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/uio.h>
22 #include <linux/sched.h>
24 #include <linux/file.h>
26 #include <linux/mman.h>
27 #include <linux/slab.h>
28 #include <linux/timer.h>
29 #include <linux/aio.h>
30 #include <linux/highmem.h>
31 #include <linux/workqueue.h>
32 #include <linux/security.h>
33 #include <linux/eventfd.h>
35 #include <asm/kmap_types.h>
36 #include <asm/uaccess.h>
37 #include <asm/mmu_context.h>
40 #define dprintk printk
42 #define dprintk(x...) do { ; } while (0)
45 /*------ sysctl variables----*/
46 static DEFINE_SPINLOCK(aio_nr_lock);
47 unsigned long aio_nr; /* current system wide number of aio requests */
48 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
49 /*----end sysctl variables---*/
51 static struct kmem_cache *kiocb_cachep;
52 static struct kmem_cache *kioctx_cachep;
54 static struct workqueue_struct *aio_wq;
56 /* Used for rare fput completion. */
57 static void aio_fput_routine(struct work_struct *);
58 static DECLARE_WORK(fput_work, aio_fput_routine);
60 static DEFINE_SPINLOCK(fput_lock);
61 static LIST_HEAD(fput_head);
63 static void aio_kick_handler(struct work_struct *);
64 static void aio_queue_work(struct kioctx *);
67 * Creates the slab caches used by the aio routines, panic on
68 * failure as this is done early during the boot sequence.
70 static int __init aio_setup(void)
72 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
73 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
75 aio_wq = create_workqueue("aio");
77 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
82 static void aio_free_ring(struct kioctx *ctx)
84 struct aio_ring_info *info = &ctx->ring_info;
87 for (i=0; i<info->nr_pages; i++)
88 put_page(info->ring_pages[i]);
90 if (info->mmap_size) {
91 down_write(&ctx->mm->mmap_sem);
92 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
93 up_write(&ctx->mm->mmap_sem);
96 if (info->ring_pages && info->ring_pages != info->internal_pages)
97 kfree(info->ring_pages);
98 info->ring_pages = NULL;
102 static int aio_setup_ring(struct kioctx *ctx)
104 struct aio_ring *ring;
105 struct aio_ring_info *info = &ctx->ring_info;
106 unsigned nr_events = ctx->max_reqs;
110 /* Compensate for the ring buffer's head/tail overlap entry */
111 nr_events += 2; /* 1 is required, 2 for good luck */
113 size = sizeof(struct aio_ring);
114 size += sizeof(struct io_event) * nr_events;
115 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
120 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
123 info->ring_pages = info->internal_pages;
124 if (nr_pages > AIO_RING_PAGES) {
125 info->ring_pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
126 if (!info->ring_pages)
130 info->mmap_size = nr_pages * PAGE_SIZE;
131 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
132 down_write(&ctx->mm->mmap_sem);
133 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
134 PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE,
136 if (IS_ERR((void *)info->mmap_base)) {
137 up_write(&ctx->mm->mmap_sem);
143 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
144 info->nr_pages = get_user_pages(current, ctx->mm,
145 info->mmap_base, nr_pages,
146 1, 0, info->ring_pages, NULL);
147 up_write(&ctx->mm->mmap_sem);
149 if (unlikely(info->nr_pages != nr_pages)) {
154 ctx->user_id = info->mmap_base;
156 info->nr = nr_events; /* trusted copy */
158 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
159 ring->nr = nr_events; /* user copy */
160 ring->id = ctx->user_id;
161 ring->head = ring->tail = 0;
162 ring->magic = AIO_RING_MAGIC;
163 ring->compat_features = AIO_RING_COMPAT_FEATURES;
164 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
165 ring->header_length = sizeof(struct aio_ring);
166 kunmap_atomic(ring, KM_USER0);
172 /* aio_ring_event: returns a pointer to the event at the given index from
173 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
175 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
176 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
177 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
179 #define aio_ring_event(info, nr, km) ({ \
180 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
181 struct io_event *__event; \
182 __event = kmap_atomic( \
183 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
184 __event += pos % AIO_EVENTS_PER_PAGE; \
188 #define put_aio_ring_event(event, km) do { \
189 struct io_event *__event = (event); \
191 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
194 static void ctx_rcu_free(struct rcu_head *head)
196 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
197 unsigned nr_events = ctx->max_reqs;
199 kmem_cache_free(kioctx_cachep, ctx);
202 spin_lock(&aio_nr_lock);
203 BUG_ON(aio_nr - nr_events > aio_nr);
205 spin_unlock(&aio_nr_lock);
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 BUG_ON(ctx->reqs_active);
217 cancel_delayed_work(&ctx->wq);
218 cancel_work_sync(&ctx->wq.work);
222 pr_debug("__put_ioctx: freeing %p\n", ctx);
223 call_rcu(&ctx->rcu_head, ctx_rcu_free);
226 #define get_ioctx(kioctx) do { \
227 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
228 atomic_inc(&(kioctx)->users); \
