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/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/percpu-refcount.h>
41 #include <asm/kmap_types.h>
42 #include <asm/uaccess.h>
44 #define AIO_RING_MAGIC 0xa10a10a1
45 #define AIO_RING_COMPAT_FEATURES 1
46 #define AIO_RING_INCOMPAT_FEATURES 0
48 unsigned id; /* kernel internal index number */
49 unsigned nr; /* number of io_events */
54 unsigned compat_features;
55 unsigned incompat_features;
56 unsigned header_length; /* size of aio_ring */
59 struct io_event io_events[0];
60 }; /* 128 bytes + ring size */
62 #define AIO_RING_PAGES 8
65 unsigned reqs_available;
69 struct percpu_ref users;
71 /* This needs improving */
72 unsigned long user_id;
73 struct hlist_node list;
75 struct __percpu kioctx_cpu *cpu;
81 /* sys_io_setup currently limits this to an unsigned int */
84 unsigned long mmap_base;
85 unsigned long mmap_size;
87 struct page **ring_pages;
90 struct rcu_head rcu_head;
91 struct work_struct rcu_work;
94 atomic_t reqs_available;
95 } ____cacheline_aligned_in_smp;
99 struct list_head active_reqs; /* used for cancellation */
100 } ____cacheline_aligned_in_smp;
103 struct mutex ring_lock;
104 wait_queue_head_t wait;
105 unsigned shadow_tail;
106 } ____cacheline_aligned_in_smp;
110 } ____cacheline_aligned_in_smp;
112 struct page *internal_pages[AIO_RING_PAGES];
115 /*------ sysctl variables----*/
116 static DEFINE_SPINLOCK(aio_nr_lock);
117 unsigned long aio_nr; /* current system wide number of aio requests */
118 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
119 /*----end sysctl variables---*/
121 static struct kmem_cache *kiocb_cachep;
122 static struct kmem_cache *kioctx_cachep;
125 * Creates the slab caches used by the aio routines, panic on
126 * failure as this is done early during the boot sequence.
128 static int __init aio_setup(void)
130 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
131 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
133 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
137 __initcall(aio_setup);
139 static void aio_free_ring(struct kioctx *ctx)
143 for (i = 0; i < ctx->nr_pages; i++)
144 put_page(ctx->ring_pages[i]);
147 vm_munmap(ctx->mmap_base, ctx->mmap_size);
149 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
150 kfree(ctx->ring_pages);
153 static int aio_setup_ring(struct kioctx *ctx)
155 struct aio_ring *ring;
156 unsigned nr_events = ctx->max_reqs;
157 struct mm_struct *mm = current->mm;
158 unsigned long size, populate;
161 nr_events = max(nr_events, num_possible_cpus() * 4);
164 /* Compensate for the ring buffer's head/tail overlap entry */
165 nr_events += 2; /* 1 is required, 2 for good luck */
167 size = sizeof(struct aio_ring);
168 size += sizeof(struct io_event) * nr_events;
169 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
174 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
177 ctx->ring_pages = ctx->internal_pages;
178 if (nr_pages > AIO_RING_PAGES) {
179 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
181 if (!ctx->ring_pages)
185 ctx->mmap_size = nr_pages * PAGE_SIZE;
186 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
187 down_write(&mm->mmap_sem);
188 ctx->mmap_base = do_mmap_pgoff(NULL, 0, ctx->mmap_size,
189 PROT_READ|PROT_WRITE,
190 MAP_ANONYMOUS|MAP_PRIVATE, 0, &populate);
191 if (IS_ERR((void *)ctx->mmap_base)) {
192 up_write(&mm->mmap_sem);
198 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
199 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
