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 /*------ sysctl variables----*/
44 static DEFINE_SPINLOCK(aio_nr_lock);
45 unsigned long aio_nr; /* current system wide number of aio requests */
46 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
47 /*----end sysctl variables---*/
49 static kmem_cache_t *kiocb_cachep;
50 static kmem_cache_t *kioctx_cachep;
52 static struct workqueue_struct *aio_wq;
54 /* Used for rare fput completion. */
55 static void aio_fput_routine(void *);
56 static DECLARE_WORK(fput_work, aio_fput_routine, NULL);
58 static DEFINE_SPINLOCK(fput_lock);
59 static LIST_HEAD(fput_head);
61 static void aio_kick_handler(void *);
62 static void aio_queue_work(struct kioctx *);
65 * Creates the slab caches used by the aio routines, panic on
66 * failure as this is done early during the boot sequence.
68 static int __init aio_setup(void)
70 kiocb_cachep = kmem_cache_create("kiocb", sizeof(struct kiocb),
71 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
72 kioctx_cachep = kmem_cache_create("kioctx", sizeof(struct kioctx),
73 0, SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
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_ANON|MAP_PRIVATE,
136 if (IS_ERR((void *)info->mmap_base)) {
137 up_write(&ctx->mm->mmap_sem);
138 printk("mmap err: %ld\n", -info->mmap_base);
144 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
145 info->nr_pages = get_user_pages(current, ctx->mm,
146 info->mmap_base, nr_pages,
147 1, 0, info->ring_pages, NULL);
148 up_write(&ctx->mm->mmap_sem);
150 if (unlikely(info->nr_pages != nr_pages)) {
155 ctx->user_id = info->mmap_base;
157 info->nr = nr_events; /* trusted copy */
159 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
160 ring->nr = nr_events; /* user copy */
161 ring->id = ctx->user_id;
162 ring->head = ring->tail = 0;
163 ring->magic = AIO_RING_MAGIC;
164 ring->compat_features = AIO_RING_COMPAT_FEATURES;
165 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
166 ring->header_length = sizeof(struct aio_ring);
167 kunmap_atomic(ring, KM_USER0);
173 /* aio_ring_event: returns a pointer to the event at the given index from
174 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
176 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
177 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
178 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
180 #define aio_ring_event(info, nr, km) ({ \
181 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
182 struct io_event *__event; \
183 __event = kmap_atomic( \
184 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
185 __event += pos % AIO_EVENTS_PER_PAGE; \
189 #define put_aio_ring_event(event, km) do { \
190 struct io_event *__event = (event); \
192 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
196 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
198 static struct kioctx *ioctx_alloc(unsigned nr_events)
200 struct mm_struct *mm;
203 /* Prevent overflows */
204 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
205 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
206 pr_debug("ENOMEM: nr_events too high\n");
207 return ERR_PTR(-EINVAL);
210 if ((unsigned long)nr_events > aio_max_nr)
211 return ERR_PTR(-EAGAIN);
213 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
215 return ERR_PTR(-ENOMEM);
217 memset(ctx, 0, sizeof(*ctx));
218 ctx->max_reqs = nr_events;
219 mm = ctx->mm = current->mm;
220 atomic_inc(&mm->mm_count);
222 atomic_set(&ctx->users, 1);
223 spin_lock_init(&ctx->ctx_lock);
224 spin_lock_init(&ctx->ring_info.ring_lock);
225 init_waitqueue_head(&ctx->wait);
227 INIT_LIST_HEAD(&ctx->active_reqs);
228 INIT_LIST_HEAD(&ctx->run_list);
229 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
231 if (aio_setup_ring(ctx) < 0)
234 /* limit the number of system wide aios */
235 spin_lock(&aio_nr_lock);
236 if (aio_nr + ctx->max_reqs > aio_max_nr ||
237 aio_nr + ctx->max_reqs < aio_nr)
240 aio_nr += ctx->max_reqs;
241 spin_unlock(&aio_nr_lock);
242 if (ctx->max_reqs == 0)
245 /* now link into global list. kludge. FIXME */
246 write_lock(&mm->ioctx_list_lock);
247 ctx->next = mm->ioctx_list;
248 mm->ioctx_list = ctx;
249 write_unlock(&mm->ioctx_list_lock);
251 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
252 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
257 return ERR_PTR(-EAGAIN);
261 kmem_cache_free(kioctx_cachep, ctx);
262 ctx = ERR_PTR(-ENOMEM);
264 dprintk("aio: error allocating ioctx %p\n", ctx);
269 * Cancels all outstanding aio requests on an aio context. Used
270 * when the processes owning a context have all exited to encourage
271 * the rapid destruction of the kioctx.
