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 = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
126 if (!info->ring_pages)
128 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
131 info->mmap_size = nr_pages * PAGE_SIZE;
132 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
133 down_write(&ctx->mm->mmap_sem);
134 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
135 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
137 if (IS_ERR((void *)info->mmap_base)) {
138 up_write(&ctx->mm->mmap_sem);
139 printk("mmap err: %ld\n", -info->mmap_base);
145 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
146 info->nr_pages = get_user_pages(current, ctx->mm,
147 info->mmap_base, nr_pages,
148 1, 0, info->ring_pages, NULL);
149 up_write(&ctx->mm->mmap_sem);
151 if (unlikely(info->nr_pages != nr_pages)) {
156 ctx->user_id = info->mmap_base;
158 info->nr = nr_events; /* trusted copy */
160 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
161 ring->nr = nr_events; /* user copy */
162 ring->id = ctx->user_id;
163 ring->head = ring->tail = 0;
164 ring->magic = AIO_RING_MAGIC;
165 ring->compat_features = AIO_RING_COMPAT_FEATURES;
166 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
167 ring->header_length = sizeof(struct aio_ring);
168 kunmap_atomic(ring, KM_USER0);
174 /* aio_ring_event: returns a pointer to the event at the given index from
175 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
177 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
178 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
179 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
181 #define aio_ring_event(info, nr, km) ({ \
182 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
183 struct io_event *__event; \
184 __event = kmap_atomic( \
185 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
186 __event += pos % AIO_EVENTS_PER_PAGE; \
190 #define put_aio_ring_event(event, km) do { \
191 struct io_event *__event = (event); \
193 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
197 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
199 static struct kioctx *ioctx_alloc(unsigned nr_events)
201 struct mm_struct *mm;
204 /* Prevent overflows */
205 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
206 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
207 pr_debug("ENOMEM: nr_events too high\n");
208 return ERR_PTR(-EINVAL);
211 if ((unsigned long)nr_events > aio_max_nr)
212 return ERR_PTR(-EAGAIN);
214 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
216 return ERR_PTR(-ENOMEM);
218 memset(ctx, 0, sizeof(*ctx));
219 ctx->max_reqs = nr_events;
220 mm = ctx->mm = current->mm;
221 atomic_inc(&mm->mm_count);
223 atomic_set(&ctx->users, 1);
224 spin_lock_init(&ctx->ctx_lock);
225 spin_lock_init(&ctx->ring_info.ring_lock);
226 init_waitqueue_head(&ctx->wait);
228 INIT_LIST_HEAD(&ctx->active_reqs);
229 INIT_LIST_HEAD(&ctx->run_list);
230 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
232 if (aio_setup_ring(ctx) < 0)
235 /* limit the number of system wide aios */
236 spin_lock(&aio_nr_lock);
237 if (aio_nr + ctx->max_reqs > aio_max_nr ||
238 aio_nr + ctx->max_reqs < aio_nr)
241 aio_nr += ctx->max_reqs;
242 spin_unlock(&aio_nr_lock);
243 if (ctx->max_reqs == 0)
246 /* now link into global list. kludge. FIXME */
247 write_lock(&mm->ioctx_list_lock);
248 ctx->next = mm->ioctx_list;
249 mm->ioctx_list = ctx;
250 write_unlock(&mm->ioctx_list_lock);
252 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
253 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
258 return ERR_PTR(-EAGAIN);
262 kmem_cache_free(kioctx_cachep, ctx);
263 ctx = ERR_PTR(-ENOMEM);
265 dprintk("aio: error allocating ioctx %p\n", ctx);
270 * Cancels all outstanding aio requests on an aio context. Used
271 * when the processes owning a context have all exited to encourage
272 * the rapid destruction of the kioctx.
