2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@redhat.com>
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>
20 #include <linux/sched.h>
22 #include <linux/file.h>
24 #include <linux/mman.h>
25 #include <linux/slab.h>
26 #include <linux/timer.h>
27 #include <linux/aio.h>
28 #include <linux/highmem.h>
29 #include <linux/workqueue.h>
30 #include <linux/security.h>
32 #include <asm/kmap_types.h>
33 #include <asm/uaccess.h>
34 #include <asm/mmu_context.h>
37 #define dprintk printk
39 #define dprintk(x...) do { ; } while (0)
42 long aio_run = 0; /* for testing only */
43 long aio_wakeups = 0; /* for testing only */
45 /*------ sysctl variables----*/
46 atomic_t aio_nr = ATOMIC_INIT(0); /* current system wide number of aio requests */
47 unsigned 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 spinlock_t fput_lock = SPIN_LOCK_UNLOCKED;
62 static void aio_kick_handler(void *);
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 info->nr_pages = nr_pages;
122 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
125 info->ring_pages = info->internal_pages;
126 if (nr_pages > AIO_RING_PAGES) {
127 info->ring_pages = kmalloc(sizeof(struct page *) * nr_pages, GFP_KERNEL);
128 if (!info->ring_pages)
130 memset(info->ring_pages, 0, sizeof(struct page *) * nr_pages);
133 info->mmap_size = nr_pages * PAGE_SIZE;
134 dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
135 down_write(&ctx->mm->mmap_sem);
136 info->mmap_base = do_mmap(NULL, 0, info->mmap_size,
137 PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE,
139 if (IS_ERR((void *)info->mmap_base)) {
140 up_write(&ctx->mm->mmap_sem);
141 printk("mmap err: %ld\n", -info->mmap_base);
147 dprintk("mmap address: 0x%08lx\n", info->mmap_base);
148 info->nr_pages = get_user_pages(current, ctx->mm,
149 info->mmap_base, nr_pages,
150 1, 0, info->ring_pages, NULL);
151 up_write(&ctx->mm->mmap_sem);
153 if (unlikely(info->nr_pages != nr_pages)) {
158 ctx->user_id = info->mmap_base;
160 info->nr = nr_events; /* trusted copy */
162 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
163 ring->nr = nr_events; /* user copy */
164 ring->id = ctx->user_id;
165 ring->head = ring->tail = 0;
166 ring->magic = AIO_RING_MAGIC;
167 ring->compat_features = AIO_RING_COMPAT_FEATURES;
168 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
169 ring->header_length = sizeof(struct aio_ring);
170 kunmap_atomic(ring, KM_USER0);
176 /* aio_ring_event: returns a pointer to the event at the given index from
177 * kmap_atomic(, km). Release the pointer with put_aio_ring_event();
179 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
180 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
181 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
183 #define aio_ring_event(info, nr, km) ({ \
184 unsigned pos = (nr) + AIO_EVENTS_OFFSET; \
185 struct io_event *__event; \
186 __event = kmap_atomic( \
187 (info)->ring_pages[pos / AIO_EVENTS_PER_PAGE], km); \
188 __event += pos % AIO_EVENTS_PER_PAGE; \
192 #define put_aio_ring_event(event, km) do { \
193 struct io_event *__event = (event); \
195 kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
199 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
201 static struct kioctx *ioctx_alloc(unsigned nr_events)
203 struct mm_struct *mm;
206 /* Prevent overflows */
207 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
208 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
209 pr_debug("ENOMEM: nr_events too high\n");
210 return ERR_PTR(-EINVAL);
213 if (nr_events > aio_max_nr)
214 return ERR_PTR(-EAGAIN);
216 ctx = kmem_cache_alloc(kioctx_cachep, GFP_KERNEL);
218 return ERR_PTR(-ENOMEM);
220 memset(ctx, 0, sizeof(*ctx));
221 ctx->max_reqs = nr_events;
222 mm = ctx->mm = current->mm;
223 atomic_inc(&mm->mm_count);
225 atomic_set(&ctx->users, 1);
226 spin_lock_init(&ctx->ctx_lock);
227 spin_lock_init(&ctx->ring_info.ring_lock);
228 init_waitqueue_head(&ctx->wait);
230 INIT_LIST_HEAD(&ctx->active_reqs);
231 INIT_LIST_HEAD(&ctx->run_list);
232 INIT_WORK(&ctx->wq, aio_kick_handler, ctx);
234 if (aio_setup_ring(ctx) < 0)
237 /* limit the number of system wide aios */
238 atomic_add(ctx->max_reqs, &aio_nr); /* undone by __put_ioctx */
239 if (unlikely(atomic_read(&aio_nr) > aio_max_nr))
242 /* now link into global list. kludge. FIXME */
243 write_lock(&mm->ioctx_list_lock);
244 ctx->next = mm->ioctx_list;
245 mm->ioctx_list = ctx;
246 write_unlock(&mm->ioctx_list_lock);
248 dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
249 ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
253 atomic_sub(ctx->max_reqs, &aio_nr);
254 ctx->max_reqs = 0; /* prevent __put_ioctx from sub'ing aio_nr */
256 return ERR_PTR(-EAGAIN);
260 kmem_cache_free(kioctx_cachep, ctx);
261 ctx = ERR_PTR(-ENOMEM);
263 dprintk("aio: error allocating ioctx %p\n", ctx);
268 * Cancels all outstanding aio requests on an aio context. Used
269 * when the processes owning a context have all exited to encourage
270 * the rapid destruction of the kioctx.
