2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
11 #include <linux/kernel.h>
12 #include <linux/init.h>
13 #include <linux/errno.h>
14 #include <linux/time.h>
15 #include <linux/aio_abi.h>
16 #include <linux/export.h>
17 #include <linux/syscalls.h>
18 #include <linux/backing-dev.h>
19 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/compat.h>
39 #include <asm/kmap_types.h>
40 #include <asm/uaccess.h>
43 #define dprintk printk
45 #define dprintk(x...) do { ; } while (0)
48 /*------ sysctl variables----*/
49 static DEFINE_SPINLOCK(aio_nr_lock);
50 unsigned long aio_nr; /* current system wide number of aio requests */
51 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
52 /*----end sysctl variables---*/
54 static struct kmem_cache *kiocb_cachep;
55 static struct kmem_cache *kioctx_cachep;
57 static struct workqueue_struct *aio_wq;
59 /* Used for rare fput completion. */
60 static void aio_fput_routine(struct work_struct *);
61 static DECLARE_WORK(fput_work, aio_fput_routine);
63 static DEFINE_SPINLOCK(fput_lock);
64 static LIST_HEAD(fput_head);
66 static void aio_kick_handler(struct work_struct *);
67 static void aio_queue_work(struct kioctx *);
70 * Creates the slab caches used by the aio routines, panic on
71 * failure as this is done early during the boot sequence.
73 static int __init aio_setup(void)
75 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
76 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
78 aio_wq = alloc_workqueue("aio", 0, 1); /* used to limit concurrency */
81 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
85 __initcall(aio_setup);
87 static void aio_free_ring(struct kioctx *ctx)
89 struct aio_ring_info *info = &ctx->ring_info;
92 for (i=0; i<info->nr_pages; i++)
93 put_page(info->ring_pages[i]);
95 if (info->mmap_size) {
96 BUG_ON(ctx->mm != current->mm);
97 vm_munmap(info->mmap_base, info->mmap_size);
100 if (info->ring_pages && info->ring_pages != info->internal_pages)
101 kfree(info->ring_pages);
102 info->ring_pages = NULL;
106 static int aio_setup_ring(struct kioctx *ctx)
108 struct aio_ring *ring;
109 struct aio_ring_info *info = &ctx->ring_info;
110 unsigned nr_events = ctx->max_reqs;
114 /* Compensate for the ring buffer's head/tail overlap entry */
115 nr_events += 2; /* 1 is required, 2 for good luck */
117 size = sizeof(struct aio_ring);
118 size += sizeof(struct io_event) * nr_events;
119 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
124 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
127 info->ring_pages = info->internal_pages;
128 if (nr_pages > AIO_RING_PAGES) {
129 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
130 if (!info->ring_pages)
134 info->mmap_size = nr_pages * PAGE_SIZE;
135 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
136 down_write(&ctx->mm->mmap_sem);
137 info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size,
138 PROT_READ|PROT_WRITE,
139 MAP_ANONYMOUS|MAP_PRIVATE, 0);
140 if (IS_ERR((void *)info->mmap_base)) {
141 up_write(&ctx->mm->mmap_sem);
147 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
148 info->nr_pages = get_user_pages(current, ctx->mm,
149 info->mmap_base, nr_pages,
150 1, 0, info->ring_pages, NULL);
151 up_write(&ctx->mm->mmap_sem);
153 if (unlikely(info->nr_pages != nr_pages)) {
158 ctx->user_id = info->mmap_base;
160 info->nr = nr_events; /* trusted copy */
162 ring = kmap_atomic(info->ring_pages[0]);
163 ring->nr = nr_events; /* user copy */
164 ring->id = ctx->user_id;
165 ring->head = ring->tail = 0;
166 ring->magic = AIO_RING_MAGIC;
167 ring->compat_features = AIO_RING_COMPAT_FEATURES;
168 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
169 ring->header_length = sizeof(struct aio_ring);
176 /* aio_ring_event: returns a pointer to the event at the given index from
177 * kmap_atomic(). Release the pointer with put_aio_ring_event();
179 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
180 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
181 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
183 #define aio_ring_event(info, nr) ({ \
184 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
185 struct io_event *__event; \
186 __event = kmap_atomic( \
187 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
188 __event += pos % AIO_EVENTS_PER_PAGE; \
192 #define put_aio_ring_event(event) do { \
193 struct io_event *__event = (event); \
195 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
198 static void ctx_rcu_free(struct rcu_head *head)
200 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
201 kmem_cache_free(kioctx_cachep, ctx);
205 * Called when the last user of an aio context has gone away,
206 * and the struct needs to be freed.
