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>
32 #include <linux/rcuref.h>
34 #include <asm/kmap_types.h>
35 #include <asm/uaccess.h>
36 #include <asm/mmu_context.h>
39 #define dprintk printk
41 #define dprintk(x...) do { ; } while (0)
44 /*------ sysctl variables----*/
45 static DEFINE_SPINLOCK(aio_nr_lock);
46 unsigned long aio_nr; /* current system wide number of aio requests */
47 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
48 /*----end sysctl variables---*/
50 static kmem_cache_t *kiocb_cachep;
51 static kmem_cache_t *kioctx_cachep;
53 static struct workqueue_struct *aio_wq;
55 /* Used for rare fput completion. */
56 static void aio_fput_routine(void *);
57 static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
59 static DEFINE_SPINLOCK(fput_lock);
60 static LIST_HEAD(fput_head);
62 static void aio_kick_handler(void *);
63 static void aio_queue_work(struct kioctx *);
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 ((unsigned long)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 spin_lock(&aio_nr_lock);
238 if (aio_nr + ctx->max_reqs > aio_max_nr ||
239 aio_nr + ctx->max_reqs < aio_nr)
242 aio_nr += ctx->max_reqs;
243 spin_unlock(&aio_nr_lock);
244 if (ctx->max_reqs == 0)
247 /* now link into global list. kludge. FIXME */
248 write_lock(&mm->ioctx_list_lock);
249 ctx->next = mm->ioctx_list;
250 mm->ioctx_list = ctx;
251 write_unlock(&mm->ioctx_list_lock);
253 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
254 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
259 return ERR_PTR(-EAGAIN);
263 kmem_cache_free(kioctx_cachep, ctx);
264 ctx = ERR_PTR(-ENOMEM);
266 dprintk("aio: error allocating ioctx %p\n", ctx);
271 * Cancels all outstanding aio requests on an aio context. Used
272 * when the processes owning a context have all exited to encourage
273 * the rapid destruction of the kioctx.
275 static void aio_cancel_all(struct kioctx *ctx)
277 int (*cancel)(struct kiocb *, struct io_event *);
279 spin_lock_irq(&ctx->ctx_lock);
281 while (!list_empty(&ctx->active_reqs)) {
282 struct list_head *pos = ctx->active_reqs.next;
283 struct kiocb *iocb = list_kiocb(pos);
284 list_del_init(&iocb->ki_list);
285 cancel = iocb->ki_cancel;
286 kiocbSetCancelled(iocb);
289 spin_unlock_irq(&ctx->ctx_lock);
291 spin_lock_irq(&ctx->ctx_lock);
294 spin_unlock_irq(&ctx->ctx_lock);
297 static void wait_for_all_aios(struct kioctx *ctx)
299 struct task_struct *tsk = current;
300 DECLARE_WAITQUEUE(wait, tsk);
302 if (!ctx->reqs_active)
305 add_wait_queue(&ctx->wait, &wait);
306 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
307 while (ctx->reqs_active) {
309 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
311 __set_task_state(tsk, TASK_RUNNING);
312 remove_wait_queue(&ctx->wait, &wait);
315 /* wait_on_sync_kiocb:
316 * Waits on the given sync kiocb to complete.
318 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
320 while (iocb->ki_users) {
321 set_current_state(TASK_UNINTERRUPTIBLE);
326 __set_current_state(TASK_RUNNING);
327 return iocb->ki_user_data;
330 /* exit_aio: called when the last user of mm goes away. At this point,
331 * there is no way for any new requests to be submited or any of the
332 * io_* syscalls to be called on the context. However, there may be
333 * outstanding requests which hold references to the context; as they
334 * go away, they will call put_ioctx and release any pinned memory
335 * associated with the request (held via struct page * references).
337 void fastcall exit_aio(struct mm_struct *mm)
339 struct kioctx *ctx = mm->ioctx_list;
340 mm->ioctx_list = NULL;
342 struct kioctx *next = ctx->next;
346 wait_for_all_aios(ctx);
348 * this is an overkill, but ensures we don't leave
349 * the ctx on the aio_wq
351 flush_workqueue(aio_wq);
353 if (1 != atomic_read(&ctx->users))
355 "exit_aio:ioctx still alive: %d %d %d\n",
356 atomic_read(&ctx->users), ctx->dead,
364 * Called when the last user of an aio context has gone away,
365 * and the struct needs to be freed.