273 static void aio_cancel_all(struct kioctx *ctx)
275 int (*cancel)(struct kiocb *, struct io_event *);
277 spin_lock_irq(&ctx->ctx_lock);
279 while (!list_empty(&ctx->active_reqs)) {
280 struct list_head *pos = ctx->active_reqs.next;
281 struct kiocb *iocb = list_kiocb(pos);
282 list_del_init(&iocb->ki_list);
283 cancel = iocb->ki_cancel;
284 kiocbSetCancelled(iocb);
287 spin_unlock_irq(&ctx->ctx_lock);
289 spin_lock_irq(&ctx->ctx_lock);
292 spin_unlock_irq(&ctx->ctx_lock);
295 static void wait_for_all_aios(struct kioctx *ctx)
297 struct task_struct *tsk = current;
298 DECLARE_WAITQUEUE(wait, tsk);
300 if (!ctx->reqs_active)
303 add_wait_queue(&ctx->wait, &wait);
304 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
305 while (ctx->reqs_active) {
307 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
309 __set_task_state(tsk, TASK_RUNNING);
310 remove_wait_queue(&ctx->wait, &wait);
313 /* wait_on_sync_kiocb:
314 * Waits on the given sync kiocb to complete.
316 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
318 while (iocb->ki_users) {
319 set_current_state(TASK_UNINTERRUPTIBLE);
324 __set_current_state(TASK_RUNNING);
325 return iocb->ki_user_data;
328 /* exit_aio: called when the last user of mm goes away. At this point,
329 * there is no way for any new requests to be submited or any of the
330 * io_* syscalls to be called on the context. However, there may be
331 * outstanding requests which hold references to the context; as they
332 * go away, they will call put_ioctx and release any pinned memory
333 * associated with the request (held via struct page * references).
335 void fastcall exit_aio(struct mm_struct *mm)
337 struct kioctx *ctx = mm->ioctx_list;
338 mm->ioctx_list = NULL;
340 struct kioctx *next = ctx->next;
344 wait_for_all_aios(ctx);
346 * this is an overkill, but ensures we don't leave
347 * the ctx on the aio_wq
349 flush_workqueue(aio_wq);
351 if (1 != atomic_read(&ctx->users))
353 "exit_aio:ioctx still alive: %d %d %d\n",
354 atomic_read(&ctx->users), ctx->dead,
362 * Called when the last user of an aio context has gone away,
363 * and the struct needs to be freed.
365 void fastcall __put_ioctx(struct kioctx *ctx)
367 unsigned nr_events = ctx->max_reqs;
369 if (unlikely(ctx->reqs_active))
372 cancel_delayed_work(&ctx->wq);
373 flush_workqueue(aio_wq);
377 pr_debug("__put_ioctx: freeing %p\n", ctx);
378 kmem_cache_free(kioctx_cachep, ctx);
381 spin_lock(&aio_nr_lock);
382 BUG_ON(aio_nr - nr_events > aio_nr);
384 spin_unlock(&aio_nr_lock);
389 * Allocate a slot for an aio request. Increments the users count
390 * of the kioctx so that the kioctx stays around until all requests are
391 * complete. Returns NULL if no requests are free.
393 * Returns with kiocb->users set to 2. The io submit code path holds
394 * an extra reference while submitting the i/o.
395 * This prevents races between the aio code path referencing the
396 * req (after submitting it) and aio_complete() freeing the req.
398 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
399 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
401 struct kiocb *req = NULL;
402 struct aio_ring *ring;
405 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
413 req->ki_cancel = NULL;
414 req->ki_retry = NULL;
417 INIT_LIST_HEAD(&req->ki_run_list);
419 /* Check if the completion queue has enough free space to
420 * accept an event from this io.
422 spin_lock_irq(&ctx->ctx_lock);
423 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
424 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
425 list_add(&req->ki_list, &ctx->active_reqs);
430 kunmap_atomic(ring, KM_USER0);
431 spin_unlock_irq(&ctx->ctx_lock);
434 kmem_cache_free(kiocb_cachep, req);
441 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
444 /* Handle a potential starvation case -- should be exceedingly rare as
445 * requests will be stuck on fput_head only if the aio_fput_routine is
446 * delayed and the requests were the last user of the struct file.
