctx_alloc(nr_events);
1417 ret = PTR_ERR(ioctx);
1418 if (!IS_ERR(ioctx)) {
1419 ret = put_user(ioctx->user_id, ctxp);
1421 if (make_fd && !ret)
1422 ret = make_aio_fd(ioctx);
1434 * Destroy the aio_context specified. May cancel any outstanding
1435 * AIOs and block on completion. Will fail with -ENOSYS if not
1436 * implemented. May fail with -EINVAL if the context pointed to
1439 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1441 struct kioctx *ioctx = lookup_ioctx(ctx);
1442 if (likely(NULL != ioctx)) {
1447 pr_debug("EINVAL: io_destroy: invalid context id\n");
1451 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1453 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1457 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1458 ssize_t this = min((ssize_t)iov->iov_len, ret);
1459 iov->iov_base += this;
1460 iov->iov_len -= this;
1461 iocb->ki_left -= this;
1463 if (iov->iov_len == 0) {
1469 /* the caller should not have done more io than what fit in
1470 * the remaining iovecs */
1471 BUG_ON(ret > 0 && iocb->ki_left == 0);
1474 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1476 struct file *file = iocb->ki_filp;
1477 struct address_space *mapping = file->f_mapping;
1478 struct inode *inode = mapping->host;
1479 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1480 unsigned long, loff_t);
1482 unsigned short opcode;
1484 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1485 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1486 rw_op = file->f_op->aio_read;
1487 opcode = IOCB_CMD_PREADV;
1489 rw_op = file->f_op->aio_write;
1490 opcode = IOCB_CMD_PWRITEV;
1493 /* This matches the pread()/pwrite() logic */
1494 if (iocb->ki_pos < 0)
1498 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1499 iocb->ki_nr_segs - iocb->ki_cur_seg,
1502 aio_advance_iovec(iocb, ret);
1504 /* retry all partial writes. retry partial reads as long as its a
1506 } while (ret > 0 && iocb->ki_left > 0 &&
1507 (opcode == IOCB_CMD_PWRITEV ||
1508 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1510 /* This means we must have transferred all that we could */
1511 /* No need to retry anymore */
1512 if ((ret == 0) || (iocb->ki_left == 0))
1513 ret = iocb->ki_nbytes - iocb->ki_left;
1515 /* If we managed to write some out we return that, rather than
1516 * the eventual error. */
1517 if (opcode == IOCB_CMD_PWRITEV
1518 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1519 && iocb->ki_nbytes - iocb->ki_left)
1520 ret = iocb->ki_nbytes - iocb->ki_left;
1525 static ssize_t aio_fdsync(struct kiocb *iocb)
1527 struct file *file = iocb->ki_filp;
1528 ssize_t ret = -EINVAL;
1530 if (file->f_op->aio_fsync)
1531 ret = file->f_op->aio_fsync(iocb, 1);
1535 static ssize_t aio_fsync(struct kiocb *iocb)
1537 struct file *file = iocb->ki_filp;
1538 ssize_t ret = -EINVAL;
1540 if (file->f_op->aio_fsync)
1541 ret = file->f_op->aio_fsync(iocb, 0);
1545 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1549 #ifdef CONFIG_COMPAT
1551 ret = compat_rw_copy_check_uvector(type,
1552 (struct compat_iovec __user *)kiocb->ki_buf,
1553 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1554 &kiocb->ki_iovec, 1);
1557 ret = rw_copy_check_uvector(type,
1558 (struct iovec __user *)kiocb->ki_buf,
1559 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1560 &kiocb->ki_iovec, 1);
1564 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1565 kiocb->ki_cur_seg = 0;
1566 /* ki_nbytes/left now reflect bytes instead of segs */
1567 kiocb->ki_nbytes = ret;
1568 kiocb->ki_left = ret;
1575 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1577 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1578 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1579 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1580 kiocb->ki_nr_segs = 1;
1581 kiocb->ki_cur_seg = 0;
1587 * Performs the initial checks and aio retry method
1588 * setup for the kiocb at the time of io submission.
