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/anon_inodes.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
64 struct io_event io_events[0];
65 }; /* 128 bytes + ring size */
67 #define AIO_RING_PAGES 8
70 unsigned reqs_available;
74 struct percpu_ref users;
77 /* This needs improving */
78 unsigned long user_id;
79 struct hlist_node list;
81 struct __percpu kioctx_cpu *cpu;
84 * For percpu reqs_available, number of slots we move to/from global
89 * This is what userspace passed to io_setup(), it's not used for
90 * anything but counting against the global max_reqs quota.
92 * The real limit is nr_events - 1, which will be larger (see
97 /* Size of ringbuffer, in units of struct io_event */
100 unsigned long mmap_base;
101 unsigned long mmap_size;
103 struct page **ring_pages;
106 struct rcu_head rcu_head;
107 struct work_struct free_work;
111 * This counts the number of available slots in the ringbuffer,
112 * so we avoid overflowing it: it's decremented (if positive)
113 * when allocating a kiocb and incremented when the resulting
114 * io_event is pulled off the ringbuffer.
116 * We batch accesses to it with a percpu version.
118 atomic_t reqs_available;
119 } ____cacheline_aligned_in_smp;
123 struct list_head active_reqs; /* used for cancellation */
124 } ____cacheline_aligned_in_smp;
127 struct mutex ring_lock;
128 wait_queue_head_t wait;
129 } ____cacheline_aligned_in_smp;
133 spinlock_t completion_lock;
134 } ____cacheline_aligned_in_smp;
136 struct page *internal_pages[AIO_RING_PAGES];
137 struct file *aio_ring_file;
140 /*------ sysctl variables----*/
141 static DEFINE_SPINLOCK(aio_nr_lock);
142 unsigned long aio_nr; /* current system wide number of aio requests */
143 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
144 /*----end sysctl variables---*/
146 static struct kmem_cache *kiocb_cachep;
147 static struct kmem_cache *kioctx_cachep;
150 * Creates the slab caches used by the aio routines, panic on
151 * failure as this is done early during the boot sequence.
153 static int __init aio_setup(void)
155 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
156 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
158 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
162 __initcall(aio_setup);
164 static void aio_free_ring(struct kioctx *ctx)
167 struct file *aio_ring_file = ctx->aio_ring_file;
169 for (i = 0; i < ctx->nr_pages; i++) {
170 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
171 page_count(ctx->ring_pages[i]));
172 put_page(ctx->ring_pages[i]);
175 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
176 kfree(ctx->ring_pages);
179 truncate_setsize(aio_ring_file->f_inode, 0);
180 pr_debug("pid(%d) i_nlink=%u d_count=%d d_unhashed=%d i_count=%d\n",
181 current->pid, aio_ring_file->f_inode->i_nlink,
182 aio_ring_file->f_path.dentry->d_count,
183 d_unhashed(aio_ring_file->f_path.dentry),
184 atomic_read(&aio_ring_file->f_inode->i_count));
186 ctx->aio_ring_file = NULL;
190 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
192 vma->vm_ops = &generic_file_vm_ops;
196 static const struct file_operations aio_ring_fops = {
197 .mmap = aio_ring_mmap,
200 static int aio_set_page_dirty(struct page *page)
205 #if IS_ENABLED(CONFIG_MIGRATION)
206 static int aio_migratepage(struct address_space *mapping, struct page *new,
207 struct page *old, enum migrate_mode mode)
209 struct kioctx *ctx = mapping->private_data;
211 unsigned idx = old->index;
214 /* Writeback must be complete */
215 BUG_ON(PageWriteback(old));
218 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
219 if (rc != MIGRATEPAGE_SUCCESS) {
226 spin_lock_irqsave(&ctx->completion_lock, flags);
227 migrate_page_copy(new, old);
228 ctx->ring_pages[idx] = new;
229 spin_unlock_irqrestore(&ctx->completion_lock, flags);
235 static const struct address_space_operations aio_ctx_aops = {
236 .set_page_dirty = aio_set_page_dirty,
237 #if IS_ENABLED(CONFIG_MIGRATION)
238 .migratepage = aio_migratepage,
242 static int aio_setup_ring(struct kioctx *ctx)
244 struct aio_ring *ring;
245 unsigned nr_events = ctx->max_reqs;
246 struct mm_struct *mm = current->mm;
247 unsigned long size, populate;
252 /* Compensate for the ring buffer's head/tail overlap entry */
253 nr_events += 2; /* 1 is required, 2 for good luck */
255 size = sizeof(struct aio_ring);
256 size += sizeof(struct io_event) * nr_events;
258 nr_pages = PFN_UP(size);
