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/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.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 */
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
65 struct io_event io_events[0];
66 }; /* 128 bytes + ring size */
68 #define AIO_RING_PAGES 8
73 struct kioctx *table[];
77 unsigned reqs_available;
81 struct percpu_ref users;
84 struct percpu_ref reqs;
86 unsigned long user_id;
88 struct __percpu kioctx_cpu *cpu;
91 * For percpu reqs_available, number of slots we move to/from global
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
99 * The real limit is nr_events - 1, which will be larger (see
104 /* Size of ringbuffer, in units of struct io_event */
107 unsigned long mmap_base;
108 unsigned long mmap_size;
110 struct page **ring_pages;
113 struct work_struct free_work;
116 * signals when all in-flight requests are done
118 struct completion *requests_done;
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
127 * We batch accesses to it with a percpu version.
129 atomic_t reqs_available;
130 } ____cacheline_aligned_in_smp;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
138 struct mutex ring_lock;
139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
144 unsigned completed_events;
145 spinlock_t completion_lock;
146 } ____cacheline_aligned_in_smp;
148 struct page *internal_pages[AIO_RING_PAGES];
149 struct file *aio_ring_file;
154 /*------ sysctl variables----*/
155 static DEFINE_SPINLOCK(aio_nr_lock);
156 unsigned long aio_nr; /* current system wide number of aio requests */
157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
158 /*----end sysctl variables---*/
160 static struct kmem_cache *kiocb_cachep;
161 static struct kmem_cache *kioctx_cachep;
163 static struct vfsmount *aio_mnt;
165 static const struct file_operations aio_ring_fops;
166 static const struct address_space_operations aio_ctx_aops;
168 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
170 struct qstr this = QSTR_INIT("[aio]", 5);
173 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
175 return ERR_CAST(inode);
177 inode->i_mapping->a_ops = &aio_ctx_aops;
178 inode->i_mapping->private_data = ctx;
179 inode->i_size = PAGE_SIZE * nr_pages;
181 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
184 return ERR_PTR(-ENOMEM);
186 path.mnt = mntget(aio_mnt);
188 d_instantiate(path.dentry, inode);
189 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
195 file->f_flags = O_RDWR;
199 static struct dentry *aio_mount(struct file_system_type *fs_type,
200 int flags, const char *dev_name, void *data)
202 static const struct dentry_operations ops = {
203 .d_dname = simple_dname,
205 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
209 * Creates the slab caches used by the aio routines, panic on
210 * failure as this is done early during the boot sequence.
212 static int __init aio_setup(void)
214 static struct file_system_type aio_fs = {
217 .kill_sb = kill_anon_super,
219 aio_mnt = kern_mount(&aio_fs);
221 panic("Failed to create aio fs mount.");
223 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
224 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
226 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
230 __initcall(aio_setup);
232 static void put_aio_ring_file(struct kioctx *ctx)
234 struct file *aio_ring_file = ctx->aio_ring_file;
236 truncate_setsize(aio_ring_file->f_inode, 0);
238 /* Prevent further access to the kioctx from migratepages */
239 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
240 aio_ring_file->f_inode->i_mapping->private_data = NULL;
241 ctx->aio_ring_file = NULL;
242 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
248 static void aio_free_ring(struct kioctx *ctx)
252 /* Disconnect the kiotx from the ring file. This prevents future
253 * accesses to the kioctx from page migration.
255 put_aio_ring_file(ctx);
257 for (i = 0; i < ctx->nr_pages; i++) {
259 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
260 page_count(ctx->ring_pages[i]));
261 page = ctx->ring_pages[i];
264 ctx->ring_pages[i] = NULL;
268 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
269 kfree(ctx->ring_pages);
270 ctx->ring_pages = NULL;
274 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
276 vma->vm_ops = &generic_file_vm_ops;
280 static const struct file_operations aio_ring_fops = {
281 .mmap = aio_ring_mmap,
284 static int aio_set_page_dirty(struct page *page)
289 #if IS_ENABLED(CONFIG_MIGRATION)
290 static int aio_migratepage(struct address_space *mapping, struct page *new,
291 struct page *old, enum migrate_mode mode)
300 /* mapping->private_lock here protects against the kioctx teardown. */
301 spin_lock(&mapping->private_lock);
302 ctx = mapping->private_data;
308 /* The ring_lock mutex. The prevents aio_read_events() from writing
309 * to the ring's head, and prevents page migration from mucking in
310 * a partially initialized kiotx.