448 req = __aio_get_req(ctx);
449 if (unlikely(NULL == req)) {
450 aio_fput_routine(NULL);
451 req = __aio_get_req(ctx);
456 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
458 assert_spin_locked(&ctx->ctx_lock);
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));
501 assert_spin_locked(&ctx->ctx_lock);
504 if (unlikely(req->ki_users < 0))
506 if (likely(req->ki_users))
508 list_del(&req->ki_list); /* remove from active_reqs */
509 req->ki_cancel = NULL;
510 req->ki_retry = NULL;
512 /* Must be done under the lock to serialise against cancellation.
513 * Call this aio_fput as it duplicates fput via the fput_work.
515 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
517 spin_lock(&fput_lock);
518 list_add(&req->ki_list, &fput_head);
519 spin_unlock(&fput_lock);
520 queue_work(aio_wq, &fput_work);
522 really_put_req(ctx, req);
527 * Returns true if this put was the last user of the kiocb,
528 * false if the request is still in use.
530 int fastcall aio_put_req(struct kiocb *req)
532 struct kioctx *ctx = req->ki_ctx;
534 spin_lock_irq(&ctx->ctx_lock);
535 ret = __aio_put_req(ctx, req);
536 spin_unlock_irq(&ctx->ctx_lock);
542 /* Lookup an ioctx id. ioctx_list is lockless for reads.
543 * FIXME: this is O(n) and is only suitable for development.
545 struct kioctx *lookup_ioctx(unsigned long ctx_id)
547 struct kioctx *ioctx;
548 struct mm_struct *mm;
551 read_lock(&mm->ioctx_list_lock);
552 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
553 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
557 read_unlock(&mm->ioctx_list_lock);
564 * Makes the calling kernel thread take on the specified
566 * Called by the retry thread execute retries within the
567 * iocb issuer's mm context, so that copy_from/to_user
568 * operations work seamlessly for aio.
569 * (Note: this routine is intended to be called only
570 * from a kernel thread context)
572 static void use_mm(struct mm_struct *mm)
574 struct mm_struct *active_mm;
575 struct task_struct *tsk = current;
578 tsk->flags |= PF_BORROWED_MM;
579 active_mm = tsk->active_mm;
580 atomic_inc(&mm->mm_count);
584 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
585 * it won't work. Update it accordingly if you change it here
587 activate_mm(active_mm, mm);
595 * Reverses the effect of use_mm, i.e. releases the
596 * specified mm context which was earlier taken on
597 * by the calling kernel thread
598 * (Note: this routine is intended to be called only
599 * from a kernel thread context)
601 * Comments: Called with ctx->ctx_lock held. This nests
602 * task_lock instead ctx_lock.
604 static void unuse_mm(struct mm_struct *mm)
606 struct task_struct *tsk = current;
609 tsk->flags &= ~PF_BORROWED_MM;
611 /* active_mm is still 'mm' */
612 enter_lazy_tlb(mm, tsk);
617 * Queue up a kiocb to be retried. Assumes that the kiocb
618 * has already been marked as kicked, and places it on
619 * the retry run list for the corresponding ioctx, if it
620 * isn't already queued. Returns 1 if it actually queued
621 * the kiocb (to tell the caller to activate the work
622 * queue to process it), or 0, if it found that it was
625 static inline int __queue_kicked_iocb(struct kiocb *iocb)
627 struct kioctx *ctx = iocb->ki_ctx;
629 assert_spin_locked(&ctx->ctx_lock);
631 if (list_empty(&iocb->ki_run_list)) {
632 list_add_tail(&iocb->ki_run_list,
640 * This is the core aio execution routine. It is
641 * invoked both for initial i/o submission and
642 * subsequent retries via the aio_kick_handler.
643 * Expects to be invoked with iocb->ki_ctx->lock
644 * already held. The lock is released and reaquired
645 * as needed during processing.
647 * Calls the iocb retry method (already setup for the
648 * iocb on initial submission) for operation specific
649 * handling, but takes care of most of common retry
650 * execution details for a given iocb. The retry method
651 * needs to be non-blocking as far as possible, to avoid
652 * holding up other iocbs waiting to be serviced by the
653 * retry kernel thread.
655 * The trickier parts in this code have to do with
656 * ensuring that only one retry instance is in progress
657 * for a given iocb at any time. Providing that guarantee
658 * simplifies the coding of individual aio operations as
659 * it avoids various potential races.