200 1, 0, ctx->ring_pages, NULL);
201 up_write(&mm->mmap_sem);
203 if (unlikely(ctx->nr_pages != nr_pages)) {
208 mm_populate(ctx->mmap_base, populate);
210 ctx->user_id = ctx->mmap_base;
211 ctx->nr = nr_events; /* trusted copy */
213 ring = kmap_atomic(ctx->ring_pages[0]);
214 ring->nr = nr_events; /* user copy */
215 ring->id = ctx->user_id;
216 ring->head = ring->tail = 0;
217 ring->magic = AIO_RING_MAGIC;
218 ring->compat_features = AIO_RING_COMPAT_FEATURES;
219 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
220 ring->header_length = sizeof(struct aio_ring);
222 flush_dcache_page(ctx->ring_pages[0]);
227 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
228 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
229 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
231 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
233 struct kioctx *ctx = req->ki_ctx;
236 spin_lock_irqsave(&ctx->ctx_lock, flags);
238 if (!req->ki_list.next)
239 list_add(&req->ki_list, &ctx->active_reqs);
241 req->ki_cancel = cancel;
243 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
245 EXPORT_SYMBOL(kiocb_set_cancel_fn);
247 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
248 struct io_event *res)
250 kiocb_cancel_fn *old, *cancel;
254 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
255 * actually has a cancel function, hence the cmpxchg()
258 cancel = ACCESS_ONCE(kiocb->ki_cancel);
260 if (!cancel || cancel == KIOCB_CANCELLED)
264 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
265 } while (cancel != old);
267 atomic_inc(&kiocb->ki_users);
268 spin_unlock_irq(&ctx->ctx_lock);
270 memset(res, 0, sizeof(*res));
271 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
272 res->data = kiocb->ki_user_data;
273 ret = cancel(kiocb, res);
275 spin_lock_irq(&ctx->ctx_lock);
280 static void free_ioctx_rcu(struct rcu_head *head)
282 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
284 free_percpu(ctx->cpu);
285 kmem_cache_free(kioctx_cachep, ctx);
289 * When this function runs, the kioctx has been removed from the "hash table"
290 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
291 * now it's safe to cancel any that need to be.
293 static void free_ioctx(struct kioctx *ctx)
295 struct aio_ring *ring;
298 unsigned cpu, head, avail;
300 spin_lock_irq(&ctx->ctx_lock);
302 while (!list_empty(&ctx->active_reqs)) {
303 req = list_first_entry(&ctx->active_reqs,
304 struct kiocb, ki_list);
306 list_del_init(&req->ki_list);
307 kiocb_cancel(ctx, req, &res);
310 spin_unlock_irq(&ctx->ctx_lock);
312 for_each_possible_cpu(cpu) {
313 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
315 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
316 kcpu->reqs_available = 0;
319 ring = kmap_atomic(ctx->ring_pages[0]);
323 while (atomic_read(&ctx->reqs_available) < ctx->nr) {
324 wait_event(ctx->wait, head != ctx->shadow_tail);
326 avail = (head < ctx->shadow_tail ? ctx->shadow_tail : ctx->nr) - head;
328 atomic_add(avail, &ctx->reqs_available);
333 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr);
337 spin_lock(&aio_nr_lock);
338 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
339 aio_nr -= ctx->max_reqs;
340 spin_unlock(&aio_nr_lock);
342 pr_debug("freeing %p\n", ctx);
345 * Here the call_rcu() is between the wait_event() for reqs_active to
346 * hit 0, and freeing the ioctx.
348 * aio_complete() decrements reqs_active, but it has to touch the ioctx
349 * after to issue a wakeup so we use rcu.