274 static void aio_cancel_all(struct kioctx *ctx)
276 int (*cancel)(struct kiocb *, struct io_event *);
278 spin_lock_irq(&ctx->ctx_lock);
280 while (!list_empty(&ctx->active_reqs)) {
281 struct list_head *pos = ctx->active_reqs.next;
282 struct kiocb *iocb = list_kiocb(pos);
283 list_del_init(&iocb->ki_list);
284 cancel = iocb->ki_cancel;
285 kiocbSetCancelled(iocb);
288 spin_unlock_irq(&ctx->ctx_lock);
290 spin_lock_irq(&ctx->ctx_lock);
293 spin_unlock_irq(&ctx->ctx_lock);
296 static void wait_for_all_aios(struct kioctx *ctx)
298 struct task_struct *tsk = current;
299 DECLARE_WAITQUEUE(wait, tsk);
301 if (!ctx->reqs_active)
304 add_wait_queue(&ctx->wait, &wait);
305 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
306 while (ctx->reqs_active) {
308 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
310 __set_task_state(tsk, TASK_RUNNING);
311 remove_wait_queue(&ctx->wait, &wait);
314 /* wait_on_sync_kiocb:
315 * Waits on the given sync kiocb to complete.
317 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
319 while (iocb->ki_users) {
320 set_current_state(TASK_UNINTERRUPTIBLE);
325 __set_current_state(TASK_RUNNING);
326 return iocb->ki_user_data;
329 /* exit_aio: called when the last user of mm goes away. At this point,
330 * there is no way for any new requests to be submited or any of the
331 * io_* syscalls to be called on the context. However, there may be
332 * outstanding requests which hold references to the context; as they
333 * go away, they will call put_ioctx and release any pinned memory
334 * associated with the request (held via struct page * references).
336 void fastcall exit_aio(struct mm_struct *mm)
338 struct kioctx *ctx = mm->ioctx_list;
339 mm->ioctx_list = NULL;
341 struct kioctx *next = ctx->next;
345 wait_for_all_aios(ctx);
347 * this is an overkill, but ensures we don't leave
348 * the ctx on the aio_wq
350 flush_workqueue(aio_wq);
352 if (1 != atomic_read(&ctx->users))
354 "exit_aio:ioctx still alive: %d %d %d\n",
355 atomic_read(&ctx->users), ctx->dead,
363 * Called when the last user of an aio context has gone away,
364 * and the struct needs to be freed.
366 void fastcall __put_ioctx(struct kioctx *ctx)
368 unsigned nr_events = ctx->max_reqs;
370 if (unlikely(ctx->reqs_active))
373 cancel_delayed_work(&ctx->wq);
374 flush_workqueue(aio_wq);
378 pr_debug("__put_ioctx: freeing %p\n", ctx);
379 kmem_cache_free(kioctx_cachep, ctx);
382 spin_lock(&aio_nr_lock);
383 BUG_ON(aio_nr - nr_events > aio_nr);
385 spin_unlock(&aio_nr_lock);
390 * Allocate a slot for an aio request. Increments the users count
391 * of the kioctx so that the kioctx stays around until all requests are
392 * complete. Returns NULL if no requests are free.
394 * Returns with kiocb->users set to 2. The io submit code path holds
395 * an extra reference while submitting the i/o.
396 * This prevents races between the aio code path referencing the
397 * req (after submitting it) and aio_complete() freeing the req.
399 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
400 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
402 struct kiocb *req = NULL;
403 struct aio_ring *ring;
406 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
414 req->ki_cancel = NULL;
415 req->ki_retry = NULL;
418 INIT_LIST_HEAD(&req->ki_run_list);
420 /* Check if the completion queue has enough free space to
421 * accept an event from this io.
423 spin_lock_irq(&ctx->ctx_lock);
424 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
425 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
426 list_add(&req->ki_list, &ctx->active_reqs);
431 kunmap_atomic(ring, KM_USER0);
432 spin_unlock_irq(&ctx->ctx_lock);
435 kmem_cache_free(kiocb_cachep, req);
442 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
445 /* Handle a potential starvation case -- should be exceedingly rare as
446 * requests will be stuck on fput_head only if the aio_fput_routine is
447 * delayed and the requests were the last user of the struct file.