208 static void __put_ioctx(struct kioctx *ctx)
210 unsigned nr_events = ctx->max_reqs;
211 BUG_ON(ctx->reqs_active);
213 cancel_delayed_work_sync(&ctx->wq);
218 spin_lock(&aio_nr_lock);
219 BUG_ON(aio_nr - nr_events > aio_nr);
221 spin_unlock(&aio_nr_lock);
223 pr_debug("__put_ioctx: freeing %p\n", ctx);
224 call_rcu(&ctx->rcu_head, ctx_rcu_free);
227 static inline int try_get_ioctx(struct kioctx *kioctx)
229 return atomic_inc_not_zero(&kioctx->users);
232 static inline void put_ioctx(struct kioctx *kioctx)
234 BUG_ON(atomic_read(&kioctx->users) <= 0);
235 if (unlikely(atomic_dec_and_test(&kioctx->users)))
240 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
242 static struct kioctx *ioctx_alloc(unsigned nr_events)
244 struct mm_struct *mm;
248 /* Prevent overflows */
249 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
250 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
251 pr_debug("ENOMEM: nr_events too high\n");
252 return ERR_PTR(-EINVAL);
255 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
256 return ERR_PTR(-EAGAIN);
258 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
260 return ERR_PTR(-ENOMEM);
262 ctx->max_reqs = nr_events;
263 mm = ctx->mm = current->mm;
264 atomic_inc(&mm->mm_count);
266 atomic_set(&ctx->users, 2);
267 spin_lock_init(&ctx->ctx_lock);
268 spin_lock_init(&ctx->ring_info.ring_lock);
269 init_waitqueue_head(&ctx->wait);
271 INIT_LIST_HEAD(&ctx->active_reqs);
272 INIT_LIST_HEAD(&ctx->run_list);
273 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
275 if (aio_setup_ring(ctx) < 0)
278 /* limit the number of system wide aios */
279 spin_lock(&aio_nr_lock);
280 if (aio_nr + nr_events > aio_max_nr ||
281 aio_nr + nr_events < aio_nr) {
282 spin_unlock(&aio_nr_lock);
285 aio_nr += ctx->max_reqs;
286 spin_unlock(&aio_nr_lock);
288 /* now link into global list. */
289 spin_lock(&mm->ioctx_lock);
290 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
291 spin_unlock(&mm->ioctx_lock);
293 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
294 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
302 kmem_cache_free(kioctx_cachep, ctx);
303 dprintk("aio: error allocating ioctx %d\n", err);
308 * Cancels all outstanding aio requests on an aio context. Used
309 * when the processes owning a context have all exited to encourage
310 * the rapid destruction of the kioctx.
312 static void kill_ctx(struct kioctx *ctx)
314 int (*cancel)(struct kiocb *, struct io_event *);
315 struct task_struct *tsk = current;
316 DECLARE_WAITQUEUE(wait, tsk);
319 spin_lock_irq(&ctx->ctx_lock);
321 while (!list_empty(&ctx->active_reqs)) {
322 struct list_head *pos = ctx->active_reqs.next;
323 struct kiocb *iocb = list_kiocb(pos);
324 list_del_init(&iocb->ki_list);
325 cancel = iocb->ki_cancel;
326 kiocbSetCancelled(iocb);
329 spin_unlock_irq(&ctx->ctx_lock);
331 spin_lock_irq(&ctx->ctx_lock);
335 if (!ctx->reqs_active)
338 add_wait_queue(&ctx->wait, &wait);
339 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
340 while (ctx->reqs_active) {
341 spin_unlock_irq(&ctx->ctx_lock);
343 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
344 spin_lock_irq(&ctx->ctx_lock);
346 __set_task_state(tsk, TASK_RUNNING);
347 remove_wait_queue(&ctx->wait, &wait);
350 spin_unlock_irq(&ctx->ctx_lock);
353 /* wait_on_sync_kiocb:
354 * Waits on the given sync kiocb to complete.
356 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
358 while (iocb->ki_users) {
359 set_current_state(TASK_UNINTERRUPTIBLE);
364 __set_current_state(TASK_RUNNING);
365 return iocb->ki_user_data;
367 EXPORT_SYMBOL(wait_on_sync_kiocb);
369 /* exit_aio: called when the last user of mm goes away. At this point,
370 * there is no way for any new requests to be submited or any of the
371 * io_* syscalls to be called on the context. However, there may be
372 * outstanding requests which hold references to the context; as they
373 * go away, they will call put_ioctx and release any pinned memory
374 * associated with the request (held via struct page * references).
376 void exit_aio(struct mm_struct *mm)
380 while (!hlist_empty(&mm->ioctx_list)) {
381 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
382 hlist_del_rcu(&ctx->list);
386 if (1 != atomic_read(&ctx->users))
388 "exit_aio:ioctx still alive: %d %d %d\n",
389 atomic_read(&ctx->users), ctx->dead,
392 * We don't need to bother with munmap() here -
393 * exit_mmap(mm) is coming and it'll unmap everything.
394 * Since aio_free_ring() uses non-zero ->mmap_size
395 * as indicator that it needs to unmap the area,
396 * just set it to 0; aio_free_ring() is the only
397 * place that uses ->mmap_size, so it's safe.
398 * That way we get all munmap done to current->mm -
399 * all other callers have ctx->mm == current->mm.
401 ctx->ring_info.mmap_size = 0;
407 * Allocate a slot for an aio request. Increments the users count
408 * of the kioctx so that the kioctx stays around until all requests are
409 * complete. Returns NULL if no requests are free.
411 * Returns with kiocb->users set to 2. The io submit code path holds
412 * an extra reference while submitting the i/o.
413 * This prevents races between the aio code path referencing the
414 * req (after submitting it) and aio_complete() freeing the req.
416 static struct kiocb *__aio_get_req(struct kioctx *ctx)
418 struct kiocb *req = NULL;
420 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
428 req->ki_cancel = NULL;
429 req->ki_retry = NULL;
432 req->ki_iovec = NULL;
433 INIT_LIST_HEAD(&req->ki_run_list);
434 req->ki_eventfd = NULL;
440 * struct kiocb's are allocated in batches to reduce the number of
441 * times the ctx lock is acquired and released.
