2 * An async IO implementation for Linux
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
72 struct kioctx *table[];
76 unsigned reqs_available;
80 struct percpu_ref users;
83 unsigned long user_id;
85 struct __percpu kioctx_cpu *cpu;
88 * For percpu reqs_available, number of slots we move to/from global
93 * This is what userspace passed to io_setup(), it's not used for
94 * anything but counting against the global max_reqs quota.
96 * The real limit is nr_events - 1, which will be larger (see
101 /* Size of ringbuffer, in units of struct io_event */
104 unsigned long mmap_base;
105 unsigned long mmap_size;
107 struct page **ring_pages;
110 struct rcu_head rcu_head;
111 struct work_struct free_work;
115 * This counts the number of available slots in the ringbuffer,
116 * so we avoid overflowing it: it's decremented (if positive)
117 * when allocating a kiocb and incremented when the resulting
118 * io_event is pulled off the ringbuffer.
120 * We batch accesses to it with a percpu version.
122 atomic_t reqs_available;
123 } ____cacheline_aligned_in_smp;
127 struct list_head active_reqs; /* used for cancellation */
128 } ____cacheline_aligned_in_smp;
131 struct mutex ring_lock;
132 wait_queue_head_t wait;
133 } ____cacheline_aligned_in_smp;
137 spinlock_t completion_lock;
138 } ____cacheline_aligned_in_smp;
140 struct page *internal_pages[AIO_RING_PAGES];
141 struct file *aio_ring_file;
146 /*------ sysctl variables----*/
147 static DEFINE_SPINLOCK(aio_nr_lock);
148 unsigned long aio_nr; /* current system wide number of aio requests */
149 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
150 /*----end sysctl variables---*/
152 static struct kmem_cache *kiocb_cachep;
153 static struct kmem_cache *kioctx_cachep;
156 * Creates the slab caches used by the aio routines, panic on
157 * failure as this is done early during the boot sequence.
159 static int __init aio_setup(void)
161 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
162 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
164 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
168 __initcall(aio_setup);
170 static void put_aio_ring_file(struct kioctx *ctx)
172 struct file *aio_ring_file = ctx->aio_ring_file;
174 truncate_setsize(aio_ring_file->f_inode, 0);
176 /* Prevent further access to the kioctx from migratepages */
177 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
178 aio_ring_file->f_inode->i_mapping->private_data = NULL;
179 ctx->aio_ring_file = NULL;
180 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
186 static void aio_free_ring(struct kioctx *ctx)
190 for (i = 0; i < ctx->nr_pages; i++) {
191 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
192 page_count(ctx->ring_pages[i]));
193 put_page(ctx->ring_pages[i]);
196 put_aio_ring_file(ctx);
198 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
199 kfree(ctx->ring_pages);
202 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
204 vma->vm_ops = &generic_file_vm_ops;
208 static const struct file_operations aio_ring_fops = {
209 .mmap = aio_ring_mmap,
212 static int aio_set_page_dirty(struct page *page)
217 #if IS_ENABLED(CONFIG_MIGRATION)
218 static int aio_migratepage(struct address_space *mapping, struct page *new,
219 struct page *old, enum migrate_mode mode)
225 /* Writeback must be complete */
226 BUG_ON(PageWriteback(old));
229 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
230 if (rc != MIGRATEPAGE_SUCCESS) {
237 /* We can potentially race against kioctx teardown here. Use the
238 * address_space's private data lock to protect the mapping's
241 spin_lock(&mapping->private_lock);
242 ctx = mapping->private_data;
245 spin_lock_irqsave(&ctx->completion_lock, flags);
246 migrate_page_copy(new, old);
248 if (idx < (pgoff_t)ctx->nr_pages)
249 ctx->ring_pages[idx] = new;
250 spin_unlock_irqrestore(&ctx->completion_lock, flags);
253 spin_unlock(&mapping->private_lock);
259 static const struct address_space_operations aio_ctx_aops = {
260 .set_page_dirty = aio_set_page_dirty,