312 if (!mutex_trylock(&ctx->ring_lock)) {
318 if (idx < (pgoff_t)ctx->nr_pages) {
319 /* Make sure the old page hasn't already been changed */
320 if (ctx->ring_pages[idx] != old)
328 /* Writeback must be complete */
329 BUG_ON(PageWriteback(old));
332 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
333 if (rc != MIGRATEPAGE_SUCCESS) {
338 /* Take completion_lock to prevent other writes to the ring buffer
339 * while the old page is copied to the new. This prevents new
340 * events from being lost.
342 spin_lock_irqsave(&ctx->completion_lock, flags);
343 migrate_page_copy(new, old);
344 BUG_ON(ctx->ring_pages[idx] != old);
345 ctx->ring_pages[idx] = new;
346 spin_unlock_irqrestore(&ctx->completion_lock, flags);
348 /* The old page is no longer accessible. */
352 mutex_unlock(&ctx->ring_lock);
354 spin_unlock(&mapping->private_lock);
359 static const struct address_space_operations aio_ctx_aops = {
360 .set_page_dirty = aio_set_page_dirty,
361 #if IS_ENABLED(CONFIG_MIGRATION)
362 .migratepage = aio_migratepage,
366 static int aio_setup_ring(struct kioctx *ctx)
368 struct aio_ring *ring;
369 unsigned nr_events = ctx->max_reqs;
370 struct mm_struct *mm = current->mm;
371 unsigned long size, unused;
376 /* Compensate for the ring buffer's head/tail overlap entry */
377 nr_events += 2; /* 1 is required, 2 for good luck */
379 size = sizeof(struct aio_ring);
380 size += sizeof(struct io_event) * nr_events;
382 nr_pages = PFN_UP(size);
386 file = aio_private_file(ctx, nr_pages);
388 ctx->aio_ring_file = NULL;
392 ctx->aio_ring_file = file;
393 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
394 / sizeof(struct io_event);
396 ctx->ring_pages = ctx->internal_pages;
397 if (nr_pages > AIO_RING_PAGES) {
398 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
400 if (!ctx->ring_pages) {
401 put_aio_ring_file(ctx);
406 for (i = 0; i < nr_pages; i++) {
408 page = find_or_create_page(file->f_inode->i_mapping,
409 i, GFP_HIGHUSER | __GFP_ZERO);
412 pr_debug("pid(%d) page[%d]->count=%d\n",
413 current->pid, i, page_count(page));
414 SetPageUptodate(page);
418 ctx->ring_pages[i] = page;
422 if (unlikely(i != nr_pages)) {
427 ctx->mmap_size = nr_pages * PAGE_SIZE;
428 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
430 down_write(&mm->mmap_sem);
431 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
432 PROT_READ | PROT_WRITE,
433 MAP_SHARED, 0, &unused);
434 up_write(&mm->mmap_sem);
435 if (IS_ERR((void *)ctx->mmap_base)) {
441 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
443 ctx->user_id = ctx->mmap_base;
444 ctx->nr_events = nr_events; /* trusted copy */
446 ring = kmap_atomic(ctx->ring_pages[0]);
447 ring->nr = nr_events; /* user copy */
449 ring->head = ring->tail = 0;
450 ring->magic = AIO_RING_MAGIC;
451 ring->compat_features = AIO_RING_COMPAT_FEATURES;
452 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
453 ring->header_length = sizeof(struct aio_ring);
455 flush_dcache_page(ctx->ring_pages[0]);
460 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
461 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
462 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
464 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
466 struct kioctx *ctx = req->ki_ctx;