262 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
264 ctx->aio_ring_file = NULL;
268 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
269 file->f_inode->i_mapping->private_data = ctx;
270 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
272 for (i = 0; i < nr_pages; i++) {
274 page = find_or_create_page(file->f_inode->i_mapping,
275 i, GFP_HIGHUSER | __GFP_ZERO);
278 pr_debug("pid(%d) page[%d]->count=%d\n",
279 current->pid, i, page_count(page));
280 SetPageUptodate(page);
284 ctx->aio_ring_file = file;
285 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
286 / sizeof(struct io_event);
288 ctx->ring_pages = ctx->internal_pages;
289 if (nr_pages > AIO_RING_PAGES) {
290 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
292 if (!ctx->ring_pages)
296 ctx->mmap_size = nr_pages * PAGE_SIZE;
297 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
299 down_write(&mm->mmap_sem);
300 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
301 PROT_READ | PROT_WRITE,
302 MAP_SHARED | MAP_POPULATE, 0, &populate);
303 if (IS_ERR((void *)ctx->mmap_base)) {
304 up_write(&mm->mmap_sem);
309 up_write(&mm->mmap_sem);
311 mm_populate(ctx->mmap_base, populate);
313 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
314 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
315 1, 0, ctx->ring_pages, NULL);
316 for (i = 0; i < ctx->nr_pages; i++)
317 put_page(ctx->ring_pages[i]);
319 if (unlikely(ctx->nr_pages != nr_pages)) {
324 ctx->user_id = ctx->mmap_base;
325 ctx->nr_events = nr_events; /* trusted copy */
327 ring = kmap_atomic(ctx->ring_pages[0]);
328 ring->nr = nr_events; /* user copy */
329 ring->id = ctx->user_id;
330 ring->head = ring->tail = 0;
331 ring->magic = AIO_RING_MAGIC;
332 ring->compat_features = AIO_RING_COMPAT_FEATURES;
333 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
334 ring->header_length = sizeof(struct aio_ring);
336 flush_dcache_page(ctx->ring_pages[0]);
341 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
342 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
343 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
345 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
347 struct kioctx *ctx = req->ki_ctx;
350 spin_lock_irqsave(&ctx->ctx_lock, flags);
352 if (!req->ki_list.next)
353 list_add(&req->ki_list, &ctx->active_reqs);
355 req->ki_cancel = cancel;
357 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
359 EXPORT_SYMBOL(kiocb_set_cancel_fn);
361 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
363 kiocb_cancel_fn *old, *cancel;
366 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
367 * actually has a cancel function, hence the cmpxchg()
370 cancel = ACCESS_ONCE(kiocb->ki_cancel);
372 if (!cancel || cancel == KIOCB_CANCELLED)
376 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
377 } while (cancel != old);
379 return cancel(kiocb);
382 static void free_ioctx_rcu(struct rcu_head *head)
384 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
386 free_percpu(ctx->cpu);
387 kmem_cache_free(kioctx_cachep, ctx);
391 * When this function runs, the kioctx has been removed from the "hash table"
392 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
393 * now it's safe to cancel any that need to be.
395 static void free_ioctx(struct work_struct *work)
397 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
398 struct aio_ring *ring;
403 spin_lock_irq(&ctx->ctx_lock);
405 while (!list_empty(&ctx->active_reqs)) {
406 req = list_first_entry(&ctx->active_reqs,
407 struct kiocb, ki_list);
409 list_del_init(&req->ki_list);
410 kiocb_cancel(ctx, req);
413 spin_unlock_irq(&ctx->ctx_lock);
415 for_each_possible_cpu(cpu) {
416 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
418 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
419 kcpu->reqs_available = 0;
423 prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
425 ring = kmap_atomic(ctx->ring_pages[0]);
426 avail = (ring->head <= ring->tail)
427 ? ring->tail - ring->head
428 : ctx->nr_events - ring->head + ring->tail;
430 atomic_add(avail, &ctx->reqs_available);
431 ring->head = ring->tail;
434 if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
439 finish_wait(&ctx->wait, &wait);
441 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
445 pr_debug("freeing %p\n", ctx);
448 * Here the call_rcu() is between the wait_event() for reqs_active to
449 * hit 0, and freeing the ioctx.