261 #if IS_ENABLED(CONFIG_MIGRATION)
262 .migratepage = aio_migratepage,
266 static int aio_setup_ring(struct kioctx *ctx)
268 struct aio_ring *ring;
269 unsigned nr_events = ctx->max_reqs;
270 struct mm_struct *mm = current->mm;
271 unsigned long size, populate;
276 /* Compensate for the ring buffer's head/tail overlap entry */
277 nr_events += 2; /* 1 is required, 2 for good luck */
279 size = sizeof(struct aio_ring);
280 size += sizeof(struct io_event) * nr_events;
282 nr_pages = PFN_UP(size);
286 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
288 ctx->aio_ring_file = NULL;
292 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
293 file->f_inode->i_mapping->private_data = ctx;
294 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
296 for (i = 0; i < nr_pages; i++) {
298 page = find_or_create_page(file->f_inode->i_mapping,
299 i, GFP_HIGHUSER | __GFP_ZERO);
302 pr_debug("pid(%d) page[%d]->count=%d\n",
303 current->pid, i, page_count(page));
304 SetPageUptodate(page);
308 ctx->aio_ring_file = file;
309 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
310 / sizeof(struct io_event);
312 ctx->ring_pages = ctx->internal_pages;
313 if (nr_pages > AIO_RING_PAGES) {
314 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
316 if (!ctx->ring_pages)
320 ctx->mmap_size = nr_pages * PAGE_SIZE;
321 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
323 down_write(&mm->mmap_sem);
324 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
325 PROT_READ | PROT_WRITE,
326 MAP_SHARED | MAP_POPULATE, 0, &populate);
327 if (IS_ERR((void *)ctx->mmap_base)) {
328 up_write(&mm->mmap_sem);
334 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
336 /* We must do this while still holding mmap_sem for write, as we
337 * need to be protected against userspace attempting to mremap()
338 * or munmap() the ring buffer.
340 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
341 1, 0, ctx->ring_pages, NULL);
343 /* Dropping the reference here is safe as the page cache will hold
344 * onto the pages for us. It is also required so that page migration
345 * can unmap the pages and get the right reference count.
347 for (i = 0; i < ctx->nr_pages; i++)
348 put_page(ctx->ring_pages[i]);
350 up_write(&mm->mmap_sem);
352 if (unlikely(ctx->nr_pages != nr_pages)) {
357 ctx->user_id = ctx->mmap_base;
358 ctx->nr_events = nr_events; /* trusted copy */
360 ring = kmap_atomic(ctx->ring_pages[0]);
361 ring->nr = nr_events; /* user copy */
363 ring->head = ring->tail = 0;
364 ring->magic = AIO_RING_MAGIC;
365 ring->compat_features = AIO_RING_COMPAT_FEATURES;
366 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
367 ring->header_length = sizeof(struct aio_ring);
369 flush_dcache_page(ctx->ring_pages[0]);
374 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
375 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
376 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
378 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
380 struct kioctx *ctx = req->ki_ctx;
383 spin_lock_irqsave(&ctx->ctx_lock, flags);
385 if (!req->ki_list.next)
386 list_add(&req->ki_list, &ctx->active_reqs);
388 req->ki_cancel = cancel;
390 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
392 EXPORT_SYMBOL(kiocb_set_cancel_fn);
394 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
396 kiocb_cancel_fn *old, *cancel;
399 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
400 * actually has a cancel function, hence the cmpxchg()
403 cancel = ACCESS_ONCE(kiocb->ki_cancel);
405 if (!cancel || cancel == KIOCB_CANCELLED)
409 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
410 } while (cancel != old);
412 return cancel(kiocb);
415 static void free_ioctx_rcu(struct rcu_head *head)
417 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
419 free_percpu(ctx->cpu);
420 kmem_cache_free(kioctx_cachep, ctx);
424 * When this function runs, the kioctx has been removed from the "hash table"
425 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
426 * now it's safe to cancel any that need to be.
