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/bio.h>
29 #include <linux/mmu_context.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/percpu-refcount.h>
42 #include <asm/kmap_types.h>
43 #include <asm/uaccess.h>
45 #define AIO_RING_MAGIC 0xa10a10a1
46 #define AIO_RING_COMPAT_FEATURES 1
47 #define AIO_RING_INCOMPAT_FEATURES 0
49 unsigned id; /* kernel internal index number */
50 unsigned nr; /* number of io_events */
55 unsigned compat_features;
56 unsigned incompat_features;
57 unsigned header_length; /* size of aio_ring */
60 struct io_event io_events[0];
61 }; /* 128 bytes + ring size */
63 #define AIO_RING_PAGES 8
66 unsigned reqs_available;
70 struct percpu_ref users;
72 /* This needs improving */
73 unsigned long user_id;
74 struct hlist_node list;
76 struct __percpu kioctx_cpu *cpu;
82 /* sys_io_setup currently limits this to an unsigned int */
85 unsigned long mmap_base;
86 unsigned long mmap_size;
88 struct page **ring_pages;
91 struct rcu_head rcu_head;
92 struct work_struct rcu_work;
95 atomic_t reqs_available;
96 } ____cacheline_aligned_in_smp;
100 struct list_head active_reqs; /* used for cancellation */
101 } ____cacheline_aligned_in_smp;
104 struct mutex ring_lock;
105 wait_queue_head_t wait;
108 * Copy of the real tail, that aio_complete uses - to reduce
109 * cacheline bouncing. The real tail will tend to be much more
110 * contended - since typically events are delivered one at a
111 * time, and then aio_read_events() slurps them up a bunch at a
112 * time - so it's helpful if aio_read_events() isn't also
113 * contending for the tail. So, aio_complete() updates
114 * shadow_tail whenever it updates tail.
116 * Also needed because tail is used as a hacky lock and isn't
117 * always the real tail.
119 unsigned shadow_tail;
120 } ____cacheline_aligned_in_smp;
124 } ____cacheline_aligned_in_smp;
126 struct page *internal_pages[AIO_RING_PAGES];
129 /*------ sysctl variables----*/
130 static DEFINE_SPINLOCK(aio_nr_lock);
131 unsigned long aio_nr; /* current system wide number of aio requests */
132 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
133 /*----end sysctl variables---*/
135 static struct kmem_cache *kiocb_cachep;
136 static struct kmem_cache *kioctx_cachep;
139 * Creates the slab caches used by the aio routines, panic on
140 * failure as this is done early during the boot sequence.
142 static int __init aio_setup(void)
144 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
145 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
147 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
151 __initcall(aio_setup);
153 static void aio_free_ring(struct kioctx *ctx)
157 for (i = 0; i < ctx->nr_pages; i++)
158 put_page(ctx->ring_pages[i]);
161 vm_munmap(ctx->mmap_base, ctx->mmap_size);
163 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
164 kfree(ctx->ring_pages);
167 static int aio_setup_ring(struct kioctx *ctx)
169 struct aio_ring *ring;
170 unsigned nr_events = ctx->max_reqs;
171 struct mm_struct *mm = current->mm;
172 unsigned long size, populate;
175 nr_events = max(nr_events, num_possible_cpus() * 4);
178 /* Compensate for the ring buffer's head/tail overlap entry */
179 nr_events += 2; /* 1 is required, 2 for good luck */
181 size = sizeof(struct aio_ring);
182 size += sizeof(struct io_event) * nr_events;
183 nr_pages = (size + PAGE_SIZE-1) >> PAGE_SHIFT;
188 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring)) / sizeof(struct io_event);
191 ctx->ring_pages = ctx->internal_pages;
192 if (nr_pages > AIO_RING_PAGES) {
193 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
195 if (!ctx->ring_pages)
199 ctx->mmap_size = nr_pages * PAGE_SIZE;
200 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
201 down_write(&mm->mmap_sem);