230 #define put_ioctx(kioctx) do { \
231 BUG_ON(atomic_read(&(kioctx)->users) <= 0); \
232 if (unlikely(atomic_dec_and_test(&(kioctx)->users))) \
233 __put_ioctx(kioctx); \
237 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
239 static struct kioctx *ioctx_alloc(unsigned nr_events)
241 struct mm_struct *mm;
245 /* Prevent overflows */
246 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
247 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
248 pr_debug("ENOMEM: nr_events too high\n");
249 return ERR_PTR(-EINVAL);
252 if ((unsigned long)nr_events > aio_max_nr)
253 return ERR_PTR(-EAGAIN);
255 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
257 return ERR_PTR(-ENOMEM);
259 ctx->max_reqs = nr_events;
260 mm = ctx->mm = current->mm;
261 atomic_inc(&mm->mm_count);
263 atomic_set(&ctx->users, 1);
264 spin_lock_init(&ctx->ctx_lock);
265 spin_lock_init(&ctx->ring_info.ring_lock);
266 init_waitqueue_head(&ctx->wait);
268 INIT_LIST_HEAD(&ctx->active_reqs);
269 INIT_LIST_HEAD(&ctx->run_list);
270 INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
272 if (aio_setup_ring(ctx) < 0)
275 /* limit the number of system wide aios */
277 spin_lock_bh(&aio_nr_lock);
278 if (aio_nr + nr_events > aio_max_nr ||
279 aio_nr + nr_events < aio_nr)
282 aio_nr += ctx->max_reqs;
283 spin_unlock_bh(&aio_nr_lock);
284 if (ctx->max_reqs || did_sync)
287 /* wait for rcu callbacks to have completed before giving up */
290 ctx->max_reqs = nr_events;
293 if (ctx->max_reqs == 0)
296 /* now link into global list. */
297 spin_lock(&mm->ioctx_lock);
298 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
299 spin_unlock(&mm->ioctx_lock);
301 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
302 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
307 return ERR_PTR(-EAGAIN);
311 kmem_cache_free(kioctx_cachep, ctx);
312 ctx = ERR_PTR(-ENOMEM);
314 dprintk("aio: error allocating ioctx %p\n", ctx);
319 * Cancels all outstanding aio requests on an aio context. Used
320 * when the processes owning a context have all exited to encourage
321 * the rapid destruction of the kioctx.
323 static void aio_cancel_all(struct kioctx *ctx)
325 int (*cancel)(struct kiocb *, struct io_event *);
327 spin_lock_irq(&ctx->ctx_lock);
329 while (!list_empty(&ctx->active_reqs)) {
330 struct list_head *pos = ctx->active_reqs.next;
331 struct kiocb *iocb = list_kiocb(pos);
332 list_del_init(&iocb->ki_list);
333 cancel = iocb->ki_cancel;
334 kiocbSetCancelled(iocb);
337 spin_unlock_irq(&ctx->ctx_lock);
339 spin_lock_irq(&ctx->ctx_lock);
342 spin_unlock_irq(&ctx->ctx_lock);
345 static void wait_for_all_aios(struct kioctx *ctx)
347 struct task_struct *tsk = current;
348 DECLARE_WAITQUEUE(wait, tsk);
350 spin_lock_irq(&ctx->ctx_lock);
351 if (!ctx->reqs_active)
354 add_wait_queue(&ctx->wait, &wait);
355 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
356 while (ctx->reqs_active) {
357 spin_unlock_irq(&ctx->ctx_lock);
359 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
360 spin_lock_irq(&ctx->ctx_lock);
362 __set_task_state(tsk, TASK_RUNNING);
363 remove_wait_queue(&ctx->wait, &wait);
366 spin_unlock_irq(&ctx->ctx_lock);
369 /* wait_on_sync_kiocb:
370 * Waits on the given sync kiocb to complete.
372 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
374 while (iocb->ki_users) {
375 set_current_state(TASK_UNINTERRUPTIBLE);
380 __set_current_state(TASK_RUNNING);
381 return iocb->ki_user_data;
384 /* exit_aio: called when the last user of mm goes away. At this point,
385 * there is no way for any new requests to be submited or any of the
386 * io_* syscalls to be called on the context. However, there may be
387 * outstanding requests which hold references to the context; as they
388 * go away, they will call put_ioctx and release any pinned memory
389 * associated with the request (held via struct page * references).
391 void exit_aio(struct mm_struct *mm)
395 while (!hlist_empty(&mm->ioctx_list)) {
396 ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
397 hlist_del_rcu(&ctx->list);
401 wait_for_all_aios(ctx);
403 * Ensure we don't leave the ctx on the aio_wq
405 cancel_work_sync(&ctx->wq.work);
407 if (1 != atomic_read(&ctx->users))
409 "exit_aio:ioctx still alive: %d %d %d\n",
410 atomic_read(&ctx->users), ctx->dead,
417 * Allocate a slot for an aio request. Increments the users count
418 * of the kioctx so that the kioctx stays around until all requests are
419 * complete. Returns NULL if no requests are free.
421 * Returns with kiocb->users set to 2. The io submit code path holds
422 * an extra reference while submitting the i/o.
423 * This prevents races between the aio code path referencing the
424 * req (after submitting it) and aio_complete() freeing the req.
426 static struct kiocb *__aio_get_req(struct kioctx *ctx)
428 struct kiocb *req = NULL;
429 struct aio_ring *ring;
432 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
440 req->ki_cancel = NULL;
441 req->ki_retry = NULL;
444 req->ki_iovec = NULL;
445 INIT_LIST_HEAD(&req->ki_run_list);
446 req->ki_eventfd = NULL;
448 /* Check if the completion queue has enough free space to
449 * accept an event from this io.