428 static void free_ioctx(struct work_struct *work)
430 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
431 struct aio_ring *ring;
436 spin_lock_irq(&ctx->ctx_lock);
438 while (!list_empty(&ctx->active_reqs)) {
439 req = list_first_entry(&ctx->active_reqs,
440 struct kiocb, ki_list);
442 list_del_init(&req->ki_list);
443 kiocb_cancel(ctx, req);
446 spin_unlock_irq(&ctx->ctx_lock);
448 for_each_possible_cpu(cpu) {
449 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
451 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
452 kcpu->reqs_available = 0;
456 prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
458 ring = kmap_atomic(ctx->ring_pages[0]);
459 avail = (ring->head <= ring->tail)
460 ? ring->tail - ring->head
461 : ctx->nr_events - ring->head + ring->tail;
463 atomic_add(avail, &ctx->reqs_available);
464 ring->head = ring->tail;
467 if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
472 finish_wait(&ctx->wait, &wait);
474 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
478 pr_debug("freeing %p\n", ctx);
481 * Here the call_rcu() is between the wait_event() for reqs_active to
482 * hit 0, and freeing the ioctx.
484 * aio_complete() decrements reqs_active, but it has to touch the ioctx
485 * after to issue a wakeup so we use rcu.
487 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
490 static void free_ioctx_ref(struct percpu_ref *ref)
492 struct kioctx *ctx = container_of(ref, struct kioctx, users);
494 INIT_WORK(&ctx->free_work, free_ioctx);
495 schedule_work(&ctx->free_work);
498 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
501 struct kioctx_table *table, *old;
502 struct aio_ring *ring;
504 spin_lock(&mm->ioctx_lock);
506 table = rcu_dereference(mm->ioctx_table);
510 for (i = 0; i < table->nr; i++)
511 if (!table->table[i]) {
513 table->table[i] = ctx;
515 spin_unlock(&mm->ioctx_lock);
517 ring = kmap_atomic(ctx->ring_pages[0]);
523 new_nr = (table ? table->nr : 1) * 4;
526 spin_unlock(&mm->ioctx_lock);
528 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
535 spin_lock(&mm->ioctx_lock);
537 old = rcu_dereference(mm->ioctx_table);
540 rcu_assign_pointer(mm->ioctx_table, table);
541 } else if (table->nr > old->nr) {
542 memcpy(table->table, old->table,
543 old->nr * sizeof(struct kioctx *));
545 rcu_assign_pointer(mm->ioctx_table, table);
555 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
557 static struct kioctx *ioctx_alloc(unsigned nr_events)
559 struct mm_struct *mm = current->mm;
564 * We keep track of the number of available ringbuffer slots, to prevent
565 * overflow (reqs_available), and we also use percpu counters for this.
567 * So since up to half the slots might be on other cpu's percpu counters
568 * and unavailable, double nr_events so userspace sees what they
569 * expected: additionally, we move req_batch slots to/from percpu
570 * counters at a time, so make sure that isn't 0:
572 nr_events = max(nr_events, num_possible_cpus() * 4);
575 /* Prevent overflows */
576 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
577 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
578 pr_debug("ENOMEM: nr_events too high\n");
579 return ERR_PTR(-EINVAL);
582 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
583 return ERR_PTR(-EAGAIN);
585 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
587 return ERR_PTR(-ENOMEM);
589 ctx->max_reqs = nr_events;
591 if (percpu_ref_init(&ctx->users, free_ioctx_ref))
594 spin_lock_init(&ctx->ctx_lock);
595 spin_lock_init(&ctx->completion_lock);
596 mutex_init(&ctx->ring_lock);
597 init_waitqueue_head(&ctx->wait);
599 INIT_LIST_HEAD(&ctx->active_reqs);
601 ctx->cpu = alloc_percpu(struct kioctx_cpu);
605 if (aio_setup_ring(ctx) < 0)
608 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
609 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
610 if (ctx->req_batch < 1)
613 /* limit the number of system wide aios */
614 spin_lock(&aio_nr_lock);
615 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
616 aio_nr + nr_events < aio_nr) {
617 spin_unlock(&aio_nr_lock);
620 aio_nr += ctx->max_reqs;
621 spin_unlock(&aio_nr_lock);
623 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
625 err = ioctx_add_table(ctx, mm);
627 goto out_cleanup_put;
629 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
630 ctx, ctx->user_id, mm, ctx->nr_events);
634 percpu_ref_put(&ctx->users);
639 free_percpu(ctx->cpu);
641 free_percpu(ctx->users.pcpu_count);
643 put_aio_ring_file(ctx);
644 kmem_cache_free(kioctx_cachep, ctx);
645 pr_debug("error allocating ioctx %d\n", err);
650 * Cancels all outstanding aio requests on an aio context. Used
651 * when the processes owning a context have all exited to encourage
652 * the rapid destruction of the kioctx.
