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 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/timer.h>
31 #include <linux/aio.h>
32 #include <linux/highmem.h>
33 #include <linux/workqueue.h>
34 #include <linux/security.h>
35 #include <linux/eventfd.h>
36 #include <linux/blkdev.h>
37 #include <linux/compat.h>
39 #include <asm/kmap_types.h>
40 #include <asm/uaccess.h>
42 #define AIO_RING_MAGIC 0xa10a10a1
43 #define AIO_RING_COMPAT_FEATURES 1
44 #define AIO_RING_INCOMPAT_FEATURES 0
46 unsigned id; /* kernel internal index number */
47 unsigned nr; /* number of io_events */
52 unsigned compat_features;
53 unsigned incompat_features;
54 unsigned header_length; /* size of aio_ring */
57 struct io_event io_events[0];
58 }; /* 128 bytes + ring size */
60 #define AIO_RING_PAGES 8
66 /* This needs improving */
67 unsigned long user_id;
68 struct hlist_node list;
71 * This is what userspace passed to io_setup(), it's not used for
72 * anything but counting against the global max_reqs quota.
74 * The real limit is nr_events - 1, which will be larger (see
79 /* Size of ringbuffer, in units of struct io_event */
82 unsigned long mmap_base;
83 unsigned long mmap_size;
85 struct page **ring_pages;
88 struct rcu_head rcu_head;
89 struct work_struct rcu_work;
93 } ____cacheline_aligned_in_smp;
97 struct list_head active_reqs; /* used for cancellation */
98 } ____cacheline_aligned_in_smp;
101 struct mutex ring_lock;
102 wait_queue_head_t wait;
103 } ____cacheline_aligned_in_smp;
107 spinlock_t completion_lock;
108 } ____cacheline_aligned_in_smp;
110 struct page *internal_pages[AIO_RING_PAGES];
113 /*------ sysctl variables----*/
114 static DEFINE_SPINLOCK(aio_nr_lock);
115 unsigned long aio_nr; /* current system wide number of aio requests */
116 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
117 /*----end sysctl variables---*/
119 static struct kmem_cache *kiocb_cachep;
120 static struct kmem_cache *kioctx_cachep;
123 * Creates the slab caches used by the aio routines, panic on
124 * failure as this is done early during the boot sequence.
126 static int __init aio_setup(void)
128 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
129 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
131 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
135 __initcall(aio_setup);
137 static void aio_free_ring(struct kioctx *ctx)
141 for (i = 0; i < ctx->nr_pages; i++)
142 put_page(ctx->ring_pages[i]);
144 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
145 kfree(ctx->ring_pages);
148 static int aio_setup_ring(struct kioctx *ctx)
150 struct aio_ring *ring;
151 unsigned nr_events = ctx->max_reqs;
152 struct mm_struct *mm = current->mm;
153 unsigned long size, populate;
156 /* Compensate for the ring buffer's head/tail overlap entry */
157 nr_events += 2; /* 1 is required, 2 for good luck */
159 size = sizeof(struct aio_ring);
160 size += sizeof(struct io_event) * nr_events;
161 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
166 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
169 ctx->ring_pages = ctx->internal_pages;
170 if (nr_pages > AIO_RING_PAGES) {
171 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
173 if (!ctx->ring_pages)
177 ctx->mmap_size = nr_pages * PAGE_SIZE;
178 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
179 down_write(&mm->mmap_sem);
180 ctx->mmap_base = do_mmap_pgoff(NULL, 0, ctx->mmap_size,
181 PROT_READ|PROT_WRITE,
182 MAP_ANONYMOUS|MAP_PRIVATE, 0, &populate);
183 if (IS_ERR((void *)ctx->mmap_base)) {
184 up_write(&mm->mmap_sem);
190 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
191 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
192 1, 0, ctx->ring_pages, NULL);
193 up_write(&mm->mmap_sem);
195 if (unlikely(ctx->nr_pages != nr_pages)) {
200 mm_populate(ctx->mmap_base, populate);
202 ctx->user_id = ctx->mmap_base;
203 ctx->nr_events = nr_events; /* trusted copy */
205 ring = kmap_atomic(ctx->ring_pages[0]);
206 ring->nr = nr_events; /* user copy */
207 ring->id = ctx->user_id;
208 ring->head = ring->tail = 0;
209 ring->magic = AIO_RING_MAGIC;
210 ring->compat_features = AIO_RING_COMPAT_FEATURES;
211 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
212 ring->header_length = sizeof(struct aio_ring);
214 flush_dcache_page(ctx->ring_pages[0]);
219 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
220 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
221 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