443 #define KIOCB_BATCH_SIZE 32L
445 struct list_head head;
446 long count; /* number of requests left to allocate */
449 static void kiocb_batch_init(struct kiocb_batch *batch, long total)
451 INIT_LIST_HEAD(&batch->head);
452 batch->count = total;
455 static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
457 struct kiocb *req, *n;
459 if (list_empty(&batch->head))
462 spin_lock_irq(&ctx->ctx_lock);
463 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
464 list_del(&req->ki_batch);
465 list_del(&req->ki_list);
466 kmem_cache_free(kiocb_cachep, req);
469 if (unlikely(!ctx->reqs_active && ctx->dead))
470 wake_up_all(&ctx->wait);
471 spin_unlock_irq(&ctx->ctx_lock);
475 * Allocate a batch of kiocbs. This avoids taking and dropping the
476 * context lock a lot during setup.
478 static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
480 unsigned short allocated, to_alloc;
482 bool called_fput = false;
483 struct kiocb *req, *n;
484 struct aio_ring *ring;
486 to_alloc = min(batch->count, KIOCB_BATCH_SIZE);
487 for (allocated = 0; allocated < to_alloc; allocated++) {
488 req = __aio_get_req(ctx);
490 /* allocation failed, go with what we've got */
492 list_add(&req->ki_batch, &batch->head);
499 spin_lock_irq(&ctx->ctx_lock);
500 ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
502 avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
504 if (avail == 0 && !called_fput) {
506 * Handle a potential starvation case. It is possible that
507 * we hold the last reference on a struct file, causing us
508 * to delay the final fput to non-irq context. In this case,
509 * ctx->reqs_active is artificially high. Calling the fput
510 * routine here may free up a slot in the event completion
511 * ring, allowing this allocation to succeed.
514 spin_unlock_irq(&ctx->ctx_lock);
515 aio_fput_routine(NULL);
520 if (avail < allocated) {
521 /* Trim back the number of requests. */
522 list_for_each_entry_safe(req, n, &batch->head, ki_batch) {
523 list_del(&req->ki_batch);
524 kmem_cache_free(kiocb_cachep, req);
525 if (--allocated <= avail)
530 batch->count -= allocated;
531 list_for_each_entry(req, &batch->head, ki_batch) {
532 list_add(&req->ki_list, &ctx->active_reqs);
537 spin_unlock_irq(&ctx->ctx_lock);
543 static inline struct kiocb *aio_get_req(struct kioctx *ctx,
544 struct kiocb_batch *batch)
548 if (list_empty(&batch->head))
549 if (kiocb_batch_refill(ctx, batch) == 0)
551 req = list_first_entry(&batch->head, struct kiocb, ki_batch);
552 list_del(&req->ki_batch);
556 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
558 assert_spin_locked(&ctx->ctx_lock);
560 if (req->ki_eventfd != NULL)
561 eventfd_ctx_put(req->ki_eventfd);
564 if (req->ki_iovec != &req->ki_inline_vec)
565 kfree(req->ki_iovec);
566 kmem_cache_free(kiocb_cachep, req);
569 if (unlikely(!ctx->reqs_active && ctx->dead))
570 wake_up_all(&ctx->wait);
573 static void aio_fput_routine(struct work_struct *data)
575 spin_lock_irq(&fput_lock);
576 while (likely(!list_empty(&fput_head))) {
577 struct kiocb *req = list_kiocb(fput_head.next);
578 struct kioctx *ctx = req->ki_ctx;
580 list_del(&req->ki_list);
581 spin_unlock_irq(&fput_lock);
583 /* Complete the fput(s) */
584 if (req->ki_filp != NULL)
587 /* Link the iocb into the context's free list */
589 spin_lock_irq(&ctx->ctx_lock);
590 really_put_req(ctx, req);
592 * at that point ctx might've been killed, but actual
595 spin_unlock_irq(&ctx->ctx_lock);
598 spin_lock_irq(&fput_lock);
600 spin_unlock_irq(&fput_lock);
604 * Returns true if this put was the last user of the request.
606 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
608 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
609 req, atomic_long_read(&req->ki_filp->f_count));
611 assert_spin_locked(&ctx->ctx_lock);
614 BUG_ON(req->ki_users < 0);
615 if (likely(req->ki_users))
617 list_del(&req->ki_list); /* remove from active_reqs */
618 req->ki_cancel = NULL;
619 req->ki_retry = NULL;
622 * Try to optimize the aio and eventfd file* puts, by avoiding to
623 * schedule work in case it is not final fput() time. In normal cases,
624 * we would not be holding the last reference to the file*, so
625 * this function will be executed w/out any aio kthread wakeup.
627 if (unlikely(!fput_atomic(req->ki_filp))) {
628 spin_lock(&fput_lock);
629 list_add(&req->ki_list, &fput_head);
630 spin_unlock(&fput_lock);
631 schedule_work(&fput_work);
634 really_put_req(ctx, req);
640 * Returns true if this put was the last user of the kiocb,
641 * false if the request is still in use.
643 int aio_put_req(struct kiocb *req)
645 struct kioctx *ctx = req->ki_ctx;
647 spin_lock_irq(&ctx->ctx_lock);
648 ret = __aio_put_req(ctx, req);
649 spin_unlock_irq(&ctx->ctx_lock);
652 EXPORT_SYMBOL(aio_put_req);
654 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
656 struct mm_struct *mm = current->mm;
657 struct kioctx *ctx, *ret = NULL;
658 struct hlist_node *n;
662 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
664 * RCU protects us against accessing freed memory but
665 * we have to be careful not to get a reference when the
666 * reference count already dropped to 0 (ctx->dead test
667 * is unreliable because of races).