449 req = __aio_get_req(ctx);
450 if (unlikely(NULL == req)) {
451 aio_fput_routine(NULL);
452 req = __aio_get_req(ctx);
457 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
459 assert_spin_locked(&ctx->ctx_lock);
463 kmem_cache_free(kiocb_cachep, req);
466 if (unlikely(!ctx->reqs_active && ctx->dead))
470 static void aio_fput_routine(void *data)
472 spin_lock_irq(&fput_lock);
473 while (likely(!list_empty(&fput_head))) {
474 struct kiocb *req = list_kiocb(fput_head.next);
475 struct kioctx *ctx = req->ki_ctx;
477 list_del(&req->ki_list);
478 spin_unlock_irq(&fput_lock);
480 /* Complete the fput */
481 __fput(req->ki_filp);
483 /* Link the iocb into the context's free list */
484 spin_lock_irq(&ctx->ctx_lock);
485 really_put_req(ctx, req);
486 spin_unlock_irq(&ctx->ctx_lock);
489 spin_lock_irq(&fput_lock);
491 spin_unlock_irq(&fput_lock);
495 * Returns true if this put was the last user of the request.
497 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
499 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
500 req, atomic_read(&req->ki_filp->f_count));
502 assert_spin_locked(&ctx->ctx_lock);
505 if (unlikely(req->ki_users < 0))
507 if (likely(req->ki_users))
509 list_del(&req->ki_list); /* remove from active_reqs */
510 req->ki_cancel = NULL;
511 req->ki_retry = NULL;
513 /* Must be done under the lock to serialise against cancellation.
514 * Call this aio_fput as it duplicates fput via the fput_work.
516 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
518 spin_lock(&fput_lock);
519 list_add(&req->ki_list, &fput_head);
520 spin_unlock(&fput_lock);
521 queue_work(aio_wq, &fput_work);
523 really_put_req(ctx, req);
528 * Returns true if this put was the last user of the kiocb,
529 * false if the request is still in use.
531 int fastcall aio_put_req(struct kiocb *req)
533 struct kioctx *ctx = req->ki_ctx;
535 spin_lock_irq(&ctx->ctx_lock);
536 ret = __aio_put_req(ctx, req);
537 spin_unlock_irq(&ctx->ctx_lock);
543 /* Lookup an ioctx id. ioctx_list is lockless for reads.
544 * FIXME: this is O(n) and is only suitable for development.
546 struct kioctx *lookup_ioctx(unsigned long ctx_id)
548 struct kioctx *ioctx;
549 struct mm_struct *mm;
552 read_lock(&mm->ioctx_list_lock);
553 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
554 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
558 read_unlock(&mm->ioctx_list_lock);
565 * Makes the calling kernel thread take on the specified
567 * Called by the retry thread execute retries within the
568 * iocb issuer's mm context, so that copy_from/to_user
569 * operations work seamlessly for aio.
570 * (Note: this routine is intended to be called only
571 * from a kernel thread context)
573 static void use_mm(struct mm_struct *mm)
575 struct mm_struct *active_mm;
576 struct task_struct *tsk = current;
579 tsk->flags |= PF_BORROWED_MM;
580 active_mm = tsk->active_mm;
581 atomic_inc(&mm->mm_count);
585 * Note that on UML this *requires* PF_BORROWED_MM to be set, otherwise
586 * it won't work. Update it accordingly if you change it here
588 activate_mm(active_mm, mm);
596 * Reverses the effect of use_mm, i.e. releases the
597 * specified mm context which was earlier taken on
598 * by the calling kernel thread
599 * (Note: this routine is intended to be called only
600 * from a kernel thread context)
602 * Comments: Called with ctx->ctx_lock held. This nests
603 * task_lock instead ctx_lock.