223 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
225 struct kioctx *ctx = req->ki_ctx;
228 spin_lock_irqsave(&ctx->ctx_lock, flags);
230 if (!req->ki_list.next)
231 list_add(&req->ki_list, &ctx->active_reqs);
233 req->ki_cancel = cancel;
235 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
237 EXPORT_SYMBOL(kiocb_set_cancel_fn);
239 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
240 struct io_event *res)
242 kiocb_cancel_fn *old, *cancel;
246 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
247 * actually has a cancel function, hence the cmpxchg()
250 cancel = ACCESS_ONCE(kiocb->ki_cancel);
252 if (!cancel || cancel == KIOCB_CANCELLED)
256 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
257 } while (cancel != old);
259 atomic_inc(&kiocb->ki_users);
260 spin_unlock_irq(&ctx->ctx_lock);
262 memset(res, 0, sizeof(*res));
263 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
264 res->data = kiocb->ki_user_data;
265 ret = cancel(kiocb, res);
267 spin_lock_irq(&ctx->ctx_lock);
272 static void free_ioctx_rcu(struct rcu_head *head)
274 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
275 kmem_cache_free(kioctx_cachep, ctx);
279 * When this function runs, the kioctx has been removed from the "hash table"
280 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
281 * now it's safe to cancel any that need to be.
283 static void free_ioctx(struct kioctx *ctx)
285 struct aio_ring *ring;
288 unsigned head, avail;
290 spin_lock_irq(&ctx->ctx_lock);
292 while (!list_empty(&ctx->active_reqs)) {
293 req = list_first_entry(&ctx->active_reqs,
294 struct kiocb, ki_list);
296 list_del_init(&req->ki_list);
297 kiocb_cancel(ctx, req, &res);
300 spin_unlock_irq(&ctx->ctx_lock);
302 ring = kmap_atomic(ctx->ring_pages[0]);
306 while (atomic_read(&ctx->reqs_active) > 0) {
307 wait_event(ctx->wait,
309 atomic_read(&ctx->reqs_active) <= 0);
311 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
313 atomic_sub(avail, &ctx->reqs_active);
315 head %= ctx->nr_events;
318 WARN_ON(atomic_read(&ctx->reqs_active) < 0);
322 pr_debug("freeing %p\n", ctx);
325 * Here the call_rcu() is between the wait_event() for reqs_active to
326 * hit 0, and freeing the ioctx.
328 * aio_complete() decrements reqs_active, but it has to touch the ioctx
329 * after to issue a wakeup so we use rcu.
331 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
334 static void put_ioctx(struct kioctx *ctx)
336 if (unlikely(atomic_dec_and_test(&ctx->users)))
341 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
343 static struct kioctx *ioctx_alloc(unsigned nr_events)
345 struct mm_struct *mm = current->mm;
349 /* Prevent overflows */
350 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
351 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
352 pr_debug("ENOMEM: nr_events too high\n");
353 return ERR_PTR(-EINVAL);
356 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
357 return ERR_PTR(-EAGAIN);
359 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
361 return ERR_PTR(-ENOMEM);
363 ctx->max_reqs = nr_events;
365 atomic_set(&ctx->users, 2);
366 atomic_set(&ctx->dead, 0);
367 spin_lock_init(&ctx->ctx_lock);
368 spin_lock_init(&ctx->completion_lock);
369 mutex_init(&ctx->ring_lock);
370 init_waitqueue_head(&ctx->wait);
372 INIT_LIST_HEAD(&ctx->active_reqs);
374 if (aio_setup_ring(ctx) < 0)
377 /* limit the number of system wide aios */
378 spin_lock(&aio_nr_lock);
379 if (aio_nr + nr_events > aio_max_nr ||
380 aio_nr + nr_events < aio_nr) {
381 spin_unlock(&aio_nr_lock);
384 aio_nr += ctx->max_reqs;
385 spin_unlock(&aio_nr_lock);
387 /* now link into global list. */
388 spin_lock(&mm->ioctx_lock);
389 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
390 spin_unlock(&mm->ioctx_lock);
392 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
393 ctx, ctx->user_id, mm, ctx->nr_events);
400 kmem_cache_free(kioctx_cachep, ctx);
401 pr_debug("error allocating ioctx %d\n", err);
405 static void kill_ioctx_work(struct work_struct *work)
407 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
409 wake_up_all(&ctx->wait);
413 static void kill_ioctx_rcu(struct rcu_head *head)
415 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
417 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
418 schedule_work(&ctx->rcu_work);
422 * Cancels all outstanding aio requests on an aio context. Used
423 * when the processes owning a context have all exited to encourage
424 * the rapid destruction of the kioctx.