669 if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
680 * Queue up a kiocb to be retried. Assumes that the kiocb
681 * has already been marked as kicked, and places it on
682 * the retry run list for the corresponding ioctx, if it
683 * isn't already queued. Returns 1 if it actually queued
684 * the kiocb (to tell the caller to activate the work
685 * queue to process it), or 0, if it found that it was
688 static inline int __queue_kicked_iocb(struct kiocb *iocb)
690 struct kioctx *ctx = iocb->ki_ctx;
692 assert_spin_locked(&ctx->ctx_lock);
694 if (list_empty(&iocb->ki_run_list)) {
695 list_add_tail(&iocb->ki_run_list,
703 * This is the core aio execution routine. It is
704 * invoked both for initial i/o submission and
705 * subsequent retries via the aio_kick_handler.
706 * Expects to be invoked with iocb->ki_ctx->lock
707 * already held. The lock is released and reacquired
708 * as needed during processing.
710 * Calls the iocb retry method (already setup for the
711 * iocb on initial submission) for operation specific
712 * handling, but takes care of most of common retry
713 * execution details for a given iocb. The retry method
714 * needs to be non-blocking as far as possible, to avoid
715 * holding up other iocbs waiting to be serviced by the
716 * retry kernel thread.
718 * The trickier parts in this code have to do with
719 * ensuring that only one retry instance is in progress
720 * for a given iocb at any time. Providing that guarantee
721 * simplifies the coding of individual aio operations as
722 * it avoids various potential races.
724 static ssize_t aio_run_iocb(struct kiocb *iocb)
726 struct kioctx *ctx = iocb->ki_ctx;
727 ssize_t (*retry)(struct kiocb *);
730 if (!(retry = iocb->ki_retry)) {
731 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
736 * We don't want the next retry iteration for this
737 * operation to start until this one has returned and
738 * updated the iocb state. However, wait_queue functions
739 * can trigger a kick_iocb from interrupt context in the
740 * meantime, indicating that data is available for the next
741 * iteration. We want to remember that and enable the
742 * next retry iteration _after_ we are through with
745 * So, in order to be able to register a "kick", but
746 * prevent it from being queued now, we clear the kick
747 * flag, but make the kick code *think* that the iocb is
748 * still on the run list until we are actually done.
749 * When we are done with this iteration, we check if
750 * the iocb was kicked in the meantime and if so, queue
754 kiocbClearKicked(iocb);
757 * This is so that aio_complete knows it doesn't need to
758 * pull the iocb off the run list (We can't just call
759 * INIT_LIST_HEAD because we don't want a kick_iocb to
760 * queue this on the run list yet)
762 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
763 spin_unlock_irq(&ctx->ctx_lock);
765 /* Quit retrying if the i/o has been cancelled */
766 if (kiocbIsCancelled(iocb)) {
768 aio_complete(iocb, ret, 0);
769 /* must not access the iocb after this */
774 * Now we are all set to call the retry method in async
779 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
781 * There's no easy way to restart the syscall since other AIO's
782 * may be already running. Just fail this IO with EINTR.
784 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
785 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
787 aio_complete(iocb, ret, 0);
790 spin_lock_irq(&ctx->ctx_lock);
792 if (-EIOCBRETRY == ret) {
794 * OK, now that we are done with this iteration
795 * and know that there is more left to go,
796 * this is where we let go so that a subsequent
797 * "kick" can start the next iteration
800 /* will make __queue_kicked_iocb succeed from here on */
801 INIT_LIST_HEAD(&iocb->ki_run_list);
802 /* we must queue the next iteration ourselves, if it
803 * has already been kicked */
804 if (kiocbIsKicked(iocb)) {
805 __queue_kicked_iocb(iocb);
808 * __queue_kicked_iocb will always return 1 here, because
809 * iocb->ki_run_list is empty at this point so it should
810 * be safe to unconditionally queue the context into the
821 * Process all pending retries queued on the ioctx
823 * Assumes it is operating within the aio issuer's mm
826 static int __aio_run_iocbs(struct kioctx *ctx)
829 struct list_head run_list;
831 assert_spin_locked(&ctx->ctx_lock);
833 list_replace_init(&ctx->run_list, &run_list);
834 while (!list_empty(&run_list)) {
835 iocb = list_entry(run_list.next, struct kiocb,
837 list_del(&iocb->ki_run_list);
839 * Hold an extra reference while retrying i/o.
841 iocb->ki_users++; /* grab extra reference */
843 __aio_put_req(ctx, iocb);
845 if (!list_empty(&ctx->run_list))
850 static void aio_queue_work(struct kioctx * ctx)
852 unsigned long timeout;
854 * if someone is waiting, get the work started right
855 * away, otherwise, use a longer delay
858 if (waitqueue_active(&ctx->wait))
862 queue_delayed_work(aio_wq, &ctx->wq, timeout);
867 * Process all pending retries queued on the ioctx
868 * run list, and keep running them until the list
870 * Assumes it is operating within the aio issuer's mm context.
872 static inline void aio_run_all_iocbs(struct kioctx *ctx)
874 spin_lock_irq(&ctx->ctx_lock);
875 while (__aio_run_iocbs(ctx))
877 spin_unlock_irq(&ctx->ctx_lock);
882 * Work queue handler triggered to process pending
883 * retries on an ioctx. Takes on the aio issuer's
884 * mm context before running the iocbs, so that
885 * copy_xxx_user operates on the issuer's address
887 * Run on aiod's context.