272 static void aio_cancel_all(struct kioctx *ctx)
274 int (*cancel)(struct kiocb *, struct io_event *);
276 spin_lock_irq(&ctx->ctx_lock);
278 while (!list_empty(&ctx->active_reqs)) {
279 struct list_head *pos = ctx->active_reqs.next;
280 struct kiocb *iocb = list_kiocb(pos);
281 list_del_init(&iocb->ki_list);
282 cancel = iocb->ki_cancel;
283 kiocbSetCancelled(iocb);
286 spin_unlock_irq(&ctx->ctx_lock);
288 spin_lock_irq(&ctx->ctx_lock);
291 spin_unlock_irq(&ctx->ctx_lock);
294 void wait_for_all_aios(struct kioctx *ctx)
296 struct task_struct *tsk = current;
297 DECLARE_WAITQUEUE(wait, tsk);
299 if (!ctx->reqs_active)
302 add_wait_queue(&ctx->wait, &wait);
303 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
304 while (ctx->reqs_active) {
306 set_task_state(tsk, TASK_UNINTERRUPTIBLE);
308 __set_task_state(tsk, TASK_RUNNING);
309 remove_wait_queue(&ctx->wait, &wait);
312 /* wait_on_sync_kiocb:
313 * Waits on the given sync kiocb to complete.
315 ssize_t fastcall wait_on_sync_kiocb(struct kiocb *iocb)
317 while (iocb->ki_users) {
318 set_current_state(TASK_UNINTERRUPTIBLE);
323 __set_current_state(TASK_RUNNING);
324 return iocb->ki_user_data;
327 /* exit_aio: called when the last user of mm goes away. At this point,
328 * there is no way for any new requests to be submited or any of the
329 * io_* syscalls to be called on the context. However, there may be
330 * outstanding requests which hold references to the context; as they
331 * go away, they will call put_ioctx and release any pinned memory
332 * associated with the request (held via struct page * references).
334 void fastcall exit_aio(struct mm_struct *mm)
336 struct kioctx *ctx = mm->ioctx_list;
337 mm->ioctx_list = NULL;
339 struct kioctx *next = ctx->next;
343 wait_for_all_aios(ctx);
345 * this is an overkill, but ensures we don't leave
346 * the ctx on the aio_wq
348 flush_workqueue(aio_wq);
350 if (1 != atomic_read(&ctx->users))
352 "exit_aio:ioctx still alive: %d %d %d\n",
353 atomic_read(&ctx->users), ctx->dead,
361 * Called when the last user of an aio context has gone away,
362 * and the struct needs to be freed.
364 void fastcall __put_ioctx(struct kioctx *ctx)
366 unsigned nr_events = ctx->max_reqs;
368 if (unlikely(ctx->reqs_active))
371 cancel_delayed_work(&ctx->wq);
372 flush_workqueue(aio_wq);
376 pr_debug("__put_ioctx: freeing %p\n", ctx);
377 kmem_cache_free(kioctx_cachep, ctx);
379 atomic_sub(nr_events, &aio_nr);
383 * Allocate a slot for an aio request. Increments the users count
384 * of the kioctx so that the kioctx stays around until all requests are
385 * complete. Returns NULL if no requests are free.
387 * Returns with kiocb->users set to 2. The io submit code path holds
388 * an extra reference while submitting the i/o.
389 * This prevents races between the aio code path referencing the
390 * req (after submitting it) and aio_complete() freeing the req.
392 static struct kiocb *FASTCALL(__aio_get_req(struct kioctx *ctx));
393 static struct kiocb fastcall *__aio_get_req(struct kioctx *ctx)
395 struct kiocb *req = NULL;
396 struct aio_ring *ring;
399 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
403 req->ki_flags = 1 << KIF_LOCKED;
407 req->ki_cancel = NULL;
408 req->ki_retry = NULL;
409 req->ki_obj.user = NULL;
412 INIT_LIST_HEAD(&req->ki_run_list);
414 /* Check if the completion queue has enough free space to
415 * accept an event from this io.
417 spin_lock_irq(&ctx->ctx_lock);
418 ring = kmap_atomic(ctx->ring_info.ring_pages[0], KM_USER0);
419 if (ctx->reqs_active < aio_ring_avail(&ctx->ring_info, ring)) {
420 list_add(&req->ki_list, &ctx->active_reqs);
425 kunmap_atomic(ring, KM_USER0);
426 spin_unlock_irq(&ctx->ctx_lock);
429 kmem_cache_free(kiocb_cachep, req);
436 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
439 /* Handle a potential starvation case -- should be exceedingly rare as
440 * requests will be stuck on fput_head only if the aio_fput_routine is
441 * delayed and the requests were the last user of the struct file.