451 spin_lock_irq(&ctx->ctx_lock);
452 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
453 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
454 list_add(&req->ki_list, &ctx->active_reqs);
458 kunmap_atomic(ring, KM_USER0);
459 spin_unlock_irq(&ctx->ctx_lock);
462 kmem_cache_free(kiocb_cachep, req);
469 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
472 /* Handle a potential starvation case -- should be exceedingly rare as
473 * requests will be stuck on fput_head only if the aio_fput_routine is
474 * delayed and the requests were the last user of the struct file.
476 req = __aio_get_req(ctx);
477 if (unlikely(NULL == req)) {
478 aio_fput_routine(NULL);
479 req = __aio_get_req(ctx);
484 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
486 assert_spin_locked(&ctx->ctx_lock);
490 if (req->ki_iovec != &req->ki_inline_vec)
491 kfree(req->ki_iovec);
492 kmem_cache_free(kiocb_cachep, req);
495 if (unlikely(!ctx->reqs_active && ctx->dead))
499 static void aio_fput_routine(struct work_struct *data)
501 spin_lock_irq(&fput_lock);
502 while (likely(!list_empty(&fput_head))) {
503 struct kiocb *req = list_kiocb(fput_head.next);
504 struct kioctx *ctx = req->ki_ctx;
506 list_del(&req->ki_list);
507 spin_unlock_irq(&fput_lock);
509 /* Complete the fput(s) */
510 if (req->ki_filp != NULL)
511 __fput(req->ki_filp);
512 if (req->ki_eventfd != NULL)
513 __fput(req->ki_eventfd);
515 /* Link the iocb into the context's free list */
516 spin_lock_irq(&ctx->ctx_lock);
517 really_put_req(ctx, req);
518 spin_unlock_irq(&ctx->ctx_lock);
521 spin_lock_irq(&fput_lock);
523 spin_unlock_irq(&fput_lock);
527 * Returns true if this put was the last user of the request.
529 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
531 int schedule_putreq = 0;
533 dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
534 req, atomic_long_read(&req->ki_filp->f_count));
536 assert_spin_locked(&ctx->ctx_lock);
539 BUG_ON(req->ki_users < 0);
540 if (likely(req->ki_users))
542 list_del(&req->ki_list); /* remove from active_reqs */
543 req->ki_cancel = NULL;
544 req->ki_retry = NULL;
547 * Try to optimize the aio and eventfd file* puts, by avoiding to
548 * schedule work in case it is not __fput() time. In normal cases,
549 * we would not be holding the last reference to the file*, so
550 * this function will be executed w/out any aio kthread wakeup.
552 if (unlikely(atomic_long_dec_and_test(&req->ki_filp->f_count)))
556 if (req->ki_eventfd != NULL) {
557 if (unlikely(atomic_long_dec_and_test(&req->ki_eventfd->f_count)))
560 req->ki_eventfd = NULL;
562 if (unlikely(schedule_putreq)) {
564 spin_lock(&fput_lock);
565 list_add(&req->ki_list, &fput_head);
566 spin_unlock(&fput_lock);
567 queue_work(aio_wq, &fput_work);
569 really_put_req(ctx, req);
574 * Returns true if this put was the last user of the kiocb,
575 * false if the request is still in use.
577 int aio_put_req(struct kiocb *req)
579 struct kioctx *ctx = req->ki_ctx;
581 spin_lock_irq(&ctx->ctx_lock);
582 ret = __aio_put_req(ctx, req);
583 spin_unlock_irq(&ctx->ctx_lock);
587 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
589 struct mm_struct *mm = current->mm;
590 struct kioctx *ctx = NULL;
591 struct hlist_node *n;
595 hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
596 if (ctx->user_id == ctx_id && !ctx->dead) {
608 * Makes the calling kernel thread take on the specified
610 * Called by the retry thread execute retries within the
611 * iocb issuer's mm context, so that copy_from/to_user
612 * operations work seamlessly for aio.
613 * (Note: this routine is intended to be called only
614 * from a kernel thread context)
616 static void use_mm(struct mm_struct *mm)
618 struct mm_struct *active_mm;
619 struct task_struct *tsk = current;
622 active_mm = tsk->active_mm;
623 atomic_inc(&mm->mm_count);
626 switch_mm(active_mm, mm, tsk);
634 * Reverses the effect of use_mm, i.e. releases the
635 * specified mm context which was earlier taken on
636 * by the calling kernel thread
637 * (Note: this routine is intended to be called only
638 * from a kernel thread context)
640 static void unuse_mm(struct mm_struct *mm)
642 struct task_struct *tsk = current;
646 /* active_mm is still 'mm' */
647 enter_lazy_tlb(mm, tsk);
652 * Queue up a kiocb to be retried. Assumes that the kiocb
653 * has already been marked as kicked, and places it on
654 * the retry run list for the corresponding ioctx, if it
655 * isn't already queued. Returns 1 if it actually queued
656 * the kiocb (to tell the caller to activate the work
657 * queue to process it), or 0, if it found that it was
660 static inline int __queue_kicked_iocb(struct kiocb *iocb)
662 struct kioctx *ctx = iocb->ki_ctx;
664 assert_spin_locked(&ctx->ctx_lock);
666 if (list_empty(&iocb->ki_run_list)) {
667 list_add_tail(&iocb->ki_run_list,
675 * This is the core aio execution routine. It is
676 * invoked both for initial i/o submission and
677 * subsequent retries via the aio_kick_handler.