351 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
354 static void put_ioctx(struct kioctx *ctx)
356 if (percpu_ref_put(&ctx->users))
361 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
363 static struct kioctx *ioctx_alloc(unsigned nr_events)
365 struct mm_struct *mm = current->mm;
369 /* Prevent overflows */
370 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
371 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
372 pr_debug("ENOMEM: nr_events too high\n");
373 return ERR_PTR(-EINVAL);
376 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
377 return ERR_PTR(-EAGAIN);
379 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
381 return ERR_PTR(-ENOMEM);
383 ctx->max_reqs = nr_events;
385 percpu_ref_init(&ctx->users);
387 percpu_ref_get(&ctx->users);
390 spin_lock_init(&ctx->ctx_lock);
391 mutex_init(&ctx->ring_lock);
392 init_waitqueue_head(&ctx->wait);
394 INIT_LIST_HEAD(&ctx->active_reqs);
396 ctx->cpu = alloc_percpu(struct kioctx_cpu);
400 if (aio_setup_ring(ctx) < 0)
403 atomic_set(&ctx->reqs_available, ctx->nr);
404 ctx->req_batch = ctx->nr / (num_possible_cpus() * 4);
405 BUG_ON(!ctx->req_batch);
407 /* limit the number of system wide aios */
408 spin_lock(&aio_nr_lock);
409 if (aio_nr + nr_events > aio_max_nr ||
410 aio_nr + nr_events < aio_nr) {
411 spin_unlock(&aio_nr_lock);
414 aio_nr += ctx->max_reqs;
415 spin_unlock(&aio_nr_lock);
417 /* now link into global list. */
418 spin_lock(&mm->ioctx_lock);
419 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
420 spin_unlock(&mm->ioctx_lock);
422 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
423 ctx, ctx->user_id, mm, ctx->nr);
430 free_percpu(ctx->cpu);
432 kmem_cache_free(kioctx_cachep, ctx);
433 pr_debug("error allocating ioctx %d\n", err);
437 static void kill_ioctx_work(struct work_struct *work)
439 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
441 wake_up_all(&ctx->wait);
445 static void kill_ioctx_rcu(struct rcu_head *head)
447 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
449 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
450 schedule_work(&ctx->rcu_work);
454 * Cancels all outstanding aio requests on an aio context. Used
455 * when the processes owning a context have all exited to encourage
456 * the rapid destruction of the kioctx.
458 static void kill_ioctx(struct kioctx *ctx)
460 if (percpu_ref_kill(&ctx->users)) {
461 hlist_del_rcu(&ctx->list);
462 /* Between hlist_del_rcu() and dropping the initial ref */
466 * We can't punt to workqueue here because put_ioctx() ->
467 * free_ioctx() will unmap the ringbuffer, and that has to be
468 * done in the original process's context. kill_ioctx_rcu/work()
469 * exist for exit_aio(), as in that path free_ioctx() won't do
472 kill_ioctx_work(&ctx->rcu_work);
476 /* wait_on_sync_kiocb:
477 * Waits on the given sync kiocb to complete.
479 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
481 while (atomic_read(&iocb->ki_users)) {
482 set_current_state(TASK_UNINTERRUPTIBLE);
483 if (!atomic_read(&iocb->ki_users))
487 __set_current_state(TASK_RUNNING);
488 return iocb->ki_user_data;
490 EXPORT_SYMBOL(wait_on_sync_kiocb);
493 * exit_aio: called when the last user of mm goes away. At this point, there is
494 * no way for any new requests to be submited or any of the io_* syscalls to be
495 * called on the context.
497 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
500 void exit_aio(struct mm_struct *mm)
503 struct hlist_node *n;
505 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
507 * We don't need to bother with munmap() here -
508 * exit_mmap(mm) is coming and it'll unmap everything.
509 * Since aio_free_ring() uses non-zero ->mmap_size
510 * as indicator that it needs to unmap the area,
511 * just set it to 0; aio_free_ring() is the only
512 * place that uses ->mmap_size, so it's safe.