367 void fastcall __put_ioctx(struct kioctx *ctx)
369 unsigned nr_events = ctx->max_reqs;
371 if (unlikely(ctx->reqs_active))
374 cancel_delayed_work(&ctx->wq);
375 flush_workqueue(aio_wq);
379 pr_debug("__put_ioctx: freeing %p\n", ctx);
380 kmem_cache_free(kioctx_cachep, ctx);
383 spin_lock(&aio_nr_lock);
384 BUG_ON(aio_nr - nr_events > aio_nr);
386 spin_unlock(&aio_nr_lock);
391 * Allocate a slot for an aio request. Increments the users count
392 * of the kioctx so that the kioctx stays around until all requests are
393 * complete. Returns NULL if no requests are free.
395 * Returns with kiocb->users set to 2. The io submit code path holds
396 * an extra reference while submitting the i/o.
397 * This prevents races between the aio code path referencing the
398 * req (after submitting it) and aio_complete() freeing the req.
400 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
401 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
403 struct kiocb *req = NULL;
404 struct aio_ring *ring;
407 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
415 req->ki_cancel = NULL;
416 req->ki_retry = NULL;
419 INIT_LIST_HEAD(&req->ki_run_list);
421 /* Check if the completion queue has enough free space to
422 * accept an event from this io.
424 spin_lock_irq(&ctx->ctx_lock);
425 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
426 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
427 list_add(&req->ki_list, &ctx->active_reqs);
432 kunmap_atomic(ring, KM_USER0);
433 spin_unlock_irq(&ctx->ctx_lock);
436 kmem_cache_free(kiocb_cachep, req);
443 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
446 /* Handle a potential starvation case -- should be exceedingly rare as
447 * requests will be stuck on fput_head only if the aio_fput_routine is
448 * delayed and the requests were the last user of the struct file.
450 req = __aio_get_req(ctx);
451 if (unlikely(NULL == req)) {
452 aio_fput_routine(NULL);
453 req = __aio_get_req(ctx);
458 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
462 kmem_cache_free(kiocb_cachep, req);
465 if (unlikely(!ctx->reqs_active && ctx->dead))
469 static void aio_fput_routine(void *data)
471 spin_lock_irq(&fput_lock);
472 while (likely(!list_empty(&fput_head))) {
473 struct kiocb *req = list_kiocb(fput_head.next);
474 struct kioctx *ctx = req->ki_ctx;
476 list_del(&req->ki_list);
477 spin_unlock_irq(&fput_lock);
479 /* Complete the fput */
480 __fput(req->ki_filp);
482 /* Link the iocb into the context's free list */
483 spin_lock_irq(&ctx->ctx_lock);
484 really_put_req(ctx, req);
485 spin_unlock_irq(&ctx->ctx_lock);
488 spin_lock_irq(&fput_lock);
490 spin_unlock_irq(&fput_lock);
494 * Returns true if this put was the last user of the request.
496 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
498 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
499 req, atomic_read(&req->ki_filp->f_count));
502 if (unlikely(req->ki_users < 0))
504 if (likely(req->ki_users))
506 list_del(&req->ki_list); /* remove from active_reqs */
507 req->ki_cancel = NULL;
508 req->ki_retry = NULL;
510 /* Must be done under the lock to serialise against cancellation.
511 * Call this aio_fput as it duplicates fput via the fput_work.
513 if (unlikely(rcuref_dec_and_test(&req->ki_filp->f_count))) {
515 spin_lock(&fput_lock);
516 list_add(&req->ki_list, &fput_head);
517 spin_unlock(&fput_lock);
518 queue_work(aio_wq, &fput_work);
520 really_put_req(ctx, req);
525 * Returns true if this put was the last user of the kiocb,
526 * false if the request is still in use.
528 int fastcall aio_put_req(struct kiocb *req)
530 struct kioctx *ctx = req->ki_ctx;
532 spin_lock_irq(&ctx->ctx_lock);
533 ret = __aio_put_req(ctx, req);
534 spin_unlock_irq(&ctx->ctx_lock);
540 /* Lookup an ioctx id. ioctx_list is lockless for reads.
541 * FIXME: this is O(n) and is only suitable for development.