661 static ssize_t aio_run_iocb(struct kiocb *iocb)
663 struct kioctx *ctx = iocb->ki_ctx;
664 ssize_t (*retry)(struct kiocb *);
667 if (iocb->ki_retried++ > 1024*1024) {
668 printk("Maximal retry count. Bytes done %Zd\n",
669 iocb->ki_nbytes - iocb->ki_left);
673 if (!(iocb->ki_retried & 0xff)) {
674 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
675 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
678 if (!(retry = iocb->ki_retry)) {
679 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
684 * We don't want the next retry iteration for this
685 * operation to start until this one has returned and
686 * updated the iocb state. However, wait_queue functions
687 * can trigger a kick_iocb from interrupt context in the
688 * meantime, indicating that data is available for the next
689 * iteration. We want to remember that and enable the
690 * next retry iteration _after_ we are through with
693 * So, in order to be able to register a "kick", but
694 * prevent it from being queued now, we clear the kick
695 * flag, but make the kick code *think* that the iocb is
696 * still on the run list until we are actually done.
697 * When we are done with this iteration, we check if
698 * the iocb was kicked in the meantime and if so, queue
702 kiocbClearKicked(iocb);
705 * This is so that aio_complete knows it doesn't need to
706 * pull the iocb off the run list (We can't just call
707 * INIT_LIST_HEAD because we don't want a kick_iocb to
708 * queue this on the run list yet)
710 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
711 spin_unlock_irq(&ctx->ctx_lock);
713 /* Quit retrying if the i/o has been cancelled */
714 if (kiocbIsCancelled(iocb)) {
716 aio_complete(iocb, ret, 0);
717 /* must not access the iocb after this */
722 * Now we are all set to call the retry method in async
723 * context. By setting this thread's io_wait context
724 * to point to the wait queue entry inside the currently
725 * running iocb for the duration of the retry, we ensure
726 * that async notification wakeups are queued by the
727 * operation instead of blocking waits, and when notified,
728 * cause the iocb to be kicked for continuation (through
729 * the aio_wake_function callback).
731 BUG_ON(current->io_wait != NULL);
732 current->io_wait = &iocb->ki_wait;
734 current->io_wait = NULL;
736 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
737 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
738 aio_complete(iocb, ret, 0);
741 spin_lock_irq(&ctx->ctx_lock);
743 if (-EIOCBRETRY == ret) {
745 * OK, now that we are done with this iteration
746 * and know that there is more left to go,
747 * this is where we let go so that a subsequent
748 * "kick" can start the next iteration
751 /* will make __queue_kicked_iocb succeed from here on */
752 INIT_LIST_HEAD(&iocb->ki_run_list);
753 /* we must queue the next iteration ourselves, if it
754 * has already been kicked */
755 if (kiocbIsKicked(iocb)) {
756 __queue_kicked_iocb(iocb);
759 * __queue_kicked_iocb will always return 1 here, because
760 * iocb->ki_run_list is empty at this point so it should
761 * be safe to unconditionally queue the context into the
772 * Process all pending retries queued on the ioctx
774 * Assumes it is operating within the aio issuer's mm
777 static int __aio_run_iocbs(struct kioctx *ctx)
780 struct list_head run_list;
782 assert_spin_locked(&ctx->ctx_lock);
784 list_replace_init(&ctx->run_list, &run_list);
785 while (!list_empty(&run_list)) {
786 iocb = list_entry(run_list.next, struct kiocb,
788 list_del(&iocb->ki_run_list);
790 * Hold an extra reference while retrying i/o.
792 iocb->ki_users++; /* grab extra reference */
794 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
797 if (!list_empty(&ctx->run_list))
802 static void aio_queue_work(struct kioctx * ctx)
804 unsigned long timeout;
806 * if someone is waiting, get the work started right
807 * away, otherwise, use a longer delay
810 if (waitqueue_active(&ctx->wait))
814 queue_delayed_work(aio_wq, &ctx->wq, timeout);
820 * Process all pending retries queued on the ioctx
822 * Assumes it is operating within the aio issuer's mm
825 static inline void aio_run_iocbs(struct kioctx *ctx)
829 spin_lock_irq(&ctx->ctx_lock);
831 requeue = __aio_run_iocbs(ctx);
832 spin_unlock_irq(&ctx->ctx_lock);
838 * just like aio_run_iocbs, but keeps running them until
839 * the list stays empty
841 static inline void aio_run_all_iocbs(struct kioctx *ctx)
843 spin_lock_irq(&ctx->ctx_lock);
844 while (__aio_run_iocbs(ctx))
846 spin_unlock_irq(&ctx->ctx_lock);
851 * Work queue handler triggered to process pending
852 * retries on an ioctx. Takes on the aio issuer's
853 * mm context before running the iocbs, so that
854 * copy_xxx_user operates on the issuer's address
856 * Run on aiod's context.