516 if (percpu_ref_kill(&ctx->users)) {
517 hlist_del_rcu(&ctx->list);
518 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
523 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
525 struct kioctx_cpu *kcpu;
528 kcpu = this_cpu_ptr(ctx->cpu);
530 kcpu->reqs_available += nr;
531 while (kcpu->reqs_available >= ctx->req_batch * 2) {
532 kcpu->reqs_available -= ctx->req_batch;
533 atomic_add(ctx->req_batch, &ctx->reqs_available);
539 static bool get_reqs_available(struct kioctx *ctx)
541 struct kioctx_cpu *kcpu;
545 kcpu = this_cpu_ptr(ctx->cpu);
547 if (!kcpu->reqs_available) {
548 int old, avail = atomic_read(&ctx->reqs_available);
551 if (avail < ctx->req_batch)
555 avail = atomic_cmpxchg(&ctx->reqs_available,
556 avail, avail - ctx->req_batch);
557 } while (avail != old);
559 kcpu->reqs_available += ctx->req_batch;
563 kcpu->reqs_available--;
570 * Allocate a slot for an aio request. Increments the ki_users count
571 * of the kioctx so that the kioctx stays around until all requests are
572 * complete. Returns NULL if no requests are free.
574 * Returns with kiocb->ki_users set to 2. The io submit code path holds
575 * an extra reference while submitting the i/o.
576 * This prevents races between the aio code path referencing the
577 * req (after submitting it) and aio_complete() freeing the req.
579 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
583 if (!get_reqs_available(ctx))
586 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
590 atomic_set(&req->ki_users, 2);
594 put_reqs_available(ctx, 1);
598 static void kiocb_free(struct kiocb *req)
602 if (req->ki_eventfd != NULL)
603 eventfd_ctx_put(req->ki_eventfd);
606 if (req->ki_iovec != &req->ki_inline_vec)
607 kfree(req->ki_iovec);
608 kmem_cache_free(kiocb_cachep, req);
611 void aio_put_req(struct kiocb *req)
613 if (atomic_dec_and_test(&req->ki_users))
616 EXPORT_SYMBOL(aio_put_req);
618 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
620 struct mm_struct *mm = current->mm;
621 struct kioctx *ctx, *ret = NULL;
625 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
626 if (ctx->user_id == ctx_id){
627 percpu_ref_get(&ctx->users);
638 * Called when the io request on the given iocb is complete.
640 void aio_complete(struct kiocb *iocb, long res, long res2)
642 struct kioctx *ctx = iocb->ki_ctx;
643 struct aio_ring *ring;
644 struct io_event *ev_page, *event;
649 * Special case handling for sync iocbs:
650 * - events go directly into the iocb for fast handling
651 * - the sync task with the iocb in its stack holds the single iocb
652 * ref, no other paths have a way to get another ref
653 * - the sync task helpfully left a reference to itself in the iocb
655 if (is_sync_kiocb(iocb)) {
656 BUG_ON(atomic_read(&iocb->ki_users) != 1);
657 iocb->ki_user_data = res;
658 atomic_set(&iocb->ki_users, 0);
659 wake_up_process(iocb->ki_obj.tsk);
664 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
665 * need to issue a wakeup after incrementing reqs_available.
669 if (iocb->ki_list.next) {
672 spin_lock_irqsave(&ctx->ctx_lock, flags);
673 list_del(&iocb->ki_list);
674 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
678 * cancelled requests don't get events, userland was given one
679 * when the event got cancelled.
681 if (unlikely(xchg(&iocb->ki_cancel,
682 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
684 * Can't use the percpu reqs_available here - could race with
687 atomic_inc(&ctx->reqs_available);
688 smp_mb__after_atomic_inc();
689 /* Still need the wake_up in case free_ioctx is waiting */
694 * Add a completion event to the ring buffer; ctx->tail is both our lock
695 * and the canonical version of the tail pointer.