678 * Expects to be invoked with iocb->ki_ctx->lock
679 * already held. The lock is released and reacquired
680 * as needed during processing.
682 * Calls the iocb retry method (already setup for the
683 * iocb on initial submission) for operation specific
684 * handling, but takes care of most of common retry
685 * execution details for a given iocb. The retry method
686 * needs to be non-blocking as far as possible, to avoid
687 * holding up other iocbs waiting to be serviced by the
688 * retry kernel thread.
690 * The trickier parts in this code have to do with
691 * ensuring that only one retry instance is in progress
692 * for a given iocb at any time. Providing that guarantee
693 * simplifies the coding of individual aio operations as
694 * it avoids various potential races.
696 static ssize_t aio_run_iocb(struct kiocb *iocb)
698 struct kioctx *ctx = iocb->ki_ctx;
699 ssize_t (*retry)(struct kiocb *);
702 if (!(retry = iocb->ki_retry)) {
703 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
708 * We don't want the next retry iteration for this
709 * operation to start until this one has returned and
710 * updated the iocb state. However, wait_queue functions
711 * can trigger a kick_iocb from interrupt context in the
712 * meantime, indicating that data is available for the next
713 * iteration. We want to remember that and enable the
714 * next retry iteration _after_ we are through with
717 * So, in order to be able to register a "kick", but
718 * prevent it from being queued now, we clear the kick
719 * flag, but make the kick code *think* that the iocb is
720 * still on the run list until we are actually done.
721 * When we are done with this iteration, we check if
722 * the iocb was kicked in the meantime and if so, queue
726 kiocbClearKicked(iocb);
729 * This is so that aio_complete knows it doesn't need to
730 * pull the iocb off the run list (We can't just call
731 * INIT_LIST_HEAD because we don't want a kick_iocb to
732 * queue this on the run list yet)
734 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
735 spin_unlock_irq(&ctx->ctx_lock);
737 /* Quit retrying if the i/o has been cancelled */
738 if (kiocbIsCancelled(iocb)) {
740 aio_complete(iocb, ret, 0);
741 /* must not access the iocb after this */
746 * Now we are all set to call the retry method in async
751 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
752 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
753 aio_complete(iocb, ret, 0);
756 spin_lock_irq(&ctx->ctx_lock);
758 if (-EIOCBRETRY == ret) {
760 * OK, now that we are done with this iteration
761 * and know that there is more left to go,
762 * this is where we let go so that a subsequent
763 * "kick" can start the next iteration
766 /* will make __queue_kicked_iocb succeed from here on */
767 INIT_LIST_HEAD(&iocb->ki_run_list);
768 /* we must queue the next iteration ourselves, if it
769 * has already been kicked */
770 if (kiocbIsKicked(iocb)) {
771 __queue_kicked_iocb(iocb);
774 * __queue_kicked_iocb will always return 1 here, because
775 * iocb->ki_run_list is empty at this point so it should
776 * be safe to unconditionally queue the context into the
787 * Process all pending retries queued on the ioctx
789 * Assumes it is operating within the aio issuer's mm
792 static int __aio_run_iocbs(struct kioctx *ctx)
795 struct list_head run_list;
797 assert_spin_locked(&ctx->ctx_lock);
799 list_replace_init(&ctx->run_list, &run_list);
800 while (!list_empty(&run_list)) {
801 iocb = list_entry(run_list.next, struct kiocb,
803 list_del(&iocb->ki_run_list);
805 * Hold an extra reference while retrying i/o.
807 iocb->ki_users++; /* grab extra reference */
809 __aio_put_req(ctx, iocb);
811 if (!list_empty(&ctx->run_list))
816 static void aio_queue_work(struct kioctx * ctx)
818 unsigned long timeout;
820 * if someone is waiting, get the work started right
821 * away, otherwise, use a longer delay
824 if (waitqueue_active(&ctx->wait))
828 queue_delayed_work(aio_wq, &ctx->wq, timeout);
834 * Process all pending retries queued on the ioctx
836 * Assumes it is operating within the aio issuer's mm
839 static inline void aio_run_iocbs(struct kioctx *ctx)
843 spin_lock_irq(&ctx->ctx_lock);
845 requeue = __aio_run_iocbs(ctx);
846 spin_unlock_irq(&ctx->ctx_lock);
852 * just like aio_run_iocbs, but keeps running them until
853 * the list stays empty
855 static inline void aio_run_all_iocbs(struct kioctx *ctx)
857 spin_lock_irq(&ctx->ctx_lock);
858 while (__aio_run_iocbs(ctx))
860 spin_unlock_irq(&ctx->ctx_lock);
865 * Work queue handler triggered to process pending
866 * retries on an ioctx. Takes on the aio issuer's
867 * mm context before running the iocbs, so that
868 * copy_xxx_user operates on the issuer's address
870 * Run on aiod's context.