443 req = __aio_get_req(ctx);
444 if (unlikely(NULL == req)) {
445 aio_fput_routine(NULL);
446 req = __aio_get_req(ctx);
451 static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
457 req->ki_obj.user = NULL;
460 kmem_cache_free(kiocb_cachep, req);
463 if (unlikely(!ctx->reqs_active && ctx->dead))
467 static void aio_fput_routine(void *data)
469 spin_lock_irq(&fput_lock);
470 while (likely(!list_empty(&fput_head))) {
471 struct kiocb *req = list_kiocb(fput_head.next);
472 struct kioctx *ctx = req->ki_ctx;
474 list_del(&req->ki_list);
475 spin_unlock_irq(&fput_lock);
477 /* Complete the fput */
478 __fput(req->ki_filp);
480 /* Link the iocb into the context's free list */
481 spin_lock_irq(&ctx->ctx_lock);
482 really_put_req(ctx, req);
483 spin_unlock_irq(&ctx->ctx_lock);
486 spin_lock_irq(&fput_lock);
488 spin_unlock_irq(&fput_lock);
492 * Returns true if this put was the last user of the request.
494 static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
496 dprintk(KERN_DEBUG "aio_put(%p): f_count=%d\n",
497 req, atomic_read(&req->ki_filp->f_count));
500 if (unlikely(req->ki_users < 0))
502 if (likely(req->ki_users))
504 list_del(&req->ki_list); /* remove from active_reqs */
505 req->ki_cancel = NULL;
506 req->ki_retry = NULL;
508 /* Must be done under the lock to serialise against cancellation.
509 * Call this aio_fput as it duplicates fput via the fput_work.
511 if (unlikely(atomic_dec_and_test(&req->ki_filp->f_count))) {
513 spin_lock(&fput_lock);
514 list_add(&req->ki_list, &fput_head);
515 spin_unlock(&fput_lock);
516 queue_work(aio_wq, &fput_work);
518 really_put_req(ctx, req);
523 * Returns true if this put was the last user of the kiocb,
524 * false if the request is still in use.
526 int fastcall aio_put_req(struct kiocb *req)
528 struct kioctx *ctx = req->ki_ctx;
530 spin_lock_irq(&ctx->ctx_lock);
531 ret = __aio_put_req(ctx, req);
532 spin_unlock_irq(&ctx->ctx_lock);
538 /* Lookup an ioctx id. ioctx_list is lockless for reads.
539 * FIXME: this is O(n) and is only suitable for development.
541 struct kioctx *lookup_ioctx(unsigned long ctx_id)
543 struct kioctx *ioctx;
544 struct mm_struct *mm;
547 read_lock(&mm->ioctx_list_lock);
548 for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
549 if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
553 read_unlock(&mm->ioctx_list_lock);
560 * Makes the calling kernel thread take on the specified
562 * Called by the retry thread execute retries within the
563 * iocb issuer's mm context, so that copy_from/to_user
564 * operations work seamlessly for aio.
565 * (Note: this routine is intended to be called only
566 * from a kernel thread context)
568 void use_mm(struct mm_struct *mm)
570 struct mm_struct *active_mm;
571 struct task_struct *tsk = current;
574 active_mm = tsk->active_mm;
575 atomic_inc(&mm->mm_count);
578 activate_mm(active_mm, mm);
586 * Reverses the effect of use_mm, i.e. releases the
587 * specified mm context which was earlier taken on
588 * by the calling kernel thread
589 * (Note: this routine is intended to be called only
590 * from a kernel thread context)
592 * Comments: Called with ctx->ctx_lock held. This nests
593 * task_lock instead ctx_lock.
595 void unuse_mm(struct mm_struct *mm)
597 struct task_struct *tsk = current;
601 /* active_mm is still 'mm' */
602 enter_lazy_tlb(mm, tsk);
607 * Queue up a kiocb to be retried. Assumes that the kiocb
608 * has already been marked as kicked, and places it on
609 * the retry run list for the corresponding ioctx, if it
610 * isn't already queued. Returns 1 if it actually queued
611 * the kiocb (to tell the caller to activate the work
612 * queue to process it), or 0, if it found that it was
615 * Should be called with the spin lock iocb->ki_ctx->ctx_lock
618 static inline int __queue_kicked_iocb(struct kiocb *iocb)
620 struct kioctx *ctx = iocb->ki_ctx;
622 if (list_empty(&iocb->ki_run_list)) {
623 list_add_tail(&iocb->ki_run_list,
632 * This is the core aio execution routine. It is
633 * invoked both for initial i/o submission and
634 * subsequent retries via the aio_kick_handler.
635 * Expects to be invoked with iocb->ki_ctx->lock
636 * already held. The lock is released and reaquired
637 * as needed during processing.
639 * Calls the iocb retry method (already setup for the
640 * iocb on initial submission) for operation specific
641 * handling, but takes care of most of common retry
642 * execution details for a given iocb. The retry method
643 * needs to be non-blocking as far as possible, to avoid
644 * holding up other iocbs waiting to be serviced by the
645 * retry kernel thread.