697 local_irq_save(flags);
698 while ((tail = xchg(&ctx->tail, UINT_MAX)) == UINT_MAX)
701 pos = tail + AIO_EVENTS_OFFSET;
703 if (++tail >= ctx->nr)
706 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
707 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
709 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
710 event->data = iocb->ki_user_data;
714 kunmap_atomic(ev_page);
715 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
717 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
718 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
721 /* after flagging the request as done, we
722 * must never even look at it again
724 smp_wmb(); /* make event visible before updating tail */
726 ctx->shadow_tail = tail;
728 ring = kmap_atomic(ctx->ring_pages[0]);
731 flush_dcache_page(ctx->ring_pages[0]);
733 /* unlock, make new tail visible before checking waitlist */
737 local_irq_restore(flags);
739 pr_debug("added to ring %p at [%u]\n", iocb, tail);
742 * Check if the user asked us to deliver the result through an
743 * eventfd. The eventfd_signal() function is safe to be called
746 if (iocb->ki_eventfd != NULL)
747 eventfd_signal(iocb->ki_eventfd, 1);
750 /* everything turned out well, dispose of the aiocb. */
753 if (waitqueue_active(&ctx->wait))
758 EXPORT_SYMBOL(aio_complete);
761 * Pull an event off of the ioctx's event ring. Returns the number of
764 static int aio_read_events_ring(struct kioctx *ctx,
765 struct io_event __user *event, long nr)
767 struct aio_ring *ring;
769 int ret = 0, copy_ret;
771 if (!mutex_trylock(&ctx->ring_lock)) {
772 __set_current_state(TASK_RUNNING);
773 mutex_lock(&ctx->ring_lock);
776 ring = kmap_atomic(ctx->ring_pages[0]);
780 pr_debug("h%u t%u m%u\n", head, ctx->shadow_tail, ctx->nr);
782 if (head == ctx->shadow_tail)
785 __set_current_state(TASK_RUNNING);
788 unsigned i = (head < ctx->shadow_tail ? ctx->shadow_tail : ctx->nr) - head;
792 if (head == ctx->shadow_tail)
795 i = min_t(int, i, nr - ret);
796 i = min_t(int, i, AIO_EVENTS_PER_PAGE -
797 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
799 pos = head + AIO_EVENTS_OFFSET;
800 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
801 pos %= AIO_EVENTS_PER_PAGE;
804 copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * i);
807 if (unlikely(copy_ret)) {
817 ring = kmap_atomic(ctx->ring_pages[0]);
820 flush_dcache_page(ctx->ring_pages[0]);
822 pr_debug("%d h%u t%u\n", ret, head, ctx->shadow_tail);
824 put_reqs_available(ctx, ret);
826 mutex_unlock(&ctx->ring_lock);
831 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
832 struct io_event __user *event, long *i)
834 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
839 if (unlikely(percpu_ref_dead(&ctx->users)))
845 return ret < 0 || *i >= min_nr;
848 static long read_events(struct kioctx *ctx, long min_nr, long nr,
849 struct io_event __user *event,
850 struct timespec __user *timeout)
852 ktime_t until = { .tv64 = KTIME_MAX };
858 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
861 until = timespec_to_ktime(ts);
864 wait_event_interruptible_hrtimeout(ctx->wait,
865 aio_read_events(ctx, min_nr, nr, event, &ret), until);
867 if (!ret && signal_pending(current))
874 * Create an aio_context capable of receiving at least nr_events.