426 static void kill_ioctx(struct kioctx *ctx)
428 if (!atomic_xchg(&ctx->dead, 1)) {
429 hlist_del_rcu(&ctx->list);
432 * It'd be more correct to do this in free_ioctx(), after all
433 * the outstanding kiocbs have finished - but by then io_destroy
434 * has already returned, so io_setup() could potentially return
435 * -EAGAIN with no ioctxs actually in use (as far as userspace
438 spin_lock(&aio_nr_lock);
439 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
440 aio_nr -= ctx->max_reqs;
441 spin_unlock(&aio_nr_lock);
444 vm_munmap(ctx->mmap_base, ctx->mmap_size);
446 /* Between hlist_del_rcu() and dropping the initial ref */
447 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
451 /* wait_on_sync_kiocb:
452 * Waits on the given sync kiocb to complete.
454 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
456 while (atomic_read(&iocb->ki_users)) {
457 set_current_state(TASK_UNINTERRUPTIBLE);
458 if (!atomic_read(&iocb->ki_users))
462 __set_current_state(TASK_RUNNING);
463 return iocb->ki_user_data;
465 EXPORT_SYMBOL(wait_on_sync_kiocb);
468 * exit_aio: called when the last user of mm goes away. At this point, there is
469 * no way for any new requests to be submited or any of the io_* syscalls to be
470 * called on the context.
472 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
475 void exit_aio(struct mm_struct *mm)
478 struct hlist_node *n;
480 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
481 if (1 != atomic_read(&ctx->users))
483 "exit_aio:ioctx still alive: %d %d %d\n",
484 atomic_read(&ctx->users),
485 atomic_read(&ctx->dead),
486 atomic_read(&ctx->reqs_active));
488 * We don't need to bother with munmap() here -
489 * exit_mmap(mm) is coming and it'll unmap everything.
490 * Since aio_free_ring() uses non-zero ->mmap_size
491 * as indicator that it needs to unmap the area,
492 * just set it to 0; aio_free_ring() is the only
493 * place that uses ->mmap_size, so it's safe.
502 * Allocate a slot for an aio request. Increments the ki_users count
503 * of the kioctx so that the kioctx stays around until all requests are
504 * complete. Returns NULL if no requests are free.
506 * Returns with kiocb->ki_users set to 2. The io submit code path holds
507 * an extra reference while submitting the i/o.
508 * This prevents races between the aio code path referencing the
509 * req (after submitting it) and aio_complete() freeing the req.
511 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
515 if (atomic_read(&ctx->reqs_active) >= ctx->nr_events)
518 if (atomic_inc_return(&ctx->reqs_active) > ctx->nr_events - 1)
521 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
525 atomic_set(&req->ki_users, 2);
530 atomic_dec(&ctx->reqs_active);
534 static void kiocb_free(struct kiocb *req)
538 if (req->ki_eventfd != NULL)
539 eventfd_ctx_put(req->ki_eventfd);
542 if (req->ki_iovec != &req->ki_inline_vec)
543 kfree(req->ki_iovec);
544 kmem_cache_free(kiocb_cachep, req);
547 void aio_put_req(struct kiocb *req)
549 if (atomic_dec_and_test(&req->ki_users))
552 EXPORT_SYMBOL(aio_put_req);
554 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
556 struct mm_struct *mm = current->mm;
557 struct kioctx *ctx, *ret = NULL;
561 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
562 if (ctx->user_id == ctx_id) {
563 atomic_inc(&ctx->users);
574 * Called when the io request on the given iocb is complete.
