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>
21 #include <linux/sched.h>
23 #include <linux/file.h>
25 #include <linux/mman.h>
26 #include <linux/slab.h>
27 #include <linux/timer.h>
28 #include <linux/aio.h>
29 #include <linux/highmem.h>
30 #include <linux/workqueue.h>
31 #include <linux/security.h>
33 #include <asm/kmap_types.h>
34 #include <asm/uaccess.h>
35 #include <asm/mmu_context.h>
38 #define dprintk printk
40 #define dprintk(x...) do { ; } while (0)
43 static long aio_run = 0; /* for testing only */
44 static long aio_wakeups = 0; /* for testing only */
46 /*------ sysctl variables----*/
47 atomic_t aio_nr = ATOMIC_INIT(0); /* current system wide number of aio requests */
48 unsigned aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
49 /*----end sysctl variables---*/
51 static kmem_cache_t *kiocb_cachep;
52 static kmem_cache_t *kioctx_cachep;
54 static struct workqueue_struct *aio_wq;
56 /* Used for rare fput completion. */
57 static void aio_fput_routine(void *);
58 static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
60 static DEFINE_SPINLOCK(fput_lock);
61 static LIST_HEAD(fput_head);
63 static void aio_kick_handler(void *);
66 * Creates the slab caches used by the aio routines, panic on
67 * failure as this is done early during the boot sequence.
69 static int __init aio_setup(void)
71 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
72 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
73 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
74 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
76 aio_wq = create_workqueue("aio");
78 pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
83 static void aio_free_ring(struct kioctx *ctx)
85 struct aio_ring_info *info = &ctx->ring_info;
88 for (i=0; i<info->nr_pages; i++)
89 put_page(info->ring_pages[i]);
91 if (info->mmap_size) {
92 down_write(&ctx->mm->mmap_sem);
93 do_munmap(ctx->mm, info->mmap_base, info->mmap_size);
94 up_write(&ctx->mm->mmap_sem);
97 if (info->ring_pages && info->ring_pages != info->internal_pages)
98 kfree(info->ring_pages);
99 info->ring_pages = NULL;
103 static int aio_setup_ring(struct kioctx *ctx)
105 struct aio_ring *ring;
106 struct aio_ring_info *info = &ctx->ring_info;
107 unsigned nr_events = ctx->max_reqs;
111 /* Compensate for the ring buffer's head/tail overlap entry */
112 nr_events += 2; /* 1 is required, 2 for good luck */
114 size = sizeof(struct aio_ring);
115 size += sizeof(struct io_event) * nr_events;
116 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
121 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
124 info->ring_pages = info->internal_pages;
125 if (nr_pages > AIO_RING_PAGES) {
126 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
127 if (!info->ring_pages)
129 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
132 info->mmap_size = nr_pages * PAGE_SIZE;
133 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
134 down_write(&ctx->mm->mmap_sem);
135 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
136 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
138 if (IS_ERR((void *)info->mmap_base)) {
139 up_write(&ctx->mm->mmap_sem);
140 printk("mmap err: %ld\n", -info->mmap_base);
146 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
147 info->nr_pages = get_user_pages(current, ctx->mm,
148 info->mmap_base, nr_pages,
149 1, 0, info->ring_pages, NULL);
150 up_write(&ctx->mm->mmap_sem);
152 if (unlikely(info->nr_pages != nr_pages)) {
157 ctx->user_id = info->mmap_base;
159 info->nr = nr_events; /* trusted copy */
161 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
162 ring->nr = nr_events; /* user copy */
163 ring->id = ctx->user_id;
164 ring->head = ring->tail = 0;
165 ring->magic = AIO_RING_MAGIC;
166 ring->compat_features = AIO_RING_COMPAT_FEATURES;
167 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
168 ring->header_length = sizeof(struct aio_ring);
169 kunmap_atomic(ring, KM_USER0);
175 /* aio_ring_event: returns a pointer to the event at the given index from
176 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
178 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
179 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
180 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
182 #define aio_ring_event(info, nr, km) ({ \
183 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
184 struct io_event *__event; \
185 __event = kmap_atomic( \
186 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
187 __event += pos % AIO_EVENTS_PER_PAGE; \
191 #define put_aio_ring_event(event, km) do { \
192 struct io_event *__event = (event); \
194 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
198 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
200 static struct kioctx *ioctx_alloc(unsigned nr_events)
202 struct mm_struct *mm;
205 /* Prevent overflows */
206 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
207 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
208 pr_debug("ENOMEM: nr_events too high\n");
209 return ERR_PTR(-EINVAL);
212 if (nr_events > aio_max_nr)
213 return ERR_PTR(-EAGAIN);
215 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
217 return ERR_PTR(-ENOMEM);
219 memset(ctx, 0, sizeof(*ctx));
220 ctx->max_reqs = nr_events;
221 mm = ctx->mm = current->mm;
222 atomic_inc(&mm->mm_count);
224 atomic_set(&ctx->users, 1);
225 spin_lock_init(&ctx->ctx_lock);
226 spin_lock_init(&ctx->ring_info.ring_lock);
227 init_waitqueue_head(&ctx->wait);
229 INIT_LIST_HEAD(&ctx->active_reqs);
230 INIT_LIST_HEAD(&ctx->run_list);
231 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
233 if (aio_setup_ring(ctx) < 0)
236 /* limit the number of system wide aios */
237 atomic_add(ctx->max_reqs, &aio_nr); /* undone by __put_ioctx */
238 if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
241 /* now link into global list. kludge. FIXME */
242 write_lock(&mm->ioctx_list_lock);
243 ctx->next = mm->ioctx_list;
244 mm->ioctx_list = ctx;
245 write_unlock(&mm->ioctx_list_lock);
247 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
248 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
252 atomic_sub(ctx->max_reqs, &aio_nr);
253 ctx->max_reqs = 0; /* prevent __put_ioctx from sub'ing aio_nr */
255 return ERR_PTR(-EAGAIN);
259 kmem_cache_free(kioctx_cachep, ctx);
260 ctx = ERR_PTR(-ENOMEM);
262 dprintk("aio: error allocating ioctx %p\n", ctx);
267 * Cancels all outstanding aio requests on an aio context. Used
268 * when the processes owning a context have all exited to encourage
269 * the rapid destruction of the kioctx.