605 static void unuse_mm(struct mm_struct *mm)
607 struct task_struct *tsk = current;
610 tsk->flags &= ~PF_BORROWED_MM;
612 /* active_mm is still 'mm' */
613 enter_lazy_tlb(mm, tsk);
618 * Queue up a kiocb to be retried. Assumes that the kiocb
619 * has already been marked as kicked, and places it on
620 * the retry run list for the corresponding ioctx, if it
621 * isn't already queued. Returns 1 if it actually queued
622 * the kiocb (to tell the caller to activate the work
623 * queue to process it), or 0, if it found that it was
626 static inline int __queue_kicked_iocb(struct kiocb *iocb)
628 struct kioctx *ctx = iocb->ki_ctx;
630 assert_spin_locked(&ctx->ctx_lock);
632 if (list_empty(&iocb->ki_run_list)) {
633 list_add_tail(&iocb->ki_run_list,
641 * This is the core aio execution routine. It is
642 * invoked both for initial i/o submission and
643 * subsequent retries via the aio_kick_handler.
644 * Expects to be invoked with iocb->ki_ctx->lock
645 * already held. The lock is released and reaquired
646 * as needed during processing.
648 * Calls the iocb retry method (already setup for the
649 * iocb on initial submission) for operation specific
650 * handling, but takes care of most of common retry
651 * execution details for a given iocb. The retry method
652 * needs to be non-blocking as far as possible, to avoid
653 * holding up other iocbs waiting to be serviced by the
654 * retry kernel thread.
656 * The trickier parts in this code have to do with
657 * ensuring that only one retry instance is in progress
658 * for a given iocb at any time. Providing that guarantee
659 * simplifies the coding of individual aio operations as
660 * it avoids various potential races.
662 static ssize_t aio_run_iocb(struct kiocb *iocb)
664 struct kioctx *ctx = iocb->ki_ctx;
665 ssize_t (*retry)(struct kiocb *);
668 if (iocb->ki_retried++ > 1024*1024) {
669 printk("Maximal retry count. Bytes done %Zd\n",
670 iocb->ki_nbytes - iocb->ki_left);
674 if (!(iocb->ki_retried & 0xff)) {
675 pr_debug("%ld retry: %d of %d\n", iocb->ki_retried,
676 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
679 if (!(retry = iocb->ki_retry)) {
680 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
685 * We don't want the next retry iteration for this
686 * operation to start until this one has returned and
687 * updated the iocb state. However, wait_queue functions
688 * can trigger a kick_iocb from interrupt context in the
689 * meantime, indicating that data is available for the next
690 * iteration. We want to remember that and enable the
691 * next retry iteration _after_ we are through with
694 * So, in order to be able to register a "kick", but
695 * prevent it from being queued now, we clear the kick
696 * flag, but make the kick code *think* that the iocb is
697 * still on the run list until we are actually done.
698 * When we are done with this iteration, we check if
699 * the iocb was kicked in the meantime and if so, queue
703 kiocbClearKicked(iocb);
706 * This is so that aio_complete knows it doesn't need to
707 * pull the iocb off the run list (We can't just call
708 * INIT_LIST_HEAD because we don't want a kick_iocb to
709 * queue this on the run list yet)
711 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
712 spin_unlock_irq(&ctx->ctx_lock);
714 /* Quit retrying if the i/o has been cancelled */
715 if (kiocbIsCancelled(iocb)) {
717 aio_complete(iocb, ret, 0);
718 /* must not access the iocb after this */
723 * Now we are all set to call the retry method in async
724 * context. By setting this thread's io_wait context
725 * to point to the wait queue entry inside the currently
726 * running iocb for the duration of the retry, we ensure
727 * that async notification wakeups are queued by the
728 * operation instead of blocking waits, and when notified,
729 * cause the iocb to be kicked for continuation (through
730 * the aio_wake_function callback).