647 * The trickier parts in this code have to do with
648 * ensuring that only one retry instance is in progress
649 * for a given iocb at any time. Providing that guarantee
650 * simplifies the coding of individual aio operations as
651 * it avoids various potential races.
653 static ssize_t aio_run_iocb(struct kiocb *iocb)
655 struct kioctx *ctx = iocb->ki_ctx;
656 ssize_t (*retry)(struct kiocb *);
659 if (iocb->ki_retried++ > 1024*1024) {
660 printk("Maximal retry count. Bytes done %Zd\n",
661 iocb->ki_nbytes - iocb->ki_left);
665 if (!(iocb->ki_retried & 0xff)) {
666 pr_debug("%ld retry: %d of %d (kick %ld, Q %ld run %ld, wake %ld)\n",
668 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes,
669 iocb->ki_kicked, iocb->ki_queued, aio_run, aio_wakeups);
672 if (!(retry = iocb->ki_retry)) {
673 printk("aio_run_iocb: iocb->ki_retry = NULL\n");
678 * We don't want the next retry iteration for this
679 * operation to start until this one has returned and
680 * updated the iocb state. However, wait_queue functions
681 * can trigger a kick_iocb from interrupt context in the
682 * meantime, indicating that data is available for the next
683 * iteration. We want to remember that and enable the
684 * next retry iteration _after_ we are through with
687 * So, in order to be able to register a "kick", but
688 * prevent it from being queued now, we clear the kick
689 * flag, but make the kick code *think* that the iocb is
690 * still on the run list until we are actually done.
691 * When we are done with this iteration, we check if
692 * the iocb was kicked in the meantime and if so, queue
696 kiocbClearKicked(iocb);
699 * This is so that aio_complete knows it doesn't need to
700 * pull the iocb off the run list (We can't just call
701 * INIT_LIST_HEAD because we don't want a kick_iocb to
702 * queue this on the run list yet)
704 iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
705 spin_unlock_irq(&ctx->ctx_lock);
707 /* Quit retrying if the i/o has been cancelled */
708 if (kiocbIsCancelled(iocb)) {
710 aio_complete(iocb, ret, 0);
711 /* must not access the iocb after this */
716 * Now we are all set to call the retry method in async
717 * context. By setting this thread's io_wait context
718 * to point to the wait queue entry inside the currently
719 * running iocb for the duration of the retry, we ensure
720 * that async notification wakeups are queued by the
721 * operation instead of blocking waits, and when notified,
722 * cause the iocb to be kicked for continuation (through
723 * the aio_wake_function callback).
725 BUG_ON(current->io_wait != NULL);
726 current->io_wait = &iocb->ki_wait;
728 current->io_wait = NULL;
730 if (-EIOCBRETRY != ret) {
731 if (-EIOCBQUEUED != ret) {
732 BUG_ON(!list_empty(&iocb->ki_wait.task_list));
733 aio_complete(iocb, ret, 0);
734 /* must not access the iocb after this */
738 * Issue an additional retry to avoid waiting forever if
739 * no waits were queued (e.g. in case of a short read).
741 if (list_empty(&iocb->ki_wait.task_list))
742 kiocbSetKicked(iocb);
745 spin_lock_irq(&ctx->ctx_lock);
747 if (-EIOCBRETRY == ret) {
749 * OK, now that we are done with this iteration
750 * and know that there is more left to go,
751 * this is where we let go so that a subsequent
752 * "kick" can start the next iteration
755 /* will make __queue_kicked_iocb succeed from here on */
756 INIT_LIST_HEAD(&iocb->ki_run_list);
757 /* we must queue the next iteration ourselves, if it
758 * has already been kicked */
759 if (kiocbIsKicked(iocb)) {
760 __queue_kicked_iocb(iocb);
768 * Process all pending retries queued on the ioctx
770 * Assumes it is operating within the aio issuer's mm
771 * context. Expects to be called with ctx->ctx_lock held
773 static int __aio_run_iocbs(struct kioctx *ctx)
779 list_splice_init(&ctx->run_list, &run_list);
780 while (!list_empty(&run_list)) {
781 iocb = list_entry(run_list.next, struct kiocb,
783 list_del(&iocb->ki_run_list);
785 * Hold an extra reference while retrying i/o.
787 iocb->ki_users++; /* grab extra reference */
789 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 void queue_kicked_iocb(struct kiocb *iocb)
883 struct kioctx *ctx = iocb->ki_ctx;
887 WARN_ON((!list_empty(&iocb->ki_wait.task_list)));
889 spin_lock_irqsave(&ctx->ctx_lock, flags);
890 run = __queue_kicked_iocb(iocb);
891 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
900 * Called typically from a wait queue callback context
901 * (aio_wake_function) to trigger a retry of the iocb.
902 * The retry is usually executed by aio workqueue
903 * threads (See aio_kick_handler).