271 static void aio_cancel_all(struct kioctx *ctx)
273 int (*cancel)(struct kiocb *, struct io_event *);
275 spin_lock_irq(&ctx->ctx_lock);
277 while (!list_empty(&ctx->active_reqs)) {
278 struct list_head *pos = ctx->active_reqs.next;
279 struct kiocb *iocb = list_kiocb(pos);
280 list_del_init(&iocb->ki_list);
281 cancel = iocb->ki_cancel;
282 kiocbSetCancelled(iocb);
285 spin_unlock_irq(&ctx->ctx_lock);
287 spin_lock_irq(&ctx->ctx_lock);
290 spin_unlock_irq(&ctx->ctx_lock);
293 static void wait_for_all_aios(struct kioctx *ctx)
295 struct task_struct *tsk = current;
296 DECLARE_WAITQUEUE(wait, tsk);
298 if (!ctx->reqs_active)
301 add_wait_queue(&ctx->wait, &wait);
302 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
303 while (ctx->reqs_active) {
305 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
307 __set_task_state(tsk, TASK_RUNNING);
308 remove_wait_queue(&ctx->wait, &wait);
311 /* wait_on_sync_kiocb:
312 * Waits on the given sync kiocb to complete.
314 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
316 while (iocb->ki_users) {
317 set_current_state(TASK_UNINTERRUPTIBLE);
322 __set_current_state(TASK_RUNNING);
323 return iocb->ki_user_data;
326 /* exit_aio: called when the last user of mm goes away. At this point,
327 * there is no way for any new requests to be submited or any of the
328 * io_* syscalls to be called on the context. However, there may be
329 * outstanding requests which hold references to the context; as they
330 * go away, they will call put_ioctx and release any pinned memory
331 * associated with the request (held via struct page * references).
333 void fastcall exit_aio(struct mm_struct *mm)
335 struct kioctx *ctx = mm->ioctx_list;
336 mm->ioctx_list = NULL;
338 struct kioctx *next = ctx->next;
342 wait_for_all_aios(ctx);
344 * this is an overkill, but ensures we don't leave
345 * the ctx on the aio_wq
347 flush_workqueue(aio_wq);
349 if (1 != atomic_read(&ctx->users))
351 "exit_aio:ioctx still alive: %d %d %d\n",
352 atomic_read(&ctx->users), ctx->dead,
360 * Called when the last user of an aio context has gone away,
361 * and the struct needs to be freed.
363 void fastcall __put_ioctx(struct kioctx *ctx)
365 unsigned nr_events = ctx->max_reqs;
367 if (unlikely(ctx->reqs_active))
370 cancel_delayed_work(&ctx->wq);
371 flush_workqueue(aio_wq);
375 pr_debug("__put_ioctx: freeing %p\n", ctx);
376 kmem_cache_free(kioctx_cachep, ctx);
378 atomic_sub(nr_events, &aio_nr);
382 * Allocate a slot for an aio request. Increments the users count
383 * of the kioctx so that the kioctx stays around until all requests are
384 * complete. Returns NULL if no requests are free.
386 * Returns with kiocb->users set to 2. The io submit code path holds
387 * an extra reference while submitting the i/o.
388 * This prevents races between the aio code path referencing the
389 * req (after submitting it) and aio_complete() freeing the req.
391 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
392 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
394 struct kiocb *req = NULL;
395 struct aio_ring *ring;
398 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
402 req->ki_flags = 1 << KIF_LOCKED;
406 req->ki_cancel = NULL;
407 req->ki_retry = NULL;
410 INIT_LIST_HEAD(&req->ki_run_list);
412 /* Check if the completion queue has enough free space to
413 * accept an event from this io.