1590 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1592 struct file *file = kiocb->ki_filp;
1595 switch (kiocb->ki_opcode) {
1596 case IOCB_CMD_PREAD:
1598 if (unlikely(!(file->f_mode & FMODE_READ)))
1601 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1604 ret = security_file_permission(file, MAY_READ);
1607 ret = aio_setup_single_vector(kiocb);
1611 if (file->f_op->aio_read)
1612 kiocb->ki_retry = aio_rw_vect_retry;
1614 case IOCB_CMD_PWRITE:
1616 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1619 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1622 ret = security_file_permission(file, MAY_WRITE);
1625 ret = aio_setup_single_vector(kiocb);
1629 if (file->f_op->aio_write)
1630 kiocb->ki_retry = aio_rw_vect_retry;
1632 case IOCB_CMD_PREADV:
1634 if (unlikely(!(file->f_mode & FMODE_READ)))
1636 ret = security_file_permission(file, MAY_READ);
1639 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1643 if (file->f_op->aio_read)
1644 kiocb->ki_retry = aio_rw_vect_retry;
1646 case IOCB_CMD_PWRITEV:
1648 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1650 ret = security_file_permission(file, MAY_WRITE);
1653 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1657 if (file->f_op->aio_write)
1658 kiocb->ki_retry = aio_rw_vect_retry;
1660 case IOCB_CMD_FDSYNC:
1662 if (file->f_op->aio_fsync)
1663 kiocb->ki_retry = aio_fdsync;
1665 case IOCB_CMD_FSYNC:
1667 if (file->f_op->aio_fsync)
1668 kiocb->ki_retry = aio_fsync;
1671 dprintk("EINVAL: io_submit: no operation provided\n");
1675 if (!kiocb->ki_retry)
1681 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1682 struct iocb *iocb, struct kiocb_batch *batch,
1689 /* enforce forwards compatibility on users */
1690 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1691 pr_debug("EINVAL: io_submit: reserve field set\n");
1695 /* prevent overflows */
1697 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1698 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1699 ((ssize_t)iocb->aio_nbytes < 0)
1701 pr_debug("EINVAL: io_submit: overflow check\n");
1705 file = fget(iocb->aio_fildes);
1706 if (unlikely(!file))
1709 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1710 if (unlikely(!req)) {
1714 req->ki_filp = file;
1715 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1717 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1718 * instance of the file* now. The file descriptor must be
1719 * an eventfd() fd, and will be signaled for each completed
1720 * event using the eventfd_signal() function.
1722 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1723 if (IS_ERR(req->ki_eventfd)) {
1724 ret = PTR_ERR(req->ki_eventfd);
1725 req->ki_eventfd = NULL;
1730 ret = put_user(req->ki_key, &user_iocb->aio_key);
1731 if (unlikely(ret)) {
1732 dprintk("EFAULT: aio_key\n");
1736 req->ki_obj.user = user_iocb;
1737 req->ki_user_data = iocb->aio_data;
1738 req->ki_pos = iocb->aio_offset;
1740 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1741 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1742 req->ki_opcode = iocb->aio_lio_opcode;
1744 ret = aio_setup_iocb(req, compat);
1749 spin_lock_irq(&ctx->ctx_lock);
1751 * We could have raced with io_destroy() and are currently holding a
1752 * reference to ctx which should be destroyed. We cannot submit IO
1753 * since ctx gets freed as soon as io_submit() puts its reference. The
1754 * check here is reliable: io_destroy() sets ctx->dead before waiting
1755 * for outstanding IO and the barrier between these two is realized by
1756 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1757 * increment ctx->reqs_active before checking for ctx->dead and the
1758 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1759 * don't see ctx->dead set here, io_destroy() waits for our IO to
1763 spin_unlock_irq(&ctx->ctx_lock);
1768 if (!list_empty(&ctx->run_list)) {
1769 /* drain the run list */
1770 while (__aio_run_iocbs(ctx))
1773 spin_unlock_irq(&ctx->ctx_lock);
1775 aio_put_req(req); /* drop extra ref to req */
1779 aio_put_req(req); /* drop extra ref to req */
1780 aio_put_req(req); /* drop i/o ref to req */
1784 long do_io_submit(aio_context_t ctx_id, long nr,
1785 struct iocb __user *__user *iocbpp, bool compat)
1790 struct blk_plug plug;
1791 struct kiocb_batch batch;
1793 if (unlikely(nr < 0))
1796 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1797 nr = LONG_MAX/sizeof(*iocbpp);
1799 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1802 ctx = lookup_ioctx(ctx_id);
1803 if (unlikely(!ctx)) {
1804 pr_debug("EINVAL: io_submit: invalid context id\n");
1808 kiocb_batch_init(&batch, nr);
1810 blk_start_plug(&plug);
1813 * AKPM: should this return a partial result if some of the IOs were
1814 * successfully submitted?