858 static void aio_kick_handler(void *data)
860 struct kioctx *ctx = data;
861 mm_segment_t oldfs = get_fs();
866 spin_lock_irq(&ctx->ctx_lock);
867 requeue =__aio_run_iocbs(ctx);
869 spin_unlock_irq(&ctx->ctx_lock);
872 * we're in a worker thread already, don't use queue_delayed_work,
875 queue_work(aio_wq, &ctx->wq);
880 * Called by kick_iocb to queue the kiocb for retry
881 * and if required activate the aio work queue to process
884 static void try_queue_kicked_iocb(struct kiocb *iocb)
886 struct kioctx *ctx = iocb->ki_ctx;
890 /* We're supposed to be the only path putting the iocb back on the run
891 * list. If we find that the iocb is *back* on a wait queue already
892 * than retry has happened before we could queue the iocb. This also
893 * means that the retry could have completed and freed our iocb, no
895 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
897 spin_lock_irqsave(&ctx->ctx_lock, flags);
898 /* set this inside the lock so that we can't race with aio_run_iocb()
899 * testing it and putting the iocb on the run list under the lock */
900 if (!kiocbTryKick(iocb))
901 run = __queue_kicked_iocb(iocb);
902 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
909 * Called typically from a wait queue callback context
910 * (aio_wake_function) to trigger a retry of the iocb.
911 * The retry is usually executed by aio workqueue
912 * threads (See aio_kick_handler).
914 void fastcall kick_iocb(struct kiocb *iocb)
916 /* sync iocbs are easy: they can only ever be executing from a
918 if (is_sync_kiocb(iocb)) {
919 kiocbSetKicked(iocb);
920 wake_up_process(iocb->ki_obj.tsk);
924 try_queue_kicked_iocb(iocb);
926 EXPORT_SYMBOL(kick_iocb);
929 * Called when the io request on the given iocb is complete.
930 * Returns true if this is the last user of the request. The
931 * only other user of the request can be the cancellation code.
933 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
935 struct kioctx *ctx = iocb->ki_ctx;
936 struct aio_ring_info *info;
937 struct aio_ring *ring;
938 struct io_event *event;
944 * Special case handling for sync iocbs:
945 * - events go directly into the iocb for fast handling
946 * - the sync task with the iocb in its stack holds the single iocb
947 * ref, no other paths have a way to get another ref
948 * - the sync task helpfully left a reference to itself in the iocb
950 if (is_sync_kiocb(iocb)) {
951 BUG_ON(iocb->ki_users != 1);
952 iocb->ki_user_data = res;
954 wake_up_process(iocb->ki_obj.tsk);
958 info = &ctx->ring_info;
960 /* add a completion event to the ring buffer.
961 * must be done holding ctx->ctx_lock to prevent
962 * other code from messing with the tail
963 * pointer since we might be called from irq
966 spin_lock_irqsave(&ctx->ctx_lock, flags);
968 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
969 list_del_init(&iocb->ki_run_list);
972 * cancelled requests don't get events, userland was given one
973 * when the event got cancelled.
975 if (kiocbIsCancelled(iocb))
978 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
981 event = aio_ring_event(info, tail, KM_IRQ0);
982 if (++tail >= info->nr)
985 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
986 event->data = iocb->ki_user_data;
990 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
991 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
994 /* after flagging the request as done, we
995 * must never even look at it again
997 smp_wmb(); /* make event visible before updating tail */
1002 put_aio_ring_event(event, KM_IRQ0);
1003 kunmap_atomic(ring, KM_IRQ1);
1005 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1007 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1008 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
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;
1118 /* needed to zero any padding within an entry (there shouldn't be
1119 * any, but C is fun!