415 spin_lock_irq(&ctx->ctx_lock);
416 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
417 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
418 list_add(&req->ki_list, &ctx->active_reqs);
423 kunmap_atomic(ring, KM_USER0);
424 spin_unlock_irq(&ctx->ctx_lock);
427 kmem_cache_free(kiocb_cachep, req);
434 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
437 /* Handle a potential starvation case -- should be exceedingly rare as
438 * requests will be stuck on fput_head only if the aio_fput_routine is
439 * delayed and the requests were the last user of the struct file.
441 req = __aio_get_req(ctx);
442 if (unlikely(NULL == req)) {
443 aio_fput_routine(NULL);
444 req = __aio_get_req(ctx);
449 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
453 kmem_cache_free(kiocb_cachep, req);
456 if (unlikely(!ctx->reqs_active && ctx->dead))
460 static void aio_fput_routine(void *data)
462 spin_lock_irq(&fput_lock);
463 while (likely(!list_empty(&fput_head))) {
464 struct kiocb *req = list_kiocb(fput_head.next);
465 struct kioctx *ctx = req->ki_ctx;
467 list_del(&req->ki_list);
468 spin_unlock_irq(&fput_lock);
470 /* Complete the fput */
471 __fput(req->ki_filp);
473 /* Link the iocb into the context's free list */
474 spin_lock_irq(&ctx->ctx_lock);
475 really_put_req(ctx, req);
476 spin_unlock_irq(&ctx->ctx_lock);
479 spin_lock_irq(&fput_lock);
481 spin_unlock_irq(&fput_lock);
485 * Returns true if this put was the last user of the request.
487 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
489 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
490 req, atomic_read(&req->ki_filp->f_count));
493 if (unlikely(req->ki_users < 0))
495 if (likely(req->ki_users))
497 list_del(&req->ki_list); /* remove from active_reqs */
498 req->ki_cancel = NULL;
499 req->ki_retry = NULL;
501 /* Must be done under the lock to serialise against cancellation.
502 * Call this aio_fput as it duplicates fput via the fput_work.
504 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
506 spin_lock(&fput_lock);
507 list_add(&req->ki_list, &fput_head);
508 spin_unlock(&fput_lock);
509 queue_work(aio_wq, &fput_work);
511 really_put_req(ctx, req);
516 * Returns true if this put was the last user of the kiocb,
517 * false if the request is still in use.
519 int fastcall aio_put_req(struct kiocb *req)
521 struct kioctx *ctx = req->ki_ctx;
523 spin_lock_irq(&ctx->ctx_lock);
524 ret = __aio_put_req(ctx, req);
525 spin_unlock_irq(&ctx->ctx_lock);
531 /* Lookup an ioctx id. ioctx_list is lockless for reads.
532 * FIXME: this is O(n) and is only suitable for development.
534 struct kioctx *lookup_ioctx(unsigned long ctx_id)
536 struct kioctx *ioctx;
537 struct mm_struct *mm;
540 read_lock(&mm->ioctx_list_lock);
541 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
542 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
546 read_unlock(&mm->ioctx_list_lock);
553 * Makes the calling kernel thread take on the specified
555 * Called by the retry thread execute retries within the
556 * iocb issuer's mm context, so that copy_from/to_user
557 * operations work seamlessly for aio.
558 * (Note: this routine is intended to be called only
559 * from a kernel thread context)
561 static void use_mm(struct mm_struct *mm)
563 struct mm_struct *active_mm;
564 struct task_struct *tsk = current;
567 tsk->flags |= PF_BORROWED_MM;
568 active_mm = tsk->active_mm;
569 atomic_inc(&mm->mm_count);
572 activate_mm(active_mm, mm);
580 * Reverses the effect of use_mm, i.e. releases the
581 * specified mm context which was earlier taken on
582 * by the calling kernel thread
583 * (Note: this routine is intended to be called only
584 * from a kernel thread context)
586 * Comments: Called with ctx->ctx_lock held. This nests
587 * task_lock instead ctx_lock.
589 static void unuse_mm(struct mm_struct *mm)
591 struct task_struct *tsk = current;
594 tsk->flags &= ~PF_BORROWED_MM;
596 /* active_mm is still 'mm' */
597 enter_lazy_tlb(mm, tsk);
602 * Queue up a kiocb to be retried. Assumes that the kiocb
603 * has already been marked as kicked, and places it on
604 * the retry run list for the corresponding ioctx, if it
605 * isn't already queued. Returns 1 if it actually queued
606 * the kiocb (to tell the caller to activate the work
607 * queue to process it), or 0, if it found that it was
610 * Should be called with the spin lock iocb->ki_ctx->ctx_lock
613 static inline int __queue_kicked_iocb(struct kiocb *iocb)
615 struct kioctx *ctx = iocb->ki_ctx;
617 if (list_empty(&iocb->ki_run_list)) {
618 list_add_tail(&iocb->ki_run_list,
627 * This is the core aio execution routine. It is
628 * invoked both for initial i/o submission and
629 * subsequent retries via the aio_kick_handler.
630 * Expects to be invoked with iocb->ki_ctx->lock
631 * already held. The lock is released and reaquired
632 * as needed during processing.