889 static void aio_kick_handler(struct work_struct *work)
891 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
892 mm_segment_t oldfs = get_fs();
893 struct mm_struct *mm;
898 spin_lock_irq(&ctx->ctx_lock);
899 requeue =__aio_run_iocbs(ctx);
901 spin_unlock_irq(&ctx->ctx_lock);
905 * we're in a worker thread already; no point using non-zero delay
908 queue_delayed_work(aio_wq, &ctx->wq, 0);
913 * Called by kick_iocb to queue the kiocb for retry
914 * and if required activate the aio work queue to process
917 static void try_queue_kicked_iocb(struct kiocb *iocb)
919 struct kioctx *ctx = iocb->ki_ctx;
923 spin_lock_irqsave(&ctx->ctx_lock, flags);
924 /* set this inside the lock so that we can't race with aio_run_iocb()
925 * testing it and putting the iocb on the run list under the lock */
926 if (!kiocbTryKick(iocb))
927 run = __queue_kicked_iocb(iocb);
928 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
935 * Called typically from a wait queue callback context
936 * to trigger a retry of the iocb.
937 * The retry is usually executed by aio workqueue
938 * threads (See aio_kick_handler).
940 void kick_iocb(struct kiocb *iocb)
942 /* sync iocbs are easy: they can only ever be executing from a
944 if (is_sync_kiocb(iocb)) {
945 kiocbSetKicked(iocb);
946 wake_up_process(iocb->ki_obj.tsk);
950 try_queue_kicked_iocb(iocb);
952 EXPORT_SYMBOL(kick_iocb);
955 * Called when the io request on the given iocb is complete.
956 * Returns true if this is the last user of the request. The
957 * only other user of the request can be the cancellation code.
959 int aio_complete(struct kiocb *iocb, long res, long res2)
961 struct kioctx *ctx = iocb->ki_ctx;
962 struct aio_ring_info *info;
963 struct aio_ring *ring;
964 struct io_event *event;
970 * Special case handling for sync iocbs:
971 * - events go directly into the iocb for fast handling
972 * - the sync task with the iocb in its stack holds the single iocb
973 * ref, no other paths have a way to get another ref
974 * - the sync task helpfully left a reference to itself in the iocb
976 if (is_sync_kiocb(iocb)) {
977 BUG_ON(iocb->ki_users != 1);
978 iocb->ki_user_data = res;
980 wake_up_process(iocb->ki_obj.tsk);
984 info = &ctx->ring_info;
986 /* add a completion event to the ring buffer.
987 * must be done holding ctx->ctx_lock to prevent
988 * other code from messing with the tail
989 * pointer since we might be called from irq
992 spin_lock_irqsave(&ctx->ctx_lock, flags);
994 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
995 list_del_init(&iocb->ki_run_list);
998 * cancelled requests don't get events, userland was given one
999 * when the event got cancelled.
1001 if (kiocbIsCancelled(iocb))
1004 ring = kmap_atomic(info->ring_pages[0]);
1007 event = aio_ring_event(info, tail);
1008 if (++tail >= info->nr)
1011 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1012 event->data = iocb->ki_user_data;
1016 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1017 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1020 /* after flagging the request as done, we
1021 * must never even look at it again
1023 smp_wmb(); /* make event visible before updating tail */
1028 put_aio_ring_event(event);
1029 kunmap_atomic(ring);
1031 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1034 * Check if the user asked us to deliver the result through an
1035 * eventfd. The eventfd_signal() function is safe to be called
1038 if (iocb->ki_eventfd != NULL)
1039 eventfd_signal(iocb->ki_eventfd, 1);
1042 /* everything turned out well, dispose of the aiocb. */
1043 ret = __aio_put_req(ctx, iocb);
1046 * We have to order our ring_info tail store above and test
1047 * of the wait list below outside the wait lock. This is
1048 * like in wake_up_bit() where clearing a bit has to be
1049 * ordered with the unlocked test.
1053 if (waitqueue_active(&ctx->wait))
1054 wake_up(&ctx->wait);
1056 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1059 EXPORT_SYMBOL(aio_complete);
1062 * Pull an event off of the ioctx's event ring. Returns the number of
1063 * events fetched (0 or 1 ;-)
1064 * FIXME: make this use cmpxchg.
1065 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1067 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1069 struct aio_ring_info *info = &ioctx->ring_info;
1070 struct aio_ring *ring;
1074 ring = kmap_atomic(info->ring_pages[0]);
1075 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1076 (unsigned long)ring->head, (unsigned long)ring->tail,
1077 (unsigned long)ring->nr);
1079 if (ring->head == ring->tail)
1082 spin_lock(&info->ring_lock);
1084 head = ring->head % info->nr;
1085 if (head != ring->tail) {
1086 struct io_event *evp = aio_ring_event(info, head);
1088 head = (head + 1) % info->nr;
1089 smp_mb(); /* finish reading the event before updatng the head */
1092 put_aio_ring_event(evp);
1094 spin_unlock(&info->ring_lock);
1097 kunmap_atomic(ring);
1098 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1099 (unsigned long)ring->head, (unsigned long)ring->tail);
1103 struct aio_timeout {
1104 struct timer_list timer;
1106 struct task_struct *p;
1109 static void timeout_func(unsigned long data)
1111 struct aio_timeout *to = (struct aio_timeout *)data;
1114 wake_up_process(to->p);
1117 static inline void init_timeout(struct aio_timeout *to)
1119 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1124 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1125 const struct timespec *ts)
1127 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1128 if (time_after(to->timer.expires, jiffies))
1129 add_timer(&to->timer);
1134 static inline void clear_timeout(struct aio_timeout *to)
1136 del_singleshot_timer_sync(&to->timer);
1139 static int read_events(struct kioctx *ctx,
1140 long min_nr, long nr,
1141 struct io_event __user *event,
1142 struct timespec __user *timeout)
1144 long start_jiffies = jiffies;
1145 struct task_struct *tsk = current;
1146 DECLARE_WAITQUEUE(wait, tsk);
1149 struct io_event ent;
1150 struct aio_timeout to;
1153 /* needed to zero any padding within an entry (there shouldn't be
1154 * any, but C is fun!