543 struct kioctx *lookup_ioctx(unsigned long ctx_id)
545 struct kioctx *ioctx;
546 struct mm_struct *mm;
549 read_lock(&mm->ioctx_list_lock);
550 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
551 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
555 read_unlock(&mm->ioctx_list_lock);
562 * Makes the calling kernel thread take on the specified
564 * Called by the retry thread execute retries within the
565 * iocb issuer's mm context, so that copy_from/to_user
566 * operations work seamlessly for aio.
567 * (Note: this routine is intended to be called only
568 * from a kernel thread context)
570 static void use_mm(struct mm_struct *mm)
572 struct mm_struct *active_mm;
573 struct task_struct *tsk = current;
576 tsk->flags |= PF_BORROWED_MM;
577 active_mm = tsk->active_mm;
578 atomic_inc(&mm->mm_count);
582 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
583 * it won't work. Update it accordingly if you change it here
585 activate_mm(active_mm, mm);
593 * Reverses the effect of use_mm, i.e. releases the
594 * specified mm context which was earlier taken on
595 * by the calling kernel thread
596 * (Note: this routine is intended to be called only
597 * from a kernel thread context)
599 * Comments: Called with ctx->ctx_lock held. This nests
600 * task_lock instead ctx_lock.
602 static void unuse_mm(struct mm_struct *mm)
604 struct task_struct *tsk = current;
607 tsk->flags &= ~PF_BORROWED_MM;
609 /* active_mm is still 'mm' */
610 enter_lazy_tlb(mm, tsk);
615 * Queue up a kiocb to be retried. Assumes that the kiocb
616 * has already been marked as kicked, and places it on
617 * the retry run list for the corresponding ioctx, if it
618 * isn't already queued. Returns 1 if it actually queued
619 * the kiocb (to tell the caller to activate the work
620 * queue to process it), or 0, if it found that it was
623 * Should be called with the spin lock iocb->ki_ctx->ctx_lock
626 static inline int __queue_kicked_iocb(struct kiocb *iocb)
628 struct kioctx *ctx = iocb->ki_ctx;
630 if (list_empty(&iocb->ki_run_list)) {
631 list_add_tail(&iocb->ki_run_list,
639 * This is the core aio execution routine. It is
640 * invoked both for initial i/o submission and
641 * subsequent retries via the aio_kick_handler.
642 * Expects to be invoked with iocb->ki_ctx->lock
643 * already held. The lock is released and reaquired
644 * as needed during processing.
646 * Calls the iocb retry method (already setup for the
647 * iocb on initial submission) for operation specific
648 * handling, but takes care of most of common retry
649 * execution details for a given iocb. The retry method
650 * needs to be non-blocking as far as possible, to avoid
651 * holding up other iocbs waiting to be serviced by the
652 * retry kernel thread.
654 * The trickier parts in this code have to do with
655 * ensuring that only one retry instance is in progress
656 * for a given iocb at any time. Providing that guarantee
657 * simplifies the coding of individual aio operations as
658 * it avoids various potential races.
660 static ssize_t aio_run_iocb(struct kiocb *iocb)
662 struct kioctx *ctx = iocb->ki_ctx;
663 ssize_t (*retry)(struct kiocb *);
666 if (iocb->ki_retried++ > 1024*1024) {
667 printk("Maximal retry count. Bytes done %Zd\n",
668 iocb->ki_nbytes - iocb->ki_left);
672 if (!(iocb->ki_retried & 0xff)) {
673 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
674 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
677 if (!(retry = iocb->ki_retry)) {
678 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
683 * We don't want the next retry iteration for this
684 * operation to start until this one has returned and
685 * updated the iocb state. However, wait_queue functions
686 * can trigger a kick_iocb from interrupt context in the
687 * meantime, indicating that data is available for the next
688 * iteration. We want to remember that and enable the
689 * next retry iteration _after_ we are through with
692 * So, in order to be able to register a "kick", but
693 * prevent it from being queued now, we clear the kick
694 * flag, but make the kick code *think* that the iocb is
695 * still on the run list until we are actually done.
696 * When we are done with this iteration, we check if
697 * the iocb was kicked in the meantime and if so, queue
701 kiocbClearKicked(iocb);
704 * This is so that aio_complete knows it doesn't need to
705 * pull the iocb off the run list (We can't just call
706 * INIT_LIST_HEAD because we don't want a kick_iocb to
707 * queue this on the run list yet)
709 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
710 spin_unlock_irq(&ctx->ctx_lock);
712 /* Quit retrying if the i/o has been cancelled */
713 if (kiocbIsCancelled(iocb)) {
715 aio_complete(iocb, ret, 0);
716 /* must not access the iocb after this */
721 * Now we are all set to call the retry method in async
722 * context. By setting this thread's io_wait context
723 * to point to the wait queue entry inside the currently
724 * running iocb for the duration of the retry, we ensure
725 * that async notification wakeups are queued by the
726 * operation instead of blocking waits, and when notified,
727 * cause the iocb to be kicked for continuation (through
728 * the aio_wake_function callback).