872 static void aio_kick_handler(struct work_struct *work)
874 struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
875 mm_segment_t oldfs = get_fs();
876 struct mm_struct *mm;
881 spin_lock_irq(&ctx->ctx_lock);
882 requeue =__aio_run_iocbs(ctx);
884 spin_unlock_irq(&ctx->ctx_lock);
888 * we're in a worker thread already, don't use queue_delayed_work,
891 queue_delayed_work(aio_wq, &ctx->wq, 0);
896 * Called by kick_iocb to queue the kiocb for retry
897 * and if required activate the aio work queue to process
900 static void try_queue_kicked_iocb(struct kiocb *iocb)
902 struct kioctx *ctx = iocb->ki_ctx;
906 /* We're supposed to be the only path putting the iocb back on the run
907 * list. If we find that the iocb is *back* on a wait queue already
908 * than retry has happened before we could queue the iocb. This also
909 * means that the retry could have completed and freed our iocb, no
911 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
913 spin_lock_irqsave(&ctx->ctx_lock, flags);
914 /* set this inside the lock so that we can't race with aio_run_iocb()
915 * testing it and putting the iocb on the run list under the lock */
916 if (!kiocbTryKick(iocb))
917 run = __queue_kicked_iocb(iocb);
918 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
925 * Called typically from a wait queue callback context
926 * (aio_wake_function) to trigger a retry of the iocb.
927 * The retry is usually executed by aio workqueue
928 * threads (See aio_kick_handler).
930 void kick_iocb(struct kiocb *iocb)
932 /* sync iocbs are easy: they can only ever be executing from a
934 if (is_sync_kiocb(iocb)) {
935 kiocbSetKicked(iocb);
936 wake_up_process(iocb->ki_obj.tsk);
940 try_queue_kicked_iocb(iocb);
942 EXPORT_SYMBOL(kick_iocb);
945 * Called when the io request on the given iocb is complete.
946 * Returns true if this is the last user of the request. The
947 * only other user of the request can be the cancellation code.
949 int aio_complete(struct kiocb *iocb, long res, long res2)
951 struct kioctx *ctx = iocb->ki_ctx;
952 struct aio_ring_info *info;
953 struct aio_ring *ring;
954 struct io_event *event;
960 * Special case handling for sync iocbs:
961 * - events go directly into the iocb for fast handling
962 * - the sync task with the iocb in its stack holds the single iocb
963 * ref, no other paths have a way to get another ref
964 * - the sync task helpfully left a reference to itself in the iocb
966 if (is_sync_kiocb(iocb)) {
967 BUG_ON(iocb->ki_users != 1);
968 iocb->ki_user_data = res;
970 wake_up_process(iocb->ki_obj.tsk);
974 info = &ctx->ring_info;
976 /* add a completion event to the ring buffer.
977 * must be done holding ctx->ctx_lock to prevent
978 * other code from messing with the tail
979 * pointer since we might be called from irq
982 spin_lock_irqsave(&ctx->ctx_lock, flags);
984 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
985 list_del_init(&iocb->ki_run_list);
988 * cancelled requests don't get events, userland was given one
989 * when the event got cancelled.
991 if (kiocbIsCancelled(iocb))
994 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
997 event = aio_ring_event(info, tail, KM_IRQ0);
998 if (++tail >= info->nr)
1001 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1002 event->data = iocb->ki_user_data;
1006 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
1007 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1010 /* after flagging the request as done, we
1011 * must never even look at it again
1013 smp_wmb(); /* make event visible before updating tail */
1018 put_aio_ring_event(event, KM_IRQ0);
1019 kunmap_atomic(ring, KM_IRQ1);
1021 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1024 * Check if the user asked us to deliver the result through an
1025 * eventfd. The eventfd_signal() function is safe to be called
1028 if (iocb->ki_eventfd != NULL)
1029 eventfd_signal(iocb->ki_eventfd, 1);
1032 /* everything turned out well, dispose of the aiocb. */
1033 ret = __aio_put_req(ctx, iocb);
1036 * We have to order our ring_info tail store above and test
1037 * of the wait list below outside the wait lock. This is
1038 * like in wake_up_bit() where clearing a bit has to be
1039 * ordered with the unlocked test.
1043 if (waitqueue_active(&ctx->wait))
1044 wake_up(&ctx->wait);
1046 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1051 * Pull an event off of the ioctx's event ring. Returns the number of
1052 * events fetched (0 or 1 ;-)
1053 * FIXME: make this use cmpxchg.
1054 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1056 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1058 struct aio_ring_info *info = &ioctx->ring_info;
1059 struct aio_ring *ring;
1063 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1064 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1065 (unsigned long)ring->head, (unsigned long)ring->tail,
1066 (unsigned long)ring->nr);
1068 if (ring->head == ring->tail)
1071 spin_lock(&info->ring_lock);
1073 head = ring->head % info->nr;
1074 if (head != ring->tail) {
1075 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1077 head = (head + 1) % info->nr;
1078 smp_mb(); /* finish reading the event before updatng the head */
1081 put_aio_ring_event(evp, KM_USER1);
1083 spin_unlock(&info->ring_lock);
1086 kunmap_atomic(ring, KM_USER0);
1087 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1088 (unsigned long)ring->head, (unsigned long)ring->tail);
1092 struct aio_timeout {
1093 struct timer_list timer;
1095 struct task_struct *p;
1098 static void timeout_func(unsigned long data)
1100 struct aio_timeout *to = (struct aio_timeout *)data;
1103 wake_up_process(to->p);
1106 static inline void init_timeout(struct aio_timeout *to)
1108 setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
1113 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1114 const struct timespec *ts)
1116 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1117 if (time_after(to->timer.expires, jiffies))
1118 add_timer(&to->timer);
1123 static inline void clear_timeout(struct aio_timeout *to)
1125 del_singleshot_timer_sync(&to->timer);
1128 static int read_events(struct kioctx *ctx,
1129 long min_nr, long nr,
1130 struct io_event __user *event,
1131 struct timespec __user *timeout)
1133 long start_jiffies = jiffies;
1134 struct task_struct *tsk = current;
1135 DECLARE_WAITQUEUE(wait, tsk);
1138 struct io_event ent;
1139 struct aio_timeout to;
1142 /* needed to zero any padding within an entry (there shouldn't be
1143 * any, but C is fun!