202 ctx->mmap_base = do_mmap_pgoff(NULL, 0, ctx->mmap_size,
203 PROT_READ|PROT_WRITE,
204 MAP_ANONYMOUS|MAP_PRIVATE, 0, &populate);
205 if (IS_ERR((void *)ctx->mmap_base)) {
206 up_write(&mm->mmap_sem);
212 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
213 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
214 1, 0, ctx->ring_pages, NULL);
215 up_write(&mm->mmap_sem);
217 if (unlikely(ctx->nr_pages != nr_pages)) {
222 mm_populate(ctx->mmap_base, populate);
224 ctx->user_id = ctx->mmap_base;
225 ctx->nr = nr_events; /* trusted copy */
227 ring = kmap_atomic(ctx->ring_pages[0]);
228 ring->nr = nr_events; /* user copy */
229 ring->id = ctx->user_id;
230 ring->head = ring->tail = 0;
231 ring->magic = AIO_RING_MAGIC;
232 ring->compat_features = AIO_RING_COMPAT_FEATURES;
233 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
234 ring->header_length = sizeof(struct aio_ring);
236 flush_dcache_page(ctx->ring_pages[0]);
241 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
242 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
243 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
245 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
247 struct kioctx *ctx = req->ki_ctx;
250 spin_lock_irqsave(&ctx->ctx_lock, flags);
252 if (!req->ki_list.next)
253 list_add(&req->ki_list, &ctx->active_reqs);
255 req->ki_cancel = cancel;
257 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
259 EXPORT_SYMBOL(kiocb_set_cancel_fn);
261 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
262 struct io_event *res)
264 kiocb_cancel_fn *old, *cancel;
268 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
269 * actually has a cancel function, hence the cmpxchg()
272 cancel = ACCESS_ONCE(kiocb->ki_cancel);
274 if (!cancel || cancel == KIOCB_CANCELLED)
278 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
279 } while (cancel != old);
281 atomic_inc(&kiocb->ki_users);
282 spin_unlock_irq(&ctx->ctx_lock);
284 memset(res, 0, sizeof(*res));
285 res->obj = (u64)(unsigned long)kiocb->ki_obj.user;
286 res->data = kiocb->ki_user_data;
287 ret = cancel(kiocb, res);
289 spin_lock_irq(&ctx->ctx_lock);
294 static void free_ioctx_rcu(struct rcu_head *head)
296 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
298 free_percpu(ctx->cpu);
299 kmem_cache_free(kioctx_cachep, ctx);
303 * When this function runs, the kioctx has been removed from the "hash table"
304 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
305 * now it's safe to cancel any that need to be.
307 static void free_ioctx(struct kioctx *ctx)
309 struct aio_ring *ring;
312 unsigned cpu, head, avail;
314 spin_lock_irq(&ctx->ctx_lock);
316 while (!list_empty(&ctx->active_reqs)) {
317 req = list_first_entry(&ctx->active_reqs,
318 struct kiocb, ki_list);
320 list_del_init(&req->ki_list);
321 kiocb_cancel(ctx, req, &res);
324 spin_unlock_irq(&ctx->ctx_lock);
326 for_each_possible_cpu(cpu) {
327 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
329 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
330 kcpu->reqs_available = 0;
333 ring = kmap_atomic(ctx->ring_pages[0]);
337 while (atomic_read(&ctx->reqs_available) < ctx->nr) {
338 wait_event(ctx->wait, head != ctx->shadow_tail);
340 avail = (head < ctx->shadow_tail ? ctx->shadow_tail : ctx->nr) - head;
342 atomic_add(avail, &ctx->reqs_available);
347 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr);
351 spin_lock(&aio_nr_lock);
352 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
353 aio_nr -= ctx->max_reqs;
354 spin_unlock(&aio_nr_lock);
356 pr_debug("freeing %p\n", ctx);
359 * Here the call_rcu() is between the wait_event() for reqs_active to
360 * hit 0, and freeing the ioctx.
362 * aio_complete() decrements reqs_active, but it has to touch the ioctx
363 * after to issue a wakeup so we use rcu.