875 * ctxp must not point to an aio_context that already exists, and
876 * must be initialized to 0 prior to the call. On successful
877 * creation of the aio_context, *ctxp is filled in with the resulting
878 * handle. May fail with -EINVAL if *ctxp is not initialized,
879 * if the specified nr_events exceeds internal limits. May fail
880 * with -EAGAIN if the specified nr_events exceeds the user's limit
881 * of available events. May fail with -ENOMEM if insufficient kernel
882 * resources are available. May fail with -EFAULT if an invalid
883 * pointer is passed for ctxp. Will fail with -ENOSYS if not
886 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
888 struct kioctx *ioctx = NULL;
892 ret = get_user(ctx, ctxp);
897 if (unlikely(ctx || nr_events == 0)) {
898 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
903 ioctx = ioctx_alloc(nr_events);
904 ret = PTR_ERR(ioctx);
905 if (!IS_ERR(ioctx)) {
906 ret = put_user(ioctx->user_id, ctxp);
917 * Destroy the aio_context specified. May cancel any outstanding
918 * AIOs and block on completion. Will fail with -ENOSYS if not
919 * implemented. May fail with -EINVAL if the context pointed to
922 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
924 struct kioctx *ioctx = lookup_ioctx(ctx);
925 if (likely(NULL != ioctx)) {
930 pr_debug("EINVAL: io_destroy: invalid context id\n");
934 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
936 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
940 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
941 ssize_t this = min((ssize_t)iov->iov_len, ret);
942 iov->iov_base += this;
943 iov->iov_len -= this;
944 iocb->ki_left -= this;
946 if (iov->iov_len == 0) {
952 /* the caller should not have done more io than what fit in
953 * the remaining iovecs */
954 BUG_ON(ret > 0 && iocb->ki_left == 0);
957 static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
959 struct file *file = iocb->ki_filp;
960 struct address_space *mapping = file->f_mapping;
961 struct inode *inode = mapping->host;
962 ssize_t (*rw_op)(struct kiocb *, const struct iovec *,
963 unsigned long, loff_t);
965 unsigned short opcode;
967 if ((iocb->ki_opcode == IOCB_CMD_PREADV) ||
968 (iocb->ki_opcode == IOCB_CMD_PREAD)) {
969 rw_op = file->f_op->aio_read;
970 opcode = IOCB_CMD_PREADV;
972 rw_op = file->f_op->aio_write;
973 opcode = IOCB_CMD_PWRITEV;
976 /* This matches the pread()/pwrite() logic */
977 if (iocb->ki_pos < 0)
981 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
982 iocb->ki_nr_segs - iocb->ki_cur_seg,
985 aio_advance_iovec(iocb, ret);
987 /* retry all partial writes. retry partial reads as long as its a
989 } while (ret > 0 && iocb->ki_left > 0 &&
990 (opcode == IOCB_CMD_PWRITEV ||
991 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
993 /* This means we must have transferred all that we could */
994 /* No need to retry anymore */
995 if ((ret == 0) || (iocb->ki_left == 0))
996 ret = iocb->ki_nbytes - iocb->ki_left;
998 /* If we managed to write some out we return that, rather than
999 * the eventual error. */
1000 if (opcode == IOCB_CMD_PWRITEV
1001 && ret < 0 && ret != -EIOCBQUEUED
1002 && iocb->ki_nbytes - iocb->ki_left)
1003 ret = iocb->ki_nbytes - iocb->ki_left;
1008 static ssize_t aio_fdsync(struct kiocb *iocb)
1010 struct file *file = iocb->ki_filp;
1011 ssize_t ret = -EINVAL;
1013 if (file->f_op->aio_fsync)
1014 ret = file->f_op->aio_fsync(iocb, 1);
1018 static ssize_t aio_fsync(struct kiocb *iocb)
1020 struct file *file = iocb->ki_filp;
1021 ssize_t ret = -EINVAL;
1023 if (file->f_op->aio_fsync)
1024 ret = file->f_op->aio_fsync(iocb, 0);
1028 static ssize_t aio_setup_vectored_rw(int type, struct kiocb *kiocb, bool compat)
1032 #ifdef CONFIG_COMPAT
1034 ret = compat_rw_copy_check_uvector(type,
1035 (struct compat_iovec __user *)kiocb->ki_buf,
1036 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1040 ret = rw_copy_check_uvector(type,
1041 (struct iovec __user *)kiocb->ki_buf,
1042 kiocb->ki_nbytes, 1, &kiocb->ki_inline_vec,
1047 ret = rw_verify_area(type, kiocb->ki_filp, &kiocb->ki_pos, ret);
1051 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1052 kiocb->ki_cur_seg = 0;
1053 /* ki_nbytes/left now reflect bytes instead of segs */
1054 kiocb->ki_nbytes = ret;
1055 kiocb->ki_left = ret;
1062 static ssize_t aio_setup_single_vector(int type, struct file * file, struct kiocb *kiocb)
1066 bytes = rw_verify_area(type, file, &kiocb->ki_pos, kiocb->ki_left);
1070 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1071 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1072 kiocb->ki_iovec->iov_len = bytes;
1073 kiocb->ki_nr_segs = 1;
1074 kiocb->ki_cur_seg = 0;
1080 * Performs the initial checks and aio retry method
1081 * setup for the kiocb at the time of io submission.