469 spin_lock_irqsave(&ctx->ctx_lock, flags);
471 if (!req->ki_list.next)
472 list_add(&req->ki_list, &ctx->active_reqs);
474 req->ki_cancel = cancel;
476 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
478 EXPORT_SYMBOL(kiocb_set_cancel_fn);
480 static int kiocb_cancel(struct kiocb *kiocb)
482 kiocb_cancel_fn *old, *cancel;
485 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
486 * actually has a cancel function, hence the cmpxchg()
489 cancel = ACCESS_ONCE(kiocb->ki_cancel);
491 if (!cancel || cancel == KIOCB_CANCELLED)
495 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
496 } while (cancel != old);
498 return cancel(kiocb);
501 static void free_ioctx(struct work_struct *work)
503 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
505 pr_debug("freeing %p\n", ctx);
508 free_percpu(ctx->cpu);
509 percpu_ref_exit(&ctx->reqs);
510 percpu_ref_exit(&ctx->users);
511 kmem_cache_free(kioctx_cachep, ctx);
514 static void free_ioctx_reqs(struct percpu_ref *ref)
516 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
518 /* At this point we know that there are no any in-flight requests */
519 if (ctx->requests_done)
520 complete(ctx->requests_done);
522 INIT_WORK(&ctx->free_work, free_ioctx);
523 schedule_work(&ctx->free_work);
527 * When this function runs, the kioctx has been removed from the "hash table"
528 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
529 * now it's safe to cancel any that need to be.
531 static void free_ioctx_users(struct percpu_ref *ref)
533 struct kioctx *ctx = container_of(ref, struct kioctx, users);
536 spin_lock_irq(&ctx->ctx_lock);
538 while (!list_empty(&ctx->active_reqs)) {
539 req = list_first_entry(&ctx->active_reqs,
540 struct kiocb, ki_list);
542 list_del_init(&req->ki_list);
546 spin_unlock_irq(&ctx->ctx_lock);
548 percpu_ref_kill(&ctx->reqs);
549 percpu_ref_put(&ctx->reqs);
552 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
555 struct kioctx_table *table, *old;
556 struct aio_ring *ring;
558 spin_lock(&mm->ioctx_lock);
559 table = rcu_dereference_raw(mm->ioctx_table);
563 for (i = 0; i < table->nr; i++)
564 if (!table->table[i]) {
566 table->table[i] = ctx;
567 spin_unlock(&mm->ioctx_lock);
569 /* While kioctx setup is in progress,
570 * we are protected from page migration
571 * changes ring_pages by ->ring_lock.
573 ring = kmap_atomic(ctx->ring_pages[0]);
579 new_nr = (table ? table->nr : 1) * 4;
580 spin_unlock(&mm->ioctx_lock);
582 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
589 spin_lock(&mm->ioctx_lock);
590 old = rcu_dereference_raw(mm->ioctx_table);
593 rcu_assign_pointer(mm->ioctx_table, table);
594 } else if (table->nr > old->nr) {
595 memcpy(table->table, old->table,
596 old->nr * sizeof(struct kioctx *));
598 rcu_assign_pointer(mm->ioctx_table, table);
607 static void aio_nr_sub(unsigned nr)
609 spin_lock(&aio_nr_lock);
610 if (WARN_ON(aio_nr - nr > aio_nr))
614 spin_unlock(&aio_nr_lock);
618 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
620 static struct kioctx *ioctx_alloc(unsigned nr_events)
622 struct mm_struct *mm = current->mm;
627 * We keep track of the number of available ringbuffer slots, to prevent
628 * overflow (reqs_available), and we also use percpu counters for this.