451 * aio_complete() decrements reqs_active, but it has to touch the ioctx
452 * after to issue a wakeup so we use rcu.
454 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
457 static void free_ioctx_ref(struct percpu_ref *ref)
459 struct kioctx *ctx = container_of(ref, struct kioctx, users);
461 INIT_WORK(&ctx->free_work, free_ioctx);
462 schedule_work(&ctx->free_work);
466 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
468 static struct kioctx *ioctx_alloc(unsigned nr_events)
470 struct mm_struct *mm = current->mm;
475 * We keep track of the number of available ringbuffer slots, to prevent
476 * overflow (reqs_available), and we also use percpu counters for this.
478 * So since up to half the slots might be on other cpu's percpu counters
479 * and unavailable, double nr_events so userspace sees what they
480 * expected: additionally, we move req_batch slots to/from percpu
481 * counters at a time, so make sure that isn't 0:
483 nr_events = max(nr_events, num_possible_cpus() * 4);
486 /* Prevent overflows */
487 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
488 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
489 pr_debug("ENOMEM: nr_events too high\n");
490 return ERR_PTR(-EINVAL);
493 if (!nr_events || (unsigned long)nr_events > aio_max_nr)
494 return ERR_PTR(-EAGAIN);
496 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
498 return ERR_PTR(-ENOMEM);
500 ctx->max_reqs = nr_events;
502 if (percpu_ref_init(&ctx->users, free_ioctx_ref))
505 spin_lock_init(&ctx->ctx_lock);
506 spin_lock_init(&ctx->completion_lock);
507 mutex_init(&ctx->ring_lock);
508 init_waitqueue_head(&ctx->wait);
510 INIT_LIST_HEAD(&ctx->active_reqs);
512 ctx->cpu = alloc_percpu(struct kioctx_cpu);
516 if (aio_setup_ring(ctx) < 0)
519 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
520 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
521 BUG_ON(!ctx->req_batch);
523 /* limit the number of system wide aios */
524 spin_lock(&aio_nr_lock);
525 if (aio_nr + nr_events > aio_max_nr ||
526 aio_nr + nr_events < aio_nr) {
527 spin_unlock(&aio_nr_lock);
530 aio_nr += ctx->max_reqs;
531 spin_unlock(&aio_nr_lock);
533 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
535 /* now link into global list. */
536 spin_lock(&mm->ioctx_lock);
537 hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
538 spin_unlock(&mm->ioctx_lock);
540 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
541 ctx, ctx->user_id, mm, ctx->nr_events);
548 free_percpu(ctx->cpu);
550 free_percpu(ctx->users.pcpu_count);
552 if (ctx->aio_ring_file)
553 fput(ctx->aio_ring_file);
554 kmem_cache_free(kioctx_cachep, ctx);
555 pr_debug("error allocating ioctx %d\n", err);
560 * Cancels all outstanding aio requests on an aio context. Used
561 * when the processes owning a context have all exited to encourage
562 * the rapid destruction of the kioctx.
564 static void kill_ioctx(struct kioctx *ctx)
566 if (!atomic_xchg(&ctx->dead, 1)) {
567 hlist_del_rcu(&ctx->list);
568 /* percpu_ref_kill() will do the necessary call_rcu() */
569 wake_up_all(&ctx->wait);
572 * It'd be more correct to do this in free_ioctx(), after all
573 * the outstanding kiocbs have finished - but by then io_destroy
574 * has already returned, so io_setup() could potentially return
575 * -EAGAIN with no ioctxs actually in use (as far as userspace
578 spin_lock(&aio_nr_lock);
579 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
580 aio_nr -= ctx->max_reqs;
581 spin_unlock(&aio_nr_lock);
584 vm_munmap(ctx->mmap_base, ctx->mmap_size);
586 percpu_ref_kill(&ctx->users);
590 /* wait_on_sync_kiocb:
591 * Waits on the given sync kiocb to complete.