1121 memset(&ent, 0, sizeof(ent));
1124 while (likely(i < nr)) {
1125 ret = aio_read_evt(ctx, &ent);
1126 if (unlikely(ret <= 0))
1129 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1130 ent.data, ent.obj, ent.res, ent.res2);
1132 /* Could we split the check in two? */
1134 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1135 dprintk("aio: lost an event due to EFAULT.\n");
1140 /* Good, event copied to userland, update counts. */
1152 /* racey check, but it gets redone */
1153 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1155 aio_run_all_iocbs(ctx);
1163 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1166 set_timeout(start_jiffies, &to, &ts);
1169 while (likely(i < nr)) {
1170 add_wait_queue_exclusive(&ctx->wait, &wait);
1172 set_task_state(tsk, TASK_INTERRUPTIBLE);
1173 ret = aio_read_evt(ctx, &ent);
1179 if (to.timed_out) /* Only check after read evt */
1182 if (signal_pending(tsk)) {
1186 /*ret = aio_read_evt(ctx, &ent);*/
1189 set_task_state(tsk, TASK_RUNNING);
1190 remove_wait_queue(&ctx->wait, &wait);
1192 if (unlikely(ret <= 0))
1196 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1197 dprintk("aio: lost an event due to EFAULT.\n");
1201 /* Good, event copied to userland, update counts. */
1212 /* Take an ioctx and remove it from the list of ioctx's. Protects
1213 * against races with itself via ->dead.
1215 static void io_destroy(struct kioctx *ioctx)
1217 struct mm_struct *mm = current->mm;
1218 struct kioctx **tmp;
1221 /* delete the entry from the list is someone else hasn't already */
1222 write_lock(&mm->ioctx_list_lock);
1223 was_dead = ioctx->dead;
1225 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1226 tmp = &(*tmp)->next)
1230 write_unlock(&mm->ioctx_list_lock);
1232 dprintk("aio_release(%p)\n", ioctx);
1233 if (likely(!was_dead))
1234 put_ioctx(ioctx); /* twice for the list */
1236 aio_cancel_all(ioctx);
1237 wait_for_all_aios(ioctx);
1238 put_ioctx(ioctx); /* once for the lookup */
1242 * Create an aio_context capable of receiving at least nr_events.
1243 * ctxp must not point to an aio_context that already exists, and
1244 * must be initialized to 0 prior to the call. On successful
1245 * creation of the aio_context, *ctxp is filled in with the resulting
1246 * handle. May fail with -EINVAL if *ctxp is not initialized,
1247 * if the specified nr_events exceeds internal limits. May fail
1248 * with -EAGAIN if the specified nr_events exceeds the user's limit
1249 * of available events. May fail with -ENOMEM if insufficient kernel
1250 * resources are available. May fail with -EFAULT if an invalid
1251 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1254 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1256 struct kioctx *ioctx = NULL;
1260 ret = get_user(ctx, ctxp);
1265 if (unlikely(ctx || nr_events == 0)) {
1266 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1271 ioctx = ioctx_alloc(nr_events);
1272 ret = PTR_ERR(ioctx);
1273 if (!IS_ERR(ioctx)) {
1274 ret = put_user(ioctx->user_id, ctxp);
1278 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1287 * Destroy the aio_context specified. May cancel any outstanding
1288 * AIOs and block on completion. Will fail with -ENOSYS if not
1289 * implemented. May fail with -EFAULT if the context pointed to
1292 asmlinkage long sys_io_destroy(aio_context_t ctx)
1294 struct kioctx *ioctx = lookup_ioctx(ctx);
1295 if (likely(NULL != ioctx)) {
1299 pr_debug("EINVAL: io_destroy: invalid context id\n");
1304 * aio_p{read,write} are the default ki_retry methods for
1305 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially
1306 * multiple calls to f_op->aio_read(). They loop around partial progress
1307 * instead of returning -EIOCBRETRY because they don't have the means to call
1310 static ssize_t aio_pread(struct kiocb *iocb)
1312 struct file *file = iocb->ki_filp;
1313 struct address_space *mapping = file->f_mapping;
1314 struct inode *inode = mapping->host;
1318 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1319 iocb->ki_left, iocb->ki_pos);
1321 * Can't just depend on iocb->ki_left to determine
1322 * whether we are done. This may have been a short read.
1325 iocb->ki_buf += ret;
1326 iocb->ki_left -= ret;
1330 * For pipes and sockets we return once we have some data; for
1331 * regular files we retry till we complete the entire read or
1332 * find that we can't read any more data (e.g short reads).