730 BUG_ON(current->io_wait != NULL);
731 current->io_wait = &iocb->ki_wait;
733 current->io_wait = NULL;
735 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
736 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
737 aio_complete(iocb, ret, 0);
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);
758 * __queue_kicked_iocb will always return 1 here, because
759 * iocb->ki_run_list is empty at this point so it should
760 * be safe to unconditionally queue the context into the
771 * Process all pending retries queued on the ioctx
773 * Assumes it is operating within the aio issuer's mm
774 * context. Expects to be called with ctx->ctx_lock held
776 static int __aio_run_iocbs(struct kioctx *ctx)
781 list_splice_init(&ctx->run_list, &run_list);
782 while (!list_empty(&run_list)) {
783 iocb = list_entry(run_list.next, struct kiocb,
785 list_del(&iocb->ki_run_list);
787 * Hold an extra reference while retrying i/o.
789 iocb->ki_users++; /* grab extra reference */
791 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
794 if (!list_empty(&ctx->run_list))
799 static void aio_queue_work(struct kioctx * ctx)
801 unsigned long timeout;
803 * if someone is waiting, get the work started right
804 * away, otherwise, use a longer delay
807 if (waitqueue_active(&ctx->wait))
811 queue_delayed_work(aio_wq, &ctx->wq, timeout);
817 * Process all pending retries queued on the ioctx
819 * Assumes it is operating within the aio issuer's mm
822 static inline void aio_run_iocbs(struct kioctx *ctx)
826 spin_lock_irq(&ctx->ctx_lock);
828 requeue = __aio_run_iocbs(ctx);
829 spin_unlock_irq(&ctx->ctx_lock);
835 * just like aio_run_iocbs, but keeps running them until
836 * the list stays empty
838 static inline void aio_run_all_iocbs(struct kioctx *ctx)
840 spin_lock_irq(&ctx->ctx_lock);
841 while (__aio_run_iocbs(ctx))
843 spin_unlock_irq(&ctx->ctx_lock);
848 * Work queue handler triggered to process pending
849 * retries on an ioctx. Takes on the aio issuer's
850 * mm context before running the iocbs, so that
851 * copy_xxx_user operates on the issuer's address
853 * Run on aiod's context.
855 static void aio_kick_handler(void *data)
857 struct kioctx *ctx = data;
858 mm_segment_t oldfs = get_fs();
863 spin_lock_irq(&ctx->ctx_lock);
864 requeue =__aio_run_iocbs(ctx);
866 spin_unlock_irq(&ctx->ctx_lock);
869 * we're in a worker thread already, don't use queue_delayed_work,
872 queue_work(aio_wq, &ctx->wq);
877 * Called by kick_iocb to queue the kiocb for retry
878 * and if required activate the aio work queue to process
881 static void try_queue_kicked_iocb(struct kiocb *iocb)
883 struct kioctx *ctx = iocb->ki_ctx;
887 /* We're supposed to be the only path putting the iocb back on the run
888 * list. If we find that the iocb is *back* on a wait queue already
889 * than retry has happened before we could queue the iocb. This also
890 * means that the retry could have completed and freed our iocb, no
892 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
894 spin_lock_irqsave(&ctx->ctx_lock, flags);
895 /* set this inside the lock so that we can't race with aio_run_iocb()
896 * testing it and putting the iocb on the run list under the lock */
897 if (!kiocbTryKick(iocb))
898 run = __queue_kicked_iocb(iocb);
899 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
906 * Called typically from a wait queue callback context
907 * (aio_wake_function) to trigger a retry of the iocb.
908 * The retry is usually executed by aio workqueue
909 * threads (See aio_kick_handler).
911 void fastcall kick_iocb(struct kiocb *iocb)
913 /* sync iocbs are easy: they can only ever be executing from a
915 if (is_sync_kiocb(iocb)) {
916 kiocbSetKicked(iocb);
917 wake_up_process(iocb->ki_obj.tsk);
921 try_queue_kicked_iocb(iocb);
923 EXPORT_SYMBOL(kick_iocb);
926 * Called when the io request on the given iocb is complete.