593 ssize_t wait_on_sync_kiocb(struct kiocb *req)
595 while (!req->ki_ctx) {
596 set_current_state(TASK_UNINTERRUPTIBLE);
601 __set_current_state(TASK_RUNNING);
602 return req->ki_user_data;
604 EXPORT_SYMBOL(wait_on_sync_kiocb);
607 * exit_aio: called when the last user of mm goes away. At this point, there is
608 * no way for any new requests to be submited or any of the io_* syscalls to be
609 * called on the context.
611 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
614 void exit_aio(struct mm_struct *mm)
617 struct hlist_node *n;
619 hlist_for_each_entry_safe(ctx, n, &mm->ioctx_list, list) {
621 * We don't need to bother with munmap() here -
622 * exit_mmap(mm) is coming and it'll unmap everything.
623 * Since aio_free_ring() uses non-zero ->mmap_size
624 * as indicator that it needs to unmap the area,
625 * just set it to 0; aio_free_ring() is the only
626 * place that uses ->mmap_size, so it's safe.
634 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
636 struct kioctx_cpu *kcpu;
639 kcpu = this_cpu_ptr(ctx->cpu);
641 kcpu->reqs_available += nr;
642 while (kcpu->reqs_available >= ctx->req_batch * 2) {
643 kcpu->reqs_available -= ctx->req_batch;
644 atomic_add(ctx->req_batch, &ctx->reqs_available);
650 static bool get_reqs_available(struct kioctx *ctx)
652 struct kioctx_cpu *kcpu;
656 kcpu = this_cpu_ptr(ctx->cpu);
658 if (!kcpu->reqs_available) {
659 int old, avail = atomic_read(&ctx->reqs_available);
662 if (avail < ctx->req_batch)
666 avail = atomic_cmpxchg(&ctx->reqs_available,
667 avail, avail - ctx->req_batch);
668 } while (avail != old);
670 kcpu->reqs_available += ctx->req_batch;
674 kcpu->reqs_available--;
681 * Allocate a slot for an aio request.
682 * Returns NULL if no requests are free.
684 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
688 if (!get_reqs_available(ctx))
691 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
698 put_reqs_available(ctx, 1);
702 static void kiocb_free(struct kiocb *req)
706 if (req->ki_eventfd != NULL)
707 eventfd_ctx_put(req->ki_eventfd);
708 kmem_cache_free(kiocb_cachep, req);
711 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
713 struct mm_struct *mm = current->mm;
714 struct kioctx *ctx, *ret = NULL;
718 hlist_for_each_entry_rcu(ctx, &mm->ioctx_list, list) {
719 if (ctx->user_id == ctx_id) {
720 percpu_ref_get(&ctx->users);
731 * Called when the io request on the given iocb is complete.
733 void aio_complete(struct kiocb *iocb, long res, long res2)
735 struct kioctx *ctx = iocb->ki_ctx;
736 struct aio_ring *ring;
737 struct io_event *ev_page, *event;
742 * Special case handling for sync iocbs:
743 * - events go directly into the iocb for fast handling
744 * - the sync task with the iocb in its stack holds the single iocb
745 * ref, no other paths have a way to get another ref
746 * - the sync task helpfully left a reference to itself in the iocb
748 if (is_sync_kiocb(iocb)) {
749 iocb->ki_user_data = res;
751 iocb->ki_ctx = ERR_PTR(-EXDEV);
752 wake_up_process(iocb->ki_obj.tsk);
757 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
758 * need to issue a wakeup after incrementing reqs_available.
762 if (iocb->ki_list.next) {
765 spin_lock_irqsave(&ctx->ctx_lock, flags);
766 list_del(&iocb->ki_list);
767 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
771 * Add a completion event to the ring buffer. Must be done holding
772 * ctx->completion_lock to prevent other code from messing with the tail
773 * pointer since we might be called from irq context.