365 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
368 static void put_ioctx(struct kioctx *ctx)
370 if (percpu_ref_put(&ctx->users))
375 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
377 static struct kioctx *ioctx_alloc(unsigned nr_events)
379 struct mm_struct *mm = current->mm;
383 /* Prevent overflows */
384 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
385 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
386 pr_debug("ENOMEM: nr_events too high\n");
387 return ERR_PTR(-EINVAL);
390 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
391 return ERR_PTR(-EAGAIN);
393 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
395 return ERR_PTR(-ENOMEM);
397 ctx->max_reqs = nr_events;
399 percpu_ref_init(&ctx->users);
401 percpu_ref_get(&ctx->users);
404 spin_lock_init(&ctx->ctx_lock);
405 mutex_init(&ctx->ring_lock);
406 init_waitqueue_head(&ctx->wait);
408 INIT_LIST_HEAD(&ctx->active_reqs);
410 ctx->cpu = alloc_percpu(struct kioctx_cpu);
414 if (aio_setup_ring(ctx) < 0)
417 atomic_set(&ctx->reqs_available, ctx->nr);
418 ctx->req_batch = ctx->nr / (num_possible_cpus() * 4);
419 BUG_ON(!ctx->req_batch);
421 /* limit the number of system wide aios */
422 spin_lock(&aio_nr_lock);
423 if (aio_nr + nr_events > aio_max_nr ||
424 aio_nr + nr_events < aio_nr) {
425 spin_unlock(&aio_nr_lock);
428 aio_nr += ctx->max_reqs;
429 spin_unlock(&aio_nr_lock);
431 /* now link into global list. */
432 spin_lock(&mm->ioctx_lock);
433 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
434 spin_unlock(&mm->ioctx_lock);
436 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
437 ctx, ctx->user_id, mm, ctx->nr);
444 free_percpu(ctx->cpu);
446 kmem_cache_free(kioctx_cachep, ctx);
447 pr_debug("error allocating ioctx %d\n", err);
451 static void kill_ioctx_work(struct work_struct *work)
453 struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
455 wake_up_all(&ctx->wait);
459 static void kill_ioctx_rcu(struct rcu_head *head)
461 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
463 INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
464 schedule_work(&ctx->rcu_work);
468 * Cancels all outstanding aio requests on an aio context. Used
469 * when the processes owning a context have all exited to encourage
470 * the rapid destruction of the kioctx.
472 static void kill_ioctx(struct kioctx *ctx)
474 if (percpu_ref_kill(&ctx->users)) {
475 hlist_del_rcu(&ctx->list);
476 /* Between hlist_del_rcu() and dropping the initial ref */
480 * We can't punt to workqueue here because put_ioctx() ->
481 * free_ioctx() will unmap the ringbuffer, and that has to be
482 * done in the original process's context. kill_ioctx_rcu/work()
483 * exist for exit_aio(), as in that path free_ioctx() won't do
486 kill_ioctx_work(&ctx->rcu_work);
490 /* wait_on_sync_kiocb:
491 * Waits on the given sync kiocb to complete.
493 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
495 while (atomic_read(&iocb->ki_users)) {
496 set_current_state(TASK_UNINTERRUPTIBLE);
497 if (!atomic_read(&iocb->ki_users))
501 __set_current_state(TASK_RUNNING);
502 return iocb->ki_user_data;
504 EXPORT_SYMBOL(wait_on_sync_kiocb);
507 * exit_aio: called when the last user of mm goes away. At this point, there is
508 * no way for any new requests to be submited or any of the io_* syscalls to be
509 * called on the context.
511 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
514 void exit_aio(struct mm_struct *mm)
517 struct hlist_node *n;
519 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
521 * We don't need to bother with munmap() here -
522 * exit_mmap(mm) is coming and it'll unmap everything.
523 * Since aio_free_ring() uses non-zero ->mmap_size
524 * as indicator that it needs to unmap the area,
525 * just set it to 0; aio_free_ring() is the only
526 * place that uses ->mmap_size, so it's safe.