1334 } while (ret > 0 && iocb->ki_left > 0 &&
1335 !S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode));
1337 /* This means we must have transferred all that we could */
1338 /* No need to retry anymore */
1339 if ((ret == 0) || (iocb->ki_left == 0))
1340 ret = iocb->ki_nbytes - iocb->ki_left;
1345 /* see aio_pread() */
1346 static ssize_t aio_pwrite(struct kiocb *iocb)
1348 struct file *file = iocb->ki_filp;
1352 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1353 iocb->ki_left, iocb->ki_pos);
1355 iocb->ki_buf += ret;
1356 iocb->ki_left -= ret;
1358 } while (ret > 0 && iocb->ki_left > 0);
1360 if ((ret == 0) || (iocb->ki_left == 0))
1361 ret = iocb->ki_nbytes - iocb->ki_left;
1366 static ssize_t aio_fdsync(struct kiocb *iocb)
1368 struct file *file = iocb->ki_filp;
1369 ssize_t ret = -EINVAL;
1371 if (file->f_op->aio_fsync)
1372 ret = file->f_op->aio_fsync(iocb, 1);
1376 static ssize_t aio_fsync(struct kiocb *iocb)
1378 struct file *file = iocb->ki_filp;
1379 ssize_t ret = -EINVAL;
1381 if (file->f_op->aio_fsync)
1382 ret = file->f_op->aio_fsync(iocb, 0);
1388 * Performs the initial checks and aio retry method
1389 * setup for the kiocb at the time of io submission.
1391 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1393 struct file *file = kiocb->ki_filp;
1396 switch (kiocb->ki_opcode) {
1397 case IOCB_CMD_PREAD:
1399 if (unlikely(!(file->f_mode & FMODE_READ)))
1402 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1405 ret = security_file_permission(file, MAY_READ);
1409 if (file->f_op->aio_read)
1410 kiocb->ki_retry = aio_pread;
1412 case IOCB_CMD_PWRITE:
1414 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1417 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1420 ret = security_file_permission(file, MAY_WRITE);
1424 if (file->f_op->aio_write)
1425 kiocb->ki_retry = aio_pwrite;
1427 case IOCB_CMD_FDSYNC:
1429 if (file->f_op->aio_fsync)
1430 kiocb->ki_retry = aio_fdsync;
1432 case IOCB_CMD_FSYNC:
1434 if (file->f_op->aio_fsync)
1435 kiocb->ki_retry = aio_fsync;
1438 dprintk("EINVAL: io_submit: no operation provided\n");
1442 if (!kiocb->ki_retry)
1449 * aio_wake_function:
1450 * wait queue callback function for aio notification,
1451 * Simply triggers a retry of the operation via kick_iocb.
1453 * This callback is specified in the wait queue entry in
1454 * a kiocb (current->io_wait points to this wait queue
1455 * entry when an aio operation executes; it is used
1456 * instead of a synchronous wait when an i/o blocking
1457 * condition is encountered during aio).
1460 * This routine is executed with the wait queue lock held.
1461 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1462 * the ioctx lock inside the wait queue lock. This is safe
1463 * because this callback isn't used for wait queues which
1464 * are nested inside ioctx lock (i.e. ctx->wait)
1466 static int aio_wake_function(wait_queue_t *wait, unsigned mode,
1467 int sync, void *key)
1469 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1471 list_del_init(&wait->task_list);
1476 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1483 /* enforce forwards compatibility on users */
1484 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1485 iocb->aio_reserved3)) {
1486 pr_debug("EINVAL: io_submit: reserve field set\n");
1490 /* prevent overflows */
1492 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1493 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1494 ((ssize_t)iocb->aio_nbytes < 0)
1496 pr_debug("EINVAL: io_submit: overflow check\n");
1500 file = fget(iocb->aio_fildes);
1501 if (unlikely(!file))
1504 req = aio_get_req(ctx); /* returns with 2 references to req */
1505 if (unlikely(!req)) {
1510 req->ki_filp = file;
1511 ret = put_user(req->ki_key, &user_iocb->aio_key);
1512 if (unlikely(ret)) {
1513 dprintk("EFAULT: aio_key\n");
1517 req->ki_obj.user = user_iocb;
1518 req->ki_user_data = iocb->aio_data;
1519 req->ki_pos = iocb->aio_offset;
1521 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1522 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1523 req->ki_opcode = iocb->aio_lio_opcode;
1524 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1525 INIT_LIST_HEAD(&req->ki_wait.task_list);
1526 req->ki_retried = 0;
1528 ret = aio_setup_iocb(req);
1533 spin_lock_irq(&ctx->ctx_lock);
1535 if (!list_empty(&ctx->run_list)) {
1536 /* drain the run list */
1537 while (__aio_run_iocbs(ctx))
1540 spin_unlock_irq(&ctx->ctx_lock);
1541 aio_put_req(req); /* drop extra ref to req */
1545 aio_put_req(req); /* drop extra ref to req */
1546 aio_put_req(req); /* drop i/o ref to req */
1551 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1552 * the number of iocbs queued. May return -EINVAL if the aio_context
1553 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1554 * *iocbpp[0] is not properly initialized, if the operation specified
1555 * is invalid for the file descriptor in the iocb. May fail with
1556 * -EFAULT if any of the data structures point to invalid data. May
1557 * fail with -EBADF if the file descriptor specified in the first
1558 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1559 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1560 * fail with -ENOSYS if not implemented.