732 BUG_ON(current->io_wait != NULL);
733 current->io_wait = &iocb->ki_wait;
735 current->io_wait = NULL;
737 if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
738 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
739 aio_complete(iocb, ret, 0);
742 spin_lock_irq(&ctx->ctx_lock);
744 if (-EIOCBRETRY == ret) {
746 * OK, now that we are done with this iteration
747 * and know that there is more left to go,
748 * this is where we let go so that a subsequent
749 * "kick" can start the next iteration
752 /* will make __queue_kicked_iocb succeed from here on */
753 INIT_LIST_HEAD(&iocb->ki_run_list);
754 /* we must queue the next iteration ourselves, if it
755 * has already been kicked */
756 if (kiocbIsKicked(iocb)) {
757 __queue_kicked_iocb(iocb);
760 * __queue_kicked_iocb will always return 1 here, because
761 * iocb->ki_run_list is empty at this point so it should
762 * be safe to unconditionally queue the context into the
773 * Process all pending retries queued on the ioctx
775 * Assumes it is operating within the aio issuer's mm
778 static int __aio_run_iocbs(struct kioctx *ctx)
783 assert_spin_locked(&ctx->ctx_lock);
785 list_splice_init(&ctx->run_list, &run_list);
786 while (!list_empty(&run_list)) {
787 iocb = list_entry(run_list.next, struct kiocb,
789 list_del(&iocb->ki_run_list);
791 * Hold an extra reference while retrying i/o.
793 iocb->ki_users++; /* grab extra reference */
795 if (__aio_put_req(ctx, iocb)) /* drop extra ref */
798 if (!list_empty(&ctx->run_list))
803 static void aio_queue_work(struct kioctx * ctx)
805 unsigned long timeout;
807 * if someone is waiting, get the work started right
808 * away, otherwise, use a longer delay
811 if (waitqueue_active(&ctx->wait))
815 queue_delayed_work(aio_wq, &ctx->wq, timeout);
821 * Process all pending retries queued on the ioctx
823 * Assumes it is operating within the aio issuer's mm
826 static inline void aio_run_iocbs(struct kioctx *ctx)
830 spin_lock_irq(&ctx->ctx_lock);
832 requeue = __aio_run_iocbs(ctx);
833 spin_unlock_irq(&ctx->ctx_lock);
839 * just like aio_run_iocbs, but keeps running them until
840 * the list stays empty
842 static inline void aio_run_all_iocbs(struct kioctx *ctx)
844 spin_lock_irq(&ctx->ctx_lock);
845 while (__aio_run_iocbs(ctx))
847 spin_unlock_irq(&ctx->ctx_lock);
852 * Work queue handler triggered to process pending
853 * retries on an ioctx. Takes on the aio issuer's
854 * mm context before running the iocbs, so that
855 * copy_xxx_user operates on the issuer's address
857 * Run on aiod's context.
859 static void aio_kick_handler(void *data)
861 struct kioctx *ctx = data;
862 mm_segment_t oldfs = get_fs();
867 spin_lock_irq(&ctx->ctx_lock);
868 requeue =__aio_run_iocbs(ctx);
870 spin_unlock_irq(&ctx->ctx_lock);
873 * we're in a worker thread already, don't use queue_delayed_work,
876 queue_work(aio_wq, &ctx->wq);
881 * Called by kick_iocb to queue the kiocb for retry
882 * and if required activate the aio work queue to process
885 static void try_queue_kicked_iocb(struct kiocb *iocb)
887 struct kioctx *ctx = iocb->ki_ctx;
891 /* We're supposed to be the only path putting the iocb back on the run
892 * list. If we find that the iocb is *back* on a wait queue already
893 * than retry has happened before we could queue the iocb. This also
894 * means that the retry could have completed and freed our iocb, no
896 BUG_ON((!list_empty(&iocb->ki_wait.task_list)));
898 spin_lock_irqsave(&ctx->ctx_lock, flags);
899 /* set this inside the lock so that we can't race with aio_run_iocb()
900 * testing it and putting the iocb on the run list under the lock */
901 if (!kiocbTryKick(iocb))
902 run = __queue_kicked_iocb(iocb);
903 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
910 * Called typically from a wait queue callback context
911 * (aio_wake_function) to trigger a retry of the iocb.