905 void fastcall kick_iocb(struct kiocb *iocb)
907 /* sync iocbs are easy: they can only ever be executing from a
909 if (is_sync_kiocb(iocb)) {
910 kiocbSetKicked(iocb);
911 wake_up_process(iocb->ki_obj.tsk);
916 /* If its already kicked we shouldn't queue it again */
917 if (!kiocbTryKick(iocb)) {
918 queue_kicked_iocb(iocb);
921 EXPORT_SYMBOL(kick_iocb);
924 * Called when the io request on the given iocb is complete.
925 * Returns true if this is the last user of the request. The
926 * only other user of the request can be the cancellation code.
928 int fastcall aio_complete(struct kiocb *iocb, long res, long res2)
930 struct kioctx *ctx = iocb->ki_ctx;
931 struct aio_ring_info *info;
932 struct aio_ring *ring;
933 struct io_event *event;
938 /* Special case handling for sync iocbs: events go directly
939 * into the iocb for fast handling. Note that this will not
940 * work if we allow sync kiocbs to be cancelled. in which
941 * case the usage count checks will have to move under ctx_lock
944 if (is_sync_kiocb(iocb)) {
947 iocb->ki_user_data = res;
948 if (iocb->ki_users == 1) {
952 spin_lock_irq(&ctx->ctx_lock);
954 ret = (0 == iocb->ki_users);
955 spin_unlock_irq(&ctx->ctx_lock);
957 /* sync iocbs put the task here for us */
958 wake_up_process(iocb->ki_obj.tsk);
962 info = &ctx->ring_info;
964 /* add a completion event to the ring buffer.
965 * must be done holding ctx->ctx_lock to prevent
966 * other code from messing with the tail
967 * pointer since we might be called from irq
970 spin_lock_irqsave(&ctx->ctx_lock, flags);
972 if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
973 list_del_init(&iocb->ki_run_list);
976 * cancelled requests don't get events, userland was given one
977 * when the event got cancelled.
979 if (kiocbIsCancelled(iocb))
982 ring = kmap_atomic(info->ring_pages[0], KM_IRQ1);
985 event = aio_ring_event(info, tail, KM_IRQ0);
986 tail = (tail + 1) % info->nr;
988 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
989 event->data = iocb->ki_user_data;
993 dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
994 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
997 /* after flagging the request as done, we
998 * must never even look at it again
1000 smp_wmb(); /* make event visible before updating tail */
1005 put_aio_ring_event(event, KM_IRQ0);
1006 kunmap_atomic(ring, KM_IRQ1);
1008 pr_debug("added to ring %p at [%lu]\n", iocb, tail);
1010 pr_debug("%ld retries: %d of %d (kicked %ld, Q %ld run %ld wake %ld)\n",
1012 iocb->ki_nbytes - iocb->ki_left, iocb->ki_nbytes,
1013 iocb->ki_kicked, iocb->ki_queued, aio_run, aio_wakeups);
1015 /* everything turned out well, dispose of the aiocb. */
1016 ret = __aio_put_req(ctx, iocb);
1018 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1020 if (waitqueue_active(&ctx->wait))
1021 wake_up(&ctx->wait);
1030 * Pull an event off of the ioctx's event ring. Returns the number of
1031 * events fetched (0 or 1 ;-)
1032 * FIXME: make this use cmpxchg.
1033 * TODO: make the ringbuffer user mmap()able (requires FIXME).
1035 static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
1037 struct aio_ring_info *info = &ioctx->ring_info;
1038 struct aio_ring *ring;
1042 ring = kmap_atomic(info->ring_pages[0], KM_USER0);
1043 dprintk("in aio_read_evt h%lu t%lu m%lu\n",
1044 (unsigned long)ring->head, (unsigned long)ring->tail,
1045 (unsigned long)ring->nr);
1047 if (ring->head == ring->tail)
1050 spin_lock(&info->ring_lock);
1052 head = ring->head % info->nr;
1053 if (head != ring->tail) {
1054 struct io_event *evp = aio_ring_event(info, head, KM_USER1);
1056 head = (head + 1) % info->nr;
1057 smp_mb(); /* finish reading the event before updatng the head */
1060 put_aio_ring_event(evp, KM_USER1);
1062 spin_unlock(&info->ring_lock);
1065 kunmap_atomic(ring, KM_USER0);
1066 dprintk("leaving aio_read_evt: %d h%lu t%lu\n", ret,
1067 (unsigned long)ring->head, (unsigned long)ring->tail);
1071 struct aio_timeout {
1072 struct timer_list timer;
1074 struct task_struct *p;
1077 static void timeout_func(unsigned long data)
1079 struct aio_timeout *to = (struct aio_timeout *)data;
1082 wake_up_process(to->p);
1085 static inline void init_timeout(struct aio_timeout *to)
1087 init_timer(&to->timer);
1088 to->timer.data = (unsigned long)to;
1089 to->timer.function = timeout_func;
1094 static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
1095 const struct timespec *ts)
1097 to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
1098 if (time_after(to->timer.expires, jiffies))
1099 add_timer(&to->timer);
1104 static inline void clear_timeout(struct aio_timeout *to)
1106 del_singleshot_timer_sync(&to->timer);
1109 static int read_events(struct kioctx *ctx,
1110 long min_nr, long nr,
1111 struct io_event __user *event,
1112 struct timespec __user *timeout)
1114 long start_jiffies = jiffies;
1115 struct task_struct *tsk = current;
1116 DECLARE_WAITQUEUE(wait, tsk);
1119 struct io_event ent;
1120 struct aio_timeout to;
1121 int event_loop = 0; /* testing only */
1124 /* needed to zero any padding within an entry (there shouldn't be
1125 * any, but C is fun!