775 spin_lock_irqsave(&ctx->completion_lock, flags);
778 pos = tail + AIO_EVENTS_OFFSET;
780 if (++tail >= ctx->nr_events)
783 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
784 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
786 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
787 event->data = iocb->ki_user_data;
791 kunmap_atomic(ev_page);
792 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
794 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
795 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
798 /* after flagging the request as done, we
799 * must never even look at it again
801 smp_wmb(); /* make event visible before updating tail */
805 ring = kmap_atomic(ctx->ring_pages[0]);
808 flush_dcache_page(ctx->ring_pages[0]);
810 spin_unlock_irqrestore(&ctx->completion_lock, flags);
812 pr_debug("added to ring %p at [%u]\n", iocb, tail);
815 * Check if the user asked us to deliver the result through an
816 * eventfd. The eventfd_signal() function is safe to be called
819 if (iocb->ki_eventfd != NULL)
820 eventfd_signal(iocb->ki_eventfd, 1);
822 /* everything turned out well, dispose of the aiocb. */
826 * We have to order our ring_info tail store above and test
827 * of the wait list below outside the wait lock. This is
828 * like in wake_up_bit() where clearing a bit has to be
829 * ordered with the unlocked test.
833 if (waitqueue_active(&ctx->wait))
838 EXPORT_SYMBOL(aio_complete);
841 * Pull an event off of the ioctx's event ring. Returns the number of
844 static long aio_read_events_ring(struct kioctx *ctx,
845 struct io_event __user *event, long nr)
847 struct aio_ring *ring;
848 unsigned head, tail, pos;
852 mutex_lock(&ctx->ring_lock);
854 ring = kmap_atomic(ctx->ring_pages[0]);
859 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
869 avail = (head <= tail ? tail : ctx->nr_events) - head;
873 avail = min(avail, nr - ret);
874 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
875 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
877 pos = head + AIO_EVENTS_OFFSET;
878 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
879 pos %= AIO_EVENTS_PER_PAGE;
882 copy_ret = copy_to_user(event + ret, ev + pos,
883 sizeof(*ev) * avail);
886 if (unlikely(copy_ret)) {
893 head %= ctx->nr_events;
896 ring = kmap_atomic(ctx->ring_pages[0]);
899 flush_dcache_page(ctx->ring_pages[0]);
901 pr_debug("%li h%u t%u\n", ret, head, tail);
903 put_reqs_available(ctx, ret);
905 mutex_unlock(&ctx->ring_lock);
910 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
911 struct io_event __user *event, long *i)
913 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
918 if (unlikely(atomic_read(&ctx->dead)))
924 return ret < 0 || *i >= min_nr;
927 static long read_events(struct kioctx *ctx, long min_nr, long nr,
928 struct io_event __user *event,
929 struct timespec __user *timeout)
931 ktime_t until = { .tv64 = KTIME_MAX };
937 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
940 until = timespec_to_ktime(ts);
944 * Note that aio_read_events() is being called as the conditional - i.e.
945 * we're calling it after prepare_to_wait() has set task state to
946 * TASK_INTERRUPTIBLE.
948 * But aio_read_events() can block, and if it blocks it's going to flip
949 * the task state back to TASK_RUNNING.
951 * This should be ok, provided it doesn't flip the state back to
952 * TASK_RUNNING and return 0 too much - that causes us to spin. That
953 * will only happen if the mutex_lock() call blocks, and we then find
954 * the ringbuffer empty. So in practice we should be ok, but it's
955 * something to be aware of when touching this code.
957 wait_event_interruptible_hrtimeout(ctx->wait,
958 aio_read_events(ctx, min_nr, nr, event, &ret), until);
960 if (!ret && signal_pending(current))
967 * Create an aio_context capable of receiving at least nr_events.