630 * So since up to half the slots might be on other cpu's percpu counters
631 * and unavailable, double nr_events so userspace sees what they
632 * expected: additionally, we move req_batch slots to/from percpu
633 * counters at a time, so make sure that isn't 0:
635 nr_events = max(nr_events, num_possible_cpus() * 4);
638 /* Prevent overflows */
639 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
640 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
641 pr_debug("ENOMEM: nr_events too high\n");
642 return ERR_PTR(-EINVAL);
645 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
646 return ERR_PTR(-EAGAIN);
648 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
650 return ERR_PTR(-ENOMEM);
652 ctx->max_reqs = nr_events;
654 spin_lock_init(&ctx->ctx_lock);
655 spin_lock_init(&ctx->completion_lock);
656 mutex_init(&ctx->ring_lock);
657 /* Protect against page migration throughout kiotx setup by keeping
658 * the ring_lock mutex held until setup is complete. */
659 mutex_lock(&ctx->ring_lock);
660 init_waitqueue_head(&ctx->wait);
662 INIT_LIST_HEAD(&ctx->active_reqs);
664 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
667 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
670 ctx->cpu = alloc_percpu(struct kioctx_cpu);
674 err = aio_setup_ring(ctx);
678 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
679 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
680 if (ctx->req_batch < 1)
683 /* limit the number of system wide aios */
684 spin_lock(&aio_nr_lock);
685 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
686 aio_nr + nr_events < aio_nr) {
687 spin_unlock(&aio_nr_lock);
691 aio_nr += ctx->max_reqs;
692 spin_unlock(&aio_nr_lock);
694 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
695 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
697 err = ioctx_add_table(ctx, mm);
701 /* Release the ring_lock mutex now that all setup is complete. */
702 mutex_unlock(&ctx->ring_lock);
704 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
705 ctx, ctx->user_id, mm, ctx->nr_events);
709 aio_nr_sub(ctx->max_reqs);
713 mutex_unlock(&ctx->ring_lock);
714 free_percpu(ctx->cpu);
715 percpu_ref_exit(&ctx->reqs);
716 percpu_ref_exit(&ctx->users);
717 kmem_cache_free(kioctx_cachep, ctx);
718 pr_debug("error allocating ioctx %d\n", err);
723 * Cancels all outstanding aio requests on an aio context. Used
724 * when the processes owning a context have all exited to encourage
725 * the rapid destruction of the kioctx.
727 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
728 struct completion *requests_done)
730 struct kioctx_table *table;
732 if (atomic_xchg(&ctx->dead, 1))
736 spin_lock(&mm->ioctx_lock);
737 table = rcu_dereference_raw(mm->ioctx_table);
738 WARN_ON(ctx != table->table[ctx->id]);
739 table->table[ctx->id] = NULL;
740 spin_unlock(&mm->ioctx_lock);
742 /* percpu_ref_kill() will do the necessary call_rcu() */
743 wake_up_all(&ctx->wait);
746 * It'd be more correct to do this in free_ioctx(), after all
747 * the outstanding kiocbs have finished - but by then io_destroy
748 * has already returned, so io_setup() could potentially return
749 * -EAGAIN with no ioctxs actually in use (as far as userspace
752 aio_nr_sub(ctx->max_reqs);
755 vm_munmap(ctx->mmap_base, ctx->mmap_size);
757 ctx->requests_done = requests_done;
758 percpu_ref_kill(&ctx->users);
762 /* wait_on_sync_kiocb:
763 * Waits on the given sync kiocb to complete.
765 ssize_t wait_on_sync_kiocb(struct kiocb *req)
767 while (!req->ki_ctx) {
768 set_current_state(TASK_UNINTERRUPTIBLE);
773 __set_current_state(TASK_RUNNING);
774 return req->ki_user_data;
776 EXPORT_SYMBOL(wait_on_sync_kiocb);
779 * exit_aio: called when the last user of mm goes away. At this point, there is
780 * no way for any new requests to be submited or any of the io_* syscalls to be
781 * called on the context.