1127 memset(&ent, 0, sizeof(ent));
1130 while (likely(i < nr)) {
1131 ret = aio_read_evt(ctx, &ent);
1132 if (unlikely(ret <= 0))
1135 dprintk("read event: %Lx %Lx %Lx %Lx\n",
1136 ent.data, ent.obj, ent.res, ent.res2);
1138 /* Could we split the check in two? */
1140 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1141 dprintk("aio: lost an event due to EFAULT.\n");
1146 /* Good, event copied to userland, update counts. */
1158 /* racey check, but it gets redone */
1159 if (!retry && unlikely(!list_empty(&ctx->run_list))) {
1161 aio_run_all_iocbs(ctx);
1169 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1172 set_timeout(start_jiffies, &to, &ts);
1175 while (likely(i < nr)) {
1176 add_wait_queue_exclusive(&ctx->wait, &wait);
1178 set_task_state(tsk, TASK_INTERRUPTIBLE);
1179 ret = aio_read_evt(ctx, &ent);
1185 if (to.timed_out) /* Only check after read evt */
1189 if (signal_pending(tsk)) {
1193 /*ret = aio_read_evt(ctx, &ent);*/
1196 set_task_state(tsk, TASK_RUNNING);
1197 remove_wait_queue(&ctx->wait, &wait);
1199 if (unlikely(ret <= 0))
1203 if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
1204 dprintk("aio: lost an event due to EFAULT.\n");
1208 /* Good, event copied to userland, update counts. */
1216 pr_debug("event loop executed %d times\n", event_loop);
1217 pr_debug("aio_run %ld\n", aio_run);
1218 pr_debug("aio_wakeups %ld\n", aio_wakeups);
1222 /* Take an ioctx and remove it from the list of ioctx's. Protects
1223 * against races with itself via ->dead.
1225 static void io_destroy(struct kioctx *ioctx)
1227 struct mm_struct *mm = current->mm;
1228 struct kioctx **tmp;
1231 /* delete the entry from the list is someone else hasn't already */
1232 write_lock(&mm->ioctx_list_lock);
1233 was_dead = ioctx->dead;
1235 for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
1236 tmp = &(*tmp)->next)
1240 write_unlock(&mm->ioctx_list_lock);
1242 dprintk("aio_release(%p)\n", ioctx);
1243 if (likely(!was_dead))
1244 put_ioctx(ioctx); /* twice for the list */
1246 aio_cancel_all(ioctx);
1247 wait_for_all_aios(ioctx);
1248 put_ioctx(ioctx); /* once for the lookup */
1252 * Create an aio_context capable of receiving at least nr_events.
1253 * ctxp must not point to an aio_context that already exists, and
1254 * must be initialized to 0 prior to the call. On successful
1255 * creation of the aio_context, *ctxp is filled in with the resulting
1256 * handle. May fail with -EINVAL if *ctxp is not initialized,
1257 * if the specified nr_events exceeds internal limits. May fail
1258 * with -EAGAIN if the specified nr_events exceeds the user's limit
1259 * of available events. May fail with -ENOMEM if insufficient kernel
1260 * resources are available. May fail with -EFAULT if an invalid
1261 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1264 asmlinkage long sys_io_setup(unsigned nr_events, aio_context_t __user *ctxp)
1266 struct kioctx *ioctx = NULL;
1270 ret = get_user(ctx, ctxp);
1275 if (unlikely(ctx || (int)nr_events <= 0)) {
1276 pr_debug("EINVAL: io_setup: ctx or nr_events > max\n");
1280 ioctx = ioctx_alloc(nr_events);
1281 ret = PTR_ERR(ioctx);
1282 if (!IS_ERR(ioctx)) {
1283 ret = put_user(ioctx->user_id, ctxp);
1295 * Destroy the aio_context specified. May cancel any outstanding
1296 * AIOs and block on completion. Will fail with -ENOSYS if not
1297 * implemented. May fail with -EFAULT if the context pointed to
1300 asmlinkage long sys_io_destroy(aio_context_t ctx)
1302 struct kioctx *ioctx = lookup_ioctx(ctx);
1303 if (likely(NULL != ioctx)) {
1307 pr_debug("EINVAL: io_destroy: invalid context id\n");
1312 * Default retry method for aio_read (also used for first time submit)
1313 * Responsible for updating iocb state as retries progress
1315 static ssize_t aio_pread(struct kiocb *iocb)
1317 struct file *file = iocb->ki_filp;
1318 struct address_space *mapping = file->f_mapping;
1319 struct inode *inode = mapping->host;
1322 ret = file->f_op->aio_read(iocb, iocb->ki_buf,
1323 iocb->ki_left, iocb->ki_pos);
1326 * Can't just depend on iocb->ki_left to determine
1327 * whether we are done. This may have been a short read.