576 void aio_complete(struct kiocb *iocb, long res, long res2)
578 struct kioctx *ctx = iocb->ki_ctx;
579 struct aio_ring *ring;
580 struct io_event *ev_page, *event;
585 * Special case handling for sync iocbs:
586 * - events go directly into the iocb for fast handling
587 * - the sync task with the iocb in its stack holds the single iocb
588 * ref, no other paths have a way to get another ref
589 * - the sync task helpfully left a reference to itself in the iocb
591 if (is_sync_kiocb(iocb)) {
592 BUG_ON(atomic_read(&iocb->ki_users) != 1);
593 iocb->ki_user_data = res;
594 atomic_set(&iocb->ki_users, 0);
595 wake_up_process(iocb->ki_obj.tsk);
600 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
601 * need to issue a wakeup after decrementing reqs_active.
605 if (iocb->ki_list.next) {
608 spin_lock_irqsave(&ctx->ctx_lock, flags);
609 list_del(&iocb->ki_list);
610 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
614 * cancelled requests don't get events, userland was given one
615 * when the event got cancelled.
617 if (unlikely(xchg(&iocb->ki_cancel,
618 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
619 atomic_dec(&ctx->reqs_active);
620 /* Still need the wake_up in case free_ioctx is waiting */
625 * Add a completion event to the ring buffer. Must be done holding
626 * ctx->ctx_lock to prevent other code from messing with the tail
627 * pointer since we might be called from irq context.
629 spin_lock_irqsave(&ctx->completion_lock, flags);
632 pos = tail + AIO_EVENTS_OFFSET;
634 if (++tail >= ctx->nr_events)
637 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
638 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
640 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
641 event->data = iocb->ki_user_data;
645 kunmap_atomic(ev_page);
646 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
648 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
649 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
652 /* after flagging the request as done, we
653 * must never even look at it again
655 smp_wmb(); /* make event visible before updating tail */
659 ring = kmap_atomic(ctx->ring_pages[0]);
662 flush_dcache_page(ctx->ring_pages[0]);
664 spin_unlock_irqrestore(&ctx->completion_lock, flags);
666 pr_debug("added to ring %p at [%u]\n", iocb, tail);
669 * Check if the user asked us to deliver the result through an
670 * eventfd. The eventfd_signal() function is safe to be called
673 if (iocb->ki_eventfd != NULL)
674 eventfd_signal(iocb->ki_eventfd, 1);
677 /* everything turned out well, dispose of the aiocb. */
681 * We have to order our ring_info tail store above and test
682 * of the wait list below outside the wait lock. This is
683 * like in wake_up_bit() where clearing a bit has to be
684 * ordered with the unlocked test.
688 if (waitqueue_active(&ctx->wait))
693 EXPORT_SYMBOL(aio_complete);
696 * Pull an event off of the ioctx's event ring. Returns the number of
699 static long aio_read_events_ring(struct kioctx *ctx,
700 struct io_event __user *event, long nr)
702 struct aio_ring *ring;
707 mutex_lock(&ctx->ring_lock);
709 ring = kmap_atomic(ctx->ring_pages[0]);
713 pr_debug("h%u t%u m%u\n", head, ctx->tail, ctx->nr_events);
715 if (head == ctx->tail)
723 avail = (head <= ctx->tail ? ctx->tail : ctx->nr_events) - head;
724 if (head == ctx->tail)
727 avail = min(avail, nr - ret);
728 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
729 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
731 pos = head + AIO_EVENTS_OFFSET;
732 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
733 pos %= AIO_EVENTS_PER_PAGE;
736 copy_ret = copy_to_user(event + ret, ev + pos,
737 sizeof(*ev) * avail);
740 if (unlikely(copy_ret)) {
747 head %= ctx->nr_events;
750 ring = kmap_atomic(ctx->ring_pages[0]);
753 flush_dcache_page(ctx->ring_pages[0]);
755 pr_debug("%li h%u t%u\n", ret, head, ctx->tail);
757 atomic_sub(ret, &ctx->reqs_active);
759 mutex_unlock(&ctx->ring_lock);
764 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
765 struct io_event __user *event, long *i)
767 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
772 if (unlikely(atomic_read(&ctx->dead)))
778 return ret < 0 || *i >= min_nr;
781 static long read_events(struct kioctx *ctx, long min_nr, long nr,
782 struct io_event __user *event,
783 struct timespec __user *timeout)
785 ktime_t until = { .tv64 = KTIME_MAX };
791 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
794 until = timespec_to_ktime(ts);
798 * Note that aio_read_events() is being called as the conditional - i.e.