1816 for (i=0; i<nr; i++) {
1817 struct iocb __user *user_iocb;
1820 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1825 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1830 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1834 blk_finish_plug(&plug);
1836 kiocb_batch_free(ctx, &batch);
1842 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1843 * the number of iocbs queued. May return -EINVAL if the aio_context
1844 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1845 * *iocbpp[0] is not properly initialized, if the operation specified
1846 * is invalid for the file descriptor in the iocb. May fail with
1847 * -EFAULT if any of the data structures point to invalid data. May
1848 * fail with -EBADF if the file descriptor specified in the first
1849 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1850 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1851 * fail with -ENOSYS if not implemented.
1853 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1854 struct iocb __user * __user *, iocbpp)
1856 return do_io_submit(ctx_id, nr, iocbpp, 0);
1860 * Finds a given iocb for cancellation.
1862 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1865 struct list_head *pos;
1867 assert_spin_locked(&ctx->ctx_lock);
1869 /* TODO: use a hash or array, this sucks. */
1870 list_for_each(pos, &ctx->active_reqs) {
1871 struct kiocb *kiocb = list_kiocb(pos);
1872 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1879 * Attempts to cancel an iocb previously passed to io_submit. If
1880 * the operation is successfully cancelled, the resulting event is
1881 * copied into the memory pointed to by result without being placed
1882 * into the completion queue and 0 is returned. May fail with
1883 * -EFAULT if any of the data structures pointed to are invalid.
1884 * May fail with -EINVAL if aio_context specified by ctx_id is
1885 * invalid. May fail with -EAGAIN if the iocb specified was not
1886 * cancelled. Will fail with -ENOSYS if not implemented.
1888 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1889 struct io_event __user *, result)
1891 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1893 struct kiocb *kiocb;
1897 ret = get_user(key, &iocb->aio_key);
1901 ctx = lookup_ioctx(ctx_id);
1905 spin_lock_irq(&ctx->ctx_lock);
1907 kiocb = lookup_kiocb(ctx, iocb, key);
1908 if (kiocb && kiocb->ki_cancel) {
1909 cancel = kiocb->ki_cancel;
1911 kiocbSetCancelled(kiocb);
1914 spin_unlock_irq(&ctx->ctx_lock);
1916 if (NULL != cancel) {
1917 struct io_event tmp;
1918 pr_debug("calling cancel\n");
1919 memset(&tmp, 0, sizeof(tmp));
1920 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1921 tmp.data = kiocb->ki_user_data;
1922 ret = cancel(kiocb, &tmp);
1924 /* Cancellation succeeded -- copy the result
1925 * into the user's buffer.
1927 if (copy_to_user(result, &tmp, sizeof(tmp)))
1939 * Attempts to read at least min_nr events and up to nr events from
1940 * the completion queue for the aio_context specified by ctx_id. If
1941 * it succeeds, the number of read events is returned. May fail with
1942 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1943 * out of range, if timeout is out of range. May fail with -EFAULT
1944 * if any of the memory specified is invalid. May return 0 or
1945 * < min_nr if the timeout specified by timeout has elapsed
1946 * before sufficient events are available, where timeout == NULL
1947 * specifies an infinite timeout. Note that the timeout pointed to by
1948 * timeout is relative and will be updated if not NULL and the
1949 * operation blocks. Will fail with -ENOSYS if not implemented.
1951 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1954 struct io_event __user *, events,
1955 struct timespec __user *, timeout)
1957 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1960 if (likely(ioctx)) {
1961 if (likely(min_nr <= nr && min_nr >= 0))
1962 ret = read_events(ioctx, min_nr, nr, events, timeout);
1966 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);