1145 memset(&ent, 0, sizeof(ent));
1148 while (likely(i < nr)) {
1149 ret = aio_read_evt(ctx, &ent);
1150 if (unlikely(ret <= 0))
1153 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1154 ent.data, ent.obj, ent.res, ent.res2);
1156 /* Could we split the check in two? */
1158 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1159 dprintk("aio: lost an event due to EFAULT.\n");
1164 /* Good, event copied to userland, update counts. */
1176 /* racey check, but it gets redone */
1177 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1179 aio_run_all_iocbs(ctx);
1187 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1190 set_timeout(start_jiffies, &to, &ts);
1193 while (likely(i < nr)) {
1194 add_wait_queue_exclusive(&ctx->wait, &wait);
1196 set_task_state(tsk, TASK_INTERRUPTIBLE);
1197 ret = aio_read_evt(ctx, &ent);
1202 if (unlikely(ctx->dead)) {
1206 if (to.timed_out) /* Only check after read evt */
1208 /* Try to only show up in io wait if there are ops
1210 if (ctx->reqs_active)
1214 if (signal_pending(tsk)) {
1218 /*ret = aio_read_evt(ctx, &ent);*/
1221 set_task_state(tsk, TASK_RUNNING);
1222 remove_wait_queue(&ctx->wait, &wait);
1224 if (unlikely(ret <= 0))
1228 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1229 dprintk("aio: lost an event due to EFAULT.\n");
1233 /* Good, event copied to userland, update counts. */
1241 destroy_timer_on_stack(&to.timer);
1245 /* Take an ioctx and remove it from the list of ioctx's. Protects
1246 * against races with itself via ->dead.
1248 static void io_destroy(struct kioctx *ioctx)
1250 struct mm_struct *mm = current->mm;
1253 /* delete the entry from the list is someone else hasn't already */
1254 spin_lock(&mm->ioctx_lock);
1255 was_dead = ioctx->dead;
1257 hlist_del_rcu(&ioctx->list);
1258 spin_unlock(&mm->ioctx_lock);
1260 dprintk("aio_release(%p)\n", ioctx);
1261 if (likely(!was_dead))
1262 put_ioctx(ioctx); /* twice for the list */
1264 aio_cancel_all(ioctx);
1265 wait_for_all_aios(ioctx);
1268 * Wake up any waiters. The setting of ctx->dead must be seen
1269 * by other CPUs at this point. Right now, we rely on the
1270 * locking done by the above calls to ensure this consistency.
1272 wake_up(&ioctx->wait);
1273 put_ioctx(ioctx); /* once for the lookup */
1277 * Create an aio_context capable of receiving at least nr_events.
1278 * ctxp must not point to an aio_context that already exists, and
1279 * must be initialized to 0 prior to the call. On successful
1280 * creation of the aio_context, *ctxp is filled in with the resulting
1281 * handle. May fail with -EINVAL if *ctxp is not initialized,
1282 * if the specified nr_events exceeds internal limits. May fail
1283 * with -EAGAIN if the specified nr_events exceeds the user's limit
1284 * of available events. May fail with -ENOMEM if insufficient kernel
1285 * resources are available. May fail with -EFAULT if an invalid
1286 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1289 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1291 struct kioctx *ioctx = NULL;
1295 ret = get_user(ctx, ctxp);
1300 if (unlikely(ctx || nr_events == 0)) {
1301 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1306 ioctx = ioctx_alloc(nr_events);
1307 ret = PTR_ERR(ioctx);
1308 if (!IS_ERR(ioctx)) {
1309 ret = put_user(ioctx->user_id, ctxp);
1313 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1322 * Destroy the aio_context specified. May cancel any outstanding
1323 * AIOs and block on completion. Will fail with -ENOSYS if not
1324 * implemented. May fail with -EFAULT if the context pointed to
1327 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1329 struct kioctx *ioctx = lookup_ioctx(ctx);
1330 if (likely(NULL != ioctx)) {
1334 pr_debug("EINVAL: io_destroy: invalid context id\n");
1338 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1340 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1344 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1345 ssize_t this = min((ssize_t)iov->iov_len, ret);
1346 iov->iov_base += this;
1347 iov->iov_len -= this;
1348 iocb->ki_left -= this;
1350 if (iov->iov_len == 0) {
1356 /* the caller should not have done more io than what fit in
1357 * the remaining iovecs */
1358 BUG_ON(ret > 0 && iocb->ki_left == 0);
1361 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
1363 struct file *file = iocb->ki_filp;
1364 struct address_space *mapping = file->f_mapping;
1365 struct inode *inode = mapping->host;
1366 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
1367 unsigned long, loff_t);
1369 unsigned short opcode;
1371 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
1372 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
1373 rw_op = file->f_op->aio_read;
1374 opcode = IOCB_CMD_PREADV;
1376 rw_op = file->f_op->aio_write;
1377 opcode = IOCB_CMD_PWRITEV;
1380 /* This matches the pread()/pwrite() logic */
1381 if (iocb->ki_pos < 0)
1385 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1386 iocb->ki_nr_segs - iocb->ki_cur_seg,
1389 aio_advance_iovec(iocb, ret);