799 * we're calling it after prepare_to_wait() has set task state to
800 * TASK_INTERRUPTIBLE.
802 * But aio_read_events() can block, and if it blocks it's going to flip
803 * the task state back to TASK_RUNNING.
805 * This should be ok, provided it doesn't flip the state back to
806 * TASK_RUNNING and return 0 too much - that causes us to spin. That
807 * will only happen if the mutex_lock() call blocks, and we then find
808 * the ringbuffer empty. So in practice we should be ok, but it's
809 * something to be aware of when touching this code.
811 wait_event_interruptible_hrtimeout(ctx->wait,
812 aio_read_events(ctx, min_nr, nr, event, &ret), until);
814 if (!ret && signal_pending(current))
821 * Create an aio_context capable of receiving at least nr_events.
822 * ctxp must not point to an aio_context that already exists, and
823 * must be initialized to 0 prior to the call. On successful
824 * creation of the aio_context, *ctxp is filled in with the resulting
825 * handle. May fail with -EINVAL if *ctxp is not initialized,
826 * if the specified nr_events exceeds internal limits. May fail
827 * with -EAGAIN if the specified nr_events exceeds the user's limit
828 * of available events. May fail with -ENOMEM if insufficient kernel
829 * resources are available. May fail with -EFAULT if an invalid
830 * pointer is passed for ctxp. Will fail with -ENOSYS if not
833 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
835 struct kioctx *ioctx = NULL;
839 ret = get_user(ctx, ctxp);
844 if (unlikely(ctx || nr_events == 0)) {
845 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
850 ioctx = ioctx_alloc(nr_events);
851 ret = PTR_ERR(ioctx);
852 if (!IS_ERR(ioctx)) {
853 ret = put_user(ioctx->user_id, ctxp);
864 * Destroy the aio_context specified. May cancel any outstanding
865 * AIOs and block on completion. Will fail with -ENOSYS if not
866 * implemented. May fail with -EINVAL if the context pointed to
869 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
871 struct kioctx *ioctx = lookup_ioctx(ctx);
872 if (likely(NULL != ioctx)) {
877 pr_debug("EINVAL: io_destroy: invalid context id\n");
881 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
883 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
887 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
888 ssize_t this = min((ssize_t)iov->iov_len, ret);
889 iov->iov_base += this;
890 iov->iov_len -= this;
891 iocb->ki_left -= this;
893 if (iov->iov_len == 0) {
899 /* the caller should not have done more io than what fit in
900 * the remaining iovecs */
901 BUG_ON(ret > 0 && iocb->ki_left == 0);
904 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
905 unsigned long, loff_t);
907 static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op)
909 struct file *file = iocb->ki_filp;
910 struct address_space *mapping = file->f_mapping;
911 struct inode *inode = mapping->host;
914 /* This matches the pread()/pwrite() logic */
915 if (iocb->ki_pos < 0)
919 file_start_write(file);
921 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
922 iocb->ki_nr_segs - iocb->ki_cur_seg,
925 aio_advance_iovec(iocb, ret);
927 /* retry all partial writes. retry partial reads as long as its a
929 } while (ret > 0 && iocb->ki_left > 0 &&
931 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
933 file_end_write(file);
935 /* This means we must have transferred all that we could */
936 /* No need to retry anymore */
937 if ((ret == 0) || (iocb->ki_left == 0))
938 ret = iocb->ki_nbytes - iocb->ki_left;
940 /* If we managed to write some out we return that, rather than
941 * the eventual error. */
943 && ret < 0 && ret != -EIOCBQUEUED
944 && iocb->ki_nbytes - iocb->ki_left)
945 ret = iocb->ki_nbytes - iocb->ki_left;
950 static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat)
954 kiocb->ki_nr_segs = kiocb->ki_nbytes;
958 ret = compat_rw_copy_check_uvector(rw,
959 (struct compat_iovec __user *)kiocb->ki_buf,
960 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
964 ret = rw_copy_check_uvector(rw,
965 (struct iovec __user *)kiocb->ki_buf,
966 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
971 /* ki_nbytes now reflect bytes instead of segs */
972 kiocb->ki_nbytes = ret;
976 static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb)
978 if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes)))
981 kiocb->ki_iovec = &kiocb->ki_inline_vec;
982 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
983 kiocb->ki_iovec->iov_len = kiocb->ki_nbytes;
984 kiocb->ki_nr_segs = 1;
990 * Performs the initial checks and aio retry method
991 * setup for the kiocb at the time of io submission.