530 if (percpu_ref_kill(&ctx->users)) {
531 hlist_del_rcu(&ctx->list);
532 call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
537 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
539 struct kioctx_cpu *kcpu;
542 kcpu = this_cpu_ptr(ctx->cpu);
544 kcpu->reqs_available += nr;
545 while (kcpu->reqs_available >= ctx->req_batch * 2) {
546 kcpu->reqs_available -= ctx->req_batch;
547 atomic_add(ctx->req_batch, &ctx->reqs_available);
553 static bool get_reqs_available(struct kioctx *ctx)
555 struct kioctx_cpu *kcpu;
559 kcpu = this_cpu_ptr(ctx->cpu);
561 if (!kcpu->reqs_available) {
562 int old, avail = atomic_read(&ctx->reqs_available);
565 if (avail < ctx->req_batch)
569 avail = atomic_cmpxchg(&ctx->reqs_available,
570 avail, avail - ctx->req_batch);
571 } while (avail != old);
573 kcpu->reqs_available += ctx->req_batch;
577 kcpu->reqs_available--;
584 * Allocate a slot for an aio request. Increments the ki_users count
585 * of the kioctx so that the kioctx stays around until all requests are
586 * complete. Returns NULL if no requests are free.
588 * Returns with kiocb->ki_users set to 2. The io submit code path holds
589 * an extra reference while submitting the i/o.
590 * This prevents races between the aio code path referencing the
591 * req (after submitting it) and aio_complete() freeing the req.
593 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
597 if (!get_reqs_available(ctx))
600 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
604 atomic_set(&req->ki_users, 2);
608 put_reqs_available(ctx, 1);
612 static void kiocb_free(struct kiocb *req)
616 if (req->ki_eventfd != NULL)
617 eventfd_ctx_put(req->ki_eventfd);
620 if (req->ki_iovec != &req->ki_inline_vec)
621 kfree(req->ki_iovec);
622 kmem_cache_free(kiocb_cachep, req);
625 void aio_put_req(struct kiocb *req)
627 if (atomic_dec_and_test(&req->ki_users))
630 EXPORT_SYMBOL(aio_put_req);
632 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
634 struct mm_struct *mm = current->mm;
635 struct kioctx *ctx, *ret = NULL;
639 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
640 if (ctx->user_id == ctx_id){
641 percpu_ref_get(&ctx->users);
651 static inline unsigned kioctx_ring_put(struct kioctx *ctx, struct kiocb *req,
654 struct io_event *ev_page, *event;
655 unsigned pos = tail + AIO_EVENTS_OFFSET;
657 if (++tail >= ctx->nr)
660 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
661 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
663 event->obj = (u64)(unsigned long)req->ki_obj.user;
664 event->data = req->ki_user_data;
665 event->res = req->ki_res;
666 event->res2 = req->ki_res2;
668 kunmap_atomic(ev_page);
669 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
671 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
672 ctx, tail, req, req->ki_obj.user, req->ki_user_data,
673 req->ki_res, req->ki_res2);
678 static inline unsigned kioctx_ring_lock(struct kioctx *ctx)
683 * ctx->tail is both our lock and the canonical version of the tail
686 while ((tail = xchg(&ctx->tail, UINT_MAX)) == UINT_MAX)
692 static inline void kioctx_ring_unlock(struct kioctx *ctx, unsigned tail)
694 struct aio_ring *ring;
700 /* make event visible before updating tail */
702 ctx->shadow_tail = tail;
704 ring = kmap_atomic(ctx->ring_pages[0]);
707 flush_dcache_page(ctx->ring_pages[0]);
709 /* unlock, make new tail visible before checking waitlist */
714 if (waitqueue_active(&ctx->wait))
718 void batch_complete_aio(struct batch_complete *batch)
720 struct kioctx *ctx = NULL;
721 struct eventfd_ctx *eventfd = NULL;
726 if (RB_EMPTY_ROOT(&batch->kiocb))
730 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
731 * need to issue a wakeup after incrementing reqs_available.