912 * The retry is usually executed by aio workqueue
913 * threads (See aio_kick_handler).
915 void fastcall kick_iocb(struct kiocb *iocb)
917 /* sync iocbs are easy: they can only ever be executing from a
919 if (is_sync_kiocb(iocb)) {
920 kiocbSetKicked(iocb);
921 wake_up_process(iocb->ki_obj.tsk);
925 try_queue_kicked_iocb(iocb);
927 EXPORT_SYMBOL(kick_iocb);
930 * Called when the io request on the given iocb is complete.
931 * Returns true if this is the last user of the request. The
932 * only other user of the request can be the cancellation code.
934 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
936 struct kioctx *ctx = iocb->ki_ctx;
937 struct aio_ring_info *info;
938 struct aio_ring *ring;
939 struct io_event *event;
945 * Special case handling for sync iocbs:
946 * - events go directly into the iocb for fast handling
947 * - the sync task with the iocb in its stack holds the single iocb
948 * ref, no other paths have a way to get another ref
949 * - the sync task helpfully left a reference to itself in the iocb
951 if (is_sync_kiocb(iocb)) {
952 BUG_ON(iocb->ki_users != 1);
953 iocb->ki_user_data = res;
955 wake_up_process(iocb->ki_obj.tsk);
959 info = &ctx->ring_info;
961 /* add a completion event to the ring buffer.
962 * must be done holding ctx->ctx_lock to prevent
963 * other code from messing with the tail
964 * pointer since we might be called from irq
967 spin_lock_irqsave(&ctx->ctx_lock, flags);
969 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
970 list_del_init(&iocb->ki_run_list);
973 * cancelled requests don't get events, userland was given one
974 * when the event got cancelled.
976 if (kiocbIsCancelled(iocb))
979 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
982 event = aio_ring_event(info, tail, KM_IRQ0);
983 if (++tail >= info->nr)
986 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
987 event->data = iocb->ki_user_data;
991 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
992 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
995 /* after flagging the request as done, we
996 * must never even look at it again
998 smp_wmb(); /* make event visible before updating tail */
1003 put_aio_ring_event(event, KM_IRQ0);
1004 kunmap_atomic(ring, KM_IRQ1);
1006 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1008 pr_debug("%ld retries: %d of %d\n", iocb->ki_retried,
1009 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes);
1011 /* everything turned out well, dispose of the aiocb. */
1012 ret = __aio_put_req(ctx, iocb);
1014 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1016 if (waitqueue_active(&ctx->wait))
1017 wake_up(&ctx->wait);
1026 * Pull an event off of the ioctx's event ring. Returns the number of
1027 * events fetched (0 or 1 ;-)
1028 * FIXME: make this use cmpxchg.