634 * Calls the iocb retry method (already setup for the
635 * iocb on initial submission) for operation specific
636 * handling, but takes care of most of common retry
637 * execution details for a given iocb. The retry method
638 * needs to be non-blocking as far as possible, to avoid
639 * holding up other iocbs waiting to be serviced by the
640 * retry kernel thread.
642 * The trickier parts in this code have to do with
643 * ensuring that only one retry instance is in progress
644 * for a given iocb at any time. Providing that guarantee
645 * simplifies the coding of individual aio operations as
646 * it avoids various potential races.
648 static ssize_t aio_run_iocb(struct kiocb *iocb)
650 struct kioctx *ctx = iocb->ki_ctx;
651 ssize_t (*retry)(struct kiocb *);
654 if (iocb->ki_retried++ > 1024*1024) {
655 printk("Maximal retry count. Bytes done %Zd\n",
656 iocb->ki_nbytes - iocb->ki_left);
660 if (!(iocb->ki_retried & 0xff)) {
661 pr_debug("%ld retry: %d of %d (kick %ld, Q %ld run %ld, wake %ld)\n",
663 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes,
664 iocb->ki_kicked, iocb->ki_queued, aio_run, aio_wakeups);
667 if (!(retry = iocb->ki_retry)) {
668 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
673 * We don't want the next retry iteration for this
674 * operation to start until this one has returned and
675 * updated the iocb state. However, wait_queue functions
676 * can trigger a kick_iocb from interrupt context in the
677 * meantime, indicating that data is available for the next
678 * iteration. We want to remember that and enable the
679 * next retry iteration _after_ we are through with
682 * So, in order to be able to register a "kick", but
683 * prevent it from being queued now, we clear the kick
684 * flag, but make the kick code *think* that the iocb is
685 * still on the run list until we are actually done.
686 * When we are done with this iteration, we check if
687 * the iocb was kicked in the meantime and if so, queue
691 kiocbClearKicked(iocb);
694 * This is so that aio_complete knows it doesn't need to
695 * pull the iocb off the run list (We can't just call
696 * INIT_LIST_HEAD because we don't want a kick_iocb to
697 * queue this on the run list yet)
699 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
700 spin_unlock_irq(&ctx->ctx_lock);
702 /* Quit retrying if the i/o has been cancelled */
703 if (kiocbIsCancelled(iocb)) {
705 aio_complete(iocb, ret, 0);
706 /* must not access the iocb after this */
711 * Now we are all set to call the retry method in async
712 * context. By setting this thread's io_wait context
713 * to point to the wait queue entry inside the currently
714 * running iocb for the duration of the retry, we ensure
715 * that async notification wakeups are queued by the
716 * operation instead of blocking waits, and when notified,
717 * cause the iocb to be kicked for continuation (through
718 * the aio_wake_function callback).
720 BUG_ON(current->io_wait != NULL);
721 current->io_wait = &iocb->ki_wait;
723 current->io_wait = NULL;
725 if (-EIOCBRETRY != ret) {
726 if (-EIOCBQUEUED != ret) {
727 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
728 aio_complete(iocb, ret, 0);
729 /* must not access the iocb after this */
733 * Issue an additional retry to avoid waiting forever if
734 * no waits were queued (e.g. in case of a short read).
736 if (list_empty(&iocb->ki_wait.task_list))
737 kiocbSetKicked(iocb);
740 spin_lock_irq(&ctx->ctx_lock);
742 if (-EIOCBRETRY == ret) {
744 * OK, now that we are done with this iteration
745 * and know that there is more left to go,
746 * this is where we let go so that a subsequent
747 * "kick" can start the next iteration
750 /* will make __queue_kicked_iocb succeed from here on */
751 INIT_LIST_HEAD(&iocb->ki_run_list);
752 /* we must queue the next iteration ourselves, if it
753 * has already been kicked */
754 if (kiocbIsKicked(iocb)) {
755 __queue_kicked_iocb(iocb);
763 * Process all pending retries queued on the ioctx
765 * Assumes it is operating within the aio issuer's mm
766 * context. Expects to be called with ctx->ctx_lock held
768 static int __aio_run_iocbs(struct kioctx *ctx)
774 list_splice_init(&ctx->run_list, &run_list);
775 while (!list_empty(&run_list)) {
776 iocb = list_entry(run_list.next, struct kiocb,
778 list_del(&iocb->ki_run_list);
780 * Hold an extra reference while retrying i/o.
782 iocb->ki_users++; /* grab extra reference */
784 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
789 if (!list_empty(&ctx->run_list))
794 static void aio_queue_work(struct kioctx * ctx)
796 unsigned long timeout;
798 * if someone is waiting, get the work started right
799 * away, otherwise, use a longer delay
802 if (waitqueue_active(&ctx->wait))
806 queue_delayed_work(aio_wq, &ctx->wq, timeout);
812 * Process all pending retries queued on the ioctx
814 * Assumes it is operating within the aio issuer's mm
817 static inline void aio_run_iocbs(struct kioctx *ctx)
821 spin_lock_irq(&ctx->ctx_lock);
823 requeue = __aio_run_iocbs(ctx);
824 spin_unlock_irq(&ctx->ctx_lock);
830 * just like aio_run_iocbs, but keeps running them until
831 * the list stays empty
833 static inline void aio_run_all_iocbs(struct kioctx *ctx)
835 spin_lock_irq(&ctx->ctx_lock);
836 while (__aio_run_iocbs(ctx))
838 spin_unlock_irq(&ctx->ctx_lock);
843 * Work queue handler triggered to process pending
844 * retries on an ioctx. Takes on the aio issuer's
845 * mm context before running the iocbs, so that
846 * copy_xxx_user operates on the issuer's address
848 * Run on aiod's context.