734 local_irq_save(flags);
736 n = rb_first(&batch->kiocb);
738 struct kiocb *req = container_of(n, struct kiocb, ki_node);
741 n->rb_right->__rb_parent_color = n->__rb_parent_color;
750 if (unlikely(xchg(&req->ki_cancel,
751 KIOCB_CANCELLED) == KIOCB_CANCELLED)) {
753 * Can't use the percpu reqs_available here - could race
756 atomic_inc(&req->ki_ctx->reqs_available);
761 if (unlikely(req->ki_eventfd != eventfd)) {
763 /* Make event visible */
764 kioctx_ring_unlock(ctx, tail);
767 eventfd_signal(eventfd, 1);
768 eventfd_ctx_put(eventfd);
771 eventfd = req->ki_eventfd;
772 req->ki_eventfd = NULL;
775 if (unlikely(req->ki_ctx != ctx)) {
776 kioctx_ring_unlock(ctx, tail);
779 tail = kioctx_ring_lock(ctx);
782 tail = kioctx_ring_put(ctx, req, tail);
786 kioctx_ring_unlock(ctx, tail);
787 local_irq_restore(flags);
791 * Check if the user asked us to deliver the result through an
792 * eventfd. The eventfd_signal() function is safe to be called
796 eventfd_signal(eventfd, 1);
797 eventfd_ctx_put(eventfd);
800 EXPORT_SYMBOL(batch_complete_aio);
802 /* aio_complete_batch
803 * Called when the io request on the given iocb is complete; @batch may be
806 void aio_complete_batch(struct kiocb *req, long res, long res2,
807 struct batch_complete *batch)
812 if (req->ki_list.next) {
813 struct kioctx *ctx = req->ki_ctx;
816 spin_lock_irqsave(&ctx->ctx_lock, flags);
817 list_del(&req->ki_list);
818 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
822 * Special case handling for sync iocbs:
823 * - events go directly into the iocb for fast handling
824 * - the sync task with the iocb in its stack holds the single iocb
825 * ref, no other paths have a way to get another ref
826 * - the sync task helpfully left a reference to itself in the iocb
828 if (is_sync_kiocb(req)) {
829 BUG_ON(atomic_read(&req->ki_users) != 1);
830 req->ki_user_data = req->ki_res;
831 atomic_set(&req->ki_users, 0);
832 wake_up_process(req->ki_obj.tsk);
836 struct rb_node **n = &batch->kiocb.rb_node, *parent = NULL;
840 t = container_of(*n, struct kiocb, ki_node);
842 res = req->ki_ctx != t->ki_ctx
843 ? req->ki_ctx < t->ki_ctx
844 : req->ki_eventfd != t->ki_eventfd
845 ? req->ki_eventfd < t->ki_eventfd
848 n = res ? &(*n)->rb_left : &(*n)->rb_right;
851 rb_link_node(&req->ki_node, parent, n);
852 rb_insert_color(&req->ki_node, &batch->kiocb);
854 struct batch_complete batch_stack;
856 memset(&req->ki_node, 0, sizeof(req->ki_node));
857 batch_stack.kiocb.rb_node = &req->ki_node;
859 batch_complete_aio(&batch_stack);
862 EXPORT_SYMBOL(aio_complete_batch);
865 * Pull an event off of the ioctx's event ring. Returns the number of
868 static long aio_read_events_ring(struct kioctx *ctx,
869 struct io_event __user *event, long nr)
871 struct aio_ring *ring;
876 mutex_lock(&ctx->ring_lock);
878 ring = kmap_atomic(ctx->ring_pages[0]);
882 pr_debug("h%u t%u m%u\n", head, ctx->shadow_tail, ctx->nr);
884 if (head == ctx->shadow_tail)
888 long avail = (head < ctx->shadow_tail
889 ? ctx->shadow_tail : ctx->nr) - head;
893 if (head == ctx->shadow_tail)
896 avail = min(avail, nr - ret);
897 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
898 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
900 pos = head + AIO_EVENTS_OFFSET;
901 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
902 pos %= AIO_EVENTS_PER_PAGE;
905 copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * avail);
908 if (unlikely(copy_ret)) {
918 ring = kmap_atomic(ctx->ring_pages[0]);
921 flush_dcache_page(ctx->ring_pages[0]);
923 pr_debug("%li h%u t%u\n", ret, head, ctx->shadow_tail);
925 put_reqs_available(ctx, ret);
927 mutex_unlock(&ctx->ring_lock);
932 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
933 struct io_event __user *event, long *i)
935 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
940 if (unlikely(percpu_ref_dead(&ctx->users)))
946 return ret < 0 || *i >= min_nr;
949 static long read_events(struct kioctx *ctx, long min_nr, long nr,
950 struct io_event __user *event,
951 struct timespec __user *timeout)
953 ktime_t until = { .tv64 = KTIME_MAX };
959 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
962 until = timespec_to_ktime(ts);
966 * Note that aio_read_events() is being called as the conditional - i.e.