1330 iocb->ki_buf += ret;
1331 iocb->ki_left -= ret;
1333 * For pipes and sockets we return once we have
1334 * some data; for regular files we retry till we
1335 * complete the entire read or find that we can't
1336 * read any more data (e.g short reads).
1338 if (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))
1342 /* This means we must have transferred all that we could */
1343 /* No need to retry anymore */
1344 if ((ret == 0) || (iocb->ki_left == 0))
1345 ret = iocb->ki_nbytes - iocb->ki_left;
1351 * Default retry method for aio_write (also used for first time submit)
1352 * Responsible for updating iocb state as retries progress
1354 static ssize_t aio_pwrite(struct kiocb *iocb)
1356 struct file *file = iocb->ki_filp;
1359 ret = file->f_op->aio_write(iocb, iocb->ki_buf,
1360 iocb->ki_left, iocb->ki_pos);
1363 iocb->ki_buf += ret;
1364 iocb->ki_left -= ret;
1369 /* This means we must have transferred all that we could */
1370 /* No need to retry anymore */
1371 if ((ret == 0) || (iocb->ki_left == 0))
1372 ret = iocb->ki_nbytes - iocb->ki_left;
1377 static ssize_t aio_fdsync(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, 1);
1387 static ssize_t aio_fsync(struct kiocb *iocb)
1389 struct file *file = iocb->ki_filp;
1390 ssize_t ret = -EINVAL;
1392 if (file->f_op->aio_fsync)
1393 ret = file->f_op->aio_fsync(iocb, 0);
1399 * Performs the initial checks and aio retry method
1400 * setup for the kiocb at the time of io submission.
1402 ssize_t aio_setup_iocb(struct kiocb *kiocb)
1404 struct file *file = kiocb->ki_filp;
1407 switch (kiocb->ki_opcode) {
1408 case IOCB_CMD_PREAD:
1410 if (unlikely(!(file->f_mode & FMODE_READ)))
1413 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1417 if (file->f_op->aio_read)
1418 kiocb->ki_retry = aio_pread;
1420 case IOCB_CMD_PWRITE:
1422 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1425 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1429 if (file->f_op->aio_write)
1430 kiocb->ki_retry = aio_pwrite;
1432 case IOCB_CMD_FDSYNC:
1434 if (file->f_op->aio_fsync)
1435 kiocb->ki_retry = aio_fdsync;
1437 case IOCB_CMD_FSYNC:
1439 if (file->f_op->aio_fsync)
1440 kiocb->ki_retry = aio_fsync;
1443 dprintk("EINVAL: io_submit: no operation provided\n");
1447 if (!kiocb->ki_retry)
1454 * aio_wake_function:
1455 * wait queue callback function for aio notification,
1456 * Simply triggers a retry of the operation via kick_iocb.
1458 * This callback is specified in the wait queue entry in
1459 * a kiocb (current->io_wait points to this wait queue
1460 * entry when an aio operation executes; it is used
1461 * instead of a synchronous wait when an i/o blocking
1462 * condition is encountered during aio).
1465 * This routine is executed with the wait queue lock held.
1466 * Since kick_iocb acquires iocb->ctx->ctx_lock, it nests
1467 * the ioctx lock inside the wait queue lock. This is safe
1468 * because this callback isn't used for wait queues which
1469 * are nested inside ioctx lock (i.e. ctx->wait)
1471 int aio_wake_function(wait_queue_t *wait, unsigned mode, int sync, void *key)
1473 struct kiocb *iocb = container_of(wait, struct kiocb, ki_wait);
1475 list_del_init(&wait->task_list);
1480 int fastcall io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1487 /* enforce forwards compatibility on users */
1488 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2 ||
1489 iocb->aio_reserved3)) {
1490 pr_debug("EINVAL: io_submit: reserve field set\n");
1494 /* prevent overflows */
1496 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1497 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1498 ((ssize_t)iocb->aio_nbytes < 0)
1500 pr_debug("EINVAL: io_submit: overflow check\n");
1504 file = fget(iocb->aio_fildes);
1505 if (unlikely(!file))
1508 req = aio_get_req(ctx); /* returns with 2 references to req */
1509 if (unlikely(!req)) {
1514 req->ki_filp = file;
1515 iocb->aio_key = req->ki_key;
1516 ret = put_user(iocb->aio_key, &user_iocb->aio_key);
1517 if (unlikely(ret)) {
1518 dprintk("EFAULT: aio_key\n");
1522 req->ki_obj.user = user_iocb;
1523 req->ki_user_data = iocb->aio_data;
1524 req->ki_pos = iocb->aio_offset;
1526 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1527 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1528 req->ki_opcode = iocb->aio_lio_opcode;
1529 init_waitqueue_func_entry(&req->ki_wait, aio_wake_function);
1530 INIT_LIST_HEAD(&req->ki_wait.task_list);
1531 req->ki_run_list.next = req->ki_run_list.prev = NULL;
1532 req->ki_retry = NULL;
1533 req->ki_retried = 0;
1539 ret = aio_setup_iocb(req);
1544 spin_lock_irq(&ctx->ctx_lock);
1545 list_add_tail(&req->ki_run_list, &ctx->run_list);
1546 /* drain the run list */
1547 while (__aio_run_iocbs(ctx))
1549 spin_unlock_irq(&ctx->ctx_lock);
1550 aio_put_req(req); /* drop extra ref to req */
1554 aio_put_req(req); /* drop extra ref to req */
1555 aio_put_req(req); /* drop i/o ref to req */
1560 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1561 * the number of iocbs queued. May return -EINVAL if the aio_context
1562 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1563 * *iocbpp[0] is not properly initialized, if the operation specified
1564 * is invalid for the file descriptor in the iocb. May fail with
1565 * -EFAULT if any of the data structures point to invalid data. May
1566 * fail with -EBADF if the file descriptor specified in the first
1567 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1568 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1569 * fail with -ENOSYS if not implemented.