783 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
786 void exit_aio(struct mm_struct *mm)
788 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
794 for (i = 0; i < table->nr; ++i) {
795 struct kioctx *ctx = table->table[i];
796 struct completion requests_done =
797 COMPLETION_INITIALIZER_ONSTACK(requests_done);
802 * We don't need to bother with munmap() here - exit_mmap(mm)
803 * is coming and it'll unmap everything. And we simply can't,
804 * this is not necessarily our ->mm.
805 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
806 * that it needs to unmap the area, just set it to 0.
809 kill_ioctx(mm, ctx, &requests_done);
811 /* Wait until all IO for the context are done. */
812 wait_for_completion(&requests_done);
815 RCU_INIT_POINTER(mm->ioctx_table, NULL);
819 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
821 struct kioctx_cpu *kcpu;
824 local_irq_save(flags);
825 kcpu = this_cpu_ptr(ctx->cpu);
826 kcpu->reqs_available += nr;
828 while (kcpu->reqs_available >= ctx->req_batch * 2) {
829 kcpu->reqs_available -= ctx->req_batch;
830 atomic_add(ctx->req_batch, &ctx->reqs_available);
833 local_irq_restore(flags);
836 static bool get_reqs_available(struct kioctx *ctx)
838 struct kioctx_cpu *kcpu;
842 local_irq_save(flags);
843 kcpu = this_cpu_ptr(ctx->cpu);
844 if (!kcpu->reqs_available) {
845 int old, avail = atomic_read(&ctx->reqs_available);
848 if (avail < ctx->req_batch)
852 avail = atomic_cmpxchg(&ctx->reqs_available,
853 avail, avail - ctx->req_batch);
854 } while (avail != old);
856 kcpu->reqs_available += ctx->req_batch;
860 kcpu->reqs_available--;
862 local_irq_restore(flags);
866 /* refill_reqs_available
867 * Updates the reqs_available reference counts used for tracking the
868 * number of free slots in the completion ring. This can be called
869 * from aio_complete() (to optimistically update reqs_available) or
870 * from aio_get_req() (the we're out of events case). It must be
871 * called holding ctx->completion_lock.
873 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
876 unsigned events_in_ring, completed;
878 /* Clamp head since userland can write to it. */
879 head %= ctx->nr_events;
881 events_in_ring = tail - head;
883 events_in_ring = ctx->nr_events - (head - tail);
885 completed = ctx->completed_events;
886 if (events_in_ring < completed)
887 completed -= events_in_ring;
894 ctx->completed_events -= completed;
895 put_reqs_available(ctx, completed);
898 /* user_refill_reqs_available
899 * Called to refill reqs_available when aio_get_req() encounters an
900 * out of space in the completion ring.
902 static void user_refill_reqs_available(struct kioctx *ctx)
904 spin_lock_irq(&ctx->completion_lock);
905 if (ctx->completed_events) {
906 struct aio_ring *ring;
909 /* Access of ring->head may race with aio_read_events_ring()
910 * here, but that's okay since whether we read the old version
911 * or the new version, and either will be valid. The important
912 * part is that head cannot pass tail since we prevent
913 * aio_complete() from updating tail by holding
914 * ctx->completion_lock. Even if head is invalid, the check
915 * against ctx->completed_events below will make sure we do the
918 ring = kmap_atomic(ctx->ring_pages[0]);
922 refill_reqs_available(ctx, head, ctx->tail);
925 spin_unlock_irq(&ctx->completion_lock);
929 * Allocate a slot for an aio request.
930 * Returns NULL if no requests are free.