1562 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1563 struct iocb __user * __user *iocbpp)
1569 if (unlikely(nr < 0))
1572 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1575 ctx = lookup_ioctx(ctx_id);
1576 if (unlikely(!ctx)) {
1577 pr_debug("EINVAL: io_submit: invalid context id\n");
1582 * AKPM: should this return a partial result if some of the IOs were
1583 * successfully submitted?
1585 for (i=0; i<nr; i++) {
1586 struct iocb __user *user_iocb;
1589 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1594 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1599 ret = io_submit_one(ctx, user_iocb, &tmp);
1609 * Finds a given iocb for cancellation.
1611 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1614 struct list_head *pos;
1616 assert_spin_locked(&ctx->ctx_lock);
1618 /* TODO: use a hash or array, this sucks. */
1619 list_for_each(pos, &ctx->active_reqs) {
1620 struct kiocb *kiocb = list_kiocb(pos);
1621 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1628 * Attempts to cancel an iocb previously passed to io_submit. If
1629 * the operation is successfully cancelled, the resulting event is
1630 * copied into the memory pointed to by result without being placed
1631 * into the completion queue and 0 is returned. May fail with
1632 * -EFAULT if any of the data structures pointed to are invalid.
1633 * May fail with -EINVAL if aio_context specified by ctx_id is
1634 * invalid. May fail with -EAGAIN if the iocb specified was not
1635 * cancelled. Will fail with -ENOSYS if not implemented.
1637 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1638 struct io_event __user *result)
1640 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1642 struct kiocb *kiocb;
1646 ret = get_user(key, &iocb->aio_key);
1650 ctx = lookup_ioctx(ctx_id);
1654 spin_lock_irq(&ctx->ctx_lock);
1656 kiocb = lookup_kiocb(ctx, iocb, key);
1657 if (kiocb && kiocb->ki_cancel) {
1658 cancel = kiocb->ki_cancel;
1660 kiocbSetCancelled(kiocb);
1663 spin_unlock_irq(&ctx->ctx_lock);
1665 if (NULL != cancel) {
1666 struct io_event tmp;
1667 pr_debug("calling cancel\n");
1668 memset(&tmp, 0, sizeof(tmp));
1669 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1670 tmp.data = kiocb->ki_user_data;
1671 ret = cancel(kiocb, &tmp);
1673 /* Cancellation succeeded -- copy the result
1674 * into the user's buffer.
1676 if (copy_to_user(result, &tmp, sizeof(tmp)))
1688 * Attempts to read at least min_nr events and up to nr events from
1689 * the completion queue for the aio_context specified by ctx_id. May
1690 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1691 * if nr is out of range, if when is out of range. May fail with
1692 * -EFAULT if any of the memory specified to is invalid. May return
1693 * 0 or < min_nr if no events are available and the timeout specified
1694 * by when has elapsed, where when == NULL specifies an infinite
1695 * timeout. Note that the timeout pointed to by when is relative and
1696 * will be updated if not NULL and the operation blocks. Will fail
1697 * with -ENOSYS if not implemented.
1699 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1702 struct io_event __user *events,
1703 struct timespec __user *timeout)
1705 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1708 if (likely(ioctx)) {
1709 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1710 ret = read_events(ioctx, min_nr, nr, events, timeout);
1717 __initcall(aio_setup);
1719 EXPORT_SYMBOL(aio_complete);
1720 EXPORT_SYMBOL(aio_put_req);
1721 EXPORT_SYMBOL(wait_on_sync_kiocb);