850 static void aio_kick_handler(void *data)
852 struct kioctx *ctx = data;
853 mm_segment_t oldfs = get_fs();
858 spin_lock_irq(&ctx->ctx_lock);
859 requeue =__aio_run_iocbs(ctx);
861 spin_unlock_irq(&ctx->ctx_lock);
864 * we're in a worker thread already, don't use queue_delayed_work,
867 queue_work(aio_wq, &ctx->wq);
872 * Called by kick_iocb to queue the kiocb for retry
873 * and if required activate the aio work queue to process
876 static void queue_kicked_iocb(struct kiocb *iocb)
878 struct kioctx *ctx = iocb->ki_ctx;
882 WARN_ON((!list_empty(&iocb->ki_wait.task_list)));
884 spin_lock_irqsave(&ctx->ctx_lock, flags);
885 run = __queue_kicked_iocb(iocb);
886 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
895 * Called typically from a wait queue callback context
896 * (aio_wake_function) to trigger a retry of the iocb.
897 * The retry is usually executed by aio workqueue
898 * threads (See aio_kick_handler).
900 void fastcall kick_iocb(struct kiocb *iocb)
902 /* sync iocbs are easy: they can only ever be executing from a
904 if (is_sync_kiocb(iocb)) {
905 kiocbSetKicked(iocb);
906 wake_up_process(iocb->ki_obj.tsk);
911 /* If its already kicked we shouldn't queue it again */
912 if (!kiocbTryKick(iocb)) {
913 queue_kicked_iocb(iocb);
916 EXPORT_SYMBOL(kick_iocb);
919 * Called when the io request on the given iocb is complete.
920 * Returns true if this is the last user of the request. The
921 * only other user of the request can be the cancellation code.
923 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
925 struct kioctx *ctx = iocb->ki_ctx;
926 struct aio_ring_info *info;
927 struct aio_ring *ring;
928 struct io_event *event;
933 /* Special case handling for sync iocbs: events go directly
934 * into the iocb for fast handling. Note that this will not
935 * work if we allow sync kiocbs to be cancelled. in which
936 * case the usage count checks will have to move under ctx_lock
939 if (is_sync_kiocb(iocb)) {
942 iocb->ki_user_data = res;
943 if (iocb->ki_users == 1) {
947 spin_lock_irq(&ctx->ctx_lock);
949 ret = (0 == iocb->ki_users);
950 spin_unlock_irq(&ctx->ctx_lock);
952 /* sync iocbs put the task here for us */
953 wake_up_process(iocb->ki_obj.tsk);
957 info = &ctx->ring_info;
959 /* add a completion event to the ring buffer.
960 * must be done holding ctx->ctx_lock to prevent
961 * other code from messing with the tail
962 * pointer since we might be called from irq
965 spin_lock_irqsave(&ctx->ctx_lock, flags);
967 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
968 list_del_init(&iocb->ki_run_list);
971 * cancelled requests don't get events, userland was given one
972 * when the event got cancelled.
974 if (kiocbIsCancelled(iocb))
977 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
980 event = aio_ring_event(info, tail, KM_IRQ0);
981 tail = (tail + 1) % info->nr;
983 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
984 event->data = iocb->ki_user_data;
988 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
989 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
992 /* after flagging the request as done, we
993 * must never even look at it again
995 smp_wmb(); /* make event visible before updating tail */
1000 put_aio_ring_event(event, KM_IRQ0);
1001 kunmap_atomic(ring, KM_IRQ1);
1003 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1005 pr_debug("%ld retries: %d of %d (kicked %ld, Q %ld run %ld wake %ld)\n",
1007 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes,
1008 iocb->ki_kicked, iocb->ki_queued, aio_run, aio_wakeups);
1010 /* everything turned out well, dispose of the aiocb. */
1011 ret = __aio_put_req(ctx, iocb);
1013 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1015 if (waitqueue_active(&ctx->wait))
1016 wake_up(&ctx->wait);
1025 * Pull an event off of the ioctx's event ring. Returns the number of
1026 * events fetched (0 or 1 ;-)
1027 * FIXME: make this use cmpxchg.