1571 asmlinkage long sys_io_submit(aio_context_t ctx_id, long nr,
1572 struct iocb __user * __user *iocbpp)
1578 if (unlikely(nr < 0))
1581 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1584 ctx = lookup_ioctx(ctx_id);
1585 if (unlikely(!ctx)) {
1586 pr_debug("EINVAL: io_submit: invalid context id\n");
1591 * AKPM: should this return a partial result if some of the IOs were
1592 * successfully submitted?
1594 for (i=0; i<nr; i++) {
1595 struct iocb __user *user_iocb;
1598 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1603 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1608 ret = io_submit_one(ctx, user_iocb, &tmp);
1618 * Finds a given iocb for cancellation.
1619 * MUST be called with ctx->ctx_lock held.
1621 struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1623 struct list_head *pos;
1624 /* TODO: use a hash or array, this sucks. */
1625 list_for_each(pos, &ctx->active_reqs) {
1626 struct kiocb *kiocb = list_kiocb(pos);
1627 if (kiocb->ki_obj.user == iocb && kiocb->ki_key == key)
1634 * Attempts to cancel an iocb previously passed to io_submit. If
1635 * the operation is successfully cancelled, the resulting event is
1636 * copied into the memory pointed to by result without being placed
1637 * into the completion queue and 0 is returned. May fail with
1638 * -EFAULT if any of the data structures pointed to are invalid.
1639 * May fail with -EINVAL if aio_context specified by ctx_id is
1640 * invalid. May fail with -EAGAIN if the iocb specified was not
1641 * cancelled. Will fail with -ENOSYS if not implemented.
1643 asmlinkage long sys_io_cancel(aio_context_t ctx_id, struct iocb __user *iocb,
1644 struct io_event __user *result)
1646 int (*cancel)(struct kiocb *iocb, struct io_event *res);
1648 struct kiocb *kiocb;
1652 ret = get_user(key, &iocb->aio_key);
1656 ctx = lookup_ioctx(ctx_id);
1660 spin_lock_irq(&ctx->ctx_lock);
1662 kiocb = lookup_kiocb(ctx, iocb, key);
1663 if (kiocb && kiocb->ki_cancel) {
1664 cancel = kiocb->ki_cancel;
1666 kiocbSetCancelled(kiocb);
1669 spin_unlock_irq(&ctx->ctx_lock);
1671 if (NULL != cancel) {
1672 struct io_event tmp;
1673 pr_debug("calling cancel\n");
1674 memset(&tmp, 0, sizeof(tmp));
1675 tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
1676 tmp.data = kiocb->ki_user_data;
1677 ret = cancel(kiocb, &tmp);
1679 /* Cancellation succeeded -- copy the result
1680 * into the user's buffer.
1682 if (copy_to_user(result, &tmp, sizeof(tmp)))
1686 printk(KERN_DEBUG "iocb has no cancel operation\n");
1694 * Attempts to read at least min_nr events and up to nr events from
1695 * the completion queue for the aio_context specified by ctx_id. May
1696 * fail with -EINVAL if ctx_id is invalid, if min_nr is out of range,
1697 * if nr is out of range, if when is out of range. May fail with
1698 * -EFAULT if any of the memory specified to is invalid. May return
1699 * 0 or < min_nr if no events are available and the timeout specified
1700 * by when has elapsed, where when == NULL specifies an infinite
1701 * timeout. Note that the timeout pointed to by when is relative and
1702 * will be updated if not NULL and the operation blocks. Will fail
1703 * with -ENOSYS if not implemented.
1705 asmlinkage long sys_io_getevents(aio_context_t ctx_id,
1708 struct io_event __user *events,
1709 struct timespec __user *timeout)
1711 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1714 if (likely(ioctx)) {
1715 if (likely(min_nr <= nr && min_nr >= 0 && nr >= 0))
1716 ret = read_events(ioctx, min_nr, nr, events, timeout);
1723 __initcall(aio_setup);
1725 EXPORT_SYMBOL(aio_complete);
1726 EXPORT_SYMBOL(aio_put_req);
1727 EXPORT_SYMBOL(wait_on_sync_kiocb);