968 * ctxp must not point to an aio_context that already exists, and
969 * must be initialized to 0 prior to the call. On successful
970 * creation of the aio_context, *ctxp is filled in with the resulting
971 * handle. May fail with -EINVAL if *ctxp is not initialized,
972 * if the specified nr_events exceeds internal limits. May fail
973 * with -EAGAIN if the specified nr_events exceeds the user's limit
974 * of available events. May fail with -ENOMEM if insufficient kernel
975 * resources are available. May fail with -EFAULT if an invalid
976 * pointer is passed for ctxp. Will fail with -ENOSYS if not
979 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
981 struct kioctx *ioctx = NULL;
985 ret = get_user(ctx, ctxp);
990 if (unlikely(ctx || nr_events == 0)) {
991 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
996 ioctx = ioctx_alloc(nr_events);
997 ret = PTR_ERR(ioctx);
998 if (!IS_ERR(ioctx)) {
999 ret = put_user(ioctx->user_id, ctxp);
1002 percpu_ref_put(&ioctx->users);
1010 * Destroy the aio_context specified. May cancel any outstanding
1011 * AIOs and block on completion. Will fail with -ENOSYS if not
1012 * implemented. May fail with -EINVAL if the context pointed to
1015 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1017 struct kioctx *ioctx = lookup_ioctx(ctx);
1018 if (likely(NULL != ioctx)) {
1020 percpu_ref_put(&ioctx->users);
1023 pr_debug("EINVAL: io_destroy: invalid context id\n");
1027 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1028 unsigned long, loff_t);
1030 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1031 int rw, char __user *buf,
1032 unsigned long *nr_segs,
1033 struct iovec **iovec,
1038 *nr_segs = kiocb->ki_nbytes;
1040 #ifdef CONFIG_COMPAT
1042 ret = compat_rw_copy_check_uvector(rw,
1043 (struct compat_iovec __user *)buf,
1044 *nr_segs, 1, *iovec, iovec);
1047 ret = rw_copy_check_uvector(rw,
1048 (struct iovec __user *)buf,
1049 *nr_segs, 1, *iovec, iovec);
1053 /* ki_nbytes now reflect bytes instead of segs */
1054 kiocb->ki_nbytes = ret;
1058 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1059 int rw, char __user *buf,
1060 unsigned long *nr_segs,
1061 struct iovec *iovec)
1063 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1066 iovec->iov_base = buf;
1067 iovec->iov_len = kiocb->ki_nbytes;
1074 * Performs the initial checks and aio retry method
1075 * setup for the kiocb at the time of io submission.
1077 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1078 char __user *buf, bool compat)
1080 struct file *file = req->ki_filp;
1082 unsigned long nr_segs;
1086 struct iovec inline_vec, *iovec = &inline_vec;
1089 case IOCB_CMD_PREAD:
1090 case IOCB_CMD_PREADV:
1093 rw_op = file->f_op->aio_read;
1096 case IOCB_CMD_PWRITE:
1097 case IOCB_CMD_PWRITEV:
1100 rw_op = file->f_op->aio_write;
1103 if (unlikely(!(file->f_mode & mode)))
1109 ret = (opcode == IOCB_CMD_PREADV ||
1110 opcode == IOCB_CMD_PWRITEV)
1111 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1113 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1118 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1120 if (iovec != &inline_vec)
1125 req->ki_nbytes = ret;
1127 /* XXX: move/kill - rw_verify_area()? */
1128 /* This matches the pread()/pwrite() logic */
1129 if (req->ki_pos < 0) {
1135 file_start_write(file);
1137 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1140 file_end_write(file);
1143 case IOCB_CMD_FDSYNC:
1144 if (!file->f_op->aio_fsync)
1147 ret = file->f_op->aio_fsync(req, 1);
1150 case IOCB_CMD_FSYNC:
1151 if (!file->f_op->aio_fsync)
1154 ret = file->f_op->aio_fsync(req, 0);
1158 pr_debug("EINVAL: no operation provided\n");
1162 if (iovec != &inline_vec)
1165 if (ret != -EIOCBQUEUED) {
1167 * There's no easy way to restart the syscall since other AIO's
1168 * may be already running. Just fail this IO with EINTR.
1170 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1171 ret == -ERESTARTNOHAND ||
1172 ret == -ERESTART_RESTARTBLOCK))
1174 aio_complete(req, ret, 0);
1180 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1181 struct iocb *iocb, bool compat)
1186 /* enforce forwards compatibility on users */
1187 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1188 pr_debug("EINVAL: reserve field set\n");
1192 /* prevent overflows */
1194 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1195 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1196 ((ssize_t)iocb->aio_nbytes < 0)
1198 pr_debug("EINVAL: io_submit: overflow check\n");
1202 req = aio_get_req(ctx);
1206 req->ki_filp = fget(iocb->aio_fildes);
1207 if (unlikely(!req->ki_filp)) {
1212 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1214 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1215 * instance of the file* now. The file descriptor must be
1216 * an eventfd() fd, and will be signaled for each completed
1217 * event using the eventfd_signal() function.