1156 memset(&ent, 0, sizeof(ent));
1159 while (likely(i < nr)) {
1160 ret = aio_read_evt(ctx, &ent);
1161 if (unlikely(ret <= 0))
1164 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1165 ent.data, ent.obj, ent.res, ent.res2);
1167 /* Could we split the check in two? */
1169 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1170 dprintk("aio: lost an event due to EFAULT.\n");
1175 /* Good, event copied to userland, update counts. */
1187 /* racey check, but it gets redone */
1188 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1190 aio_run_all_iocbs(ctx);
1198 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1201 set_timeout(start_jiffies, &to, &ts);
1204 while (likely(i < nr)) {
1205 add_wait_queue_exclusive(&ctx->wait, &wait);
1207 set_task_state(tsk, TASK_INTERRUPTIBLE);
1208 ret = aio_read_evt(ctx, &ent);
1213 if (unlikely(ctx->dead)) {
1217 if (to.timed_out) /* Only check after read evt */
1219 /* Try to only show up in io wait if there are ops
1221 if (ctx->reqs_active)
1225 if (signal_pending(tsk)) {
1229 /*ret = aio_read_evt(ctx, &ent);*/
1232 set_task_state(tsk, TASK_RUNNING);
1233 remove_wait_queue(&ctx->wait, &wait);
1235 if (unlikely(ret <= 0))
1239 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1240 dprintk("aio: lost an event due to EFAULT.\n");
1244 /* Good, event copied to userland, update counts. */
1252 destroy_timer_on_stack(&to.timer);
1256 /* Take an ioctx and remove it from the list of ioctx's. Protects
1257 * against races with itself via ->dead.
1259 static void io_destroy(struct kioctx *ioctx)
1261 struct mm_struct *mm = current->mm;
1264 /* delete the entry from the list is someone else hasn't already */
1265 spin_lock(&mm->ioctx_lock);
1266 was_dead = ioctx->dead;
1268 hlist_del_rcu(&ioctx->list);
1269 spin_unlock(&mm->ioctx_lock);
1271 dprintk("aio_release(%p)\n", ioctx);
1272 if (likely(!was_dead))
1273 put_ioctx(ioctx); /* twice for the list */
1278 * Wake up any waiters. The setting of ctx->dead must be seen
1279 * by other CPUs at this point. Right now, we rely on the
1280 * locking done by the above calls to ensure this consistency.
1282 wake_up_all(&ioctx->wait);
1286 * Create an aio_context capable of receiving at least nr_events.
1287 * ctxp must not point to an aio_context that already exists, and
1288 * must be initialized to 0 prior to the call. On successful
1289 * creation of the aio_context, *ctxp is filled in with the resulting
1290 * handle. May fail with -EINVAL if *ctxp is not initialized,
1291 * if the specified nr_events exceeds internal limits. May fail
1292 * with -EAGAIN if the specified nr_events exceeds the user's limit
1293 * of available events. May fail with -ENOMEM if insufficient kernel
1294 * resources are available. May fail with -EFAULT if an invalid
1295 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1298 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1300 struct kioctx *ioctx = NULL;
1304 ret = get_user(ctx, ctxp);
1309 if (unlikely(ctx || nr_events == 0)) {
1310 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1315 ioctx = ioctx_alloc(nr_events);
1316 ret = PTR_ERR(ioctx);
1317 if (!IS_ERR(ioctx)) {
1318 ret = put_user(ioctx->user_id, ctxp);
1329 * Destroy the aio_context specified. May cancel any outstanding
1330 * AIOs and block on completion. Will fail with -ENOSYS if not
1331 * implemented. May fail with -EINVAL if the context pointed to
1334 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1336 struct kioctx *ioctx = lookup_ioctx(ctx);
1337 if (likely(NULL != ioctx)) {
1342 pr_debug("EINVAL: io_destroy: invalid context id\n");
1346 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1348 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1352 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1353 ssize_t this = min((ssize_t)iov->iov_len, ret);
1354 iov->iov_base += this;
1355 iov->iov_len -= this;
1356 iocb->ki_left -= this;
1358 if (iov->iov_len == 0) {
1364 /* the caller should not have done more io than what fit in
1365 * the remaining iovecs */
1366 BUG_ON(ret > 0 && iocb->ki_left == 0);
1369 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1371 struct file *file = iocb->ki_filp;
1372 struct address_space *mapping = file->f_mapping;
1373 struct inode *inode = mapping->host;
1374 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1375 unsigned long, loff_t);
1377 unsigned short opcode;
1379 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1380 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1381 rw_op = file->f_op->aio_read;
1382 opcode = IOCB_CMD_PREADV;
1384 rw_op = file->f_op->aio_write;
1385 opcode = IOCB_CMD_PWRITEV;
1388 /* This matches the pread()/pwrite() logic */