927 * Returns true if this is the last user of the request. The
928 * only other user of the request can be the cancellation code.
930 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
932 struct kioctx *ctx = iocb->ki_ctx;
933 struct aio_ring_info *info;
934 struct aio_ring *ring;
935 struct io_event *event;
941 * Special case handling for sync iocbs:
942 * - events go directly into the iocb for fast handling
943 * - the sync task with the iocb in its stack holds the single iocb
944 * ref, no other paths have a way to get another ref
945 * - the sync task helpfully left a reference to itself in the iocb
947 if (is_sync_kiocb(iocb)) {
948 BUG_ON(iocb->ki_users != 1);
949 iocb->ki_user_data = res;
951 wake_up_process(iocb->ki_obj.tsk);
955 info = &ctx->ring_info;
957 /* add a completion event to the ring buffer.
958 * must be done holding ctx->ctx_lock to prevent
959 * other code from messing with the tail
960 * pointer since we might be called from irq
963 spin_lock_irqsave(&ctx->ctx_lock, flags);
965 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
966 list_del_init(&iocb->ki_run_list);
969 * cancelled requests don't get events, userland was given one
970 * when the event got cancelled.
972 if (kiocbIsCancelled(iocb))
975 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
978 event = aio_ring_event(info, tail, KM_IRQ0);
979 if (++tail >= info->nr)
982 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
983 event->data = iocb->ki_user_data;
987 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
988 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
991 /* after flagging the request as done, we
992 * must never even look at it again
994 smp_wmb(); /* make event visible before updating tail */
999 put_aio_ring_event(event, KM_IRQ0);
1000 kunmap_atomic(ring, KM_IRQ1);
1002 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1004 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1005 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
1007 /* everything turned out well, dispose of the aiocb. */
1008 ret = __aio_put_req(ctx, iocb);
1010 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1012 if (waitqueue_active(&ctx->wait))
1013 wake_up(&ctx->wait);
1022 * Pull an event off of the ioctx's event ring. Returns the number of
1023 * events fetched (0 or 1 ;-)
1024 * FIXME: make this use cmpxchg.
1025 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1027 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1029 struct aio_ring_info *info = &ioctx->ring_info;
1030 struct aio_ring *ring;
1034 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1035 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1036 (unsigned long)ring->head, (unsigned long)ring->tail,
1037 (unsigned long)ring->nr);
1039 if (ring->head == ring->tail)
1042 spin_lock(&info->ring_lock);
1044 head = ring->head % info->nr;
1045 if (head != ring->tail) {
1046 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1048 head = (head + 1) % info->nr;
1049 smp_mb(); /* finish reading the event before updatng the head */
1052 put_aio_ring_event(evp, KM_USER1);
1054 spin_unlock(&info->ring_lock);
1057 kunmap_atomic(ring, KM_USER0);
1058 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1059 (unsigned long)ring->head, (unsigned long)ring->tail);
1063 struct aio_timeout {
1064 struct timer_list timer;
1066 struct task_struct *p;
1069 static void timeout_func(unsigned long data)
1071 struct aio_timeout *to = (struct aio_timeout *)data;
1074 wake_up_process(to->p);
1077 static inline void init_timeout(struct aio_timeout *to)
1079 init_timer(&to->timer);
1080 to->timer.data = (unsigned long)to;
1081 to->timer.function = timeout_func;
1086 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1087 const struct timespec *ts)
1089 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1090 if (time_after(to->timer.expires, jiffies))
1091 add_timer(&to->timer);
1096 static inline void clear_timeout(struct aio_timeout *to)
1098 del_singleshot_timer_sync(&to->timer);
1101 static int read_events(struct kioctx *ctx,
1102 long min_nr, long nr,
1103 struct io_event __user *event,
1104 struct timespec __user *timeout)
1106 long start_jiffies = jiffies;
1107 struct task_struct *tsk = current;
1108 DECLARE_WAITQUEUE(wait, tsk);
1111 struct io_event ent;
1112 struct aio_timeout to;
1115 /* needed to zero any padding within an entry (there shouldn't be
1116 * any, but C is fun!