654 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
656 if (!atomic_xchg(&ctx->dead, 1)) {
657 struct kioctx_table *table;
659 spin_lock(&mm->ioctx_lock);
661 table = rcu_dereference(mm->ioctx_table);
663 WARN_ON(ctx != table->table[ctx->id]);
664 table->table[ctx->id] = NULL;
666 spin_unlock(&mm->ioctx_lock);
668 /* percpu_ref_kill() will do the necessary call_rcu() */
669 wake_up_all(&ctx->wait);
672 * It'd be more correct to do this in free_ioctx(), after all
673 * the outstanding kiocbs have finished - but by then io_destroy
674 * has already returned, so io_setup() could potentially return
675 * -EAGAIN with no ioctxs actually in use (as far as userspace
678 spin_lock(&aio_nr_lock);
679 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
680 aio_nr -= ctx->max_reqs;
681 spin_unlock(&aio_nr_lock);
684 vm_munmap(ctx->mmap_base, ctx->mmap_size);
686 percpu_ref_kill(&ctx->users);
690 /* wait_on_sync_kiocb:
691 * Waits on the given sync kiocb to complete.
693 ssize_t wait_on_sync_kiocb(struct kiocb *req)
695 while (!req->ki_ctx) {
696 set_current_state(TASK_UNINTERRUPTIBLE);
701 __set_current_state(TASK_RUNNING);
702 return req->ki_user_data;
704 EXPORT_SYMBOL(wait_on_sync_kiocb);
707 * exit_aio: called when the last user of mm goes away. At this point, there is
708 * no way for any new requests to be submited or any of the io_* syscalls to be
709 * called on the context.
711 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
714 void exit_aio(struct mm_struct *mm)
716 struct kioctx_table *table;
722 table = rcu_dereference(mm->ioctx_table);
725 if (!table || i >= table->nr) {
727 rcu_assign_pointer(mm->ioctx_table, NULL);
733 ctx = table->table[i++];
739 * We don't need to bother with munmap() here -
740 * exit_mmap(mm) is coming and it'll unmap everything.
741 * Since aio_free_ring() uses non-zero ->mmap_size
742 * as indicator that it needs to unmap the area,
743 * just set it to 0; aio_free_ring() is the only
744 * place that uses ->mmap_size, so it's safe.
752 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
754 struct kioctx_cpu *kcpu;
757 kcpu = this_cpu_ptr(ctx->cpu);
759 kcpu->reqs_available += nr;
760 while (kcpu->reqs_available >= ctx->req_batch * 2) {
761 kcpu->reqs_available -= ctx->req_batch;
762 atomic_add(ctx->req_batch, &ctx->reqs_available);
768 static bool get_reqs_available(struct kioctx *ctx)
770 struct kioctx_cpu *kcpu;
774 kcpu = this_cpu_ptr(ctx->cpu);
776 if (!kcpu->reqs_available) {
777 int old, avail = atomic_read(&ctx->reqs_available);
780 if (avail < ctx->req_batch)
784 avail = atomic_cmpxchg(&ctx->reqs_available,
785 avail, avail - ctx->req_batch);
786 } while (avail != old);
788 kcpu->reqs_available += ctx->req_batch;
792 kcpu->reqs_available--;
799 * Allocate a slot for an aio request.