1028 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1030 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1032 struct aio_ring_info *info = &ioctx->ring_info;
1033 struct aio_ring *ring;
1037 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1038 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1039 (unsigned long)ring->head, (unsigned long)ring->tail,
1040 (unsigned long)ring->nr);
1042 if (ring->head == ring->tail)
1045 spin_lock(&info->ring_lock);
1047 head = ring->head % info->nr;
1048 if (head != ring->tail) {
1049 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1051 head = (head + 1) % info->nr;
1052 smp_mb(); /* finish reading the event before updatng the head */
1055 put_aio_ring_event(evp, KM_USER1);
1057 spin_unlock(&info->ring_lock);
1060 kunmap_atomic(ring, KM_USER0);
1061 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1062 (unsigned long)ring->head, (unsigned long)ring->tail);
1066 struct aio_timeout {
1067 struct timer_list timer;
1069 struct task_struct *p;
1072 static void timeout_func(unsigned long data)
1074 struct aio_timeout *to = (struct aio_timeout *)data;
1077 wake_up_process(to->p);
1080 static inline void init_timeout(struct aio_timeout *to)
1082 init_timer(&to->timer);
1083 to->timer.data = (unsigned long)to;
1084 to->timer.function = timeout_func;
1089 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1090 const struct timespec *ts)
1092 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1093 if (time_after(to->timer.expires, jiffies))
1094 add_timer(&to->timer);
1099 static inline void clear_timeout(struct aio_timeout *to)
1101 del_singleshot_timer_sync(&to->timer);
1104 static int read_events(struct kioctx *ctx,
1105 long min_nr, long nr,
1106 struct io_event __user *event,
1107 struct timespec __user *timeout)
1109 long start_jiffies = jiffies;
1110 struct task_struct *tsk = current;
1111 DECLARE_WAITQUEUE(wait, tsk);
1114 struct io_event ent;
1115 struct aio_timeout to;
1116 int event_loop = 0; /* testing only */
1119 /* needed to zero any padding within an entry (there shouldn't be
1120 * any, but C is fun!
1122 memset(&ent, 0, sizeof(ent));
1125 while (likely(i < nr)) {
1126 ret = aio_read_evt(ctx, &ent);
1127 if (unlikely(ret <= 0))
1130 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1131 ent.data, ent.obj, ent.res, ent.res2);
1133 /* Could we split the check in two? */
1135 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1136 dprintk("aio: lost an event due to EFAULT.\n");
1141 /* Good, event copied to userland, update counts. */
1153 /* racey check, but it gets redone */
1154 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1156 aio_run_all_iocbs(ctx);
1164 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1167 set_timeout(start_jiffies, &to, &ts);
1170 while (likely(i < nr)) {
1171 add_wait_queue_exclusive(&ctx->wait, &wait);
1173 set_task_state(tsk, TASK_INTERRUPTIBLE);
1174 ret = aio_read_evt(ctx, &ent);
1180 if (to.timed_out) /* Only check after read evt */
1184 if (signal_pending(tsk)) {
1188 /*ret = aio_read_evt(ctx, &ent);*/
1191 set_task_state(tsk, TASK_RUNNING);
1192 remove_wait_queue(&ctx->wait, &wait);
1194 if (unlikely(ret <= 0))
1198 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1199 dprintk("aio: lost an event due to EFAULT.\n");
1203 /* Good, event copied to userland, update counts. */
1211 pr_debug("event loop executed %d times\n", event_loop);
1212 pr_debug("aio_run %ld\n", aio_run);
1213 pr_debug("aio_wakeups %ld\n", aio_wakeups);
1217 /* Take an ioctx and remove it from the list of ioctx's. Protects
1218 * against races with itself via ->dead.
1220 static void io_destroy(struct kioctx *ioctx)
1222 struct mm_struct *mm = current->mm;
1223 struct kioctx **tmp;
1226 /* delete the entry from the list is someone else hasn't already */
1227 write_lock(&mm->ioctx_list_lock);
1228 was_dead = ioctx->dead;
1230 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1231 tmp = &(*tmp)->next)
1235 write_unlock(&mm->ioctx_list_lock);
1237 dprintk("aio_release(%p)\n", ioctx);
1238 if (likely(!was_dead))
1239 put_ioctx(ioctx); /* twice for the list */
1241 aio_cancel_all(ioctx);
1242 wait_for_all_aios(ioctx);
1243 put_ioctx(ioctx); /* once for the lookup */
1247 * Create an aio_context capable of receiving at least nr_events.