1118 memset(&ent, 0, sizeof(ent));
1121 while (likely(i < nr)) {
1122 ret = aio_read_evt(ctx, &ent);
1123 if (unlikely(ret <= 0))
1126 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1127 ent.data, ent.obj, ent.res, ent.res2);
1129 /* Could we split the check in two? */
1131 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1132 dprintk("aio: lost an event due to EFAULT.\n");
1137 /* Good, event copied to userland, update counts. */
1149 /* racey check, but it gets redone */
1150 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1152 aio_run_all_iocbs(ctx);
1160 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1163 set_timeout(start_jiffies, &to, &ts);
1166 while (likely(i < nr)) {
1167 add_wait_queue_exclusive(&ctx->wait, &wait);
1169 set_task_state(tsk, TASK_INTERRUPTIBLE);
1170 ret = aio_read_evt(ctx, &ent);
1176 if (to.timed_out) /* Only check after read evt */
1179 if (signal_pending(tsk)) {
1183 /*ret = aio_read_evt(ctx, &ent);*/
1186 set_task_state(tsk, TASK_RUNNING);
1187 remove_wait_queue(&ctx->wait, &wait);
1189 if (unlikely(ret <= 0))
1193 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1194 dprintk("aio: lost an event due to EFAULT.\n");
1198 /* Good, event copied to userland, update counts. */
1209 /* Take an ioctx and remove it from the list of ioctx's. Protects
1210 * against races with itself via ->dead.
1212 static void io_destroy(struct kioctx *ioctx)
1214 struct mm_struct *mm = current->mm;
1215 struct kioctx **tmp;
1218 /* delete the entry from the list is someone else hasn't already */
1219 write_lock(&mm->ioctx_list_lock);
1220 was_dead = ioctx->dead;
1222 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1223 tmp = &(*tmp)->next)
1227 write_unlock(&mm->ioctx_list_lock);
1229 dprintk("aio_release(%p)\n", ioctx);
1230 if (likely(!was_dead))
1231 put_ioctx(ioctx); /* twice for the list */
1233 aio_cancel_all(ioctx);
1234 wait_for_all_aios(ioctx);
1235 put_ioctx(ioctx); /* once for the lookup */
1239 * Create an aio_context capable of receiving at least nr_events.
1240 * ctxp must not point to an aio_context that already exists, and
1241 * must be initialized to 0 prior to the call. On successful
1242 * creation of the aio_context, *ctxp is filled in with the resulting
1243 * handle. May fail with -EINVAL if *ctxp is not initialized,
1244 * if the specified nr_events exceeds internal limits. May fail
1245 * with -EAGAIN if the specified nr_events exceeds the user's limit
1246 * of available events. May fail with -ENOMEM if insufficient kernel
1247 * resources are available. May fail with -EFAULT if an invalid
1248 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1251 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1253 struct kioctx *ioctx = NULL;
1257 ret = get_user(ctx, ctxp);
1262 if (unlikely(ctx || nr_events == 0)) {
1263 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1268 ioctx = ioctx_alloc(nr_events);
1269 ret = PTR_ERR(ioctx);
1270 if (!IS_ERR(ioctx)) {
1271 ret = put_user(ioctx->user_id, ctxp);
1275 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1284 * Destroy the aio_context specified. May cancel any outstanding
1285 * AIOs and block on completion. Will fail with -ENOSYS if not
1286 * implemented. May fail with -EFAULT if the context pointed to
1289 asmlinkage long sys_io_destroy(aio_context_t ctx)
1291 struct kioctx *ioctx = lookup_ioctx(ctx);
1292 if (likely(NULL != ioctx)) {
1296 pr_debug("EINVAL: io_destroy: invalid context id\n");
1301 * aio_p{read,write} are the default ki_retry methods for
1302 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially
1303 * multiple calls to f_op->aio_read(). They loop around partial progress
1304 * instead of returning -EIOCBRETRY because they don't have the means to call
1307 static ssize_t aio_pread(struct kiocb *iocb)
1309 struct file *file = iocb->ki_filp;
1310 struct address_space *mapping = file->f_mapping;
1311 struct inode *inode = mapping->host;
1315 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1316 iocb->ki_left, iocb->ki_pos);
1318 * Can't just depend on iocb->ki_left to determine
1319 * whether we are done. This may have been a short read.
1322 iocb->ki_buf += ret;
1323 iocb->ki_left -= ret;
1327 * For pipes and sockets we return once we have some data; for
1328 * regular files we retry till we complete the entire read or
1329 * find that we can't read any more data (e.g short reads).