800 * Returns NULL if no requests are free.
802 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
806 if (!get_reqs_available(ctx))
809 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
816 put_reqs_available(ctx, 1);
820 static void kiocb_free(struct kiocb *req)
824 if (req->ki_eventfd != NULL)
825 eventfd_ctx_put(req->ki_eventfd);
826 kmem_cache_free(kiocb_cachep, req);
829 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
831 struct aio_ring __user *ring = (void __user *)ctx_id;
832 struct mm_struct *mm = current->mm;
833 struct kioctx *ctx, *ret = NULL;
834 struct kioctx_table *table;
837 if (get_user(id, &ring->id))
841 table = rcu_dereference(mm->ioctx_table);
843 if (!table || id >= table->nr)
846 ctx = table->table[id];
847 if (ctx && ctx->user_id == ctx_id) {
848 percpu_ref_get(&ctx->users);
857 * Called when the io request on the given iocb is complete.
859 void aio_complete(struct kiocb *iocb, long res, long res2)
861 struct kioctx *ctx = iocb->ki_ctx;
862 struct aio_ring *ring;
863 struct io_event *ev_page, *event;
868 * Special case handling for sync iocbs:
869 * - events go directly into the iocb for fast handling
870 * - the sync task with the iocb in its stack holds the single iocb
871 * ref, no other paths have a way to get another ref
872 * - the sync task helpfully left a reference to itself in the iocb
874 if (is_sync_kiocb(iocb)) {
875 iocb->ki_user_data = res;
877 iocb->ki_ctx = ERR_PTR(-EXDEV);
878 wake_up_process(iocb->ki_obj.tsk);
883 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
884 * need to issue a wakeup after incrementing reqs_available.
888 if (iocb->ki_list.next) {
891 spin_lock_irqsave(&ctx->ctx_lock, flags);
892 list_del(&iocb->ki_list);
893 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
897 * Add a completion event to the ring buffer. Must be done holding
898 * ctx->completion_lock to prevent other code from messing with the tail
899 * pointer since we might be called from irq context.
901 spin_lock_irqsave(&ctx->completion_lock, flags);
904 pos = tail + AIO_EVENTS_OFFSET;
906 if (++tail >= ctx->nr_events)
909 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
910 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
912 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
913 event->data = iocb->ki_user_data;
917 kunmap_atomic(ev_page);
918 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
920 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
921 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
924 /* after flagging the request as done, we
925 * must never even look at it again
927 smp_wmb(); /* make event visible before updating tail */
931 ring = kmap_atomic(ctx->ring_pages[0]);
934 flush_dcache_page(ctx->ring_pages[0]);
936 spin_unlock_irqrestore(&ctx->completion_lock, flags);
938 pr_debug("added to ring %p at [%u]\n", iocb, tail);
941 * Check if the user asked us to deliver the result through an
942 * eventfd. The eventfd_signal() function is safe to be called
945 if (iocb->ki_eventfd != NULL)
946 eventfd_signal(iocb->ki_eventfd, 1);
948 /* everything turned out well, dispose of the aiocb. */
952 * We have to order our ring_info tail store above and test
953 * of the wait list below outside the wait lock. This is
954 * like in wake_up_bit() where clearing a bit has to be
955 * ordered with the unlocked test.