1083 static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
1085 struct file *file = kiocb->ki_filp;
1088 switch (kiocb->ki_opcode) {
1089 case IOCB_CMD_PREAD:
1091 if (unlikely(!(file->f_mode & FMODE_READ)))
1094 if (unlikely(!access_ok(VERIFY_WRITE, kiocb->ki_buf,
1097 ret = aio_setup_single_vector(READ, file, kiocb);
1101 if (file->f_op->aio_read)
1102 kiocb->ki_retry = aio_rw_vect_retry;
1104 case IOCB_CMD_PWRITE:
1106 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1109 if (unlikely(!access_ok(VERIFY_READ, kiocb->ki_buf,
1112 ret = aio_setup_single_vector(WRITE, file, kiocb);
1116 if (file->f_op->aio_write)
1117 kiocb->ki_retry = aio_rw_vect_retry;
1119 case IOCB_CMD_PREADV:
1121 if (unlikely(!(file->f_mode & FMODE_READ)))
1123 ret = aio_setup_vectored_rw(READ, kiocb, compat);
1127 if (file->f_op->aio_read)
1128 kiocb->ki_retry = aio_rw_vect_retry;
1130 case IOCB_CMD_PWRITEV:
1132 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1134 ret = aio_setup_vectored_rw(WRITE, kiocb, compat);
1138 if (file->f_op->aio_write)
1139 kiocb->ki_retry = aio_rw_vect_retry;
1141 case IOCB_CMD_FDSYNC:
1143 if (file->f_op->aio_fsync)
1144 kiocb->ki_retry = aio_fdsync;
1146 case IOCB_CMD_FSYNC:
1148 if (file->f_op->aio_fsync)
1149 kiocb->ki_retry = aio_fsync;
1152 pr_debug("EINVAL: no operation provided\n");
1156 if (!kiocb->ki_retry)
1162 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1163 struct iocb *iocb, bool compat)
1168 /* enforce forwards compatibility on users */
1169 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1170 pr_debug("EINVAL: reserve field set\n");
1174 /* prevent overflows */
1176 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1177 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1178 ((ssize_t)iocb->aio_nbytes < 0)
1180 pr_debug("EINVAL: io_submit: overflow check\n");
1184 req = aio_get_req(ctx); /* returns with 2 references to req */
1188 req->ki_filp = fget(iocb->aio_fildes);
1189 if (unlikely(!req->ki_filp)) {
1194 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1196 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1197 * instance of the file* now. The file descriptor must be
1198 * an eventfd() fd, and will be signaled for each completed
1199 * event using the eventfd_signal() function.
1201 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1202 if (IS_ERR(req->ki_eventfd)) {
1203 ret = PTR_ERR(req->ki_eventfd);
1204 req->ki_eventfd = NULL;
1209 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1210 if (unlikely(ret)) {
1211 pr_debug("EFAULT: aio_key\n");
1215 req->ki_obj.user = user_iocb;
1216 req->ki_user_data = iocb->aio_data;
1217 req->ki_pos = iocb->aio_offset;
1219 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1220 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1221 req->ki_opcode = iocb->aio_lio_opcode;
1223 ret = aio_setup_iocb(req, compat);
1227 ret = req->ki_retry(req);
1228 if (ret != -EIOCBQUEUED) {
1230 * There's no easy way to restart the syscall since other AIO's
1231 * may be already running. Just fail this IO with EINTR.