1389 if (iocb->ki_pos < 0)
1393 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1394 iocb->ki_nr_segs - iocb->ki_cur_seg,
1397 aio_advance_iovec(iocb, ret);
1399 /* retry all partial writes. retry partial reads as long as its a
1401 } while (ret > 0 && iocb->ki_left > 0 &&
1402 (opcode == IOCB_CMD_PWRITEV ||
1403 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1405 /* This means we must have transferred all that we could */
1406 /* No need to retry anymore */
1407 if ((ret == 0) || (iocb->ki_left == 0))
1408 ret = iocb->ki_nbytes - iocb->ki_left;
1410 /* If we managed to write some out we return that, rather than
1411 * the eventual error. */
1412 if (opcode == IOCB_CMD_PWRITEV
1413 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1414 && iocb->ki_nbytes - iocb->ki_left)
1415 ret = iocb->ki_nbytes - iocb->ki_left;
1420 static ssize_t aio_fdsync(struct kiocb *iocb)
1422 struct file *file = iocb->ki_filp;
1423 ssize_t ret = -EINVAL;
1425 if (file->f_op->aio_fsync)
1426 ret = file->f_op->aio_fsync(iocb, 1);
1430 static ssize_t aio_fsync(struct kiocb *iocb)
1432 struct file *file = iocb->ki_filp;
1433 ssize_t ret = -EINVAL;
1435 if (file->f_op->aio_fsync)
1436 ret = file->f_op->aio_fsync(iocb, 0);
1440 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1444 #ifdef CONFIG_COMPAT
1446 ret = compat_rw_copy_check_uvector(type,
1447 (struct compat_iovec __user *)kiocb->ki_buf,
1448 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1452 ret = rw_copy_check_uvector(type,
1453 (struct iovec __user *)kiocb->ki_buf,
1454 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1459 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1463 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1464 kiocb->ki_cur_seg = 0;
1465 /* ki_nbytes/left now reflect bytes instead of segs */
1466 kiocb->ki_nbytes = ret;
1467 kiocb->ki_left = ret;
1474 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1478 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1482 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1483 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1484 kiocb->ki_iovec->iov_len = bytes;
1485 kiocb->ki_nr_segs = 1;
1486 kiocb->ki_cur_seg = 0;
1492 * Performs the initial checks and aio retry method
1493 * setup for the kiocb at the time of io submission.
1495 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1497 struct file *file = kiocb->ki_filp;
1500 switch (kiocb->ki_opcode) {
1501 case IOCB_CMD_PREAD:
1503 if (unlikely(!(file->f_mode & FMODE_READ)))
1506 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1509 ret = aio_setup_single_vector(READ, file, kiocb);
1513 if (file->f_op->aio_read)
1514 kiocb->ki_retry = aio_rw_vect_retry;
1516 case IOCB_CMD_PWRITE:
1518 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1521 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1524 ret = aio_setup_single_vector(WRITE, file, kiocb);
1528 if (file->f_op->aio_write)
1529 kiocb->ki_retry = aio_rw_vect_retry;
1531 case IOCB_CMD_PREADV:
1533 if (unlikely(!(file->f_mode & FMODE_READ)))
1535 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1539 if (file->f_op->aio_read)
1540 kiocb->ki_retry = aio_rw_vect_retry;
1542 case IOCB_CMD_PWRITEV:
1544 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1546 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1550 if (file->f_op->aio_write)
1551 kiocb->ki_retry = aio_rw_vect_retry;
1553 case IOCB_CMD_FDSYNC:
1555 if (file->f_op->aio_fsync)
1556 kiocb->ki_retry = aio_fdsync;
1558 case IOCB_CMD_FSYNC:
1560 if (file->f_op->aio_fsync)
1561 kiocb->ki_retry = aio_fsync;
1564 dprintk("EINVAL: io_submit: no operation provided\n");
1568 if (!kiocb->ki_retry)
1574 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1575 struct iocb *iocb, struct kiocb_batch *batch,
1582 /* enforce forwards compatibility on users */
1583 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1584 pr_debug("EINVAL: io_submit: reserve field set\n");
1588 /* prevent overflows */
1590 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1591 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1592 ((ssize_t)iocb->aio_nbytes < 0)
1594 pr_debug("EINVAL: io_submit: overflow check\n");
1598 file = fget(iocb->aio_fildes);
1599 if (unlikely(!file))
1602 req = aio_get_req(ctx, batch); /* returns with 2 references to req */
1603 if (unlikely(!req)) {
1607 req->ki_filp = file;
1608 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1610 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1611 * instance of the file* now. The file descriptor must be
1612 * an eventfd() fd, and will be signaled for each completed
1613 * event using the eventfd_signal() function.