1219 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1220 if (IS_ERR(req->ki_eventfd)) {
1221 ret = PTR_ERR(req->ki_eventfd);
1222 req->ki_eventfd = NULL;
1227 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1228 if (unlikely(ret)) {
1229 pr_debug("EFAULT: aio_key\n");
1233 req->ki_obj.user = user_iocb;
1234 req->ki_user_data = iocb->aio_data;
1235 req->ki_pos = iocb->aio_offset;
1236 req->ki_nbytes = iocb->aio_nbytes;
1238 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1239 (char __user *)(unsigned long)iocb->aio_buf,
1246 put_reqs_available(ctx, 1);
1251 long do_io_submit(aio_context_t ctx_id, long nr,
1252 struct iocb __user *__user *iocbpp, bool compat)
1257 struct blk_plug plug;
1259 if (unlikely(nr < 0))
1262 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1263 nr = LONG_MAX/sizeof(*iocbpp);
1265 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1268 ctx = lookup_ioctx(ctx_id);
1269 if (unlikely(!ctx)) {
1270 pr_debug("EINVAL: invalid context id\n");
1274 blk_start_plug(&plug);
1277 * AKPM: should this return a partial result if some of the IOs were
1278 * successfully submitted?
1280 for (i=0; i<nr; i++) {
1281 struct iocb __user *user_iocb;
1284 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1289 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1294 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1298 blk_finish_plug(&plug);
1300 percpu_ref_put(&ctx->users);
1305 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1306 * the number of iocbs queued. May return -EINVAL if the aio_context
1307 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1308 * *iocbpp[0] is not properly initialized, if the operation specified
1309 * is invalid for the file descriptor in the iocb. May fail with
1310 * -EFAULT if any of the data structures point to invalid data. May
1311 * fail with -EBADF if the file descriptor specified in the first
1312 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1313 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1314 * fail with -ENOSYS if not implemented.
1316 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1317 struct iocb __user * __user *, iocbpp)
1319 return do_io_submit(ctx_id, nr, iocbpp, 0);
1323 * Finds a given iocb for cancellation.
1325 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1328 struct list_head *pos;
1330 assert_spin_locked(&ctx->ctx_lock);
1332 if (key != KIOCB_KEY)
1335 /* TODO: use a hash or array, this sucks. */
1336 list_for_each(pos, &ctx->active_reqs) {
1337 struct kiocb *kiocb = list_kiocb(pos);
1338 if (kiocb->ki_obj.user == iocb)
1345 * Attempts to cancel an iocb previously passed to io_submit. If
1346 * the operation is successfully cancelled, the resulting event is
1347 * copied into the memory pointed to by result without being placed
1348 * into the completion queue and 0 is returned. May fail with
1349 * -EFAULT if any of the data structures pointed to are invalid.
1350 * May fail with -EINVAL if aio_context specified by ctx_id is
1351 * invalid. May fail with -EAGAIN if the iocb specified was not
1352 * cancelled. Will fail with -ENOSYS if not implemented.
1354 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1355 struct io_event __user *, result)
1358 struct kiocb *kiocb;
1362 ret = get_user(key, &iocb->aio_key);
1366 ctx = lookup_ioctx(ctx_id);
1370 spin_lock_irq(&ctx->ctx_lock);
1372 kiocb = lookup_kiocb(ctx, iocb, key);
1374 ret = kiocb_cancel(ctx, kiocb);
1378 spin_unlock_irq(&ctx->ctx_lock);
1382 * The result argument is no longer used - the io_event is
1383 * always delivered via the ring buffer. -EINPROGRESS indicates
1384 * cancellation is progress:
1389 percpu_ref_put(&ctx->users);
1395 * Attempts to read at least min_nr events and up to nr events from
1396 * the completion queue for the aio_context specified by ctx_id. If
1397 * it succeeds, the number of read events is returned. May fail with
1398 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1399 * out of range, if timeout is out of range. May fail with -EFAULT
1400 * if any of the memory specified is invalid. May return 0 or
1401 * < min_nr if the timeout specified by timeout has elapsed
1402 * before sufficient events are available, where timeout == NULL
1403 * specifies an infinite timeout. Note that the timeout pointed to by
1404 * timeout is relative. Will fail with -ENOSYS if not implemented.
1406 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1409 struct io_event __user *, events,
1410 struct timespec __user *, timeout)
1412 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1415 if (likely(ioctx)) {
1416 if (likely(min_nr <= nr && min_nr >= 0))
1417 ret = read_events(ioctx, min_nr, nr, events, timeout);
1418 percpu_ref_put(&ioctx->users);