993 static ssize_t aio_run_iocb(struct kiocb *req, bool compat)
995 struct file *file = req->ki_filp;
1001 switch (req->ki_opcode) {
1002 case IOCB_CMD_PREAD:
1003 case IOCB_CMD_PREADV:
1006 rw_op = file->f_op->aio_read;
1009 case IOCB_CMD_PWRITE:
1010 case IOCB_CMD_PWRITEV:
1013 rw_op = file->f_op->aio_write;
1016 if (unlikely(!(file->f_mode & mode)))
1022 ret = (req->ki_opcode == IOCB_CMD_PREADV ||
1023 req->ki_opcode == IOCB_CMD_PWRITEV)
1024 ? aio_setup_vectored_rw(rw, req, compat)
1025 : aio_setup_single_vector(rw, req);
1029 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1033 req->ki_nbytes = ret;
1036 ret = aio_rw_vect_retry(req, rw, rw_op);
1039 case IOCB_CMD_FDSYNC:
1040 if (!file->f_op->aio_fsync)
1043 ret = file->f_op->aio_fsync(req, 1);
1046 case IOCB_CMD_FSYNC:
1047 if (!file->f_op->aio_fsync)
1050 ret = file->f_op->aio_fsync(req, 0);
1054 pr_debug("EINVAL: no operation provided\n");
1058 if (ret != -EIOCBQUEUED) {
1060 * There's no easy way to restart the syscall since other AIO's
1061 * may be already running. Just fail this IO with EINTR.
1063 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1064 ret == -ERESTARTNOHAND ||
1065 ret == -ERESTART_RESTARTBLOCK))
1067 aio_complete(req, ret, 0);
1073 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1074 struct iocb *iocb, bool compat)
1079 /* enforce forwards compatibility on users */
1080 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1081 pr_debug("EINVAL: reserve field set\n");
1085 /* prevent overflows */
1087 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1088 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1089 ((ssize_t)iocb->aio_nbytes < 0)
1091 pr_debug("EINVAL: io_submit: overflow check\n");
1095 req = aio_get_req(ctx);
1099 req->ki_filp = fget(iocb->aio_fildes);
1100 if (unlikely(!req->ki_filp)) {
1105 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1107 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1108 * instance of the file* now. The file descriptor must be
1109 * an eventfd() fd, and will be signaled for each completed
1110 * event using the eventfd_signal() function.