1391 /* retry all partial writes. retry partial reads as long as its a
1393 } while (ret > 0 && iocb->ki_left > 0 &&
1394 (opcode == IOCB_CMD_PWRITEV ||
1395 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1397 /* This means we must have transferred all that we could */
1398 /* No need to retry anymore */
1399 if ((ret == 0) || (iocb->ki_left == 0))
1400 ret = iocb->ki_nbytes - iocb->ki_left;
1402 /* If we managed to write some out we return that, rather than
1403 * the eventual error. */
1404 if (opcode == IOCB_CMD_PWRITEV
1405 && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
1406 && iocb->ki_nbytes - iocb->ki_left)
1407 ret = iocb->ki_nbytes - iocb->ki_left;
1412 static ssize_t aio_fdsync(struct kiocb *iocb)
1414 struct file *file = iocb->ki_filp;
1415 ssize_t ret = -EINVAL;
1417 if (file->f_op->aio_fsync)
1418 ret = file->f_op->aio_fsync(iocb, 1);
1422 static ssize_t aio_fsync(struct kiocb *iocb)
1424 struct file *file = iocb->ki_filp;
1425 ssize_t ret = -EINVAL;
1427 if (file->f_op->aio_fsync)
1428 ret = file->f_op->aio_fsync(iocb, 0);
1432 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb)
1436 ret = rw_copy_check_uvector(type, (struct iovec __user *)kiocb->ki_buf,
1437 kiocb->ki_nbytes, 1,
1438 &kiocb->ki_inline_vec, &kiocb->ki_iovec);
1442 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1443 kiocb->ki_cur_seg = 0;
1444 /* ki_nbytes/left now reflect bytes instead of segs */
1445 kiocb->ki_nbytes = ret;
1446 kiocb->ki_left = ret;
1453 static ssize_t aio_setup_single_vector(struct kiocb *kiocb)
1455 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1456 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1457 kiocb->ki_iovec->iov_len = kiocb->ki_left;
1458 kiocb->ki_nr_segs = 1;
1459 kiocb->ki_cur_seg = 0;
1465 * Performs the initial checks and aio retry method
1466 * setup for the kiocb at the time of io submission.
1468 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1470 struct file *file = kiocb->ki_filp;
1473 switch (kiocb->ki_opcode) {
1474 case IOCB_CMD_PREAD:
1476 if (unlikely(!(file->f_mode & FMODE_READ)))
1479 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1482 ret = security_file_permission(file, MAY_READ);
1485 ret = aio_setup_single_vector(kiocb);
1489 if (file->f_op->aio_read)
1490 kiocb->ki_retry = aio_rw_vect_retry;
1492 case IOCB_CMD_PWRITE:
1494 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1497 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1500 ret = security_file_permission(file, MAY_WRITE);
1503 ret = aio_setup_single_vector(kiocb);
1507 if (file->f_op->aio_write)
1508 kiocb->ki_retry = aio_rw_vect_retry;
1510 case IOCB_CMD_PREADV:
1512 if (unlikely(!(file->f_mode & FMODE_READ)))
1514 ret = security_file_permission(file, MAY_READ);
1517 ret = aio_setup_vectored_rw(READ, kiocb);
1521 if (file->f_op->aio_read)
1522 kiocb->ki_retry = aio_rw_vect_retry;
1524 case IOCB_CMD_PWRITEV:
1526 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1528 ret = security_file_permission(file, MAY_WRITE);
1531 ret = aio_setup_vectored_rw(WRITE, kiocb);
1535 if (file->f_op->aio_write)
1536 kiocb->ki_retry = aio_rw_vect_retry;
1538 case IOCB_CMD_FDSYNC:
1540 if (file->f_op->aio_fsync)
1541 kiocb->ki_retry = aio_fdsync;
1543 case IOCB_CMD_FSYNC:
1545 if (file->f_op->aio_fsync)
1546 kiocb->ki_retry = aio_fsync;
1549 dprintk("EINVAL: io_submit: no operation provided\n");
1553 if (!kiocb->ki_retry)
1560 * aio_wake_function:
1561 * wait queue callback function for aio notification,
1562 * Simply triggers a retry of the operation via kick_iocb.
1564 * This callback is specified in the wait queue entry in
1568 * This routine is executed with the wait queue lock held.
1569 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1570 * the ioctx lock inside the wait queue lock. This is safe
1571 * because this callback isn't used for wait queues which
1572 * are nested inside ioctx lock (i.e. ctx->wait)
1574 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1575 int sync, void *key)
1577 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1579 list_del_init(&wait->task_list);
1584 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1591 /* enforce forwards compatibility on users */
1592 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1593 pr_debug("EINVAL: io_submit: reserve field set\n");
1597 /* prevent overflows */
1599 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1600 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1601 ((ssize_t)iocb->aio_nbytes < 0)
1603 pr_debug("EINVAL: io_submit: overflow check\n");
1607 file = fget(iocb->aio_fildes);
1608 if (unlikely(!file))
1611 req = aio_get_req(ctx); /* returns with 2 references to req */
1612 if (unlikely(!req)) {
1616 req->ki_filp = file;
1617 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1619 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1620 * instance of the file* now. The file descriptor must be
1621 * an eventfd() fd, and will be signaled for each completed
1622 * event using the eventfd_signal() function.