967 * we're calling it after prepare_to_wait() has set task state to
968 * TASK_INTERRUPTIBLE.
970 * But aio_read_events() can block, and if it blocks it's going to flip
971 * the task state back to TASK_RUNNING.
973 * This should be ok, provided it doesn't flip the state back to
974 * TASK_RUNNING and return 0 too much - that causes us to spin. That
975 * will only happen if the mutex_lock() call blocks, and we then find
976 * the ringbuffer empty. So in practice we should be ok, but it's
977 * something to be aware of when touching this code.
979 wait_event_interruptible_hrtimeout(ctx->wait,
980 aio_read_events(ctx, min_nr, nr, event, &ret), until);
982 if (!ret && signal_pending(current))
989 * Create an aio_context capable of receiving at least nr_events.
990 * ctxp must not point to an aio_context that already exists, and
991 * must be initialized to 0 prior to the call. On successful
992 * creation of the aio_context, *ctxp is filled in with the resulting
993 * handle. May fail with -EINVAL if *ctxp is not initialized,
994 * if the specified nr_events exceeds internal limits. May fail
995 * with -EAGAIN if the specified nr_events exceeds the user's limit
996 * of available events. May fail with -ENOMEM if insufficient kernel
997 * resources are available. May fail with -EFAULT if an invalid
998 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1001 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1003 struct kioctx *ioctx = NULL;
1007 ret = get_user(ctx, ctxp);
1012 if (unlikely(ctx || nr_events == 0)) {
1013 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1018 ioctx = ioctx_alloc(nr_events);
1019 ret = PTR_ERR(ioctx);
1020 if (!IS_ERR(ioctx)) {
1021 ret = put_user(ioctx->user_id, ctxp);
1032 * Destroy the aio_context specified. May cancel any outstanding
1033 * AIOs and block on completion. Will fail with -ENOSYS if not
1034 * implemented. May fail with -EINVAL if the context pointed to
1037 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1039 struct kioctx *ioctx = lookup_ioctx(ctx);
1040 if (likely(NULL != ioctx)) {
1045 pr_debug("EINVAL: io_destroy: invalid context id\n");
1049 static void aio_advance_iovec(struct kiocb *iocb, ssize_t ret)
1051 struct iovec *iov = &iocb->ki_iovec[iocb->ki_cur_seg];
1055 while (iocb->ki_cur_seg < iocb->ki_nr_segs && ret > 0) {
1056 ssize_t this = min((ssize_t)iov->iov_len, ret);
1057 iov->iov_base += this;
1058 iov->iov_len -= this;
1059 iocb->ki_left -= this;
1061 if (iov->iov_len == 0) {
1067 /* the caller should not have done more io than what fit in
1068 * the remaining iovecs */
1069 BUG_ON(ret > 0 && iocb->ki_left == 0);
1072 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1073 unsigned long, loff_t);
1075 static ssize_t aio_rw_vect_retry(struct kiocb *iocb, int rw, aio_rw_op *rw_op)
1077 struct file *file = iocb->ki_filp;
1078 struct address_space *mapping = file->f_mapping;
1079 struct inode *inode = mapping->host;
1082 /* This matches the pread()/pwrite() logic */
1083 if (iocb->ki_pos < 0)
1087 ret = rw_op(iocb, &iocb->ki_iovec[iocb->ki_cur_seg],
1088 iocb->ki_nr_segs - iocb->ki_cur_seg,
1091 aio_advance_iovec(iocb, ret);
1093 /* retry all partial writes. retry partial reads as long as its a
1095 } while (ret > 0 && iocb->ki_left > 0 &&
1097 (!S_ISFIFO(inode->i_mode) && !S_ISSOCK(inode->i_mode))));
1099 /* This means we must have transferred all that we could */