1029 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1031 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1033 struct aio_ring_info *info = &ioctx->ring_info;
1034 struct aio_ring *ring;
1038 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1039 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1040 (unsigned long)ring->head, (unsigned long)ring->tail,
1041 (unsigned long)ring->nr);
1043 if (ring->head == ring->tail)
1046 spin_lock(&info->ring_lock);
1048 head = ring->head % info->nr;
1049 if (head != ring->tail) {
1050 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1052 head = (head + 1) % info->nr;
1053 smp_mb(); /* finish reading the event before updatng the head */
1056 put_aio_ring_event(evp, KM_USER1);
1058 spin_unlock(&info->ring_lock);
1061 kunmap_atomic(ring, KM_USER0);
1062 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1063 (unsigned long)ring->head, (unsigned long)ring->tail);
1067 struct aio_timeout {
1068 struct timer_list timer;
1070 struct task_struct *p;
1073 static void timeout_func(unsigned long data)
1075 struct aio_timeout *to = (struct aio_timeout *)data;
1078 wake_up_process(to->p);
1081 static inline void init_timeout(struct aio_timeout *to)
1083 init_timer(&to->timer);
1084 to->timer.data = (unsigned long)to;
1085 to->timer.function = timeout_func;
1090 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1091 const struct timespec *ts)
1093 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1094 if (time_after(to->timer.expires, jiffies))
1095 add_timer(&to->timer);
1100 static inline void clear_timeout(struct aio_timeout *to)
1102 del_singleshot_timer_sync(&to->timer);
1105 static int read_events(struct kioctx *ctx,
1106 long min_nr, long nr,
1107 struct io_event __user *event,
1108 struct timespec __user *timeout)
1110 long start_jiffies = jiffies;
1111 struct task_struct *tsk = current;
1112 DECLARE_WAITQUEUE(wait, tsk);
1115 struct io_event ent;
1116 struct aio_timeout to;
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 */
1183 if (signal_pending(tsk)) {
1187 /*ret = aio_read_evt(ctx, &ent);*/
1190 set_task_state(tsk, TASK_RUNNING);
1191 remove_wait_queue(&ctx->wait, &wait);
1193 if (unlikely(ret <= 0))
1197 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1198 dprintk("aio: lost an event due to EFAULT.\n");
1202 /* Good, event copied to userland, update counts. */
1213 /* Take an ioctx and remove it from the list of ioctx's. Protects
1214 * against races with itself via ->dead.
1216 static void io_destroy(struct kioctx *ioctx)
1218 struct mm_struct *mm = current->mm;
1219 struct kioctx **tmp;
1222 /* delete the entry from the list is someone else hasn't already */
1223 write_lock(&mm->ioctx_list_lock);
1224 was_dead = ioctx->dead;
1226 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1227 tmp = &(*tmp)->next)
1231 write_unlock(&mm->ioctx_list_lock);
1233 dprintk("aio_release(%p)\n", ioctx);
1234 if (likely(!was_dead))
1235 put_ioctx(ioctx); /* twice for the list */
1237 aio_cancel_all(ioctx);
1238 wait_for_all_aios(ioctx);
1239 put_ioctx(ioctx); /* once for the lookup */
1243 * Create an aio_context capable of receiving at least nr_events.
1244 * ctxp must not point to an aio_context that already exists, and
1245 * must be initialized to 0 prior to the call. On successful
1246 * creation of the aio_context, *ctxp is filled in with the resulting
1247 * handle. May fail with -EINVAL if *ctxp is not initialized,
1248 * if the specified nr_events exceeds internal limits. May fail
1249 * with -EAGAIN if the specified nr_events exceeds the user's limit
1250 * of available events. May fail with -ENOMEM if insufficient kernel
1251 * resources are available. May fail with -EFAULT if an invalid
1252 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1255 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1257 struct kioctx *ioctx = NULL;
1261 ret = get_user(ctx, ctxp);
1266 if (unlikely(ctx || nr_events == 0)) {
1267 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1272 ioctx = ioctx_alloc(nr_events);
1273 ret = PTR_ERR(ioctx);
1274 if (!IS_ERR(ioctx)) {
1275 ret = put_user(ioctx->user_id, ctxp);
1279 get_ioctx(ioctx); /* io_destroy() expects us to hold a ref */
1288 * Destroy the aio_context specified. May cancel any outstanding
1289 * AIOs and block on completion. Will fail with -ENOSYS if not
1290 * implemented. May fail with -EFAULT if the context pointed to
1293 asmlinkage long sys_io_destroy(aio_context_t ctx)
1295 struct kioctx *ioctx = lookup_ioctx(ctx);
1296 if (likely(NULL != ioctx)) {
1300 pr_debug("EINVAL: io_destroy: invalid context id\n");
1305 * aio_p{read,write} are the default ki_retry methods for
1306 * IO_CMD_P{READ,WRITE}. They maintains kiocb retry state around potentially
1307 * multiple calls to f_op->aio_read(). They loop around partial progress
1308 * instead of returning -EIOCBRETRY because they don't have the means to call
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;
1319 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1320 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;
1331 * For pipes and sockets we return once we have some data; for
1332 * regular files we retry till we complete the entire read or
1333 * find that we can't read any more data (e.g short reads).