1615 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1616 if (IS_ERR(req->ki_eventfd)) {
1617 ret = PTR_ERR(req->ki_eventfd);
1618 req->ki_eventfd = NULL;
1623 ret = put_user(req->ki_key, &user_iocb->aio_key);
1624 if (unlikely(ret)) {
1625 dprintk("EFAULT: aio_key\n");
1629 req->ki_obj.user = user_iocb;
1630 req->ki_user_data = iocb->aio_data;
1631 req->ki_pos = iocb->aio_offset;
1633 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1634 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1635 req->ki_opcode = iocb->aio_lio_opcode;
1637 ret = aio_setup_iocb(req, compat);
1642 spin_lock_irq(&ctx->ctx_lock);
1644 * We could have raced with io_destroy() and are currently holding a
1645 * reference to ctx which should be destroyed. We cannot submit IO
1646 * since ctx gets freed as soon as io_submit() puts its reference. The
1647 * check here is reliable: io_destroy() sets ctx->dead before waiting
1648 * for outstanding IO and the barrier between these two is realized by
1649 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock. Analogously we
1650 * increment ctx->reqs_active before checking for ctx->dead and the
1651 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
1652 * don't see ctx->dead set here, io_destroy() waits for our IO to
1656 spin_unlock_irq(&ctx->ctx_lock);
1661 if (!list_empty(&ctx->run_list)) {
1662 /* drain the run list */
1663 while (__aio_run_iocbs(ctx))
1666 spin_unlock_irq(&ctx->ctx_lock);
1668 aio_put_req(req); /* drop extra ref to req */
1672 aio_put_req(req); /* drop extra ref to req */
1673 aio_put_req(req); /* drop i/o ref to req */
1677 long do_io_submit(aio_context_t ctx_id, long nr,
1678 struct iocb __user *__user *iocbpp, bool compat)
1683 struct blk_plug plug;
1684 struct kiocb_batch batch;
1686 if (unlikely(nr < 0))
1689 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1690 nr = LONG_MAX/sizeof(*iocbpp);
1692 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1695 ctx = lookup_ioctx(ctx_id);
1696 if (unlikely(!ctx)) {
1697 pr_debug("EINVAL: io_submit: invalid context id\n");
1701 kiocb_batch_init(&batch, nr);
1703 blk_start_plug(&plug);
1706 * AKPM: should this return a partial result if some of the IOs were
1707 * successfully submitted?
1709 for (i=0; i<nr; i++) {
1710 struct iocb __user *user_iocb;
1713 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1718 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1723 ret = io_submit_one(ctx, user_iocb, &tmp, &batch, compat);
1727 blk_finish_plug(&plug);
1729 kiocb_batch_free(ctx, &batch);
1735 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1736 * the number of iocbs queued. May return -EINVAL if the aio_context
1737 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1738 * *iocbpp[0] is not properly initialized, if the operation specified
1739 * is invalid for the file descriptor in the iocb. May fail with
1740 * -EFAULT if any of the data structures point to invalid data. May
1741 * fail with -EBADF if the file descriptor specified in the first
1742 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1743 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1744 * fail with -ENOSYS if not implemented.
1746 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1747 struct iocb __user * __user *, iocbpp)
1749 return do_io_submit(ctx_id, nr, iocbpp, 0);
1753 * Finds a given iocb for cancellation.
1755 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1758 struct list_head *pos;
1760 assert_spin_locked(&ctx->ctx_lock);
1762 /* TODO: use a hash or array, this sucks. */
1763 list_for_each(pos, &ctx->active_reqs) {
1764 struct kiocb *kiocb = list_kiocb(pos);
1765 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1772 * Attempts to cancel an iocb previously passed to io_submit. If
1773 * the operation is successfully cancelled, the resulting event is
1774 * copied into the memory pointed to by result without being placed
1775 * into the completion queue and 0 is returned. May fail with
1776 * -EFAULT if any of the data structures pointed to are invalid.
1777 * May fail with -EINVAL if aio_context specified by ctx_id is
1778 * invalid. May fail with -EAGAIN if the iocb specified was not
1779 * cancelled. Will fail with -ENOSYS if not implemented.
1781 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1782 struct io_event __user *, result)
1784 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1786 struct kiocb *kiocb;
1790 ret = get_user(key, &iocb->aio_key);
1794 ctx = lookup_ioctx(ctx_id);
1798 spin_lock_irq(&ctx->ctx_lock);
1800 kiocb = lookup_kiocb(ctx, iocb, key);
1801 if (kiocb && kiocb->ki_cancel) {
1802 cancel = kiocb->ki_cancel;
1804 kiocbSetCancelled(kiocb);
1807 spin_unlock_irq(&ctx->ctx_lock);
1809 if (NULL != cancel) {
1810 struct io_event tmp;
1811 pr_debug("calling cancel\n");
1812 memset(&tmp, 0, sizeof(tmp));
1813 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1814 tmp.data = kiocb->ki_user_data;
1815 ret = cancel(kiocb, &tmp);
1817 /* Cancellation succeeded -- copy the result
1818 * into the user's buffer.
1820 if (copy_to_user(result, &tmp, sizeof(tmp)))
1832 * Attempts to read at least min_nr events and up to nr events from
1833 * the completion queue for the aio_context specified by ctx_id. If
1834 * it succeeds, the number of read events is returned. May fail with
1835 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1836 * out of range, if timeout is out of range. May fail with -EFAULT
1837 * if any of the memory specified is invalid. May return 0 or
1838 * < min_nr if the timeout specified by timeout has elapsed
1839 * before sufficient events are available, where timeout == NULL
1840 * specifies an infinite timeout. Note that the timeout pointed to by
1841 * timeout is relative and will be updated if not NULL and the
1842 * operation blocks. Will fail with -ENOSYS if not implemented.
1844 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1847 struct io_event __user *, events,
1848 struct timespec __user *, timeout)
1850 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1853 if (likely(ioctx)) {
1854 if (likely(min_nr <= nr && min_nr >= 0))
1855 ret = read_events(ioctx, min_nr, nr, events, timeout);
1859 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);