1331 } while (ret > 0 && iocb->ki_left > 0 &&
1332 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode));
1334 /* This means we must have transferred all that we could */
1335 /* No need to retry anymore */
1336 if ((ret == 0) || (iocb->ki_left == 0))
1337 ret = iocb->ki_nbytes - iocb->ki_left;
1342 /* see aio_pread() */
1343 static ssize_t aio_pwrite(struct kiocb *iocb)
1345 struct file *file = iocb->ki_filp;
1349 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1350 iocb->ki_left, iocb->ki_pos);
1352 iocb->ki_buf += ret;
1353 iocb->ki_left -= ret;
1355 } while (ret > 0 && iocb->ki_left > 0);
1357 if ((ret == 0) || (iocb->ki_left == 0))
1358 ret = iocb->ki_nbytes - iocb->ki_left;
1363 static ssize_t aio_fdsync(struct kiocb *iocb)
1365 struct file *file = iocb->ki_filp;
1366 ssize_t ret = -EINVAL;
1368 if (file->f_op->aio_fsync)
1369 ret = file->f_op->aio_fsync(iocb, 1);
1373 static ssize_t aio_fsync(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, 0);
1385 * Performs the initial checks and aio retry method
1386 * setup for the kiocb at the time of io submission.
1388 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1390 struct file *file = kiocb->ki_filp;
1393 switch (kiocb->ki_opcode) {
1394 case IOCB_CMD_PREAD:
1396 if (unlikely(!(file->f_mode & FMODE_READ)))
1399 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1402 ret = security_file_permission(file, MAY_READ);
1406 if (file->f_op->aio_read)
1407 kiocb->ki_retry = aio_pread;
1409 case IOCB_CMD_PWRITE:
1411 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1414 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1417 ret = security_file_permission(file, MAY_WRITE);
1421 if (file->f_op->aio_write)
1422 kiocb->ki_retry = aio_pwrite;
1424 case IOCB_CMD_FDSYNC:
1426 if (file->f_op->aio_fsync)
1427 kiocb->ki_retry = aio_fdsync;
1429 case IOCB_CMD_FSYNC:
1431 if (file->f_op->aio_fsync)
1432 kiocb->ki_retry = aio_fsync;
1435 dprintk("EINVAL: io_submit: no operation provided\n");
1439 if (!kiocb->ki_retry)
1446 * aio_wake_function:
1447 * wait queue callback function for aio notification,
1448 * Simply triggers a retry of the operation via kick_iocb.
1450 * This callback is specified in the wait queue entry in
1451 * a kiocb (current->io_wait points to this wait queue
1452 * entry when an aio operation executes; it is used
1453 * instead of a synchronous wait when an i/o blocking
1454 * condition is encountered during aio).
1457 * This routine is executed with the wait queue lock held.
1458 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1459 * the ioctx lock inside the wait queue lock. This is safe
1460 * because this callback isn't used for wait queues which
1461 * are nested inside ioctx lock (i.e. ctx->wait)
1463 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1464 int sync, void *key)
1466 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1468 list_del_init(&wait->task_list);
1473 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1480 /* enforce forwards compatibility on users */
1481 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1482 iocb->aio_reserved3)) {
1483 pr_debug("EINVAL: io_submit: reserve field set\n");
1487 /* prevent overflows */
1489 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1490 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1491 ((ssize_t)iocb->aio_nbytes < 0)
1493 pr_debug("EINVAL: io_submit: overflow check\n");
1497 file = fget(iocb->aio_fildes);
1498 if (unlikely(!file))
1501 req = aio_get_req(ctx); /* returns with 2 references to req */
1502 if (unlikely(!req)) {
1507 req->ki_filp = file;
1508 ret = put_user(req->ki_key, &user_iocb->aio_key);
1509 if (unlikely(ret)) {
1510 dprintk("EFAULT: aio_key\n");
1514 req->ki_obj.user = user_iocb;
1515 req->ki_user_data = iocb->aio_data;
1516 req->ki_pos = iocb->aio_offset;
1518 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1519 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1520 req->ki_opcode = iocb->aio_lio_opcode;
1521 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1522 INIT_LIST_HEAD(&req->ki_wait.task_list);
1523 req->ki_retried = 0;
1525 ret = aio_setup_iocb(req);
1530 spin_lock_irq(&ctx->ctx_lock);
1532 if (!list_empty(&ctx->run_list)) {
1533 /* drain the run list */
1534 while (__aio_run_iocbs(ctx))
1537 spin_unlock_irq(&ctx->ctx_lock);
1538 aio_put_req(req); /* drop extra ref to req */
1542 aio_put_req(req); /* drop extra ref to req */
1543 aio_put_req(req); /* drop i/o ref to req */
1548 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1549 * the number of iocbs queued. May return -EINVAL if the aio_context
1550 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1551 * *iocbpp[0] is not properly initialized, if the operation specified
1552 * is invalid for the file descriptor in the iocb. May fail with
1553 * -EFAULT if any of the data structures point to invalid data. May
1554 * fail with -EBADF if the file descriptor specified in the first
1555 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1556 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1557 * fail with -ENOSYS if not implemented.