1335 } while (ret > 0 && iocb->ki_left > 0 &&
1336 !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;
1346 /* see aio_pread() */
1347 static ssize_t aio_pwrite(struct kiocb *iocb)
1349 struct file *file = iocb->ki_filp;
1353 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1354 iocb->ki_left, iocb->ki_pos);
1356 iocb->ki_buf += ret;
1357 iocb->ki_left -= ret;
1359 } while (ret > 0 && iocb->ki_left > 0);
1361 if ((ret == 0) || (iocb->ki_left == 0))
1362 ret = iocb->ki_nbytes - iocb->ki_left;
1367 static ssize_t aio_fdsync(struct kiocb *iocb)
1369 struct file *file = iocb->ki_filp;
1370 ssize_t ret = -EINVAL;
1372 if (file->f_op->aio_fsync)
1373 ret = file->f_op->aio_fsync(iocb, 1);
1377 static ssize_t aio_fsync(struct kiocb *iocb)
1379 struct file *file = iocb->ki_filp;
1380 ssize_t ret = -EINVAL;
1382 if (file->f_op->aio_fsync)
1383 ret = file->f_op->aio_fsync(iocb, 0);
1389 * Performs the initial checks and aio retry method
1390 * setup for the kiocb at the time of io submission.
1392 static ssize_t aio_setup_iocb(struct kiocb *kiocb)
1394 struct file *file = kiocb->ki_filp;
1397 switch (kiocb->ki_opcode) {
1398 case IOCB_CMD_PREAD:
1400 if (unlikely(!(file->f_mode & FMODE_READ)))
1403 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1406 ret = security_file_permission(file, MAY_READ);
1410 if (file->f_op->aio_read)
1411 kiocb->ki_retry = aio_pread;
1413 case IOCB_CMD_PWRITE:
1415 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1418 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1421 ret = security_file_permission(file, MAY_WRITE);
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;
1529 ret = aio_setup_iocb(req);
1534 spin_lock_irq(&ctx->ctx_lock);
1536 if (!list_empty(&ctx->run_list)) {
1537 /* drain the run list */
1538 while (__aio_run_iocbs(ctx))
1541 spin_unlock_irq(&ctx->ctx_lock);
1542 aio_put_req(req); /* drop extra ref to req */
1546 aio_put_req(req); /* drop extra ref to req */
1547 aio_put_req(req); /* drop i/o ref to req */
1552 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1553 * the number of iocbs queued. May return -EINVAL if the aio_context
1554 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1555 * *iocbpp[0] is not properly initialized, if the operation specified
1556 * is invalid for the file descriptor in the iocb. May fail with
1557 * -EFAULT if any of the data structures point to invalid data. May
1558 * fail with -EBADF if the file descriptor specified in the first
1559 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1560 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1561 * fail with -ENOSYS if not implemented.
1563 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1564 struct iocb __user * __user *iocbpp)
1570 if (unlikely(nr < 0))
1573 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1576 ctx = lookup_ioctx(ctx_id);
1577 if (unlikely(!ctx)) {
1578 pr_debug("EINVAL: io_submit: invalid context id\n");
1583 * AKPM: should this return a partial result if some of the IOs were
1584 * successfully submitted?
1586 for (i=0; i<nr; i++) {
1587 struct iocb __user *user_iocb;
1590 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1595 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1600 ret = io_submit_one(ctx, user_iocb, &tmp);
1610 * Finds a given iocb for cancellation.
1612 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1615 struct list_head *pos;
1617 assert_spin_locked(&ctx->ctx_lock);
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)))
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);