959 if (waitqueue_active(&ctx->wait))
964 EXPORT_SYMBOL(aio_complete);
967 * Pull an event off of the ioctx's event ring. Returns the number of
970 static long aio_read_events_ring(struct kioctx *ctx,
971 struct io_event __user *event, long nr)
973 struct aio_ring *ring;
974 unsigned head, tail, pos;
978 mutex_lock(&ctx->ring_lock);
980 ring = kmap_atomic(ctx->ring_pages[0]);
985 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
995 avail = (head <= tail ? tail : ctx->nr_events) - head;
999 avail = min(avail, nr - ret);
1000 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1001 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1003 pos = head + AIO_EVENTS_OFFSET;
1004 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1005 pos %= AIO_EVENTS_PER_PAGE;
1008 copy_ret = copy_to_user(event + ret, ev + pos,
1009 sizeof(*ev) * avail);
1012 if (unlikely(copy_ret)) {
1019 head %= ctx->nr_events;
1022 ring = kmap_atomic(ctx->ring_pages[0]);
1024 kunmap_atomic(ring);
1025 flush_dcache_page(ctx->ring_pages[0]);
1027 pr_debug("%li h%u t%u\n", ret, head, tail);
1029 put_reqs_available(ctx, ret);
1031 mutex_unlock(&ctx->ring_lock);
1036 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1037 struct io_event __user *event, long *i)
1039 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1044 if (unlikely(atomic_read(&ctx->dead)))
1050 return ret < 0 || *i >= min_nr;
1053 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1054 struct io_event __user *event,
1055 struct timespec __user *timeout)
1057 ktime_t until = { .tv64 = KTIME_MAX };
1063 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1066 until = timespec_to_ktime(ts);
1070 * Note that aio_read_events() is being called as the conditional - i.e.
1071 * we're calling it after prepare_to_wait() has set task state to
1072 * TASK_INTERRUPTIBLE.
1074 * But aio_read_events() can block, and if it blocks it's going to flip
1075 * the task state back to TASK_RUNNING.
1077 * This should be ok, provided it doesn't flip the state back to
1078 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1079 * will only happen if the mutex_lock() call blocks, and we then find
1080 * the ringbuffer empty. So in practice we should be ok, but it's
1081 * something to be aware of when touching this code.
1083 wait_event_interruptible_hrtimeout(ctx->wait,
1084 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1086 if (!ret && signal_pending(current))
1093 * Create an aio_context capable of receiving at least nr_events.
1094 * ctxp must not point to an aio_context that already exists, and
1095 * must be initialized to 0 prior to the call. On successful
1096 * creation of the aio_context, *ctxp is filled in with the resulting
1097 * handle. May fail with -EINVAL if *ctxp is not initialized,
1098 * if the specified nr_events exceeds internal limits. May fail
1099 * with -EAGAIN if the specified nr_events exceeds the user's limit
1100 * of available events. May fail with -ENOMEM if insufficient kernel
1101 * resources are available. May fail with -EFAULT if an invalid
1102 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1105 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1107 struct kioctx *ioctx = NULL;
1111 ret = get_user(ctx, ctxp);
1116 if (unlikely(ctx || nr_events == 0)) {
1117 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1122 ioctx = ioctx_alloc(nr_events);
1123 ret = PTR_ERR(ioctx);
1124 if (!IS_ERR(ioctx)) {
1125 ret = put_user(ioctx->user_id, ctxp);
1127 kill_ioctx(current->mm, ioctx);
1128 percpu_ref_put(&ioctx->users);
1136 * Destroy the aio_context specified. May cancel any outstanding
1137 * AIOs and block on completion. Will fail with -ENOSYS if not
1138 * implemented. May fail with -EINVAL if the context pointed to
1141 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1143 struct kioctx *ioctx = lookup_ioctx(ctx);
1144 if (likely(NULL != ioctx)) {
1145 kill_ioctx(current->mm, ioctx);
1146 percpu_ref_put(&ioctx->users);
1149 pr_debug("EINVAL: io_destroy: invalid context id\n");
1153 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1154 unsigned long, loff_t);
1156 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1157 int rw, char __user *buf,
1158 unsigned long *nr_segs,
1159 struct iovec **iovec,
1164 *nr_segs = kiocb->ki_nbytes;
1166 #ifdef CONFIG_COMPAT
1168 ret = compat_rw_copy_check_uvector(rw,
1169 (struct compat_iovec __user *)buf,
1170 *nr_segs, 1, *iovec, iovec);
1173 ret = rw_copy_check_uvector(rw,
1174 (struct iovec __user *)buf,
1175 *nr_segs, 1, *iovec, iovec);
1179 /* ki_nbytes now reflect bytes instead of segs */
1180 kiocb->ki_nbytes = ret;
1184 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1185 int rw, char __user *buf,
1186 unsigned long *nr_segs,
1187 struct iovec *iovec)
1189 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1192 iovec->iov_base = buf;
1193 iovec->iov_len = kiocb->ki_nbytes;
1200 * Performs the initial checks and aio retry method
1201 * setup for the kiocb at the time of io submission.