1112 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1113 if (IS_ERR(req->ki_eventfd)) {
1114 ret = PTR_ERR(req->ki_eventfd);
1115 req->ki_eventfd = NULL;
1120 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1121 if (unlikely(ret)) {
1122 pr_debug("EFAULT: aio_key\n");
1126 req->ki_obj.user = user_iocb;
1127 req->ki_user_data = iocb->aio_data;
1128 req->ki_pos = iocb->aio_offset;
1130 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1131 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1132 req->ki_opcode = iocb->aio_lio_opcode;
1134 ret = aio_run_iocb(req, compat);
1138 aio_put_req(req); /* drop extra ref to req */
1141 atomic_dec(&ctx->reqs_active);
1142 aio_put_req(req); /* drop extra ref to req */
1143 aio_put_req(req); /* drop i/o ref to req */
1147 long do_io_submit(aio_context_t ctx_id, long nr,
1148 struct iocb __user *__user *iocbpp, bool compat)
1153 struct blk_plug plug;
1155 if (unlikely(nr < 0))
1158 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1159 nr = LONG_MAX/sizeof(*iocbpp);
1161 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1164 ctx = lookup_ioctx(ctx_id);
1165 if (unlikely(!ctx)) {
1166 pr_debug("EINVAL: invalid context id\n");
1170 blk_start_plug(&plug);
1173 * AKPM: should this return a partial result if some of the IOs were
1174 * successfully submitted?
1176 for (i=0; i<nr; i++) {
1177 struct iocb __user *user_iocb;
1180 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1185 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1190 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1194 blk_finish_plug(&plug);
1201 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1202 * the number of iocbs queued. May return -EINVAL if the aio_context
1203 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1204 * *iocbpp[0] is not properly initialized, if the operation specified
1205 * is invalid for the file descriptor in the iocb. May fail with
1206 * -EFAULT if any of the data structures point to invalid data. May
1207 * fail with -EBADF if the file descriptor specified in the first
1208 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1209 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1210 * fail with -ENOSYS if not implemented.
1212 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1213 struct iocb __user * __user *, iocbpp)
1215 return do_io_submit(ctx_id, nr, iocbpp, 0);
1219 * Finds a given iocb for cancellation.
1221 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1224 struct list_head *pos;
1226 assert_spin_locked(&ctx->ctx_lock);
1228 if (key != KIOCB_KEY)
1231 /* TODO: use a hash or array, this sucks. */
1232 list_for_each(pos, &ctx->active_reqs) {
1233 struct kiocb *kiocb = list_kiocb(pos);
1234 if (kiocb->ki_obj.user == iocb)
1241 * Attempts to cancel an iocb previously passed to io_submit. If
1242 * the operation is successfully cancelled, the resulting event is
1243 * copied into the memory pointed to by result without being placed
1244 * into the completion queue and 0 is returned. May fail with
1245 * -EFAULT if any of the data structures pointed to are invalid.
1246 * May fail with -EINVAL if aio_context specified by ctx_id is
1247 * invalid. May fail with -EAGAIN if the iocb specified was not
1248 * cancelled. Will fail with -ENOSYS if not implemented.
1250 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1251 struct io_event __user *, result)
1253 struct io_event res;
1255 struct kiocb *kiocb;
1259 ret = get_user(key, &iocb->aio_key);
1263 ctx = lookup_ioctx(ctx_id);
1267 spin_lock_irq(&ctx->ctx_lock);
1269 kiocb = lookup_kiocb(ctx, iocb, key);
1271 ret = kiocb_cancel(ctx, kiocb, &res);
1275 spin_unlock_irq(&ctx->ctx_lock);
1278 /* Cancellation succeeded -- copy the result
1279 * into the user's buffer.
1281 if (copy_to_user(result, &res, sizeof(res)))
1291 * Attempts to read at least min_nr events and up to nr events from
1292 * the completion queue for the aio_context specified by ctx_id. If
1293 * it succeeds, the number of read events is returned. May fail with
1294 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1295 * out of range, if timeout is out of range. May fail with -EFAULT
1296 * if any of the memory specified is invalid. May return 0 or
1297 * < min_nr if the timeout specified by timeout has elapsed
1298 * before sufficient events are available, where timeout == NULL
1299 * specifies an infinite timeout. Note that the timeout pointed to by
1300 * timeout is relative. Will fail with -ENOSYS if not implemented.
1302 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1305 struct io_event __user *, events,
1306 struct timespec __user *, timeout)
1308 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1311 if (likely(ioctx)) {
1312 if (likely(min_nr <= nr && min_nr >= 0))
1313 ret = read_events(ioctx, min_nr, nr, events, timeout);