1100 /* No need to retry anymore */
1101 if ((ret == 0) || (iocb->ki_left == 0))
1102 ret = iocb->ki_nbytes - iocb->ki_left;
1104 /* If we managed to write some out we return that, rather than
1105 * the eventual error. */
1107 && ret < 0 && ret != -EIOCBQUEUED
1108 && iocb->ki_nbytes - iocb->ki_left)
1109 ret = iocb->ki_nbytes - iocb->ki_left;
1114 static ssize_t aio_setup_vectored_rw(int rw, struct kiocb *kiocb, bool compat)
1118 kiocb->ki_nr_segs = kiocb->ki_nbytes;
1120 #ifdef CONFIG_COMPAT
1122 ret = compat_rw_copy_check_uvector(rw,
1123 (struct compat_iovec __user *)kiocb->ki_buf,
1124 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1128 ret = rw_copy_check_uvector(rw,
1129 (struct iovec __user *)kiocb->ki_buf,
1130 kiocb->ki_nr_segs, 1, &kiocb->ki_inline_vec,
1135 /* ki_nbytes now reflect bytes instead of segs */
1136 kiocb->ki_nbytes = ret;
1140 static ssize_t aio_setup_single_vector(int rw, struct kiocb *kiocb)
1142 if (unlikely(!access_ok(!rw, kiocb->ki_buf, kiocb->ki_nbytes)))
1145 kiocb->ki_iovec = &kiocb->ki_inline_vec;
1146 kiocb->ki_iovec->iov_base = kiocb->ki_buf;
1147 kiocb->ki_iovec->iov_len = kiocb->ki_nbytes;
1148 kiocb->ki_nr_segs = 1;
1154 * Performs the initial checks and aio retry method
1155 * setup for the kiocb at the time of io submission.
1157 static ssize_t aio_run_iocb(struct kiocb *req, bool compat)
1159 struct file *file = req->ki_filp;
1165 switch (req->ki_opcode) {
1166 case IOCB_CMD_PREAD:
1167 case IOCB_CMD_PREADV:
1170 rw_op = file->f_op->aio_read;
1173 case IOCB_CMD_PWRITE:
1174 case IOCB_CMD_PWRITEV:
1177 rw_op = file->f_op->aio_write;
1180 if (unlikely(!(file->f_mode & mode)))
1186 ret = (req->ki_opcode == IOCB_CMD_PREADV ||
1187 req->ki_opcode == IOCB_CMD_PWRITEV)
1188 ? aio_setup_vectored_rw(rw, req, compat)
1189 : aio_setup_single_vector(rw, req);
1193 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1197 req->ki_nbytes = ret;
1200 ret = aio_rw_vect_retry(req, rw, rw_op);
1203 case IOCB_CMD_FDSYNC:
1204 if (!file->f_op->aio_fsync)
1207 ret = file->f_op->aio_fsync(req, 1);
1210 case IOCB_CMD_FSYNC:
1211 if (!file->f_op->aio_fsync)
1214 ret = file->f_op->aio_fsync(req, 0);
1218 pr_debug("EINVAL: no operation provided\n");
1222 if (ret != -EIOCBQUEUED) {
1224 * There's no easy way to restart the syscall since other AIO's
1225 * may be already running. Just fail this IO with EINTR.
1227 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1228 ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
1230 aio_complete(req, ret, 0);
1236 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1237 struct iocb *iocb, bool compat)
1242 /* enforce forwards compatibility on users */
1243 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1244 pr_debug("EINVAL: reserve field set\n");
1248 /* prevent overflows */
1250 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1251 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1252 ((ssize_t)iocb->aio_nbytes < 0)
1254 pr_debug("EINVAL: io_submit: overflow check\n");
1258 req = aio_get_req(ctx);
1262 req->ki_filp = fget(iocb->aio_fildes);
1263 if (unlikely(!req->ki_filp)) {
1268 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1270 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1271 * instance of the file* now. The file descriptor must be
1272 * an eventfd() fd, and will be signaled for each completed
1273 * event using the eventfd_signal() function.