1248 * ctxp must not point to an aio_context that already exists, and
1249 * must be initialized to 0 prior to the call. On successful
1250 * creation of the aio_context, *ctxp is filled in with the resulting
1251 * handle. May fail with -EINVAL if *ctxp is not initialized,
1252 * if the specified nr_events exceeds internal limits. May fail
1253 * with -EAGAIN if the specified nr_events exceeds the user's limit
1254 * of available events. May fail with -ENOMEM if insufficient kernel
1255 * resources are available. May fail with -EFAULT if an invalid
1256 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1259 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1261 struct kioctx *ioctx = NULL;
1265 ret = get_user(ctx, ctxp);
1270 if (unlikely(ctx || (int)nr_events <= 0)) {
1271 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1275 ioctx = ioctx_alloc(nr_events);
1276 ret = PTR_ERR(ioctx);
1277 if (!IS_ERR(ioctx)) {
1278 ret = put_user(ioctx->user_id, ctxp);
1282 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1291 * Destroy the aio_context specified. May cancel any outstanding
1292 * AIOs and block on completion. Will fail with -ENOSYS if not
1293 * implemented. May fail with -EFAULT if the context pointed to
1296 asmlinkage long sys_io_destroy(aio_context_t ctx)
1298 struct kioctx *ioctx = lookup_ioctx(ctx);
1299 if (likely(NULL != ioctx)) {
1303 pr_debug("EINVAL: io_destroy: invalid context id\n");
1308 * Default retry method for aio_read (also used for first time submit)
1309 * Responsible for updating iocb state as retries progress
1311 static ssize_t aio_pread(struct kiocb *iocb)
1313 struct file *file = iocb->ki_filp;
1314 struct address_space *mapping = file->f_mapping;
1315 struct inode *inode = mapping->host;
1318 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1319 iocb->ki_left, iocb->ki_pos);
1322 * Can't just depend on iocb->ki_left to determine
1323 * whether we are done. This may have been a short read.
1326 iocb->ki_buf += ret;
1327 iocb->ki_left -= ret;
1329 * For pipes and sockets we return once we have
1330 * some data; for regular files we retry till we
1331 * complete the entire read or find that we can't
1332 * read any more data (e.g short reads).
1334 if (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))
1338 /* This means we must have transferred all that we could */
1339 /* No need to retry anymore */
1340 if ((ret == 0) || (iocb->ki_left == 0))
1341 ret = iocb->ki_nbytes - iocb->ki_left;
1347 * Default retry method for aio_write (also used for first time submit)
1348 * Responsible for updating iocb state as retries progress
1350 static ssize_t aio_pwrite(struct kiocb *iocb)
1352 struct file *file = iocb->ki_filp;
1355 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1356 iocb->ki_left, iocb->ki_pos);
1359 iocb->ki_buf += ret;
1360 iocb->ki_left -= ret;
1365 /* This means we must have transferred all that we could */
1366 /* No need to retry anymore */
1367 if ((ret == 0) || (iocb->ki_left == 0))
1368 ret = iocb->ki_nbytes - iocb->ki_left;
1373 static ssize_t aio_fdsync(struct kiocb *iocb)
1375 struct file *file = iocb->ki_filp;
1376 ssize_t ret = -EINVAL;
1378 if (file->f_op->aio_fsync)
1379 ret = file->f_op->aio_fsync(iocb, 1);
1383 static ssize_t aio_fsync(struct kiocb *iocb)
1385 struct file *file = iocb->ki_filp;
1386 ssize_t ret = -EINVAL;
1388 if (file->f_op->aio_fsync)
1389 ret = file->f_op->aio_fsync(iocb, 0);
1395 * Performs the initial checks and aio retry method
1396 * setup for the kiocb at the time of io submission.
1398 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1400 struct file *file = kiocb->ki_filp;
1403 switch (kiocb->ki_opcode) {
1404 case IOCB_CMD_PREAD:
1406 if (unlikely(!(file->f_mode & FMODE_READ)))
1409 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1413 if (file->f_op->aio_read)
1414 kiocb->ki_retry = aio_pread;
1416 case IOCB_CMD_PWRITE:
1418 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1421 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1425 if (file->f_op->aio_write)
1426 kiocb->ki_retry = aio_pwrite;
1428 case IOCB_CMD_FDSYNC:
1430 if (file->f_op->aio_fsync)
1431 kiocb->ki_retry = aio_fdsync;
1433 case IOCB_CMD_FSYNC:
1435 if (file->f_op->aio_fsync)
1436 kiocb->ki_retry = aio_fsync;
1439 dprintk("EINVAL: io_submit: no operation provided\n");
1443 if (!kiocb->ki_retry)
1450 * aio_wake_function:
1451 * wait queue callback function for aio notification,
1452 * Simply triggers a retry of the operation via kick_iocb.
1454 * This callback is specified in the wait queue entry in
1455 * a kiocb (current->io_wait points to this wait queue
1456 * entry when an aio operation executes; it is used
1457 * instead of a synchronous wait when an i/o blocking
1458 * condition is encountered during aio).
1461 * This routine is executed with the wait queue lock held.