1624 req->ki_eventfd = eventfd_fget((int) iocb->aio_resfd);
1625 if (IS_ERR(req->ki_eventfd)) {
1626 ret = PTR_ERR(req->ki_eventfd);
1627 req->ki_eventfd = NULL;
1632 ret = put_user(req->ki_key, &user_iocb->aio_key);
1633 if (unlikely(ret)) {
1634 dprintk("EFAULT: aio_key\n");
1638 req->ki_obj.user = user_iocb;
1639 req->ki_user_data = iocb->aio_data;
1640 req->ki_pos = iocb->aio_offset;
1642 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1643 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1644 req->ki_opcode = iocb->aio_lio_opcode;
1645 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1646 INIT_LIST_HEAD(&req->ki_wait.task_list);
1648 ret = aio_setup_iocb(req);
1653 spin_lock_irq(&ctx->ctx_lock);
1655 if (!list_empty(&ctx->run_list)) {
1656 /* drain the run list */
1657 while (__aio_run_iocbs(ctx))
1660 spin_unlock_irq(&ctx->ctx_lock);
1661 aio_put_req(req); /* drop extra ref to req */
1665 aio_put_req(req); /* drop extra ref to req */
1666 aio_put_req(req); /* drop i/o ref to req */
1671 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1672 * the number of iocbs queued. May return -EINVAL if the aio_context
1673 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1674 * *iocbpp[0] is not properly initialized, if the operation specified
1675 * is invalid for the file descriptor in the iocb. May fail with
1676 * -EFAULT if any of the data structures point to invalid data. May
1677 * fail with -EBADF if the file descriptor specified in the first
1678 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1679 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1680 * fail with -ENOSYS if not implemented.
1682 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1683 struct iocb __user * __user *, iocbpp)
1689 if (unlikely(nr < 0))
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");
1702 * AKPM: should this return a partial result if some of the IOs were
1703 * successfully submitted?
1705 for (i=0; i<nr; i++) {
1706 struct iocb __user *user_iocb;
1709 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1714 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1719 ret = io_submit_one(ctx, user_iocb, &tmp);
1729 * Finds a given iocb for cancellation.
1731 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1734 struct list_head *pos;
1736 assert_spin_locked(&ctx->ctx_lock);
1738 /* TODO: use a hash or array, this sucks. */
1739 list_for_each(pos, &ctx->active_reqs) {
1740 struct kiocb *kiocb = list_kiocb(pos);
1741 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1748 * Attempts to cancel an iocb previously passed to io_submit. If
1749 * the operation is successfully cancelled, the resulting event is
1750 * copied into the memory pointed to by result without being placed
1751 * into the completion queue and 0 is returned. May fail with
1752 * -EFAULT if any of the data structures pointed to are invalid.
1753 * May fail with -EINVAL if aio_context specified by ctx_id is
1754 * invalid. May fail with -EAGAIN if the iocb specified was not
1755 * cancelled. Will fail with -ENOSYS if not implemented.
1757 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1758 struct io_event __user *, result)
1760 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1762 struct kiocb *kiocb;
1766 ret = get_user(key, &iocb->aio_key);
1770 ctx = lookup_ioctx(ctx_id);
1774 spin_lock_irq(&ctx->ctx_lock);
1776 kiocb = lookup_kiocb(ctx, iocb, key);
1777 if (kiocb && kiocb->ki_cancel) {
1778 cancel = kiocb->ki_cancel;
1780 kiocbSetCancelled(kiocb);
1783 spin_unlock_irq(&ctx->ctx_lock);
1785 if (NULL != cancel) {
1786 struct io_event tmp;
1787 pr_debug("calling cancel\n");
1788 memset(&tmp, 0, sizeof(tmp));
1789 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1790 tmp.data = kiocb->ki_user_data;
1791 ret = cancel(kiocb, &tmp);
1793 /* Cancellation succeeded -- copy the result
1794 * into the user's buffer.
1796 if (copy_to_user(result, &tmp, sizeof(tmp)))
1808 * Attempts to read at least min_nr events and up to nr events from
1809 * the completion queue for the aio_context specified by ctx_id. May
1810 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1811 * if nr is out of range, if when is out of range. May fail with
1812 * -EFAULT if any of the memory specified to is invalid. May return
1813 * 0 or < min_nr if no events are available and the timeout specified
1814 * by when has elapsed, where when == NULL specifies an infinite
1815 * timeout. Note that the timeout pointed to by when is relative and
1816 * will be updated if not NULL and the operation blocks. Will fail
1817 * with -ENOSYS if not implemented.
1819 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1822 struct io_event __user *, events,
1823 struct timespec __user *, timeout)
1825 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1828 if (likely(ioctx)) {
1829 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1830 ret = read_events(ioctx, min_nr, nr, events, timeout);
1834 asmlinkage_protect(5, ret, ctx_id, min_nr, nr, events, timeout);
1838 __initcall(aio_setup);
1840 EXPORT_SYMBOL(aio_complete);
1841 EXPORT_SYMBOL(aio_put_req);
1842 EXPORT_SYMBOL(wait_on_sync_kiocb);