1233 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1234 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
1236 aio_complete(req, ret, 0);
1239 aio_put_req(req); /* drop extra ref to req */
1243 put_reqs_available(ctx, 1);
1244 aio_put_req(req); /* drop extra ref to req */
1245 aio_put_req(req); /* drop i/o ref to req */
1249 long do_io_submit(aio_context_t ctx_id, long nr,
1250 struct iocb __user *__user *iocbpp, bool compat)
1255 struct blk_plug plug;
1257 if (unlikely(nr < 0))
1260 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1261 nr = LONG_MAX/sizeof(*iocbpp);
1263 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1266 ctx = lookup_ioctx(ctx_id);
1267 if (unlikely(!ctx)) {
1268 pr_debug("EINVAL: invalid context id\n");
1272 blk_start_plug(&plug);
1275 * AKPM: should this return a partial result if some of the IOs were
1276 * successfully submitted?
1278 for (i=0; i<nr; i++) {
1279 struct iocb __user *user_iocb;
1282 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1287 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1292 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1296 blk_finish_plug(&plug);
1303 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1304 * the number of iocbs queued. May return -EINVAL if the aio_context
1305 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1306 * *iocbpp[0] is not properly initialized, if the operation specified
1307 * is invalid for the file descriptor in the iocb. May fail with
1308 * -EFAULT if any of the data structures point to invalid data. May
1309 * fail with -EBADF if the file descriptor specified in the first
1310 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1311 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1312 * fail with -ENOSYS if not implemented.
1314 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1315 struct iocb __user * __user *, iocbpp)
1317 return do_io_submit(ctx_id, nr, iocbpp, 0);
1321 * Finds a given iocb for cancellation.
1323 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1326 struct list_head *pos;
1328 assert_spin_locked(&ctx->ctx_lock);
1330 if (key != KIOCB_KEY)
1333 /* TODO: use a hash or array, this sucks. */
1334 list_for_each(pos, &ctx->active_reqs) {
1335 struct kiocb *kiocb = list_kiocb(pos);
1336 if (kiocb->ki_obj.user == iocb)
1343 * Attempts to cancel an iocb previously passed to io_submit. If
1344 * the operation is successfully cancelled, the resulting event is
1345 * copied into the memory pointed to by result without being placed
1346 * into the completion queue and 0 is returned. May fail with
1347 * -EFAULT if any of the data structures pointed to are invalid.
1348 * May fail with -EINVAL if aio_context specified by ctx_id is
1349 * invalid. May fail with -EAGAIN if the iocb specified was not
1350 * cancelled. Will fail with -ENOSYS if not implemented.
1352 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1353 struct io_event __user *, result)
1355 struct io_event res;
1357 struct kiocb *kiocb;
1361 ret = get_user(key, &iocb->aio_key);
1365 ctx = lookup_ioctx(ctx_id);
1369 spin_lock_irq(&ctx->ctx_lock);
1371 kiocb = lookup_kiocb(ctx, iocb, key);
1373 ret = kiocb_cancel(ctx, kiocb, &res);
1377 spin_unlock_irq(&ctx->ctx_lock);
1380 /* Cancellation succeeded -- copy the result
1381 * into the user's buffer.
1383 if (copy_to_user(result, &res, sizeof(res)))
1393 * Attempts to read at least min_nr events and up to nr events from
1394 * the completion queue for the aio_context specified by ctx_id. If
1395 * it succeeds, the number of read events is returned. May fail with
1396 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1397 * out of range, if timeout is out of range. May fail with -EFAULT
1398 * if any of the memory specified is invalid. May return 0 or
1399 * < min_nr if the timeout specified by timeout has elapsed
1400 * before sufficient events are available, where timeout == NULL
1401 * specifies an infinite timeout. Note that the timeout pointed to by
1402 * timeout is relative and will be updated if not NULL and the
1403 * operation blocks. Will fail with -ENOSYS if not implemented.
1405 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1408 struct io_event __user *, events,
1409 struct timespec __user *, timeout)
1411 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1414 if (likely(ioctx)) {
1415 if (likely(min_nr <= nr && min_nr >= 0))
1416 ret = read_events(ioctx, min_nr, nr, events, timeout);