1203 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1204 char __user *buf, bool compat)
1206 struct file *file = req->ki_filp;
1208 unsigned long nr_segs;
1212 struct iovec inline_vec, *iovec = &inline_vec;
1215 case IOCB_CMD_PREAD:
1216 case IOCB_CMD_PREADV:
1219 rw_op = file->f_op->aio_read;
1222 case IOCB_CMD_PWRITE:
1223 case IOCB_CMD_PWRITEV:
1226 rw_op = file->f_op->aio_write;
1229 if (unlikely(!(file->f_mode & mode)))
1235 ret = (opcode == IOCB_CMD_PREADV ||
1236 opcode == IOCB_CMD_PWRITEV)
1237 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1239 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1244 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1246 if (iovec != &inline_vec)
1251 req->ki_nbytes = ret;
1253 /* XXX: move/kill - rw_verify_area()? */
1254 /* This matches the pread()/pwrite() logic */
1255 if (req->ki_pos < 0) {
1261 file_start_write(file);
1263 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1266 file_end_write(file);
1269 case IOCB_CMD_FDSYNC:
1270 if (!file->f_op->aio_fsync)
1273 ret = file->f_op->aio_fsync(req, 1);
1276 case IOCB_CMD_FSYNC:
1277 if (!file->f_op->aio_fsync)
1280 ret = file->f_op->aio_fsync(req, 0);
1284 pr_debug("EINVAL: no operation provided\n");
1288 if (iovec != &inline_vec)
1291 if (ret != -EIOCBQUEUED) {
1293 * There's no easy way to restart the syscall since other AIO's
1294 * may be already running. Just fail this IO with EINTR.
1296 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1297 ret == -ERESTARTNOHAND ||
1298 ret == -ERESTART_RESTARTBLOCK))
1300 aio_complete(req, ret, 0);
1306 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1307 struct iocb *iocb, bool compat)
1312 /* enforce forwards compatibility on users */
1313 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1314 pr_debug("EINVAL: reserve field set\n");
1318 /* prevent overflows */
1320 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1321 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1322 ((ssize_t)iocb->aio_nbytes < 0)
1324 pr_debug("EINVAL: io_submit: overflow check\n");
1328 req = aio_get_req(ctx);
1332 req->ki_filp = fget(iocb->aio_fildes);
1333 if (unlikely(!req->ki_filp)) {
1338 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1340 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1341 * instance of the file* now. The file descriptor must be
1342 * an eventfd() fd, and will be signaled for each completed
1343 * event using the eventfd_signal() function.
1345 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1346 if (IS_ERR(req->ki_eventfd)) {
1347 ret = PTR_ERR(req->ki_eventfd);
1348 req->ki_eventfd = NULL;
1353 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1354 if (unlikely(ret)) {
1355 pr_debug("EFAULT: aio_key\n");
1359 req->ki_obj.user = user_iocb;
1360 req->ki_user_data = iocb->aio_data;
1361 req->ki_pos = iocb->aio_offset;
1362 req->ki_nbytes = iocb->aio_nbytes;
1364 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1365 (char __user *)(unsigned long)iocb->aio_buf,
1372 put_reqs_available(ctx, 1);
1377 long do_io_submit(aio_context_t ctx_id, long nr,
1378 struct iocb __user *__user *iocbpp, bool compat)
1383 struct blk_plug plug;
1385 if (unlikely(nr < 0))
1388 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1389 nr = LONG_MAX/sizeof(*iocbpp);
1391 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1394 ctx = lookup_ioctx(ctx_id);
1395 if (unlikely(!ctx)) {
1396 pr_debug("EINVAL: invalid context id\n");
1400 blk_start_plug(&plug);
1403 * AKPM: should this return a partial result if some of the IOs were
1404 * successfully submitted?