1559 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1560 struct iocb __user * __user *iocbpp)
1566 if (unlikely(nr < 0))
1569 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1572 ctx = lookup_ioctx(ctx_id);
1573 if (unlikely(!ctx)) {
1574 pr_debug("EINVAL: io_submit: invalid context id\n");
1579 * AKPM: should this return a partial result if some of the IOs were
1580 * successfully submitted?
1582 for (i=0; i<nr; i++) {
1583 struct iocb __user *user_iocb;
1586 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1591 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1596 ret = io_submit_one(ctx, user_iocb, &tmp);
1606 * Finds a given iocb for cancellation.
1607 * MUST be called with ctx->ctx_lock held.
1609 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1612 struct list_head *pos;
1613 /* TODO: use a hash or array, this sucks. */
1614 list_for_each(pos, &ctx->active_reqs) {
1615 struct kiocb *kiocb = list_kiocb(pos);
1616 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1623 * Attempts to cancel an iocb previously passed to io_submit. If
1624 * the operation is successfully cancelled, the resulting event is
1625 * copied into the memory pointed to by result without being placed
1626 * into the completion queue and 0 is returned. May fail with
1627 * -EFAULT if any of the data structures pointed to are invalid.
1628 * May fail with -EINVAL if aio_context specified by ctx_id is
1629 * invalid. May fail with -EAGAIN if the iocb specified was not
1630 * cancelled. Will fail with -ENOSYS if not implemented.
1632 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1633 struct io_event __user *result)
1635 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1637 struct kiocb *kiocb;
1641 ret = get_user(key, &iocb->aio_key);
1645 ctx = lookup_ioctx(ctx_id);
1649 spin_lock_irq(&ctx->ctx_lock);
1651 kiocb = lookup_kiocb(ctx, iocb, key);
1652 if (kiocb && kiocb->ki_cancel) {
1653 cancel = kiocb->ki_cancel;
1655 kiocbSetCancelled(kiocb);
1658 spin_unlock_irq(&ctx->ctx_lock);
1660 if (NULL != cancel) {
1661 struct io_event tmp;
1662 pr_debug("calling cancel\n");
1663 memset(&tmp, 0, sizeof(tmp));
1664 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1665 tmp.data = kiocb->ki_user_data;
1666 ret = cancel(kiocb, &tmp);
1668 /* Cancellation succeeded -- copy the result
1669 * into the user's buffer.
1671 if (copy_to_user(result, &tmp, sizeof(tmp)))
1683 * Attempts to read at least min_nr events and up to nr events from
1684 * the completion queue for the aio_context specified by ctx_id. May
1685 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1686 * if nr is out of range, if when is out of range. May fail with
1687 * -EFAULT if any of the memory specified to is invalid. May return
1688 * 0 or < min_nr if no events are available and the timeout specified
1689 * by when has elapsed, where when == NULL specifies an infinite
1690 * timeout. Note that the timeout pointed to by when is relative and
1691 * will be updated if not NULL and the operation blocks. Will fail
1692 * with -ENOSYS if not implemented.
1694 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1697 struct io_event __user *events,
1698 struct timespec __user *timeout)
1700 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1703 if (likely(ioctx)) {
1704 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1705 ret = read_events(ioctx, min_nr, nr, events, timeout);
1712 __initcall(aio_setup);
1714 EXPORT_SYMBOL(aio_complete);
1715 EXPORT_SYMBOL(aio_put_req);
1716 EXPORT_SYMBOL(wait_on_sync_kiocb);