1275 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1276 if (IS_ERR(req->ki_eventfd)) {
1277 ret = PTR_ERR(req->ki_eventfd);
1278 req->ki_eventfd = NULL;
1283 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1284 if (unlikely(ret)) {
1285 pr_debug("EFAULT: aio_key\n");
1289 req->ki_obj.user = user_iocb;
1290 req->ki_user_data = iocb->aio_data;
1291 req->ki_pos = iocb->aio_offset;
1293 req->ki_buf = (char __user *)(unsigned long)iocb->aio_buf;
1294 req->ki_left = req->ki_nbytes = iocb->aio_nbytes;
1295 req->ki_opcode = iocb->aio_lio_opcode;
1297 ret = aio_run_iocb(req, compat);
1301 aio_put_req(req); /* drop extra ref to req */
1304 put_reqs_available(ctx, 1);
1305 aio_put_req(req); /* drop extra ref to req */
1306 aio_put_req(req); /* drop i/o ref to req */
1310 long do_io_submit(aio_context_t ctx_id, long nr,
1311 struct iocb __user *__user *iocbpp, bool compat)
1316 struct blk_plug plug;
1318 if (unlikely(nr < 0))
1321 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1322 nr = LONG_MAX/sizeof(*iocbpp);
1324 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1327 ctx = lookup_ioctx(ctx_id);
1328 if (unlikely(!ctx)) {
1329 pr_debug("EINVAL: invalid context id\n");
1333 blk_start_plug(&plug);
1336 * AKPM: should this return a partial result if some of the IOs were
1337 * successfully submitted?
1339 for (i=0; i<nr; i++) {
1340 struct iocb __user *user_iocb;
1343 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1348 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1353 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1357 blk_finish_plug(&plug);
1364 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1365 * the number of iocbs queued. May return -EINVAL if the aio_context
1366 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1367 * *iocbpp[0] is not properly initialized, if the operation specified
1368 * is invalid for the file descriptor in the iocb. May fail with
1369 * -EFAULT if any of the data structures point to invalid data. May
1370 * fail with -EBADF if the file descriptor specified in the first
1371 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1372 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1373 * fail with -ENOSYS if not implemented.
1375 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1376 struct iocb __user * __user *, iocbpp)
1378 return do_io_submit(ctx_id, nr, iocbpp, 0);
1382 * Finds a given iocb for cancellation.
1384 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1387 struct list_head *pos;
1389 assert_spin_locked(&ctx->ctx_lock);
1391 if (key != KIOCB_KEY)
1394 /* TODO: use a hash or array, this sucks. */
1395 list_for_each(pos, &ctx->active_reqs) {
1396 struct kiocb *kiocb = list_kiocb(pos);
1397 if (kiocb->ki_obj.user == iocb)
1404 * Attempts to cancel an iocb previously passed to io_submit. If
1405 * the operation is successfully cancelled, the resulting event is
1406 * copied into the memory pointed to by result without being placed
1407 * into the completion queue and 0 is returned. May fail with
1408 * -EFAULT if any of the data structures pointed to are invalid.
1409 * May fail with -EINVAL if aio_context specified by ctx_id is
1410 * invalid. May fail with -EAGAIN if the iocb specified was not
1411 * cancelled. Will fail with -ENOSYS if not implemented.
1413 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1414 struct io_event __user *, result)
1416 struct io_event res;
1418 struct kiocb *kiocb;
1422 ret = get_user(key, &iocb->aio_key);
1426 ctx = lookup_ioctx(ctx_id);
1430 spin_lock_irq(&ctx->ctx_lock);
1432 kiocb = lookup_kiocb(ctx, iocb, key);
1434 ret = kiocb_cancel(ctx, kiocb, &res);
1438 spin_unlock_irq(&ctx->ctx_lock);
1441 /* Cancellation succeeded -- copy the result
1442 * into the user's buffer.
1444 if (copy_to_user(result, &res, sizeof(res)))
1454 * Attempts to read at least min_nr events and up to nr events from
1455 * the completion queue for the aio_context specified by ctx_id. If
1456 * it succeeds, the number of read events is returned. May fail with
1457 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1458 * out of range, if timeout is out of range. May fail with -EFAULT
1459 * if any of the memory specified is invalid. May return 0 or
1460 * < min_nr if the timeout specified by timeout has elapsed
1461 * before sufficient events are available, where timeout == NULL
1462 * specifies an infinite timeout. Note that the timeout pointed to by
1463 * timeout is relative and will be updated if not NULL and the
1464 * operation blocks. Will fail with -ENOSYS if not implemented.
1466 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1469 struct io_event __user *, events,
1470 struct timespec __user *, timeout)
1472 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1475 if (likely(ioctx)) {
1476 if (likely(min_nr <= nr && min_nr >= 0))
1477 ret = read_events(ioctx, min_nr, nr, events, timeout);