1406 for (i=0; i<nr; i++) {
1407 struct iocb __user *user_iocb;
1410 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1415 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1420 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1424 blk_finish_plug(&plug);
1426 percpu_ref_put(&ctx->users);
1431 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1432 * the number of iocbs queued. May return -EINVAL if the aio_context
1433 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1434 * *iocbpp[0] is not properly initialized, if the operation specified
1435 * is invalid for the file descriptor in the iocb. May fail with
1436 * -EFAULT if any of the data structures point to invalid data. May
1437 * fail with -EBADF if the file descriptor specified in the first
1438 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1439 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1440 * fail with -ENOSYS if not implemented.
1442 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1443 struct iocb __user * __user *, iocbpp)
1445 return do_io_submit(ctx_id, nr, iocbpp, 0);
1449 * Finds a given iocb for cancellation.
1451 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1454 struct list_head *pos;
1456 assert_spin_locked(&ctx->ctx_lock);
1458 if (key != KIOCB_KEY)
1461 /* TODO: use a hash or array, this sucks. */
1462 list_for_each(pos, &ctx->active_reqs) {
1463 struct kiocb *kiocb = list_kiocb(pos);
1464 if (kiocb->ki_obj.user == iocb)
1471 * Attempts to cancel an iocb previously passed to io_submit. If
1472 * the operation is successfully cancelled, the resulting event is
1473 * copied into the memory pointed to by result without being placed
1474 * into the completion queue and 0 is returned. May fail with
1475 * -EFAULT if any of the data structures pointed to are invalid.
1476 * May fail with -EINVAL if aio_context specified by ctx_id is
1477 * invalid. May fail with -EAGAIN if the iocb specified was not
1478 * cancelled. Will fail with -ENOSYS if not implemented.
1480 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1481 struct io_event __user *, result)
1484 struct kiocb *kiocb;
1488 ret = get_user(key, &iocb->aio_key);
1492 ctx = lookup_ioctx(ctx_id);
1496 spin_lock_irq(&ctx->ctx_lock);
1498 kiocb = lookup_kiocb(ctx, iocb, key);
1500 ret = kiocb_cancel(ctx, kiocb);
1504 spin_unlock_irq(&ctx->ctx_lock);
1508 * The result argument is no longer used - the io_event is
1509 * always delivered via the ring buffer. -EINPROGRESS indicates
1510 * cancellation is progress:
1515 percpu_ref_put(&ctx->users);
1521 * Attempts to read at least min_nr events and up to nr events from
1522 * the completion queue for the aio_context specified by ctx_id. If
1523 * it succeeds, the number of read events is returned. May fail with
1524 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1525 * out of range, if timeout is out of range. May fail with -EFAULT
1526 * if any of the memory specified is invalid. May return 0 or
1527 * < min_nr if the timeout specified by timeout has elapsed
1528 * before sufficient events are available, where timeout == NULL
1529 * specifies an infinite timeout. Note that the timeout pointed to by
1530 * timeout is relative. Will fail with -ENOSYS if not implemented.
1532 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1535 struct io_event __user *, events,
1536 struct timespec __user *, timeout)
1538 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1541 if (likely(ioctx)) {
1542 if (likely(min_nr <= nr && min_nr >= 0))
1543 ret = read_events(ioctx, min_nr, nr, events, timeout);
1544 percpu_ref_put(&ioctx->users);