1462 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1463 * the ioctx lock inside the wait queue lock. This is safe
1464 * because this callback isn't used for wait queues which
1465 * are nested inside ioctx lock (i.e. ctx->wait)
1467 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1468 int sync, void *key)
1470 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1472 list_del_init(&wait->task_list);
1477 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1484 /* enforce forwards compatibility on users */
1485 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1486 iocb->aio_reserved3)) {
1487 pr_debug("EINVAL: io_submit: reserve field set\n");
1491 /* prevent overflows */
1493 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1494 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1495 ((ssize_t)iocb->aio_nbytes < 0)
1497 pr_debug("EINVAL: io_submit: overflow check\n");
1501 file = fget(iocb->aio_fildes);
1502 if (unlikely(!file))
1505 req = aio_get_req(ctx); /* returns with 2 references to req */
1506 if (unlikely(!req)) {
1511 req->ki_filp = file;
1512 ret = put_user(req->ki_key, &user_iocb->aio_key);
1513 if (unlikely(ret)) {
1514 dprintk("EFAULT: aio_key\n");
1518 req->ki_obj.user = user_iocb;
1519 req->ki_user_data = iocb->aio_data;
1520 req->ki_pos = iocb->aio_offset;
1522 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1523 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1524 req->ki_opcode = iocb->aio_lio_opcode;
1525 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1526 INIT_LIST_HEAD(&req->ki_wait.task_list);
1527 req->ki_retried = 0;
1533 ret = aio_setup_iocb(req);
1538 spin_lock_irq(&ctx->ctx_lock);
1539 list_add_tail(&req->ki_run_list, &ctx->run_list);
1540 /* drain the run list */
1541 while (__aio_run_iocbs(ctx))
1543 spin_unlock_irq(&ctx->ctx_lock);
1544 aio_put_req(req); /* drop extra ref to req */
1548 aio_put_req(req); /* drop extra ref to req */
1549 aio_put_req(req); /* drop i/o ref to req */
1554 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1555 * the number of iocbs queued. May return -EINVAL if the aio_context
1556 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1557 * *iocbpp[0] is not properly initialized, if the operation specified
1558 * is invalid for the file descriptor in the iocb. May fail with
1559 * -EFAULT if any of the data structures point to invalid data. May
1560 * fail with -EBADF if the file descriptor specified in the first
1561 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1562 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1563 * fail with -ENOSYS if not implemented.
1565 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1566 struct iocb __user * __user *iocbpp)
1572 if (unlikely(nr < 0))
1575 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1578 ctx = lookup_ioctx(ctx_id);
1579 if (unlikely(!ctx)) {
1580 pr_debug("EINVAL: io_submit: invalid context id\n");
1585 * AKPM: should this return a partial result if some of the IOs were
1586 * successfully submitted?
1588 for (i=0; i<nr; i++) {
1589 struct iocb __user *user_iocb;
1592 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1597 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1602 ret = io_submit_one(ctx, user_iocb, &tmp);
1612 * Finds a given iocb for cancellation.
1613 * MUST be called with ctx->ctx_lock held.
1615 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1618 struct list_head *pos;
1619 /* TODO: use a hash or array, this sucks. */
1620 list_for_each(pos, &ctx->active_reqs) {
1621 struct kiocb *kiocb = list_kiocb(pos);
1622 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1629 * Attempts to cancel an iocb previously passed to io_submit. If
1630 * the operation is successfully cancelled, the resulting event is
1631 * copied into the memory pointed to by result without being placed
1632 * into the completion queue and 0 is returned. May fail with
1633 * -EFAULT if any of the data structures pointed to are invalid.
1634 * May fail with -EINVAL if aio_context specified by ctx_id is
1635 * invalid. May fail with -EAGAIN if the iocb specified was not
1636 * cancelled. Will fail with -ENOSYS if not implemented.
1638 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1639 struct io_event __user *result)
1641 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1643 struct kiocb *kiocb;
1647 ret = get_user(key, &iocb->aio_key);
1651 ctx = lookup_ioctx(ctx_id);
1655 spin_lock_irq(&ctx->ctx_lock);
1657 kiocb = lookup_kiocb(ctx, iocb, key);
1658 if (kiocb && kiocb->ki_cancel) {
1659 cancel = kiocb->ki_cancel;
1661 kiocbSetCancelled(kiocb);
1664 spin_unlock_irq(&ctx->ctx_lock);
1666 if (NULL != cancel) {
1667 struct io_event tmp;
1668 pr_debug("calling cancel\n");
1669 memset(&tmp, 0, sizeof(tmp));
1670 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1671 tmp.data = kiocb->ki_user_data;
1672 ret = cancel(kiocb, &tmp);
1674 /* Cancellation succeeded -- copy the result
1675 * into the user's buffer.
1677 if (copy_to_user(result, &tmp, sizeof(tmp)))
1681 printk(KERN_DEBUG "iocb has no cancel operation\n");
1689 * Attempts to read at least min_nr events and up to nr events from
1690 * the completion queue for the aio_context specified by ctx_id. May
1691 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1692 * if nr is out of range, if when is out of range. May fail with
1693 * -EFAULT if any of the memory specified to is invalid. May return
1694 * 0 or < min_nr if no events are available and the timeout specified
1695 * by when has elapsed, where when == NULL specifies an infinite
1696 * timeout. Note that the timeout pointed to by when is relative and
1697 * will be updated if not NULL and the operation blocks. Will fail
1698 * with -ENOSYS if not implemented.
1700 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1703 struct io_event __user *events,
1704 struct timespec __user *timeout)
1706 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1709 if (likely(ioctx)) {
1710 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1711 ret = read_events(ioctx, min_nr, nr, events, timeout);
1718 __initcall(aio_setup);
1720 EXPORT_SYMBOL(aio_complete);
1721 EXPORT_SYMBOL(aio_put_req);
1722 EXPORT_SYMBOL(wait_on_sync_kiocb);