932 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
936 if (!get_reqs_available(ctx)) {
937 user_refill_reqs_available(ctx);
938 if (!get_reqs_available(ctx))
942 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
946 percpu_ref_get(&ctx->reqs);
951 put_reqs_available(ctx, 1);
955 static void kiocb_free(struct kiocb *req)
959 if (req->ki_eventfd != NULL)
960 eventfd_ctx_put(req->ki_eventfd);
961 kmem_cache_free(kiocb_cachep, req);
964 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
966 struct aio_ring __user *ring = (void __user *)ctx_id;
967 struct mm_struct *mm = current->mm;
968 struct kioctx *ctx, *ret = NULL;
969 struct kioctx_table *table;
972 if (get_user(id, &ring->id))
976 table = rcu_dereference(mm->ioctx_table);
978 if (!table || id >= table->nr)
981 ctx = table->table[id];
982 if (ctx && ctx->user_id == ctx_id) {
983 percpu_ref_get(&ctx->users);
992 * Called when the io request on the given iocb is complete.
994 void aio_complete(struct kiocb *iocb, long res, long res2)
996 struct kioctx *ctx = iocb->ki_ctx;
997 struct aio_ring *ring;
998 struct io_event *ev_page, *event;
999 unsigned tail, pos, head;
1000 unsigned long flags;
1003 * Special case handling for sync iocbs:
1004 * - events go directly into the iocb for fast handling
1005 * - the sync task with the iocb in its stack holds the single iocb
1006 * ref, no other paths have a way to get another ref
1007 * - the sync task helpfully left a reference to itself in the iocb
1009 if (is_sync_kiocb(iocb)) {
1010 iocb->ki_user_data = res;
1012 iocb->ki_ctx = ERR_PTR(-EXDEV);
1013 wake_up_process(iocb->ki_obj.tsk);
1017 if (iocb->ki_list.next) {
1018 unsigned long flags;
1020 spin_lock_irqsave(&ctx->ctx_lock, flags);
1021 list_del(&iocb->ki_list);
1022 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1026 * Add a completion event to the ring buffer. Must be done holding
1027 * ctx->completion_lock to prevent other code from messing with the tail
1028 * pointer since we might be called from irq context.
1030 spin_lock_irqsave(&ctx->completion_lock, flags);
1033 pos = tail + AIO_EVENTS_OFFSET;
1035 if (++tail >= ctx->nr_events)
1038 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1039 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1041 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1042 event->data = iocb->ki_user_data;
1046 kunmap_atomic(ev_page);
1047 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1049 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1050 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1053 /* after flagging the request as done, we
1054 * must never even look at it again
1056 smp_wmb(); /* make event visible before updating tail */
1060 ring = kmap_atomic(ctx->ring_pages[0]);
1063 kunmap_atomic(ring);
1064 flush_dcache_page(ctx->ring_pages[0]);
1066 ctx->completed_events++;
1067 if (ctx->completed_events > 1)
1068 refill_reqs_available(ctx, head, tail);
1069 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1071 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1074 * Check if the user asked us to deliver the result through an
1075 * eventfd. The eventfd_signal() function is safe to be called
1078 if (iocb->ki_eventfd != NULL)
1079 eventfd_signal(iocb->ki_eventfd, 1);
1081 /* everything turned out well, dispose of the aiocb. */
1085 * We have to order our ring_info tail store above and test
1086 * of the wait list below outside the wait lock. This is
1087 * like in wake_up_bit() where clearing a bit has to be
1088 * ordered with the unlocked test.
1092 if (waitqueue_active(&ctx->wait))
1093 wake_up(&ctx->wait);
1095 percpu_ref_put(&ctx->reqs);
1097 EXPORT_SYMBOL(aio_complete);
1099 /* aio_read_events_ring
1100 * Pull an event off of the ioctx's event ring. Returns the number of
1103 static long aio_read_events_ring(struct kioctx *ctx,
1104 struct io_event __user *event, long nr)
1106 struct aio_ring *ring;
1107 unsigned head, tail, pos;
1111 mutex_lock(&ctx->ring_lock);
1113 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1114 ring = kmap_atomic(ctx->ring_pages[0]);
1117 kunmap_atomic(ring);
1120 * Ensure that once we've read the current tail pointer, that
1121 * we also see the events that were stored up to the tail.
1125 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1130 head %= ctx->nr_events;
1131 tail %= ctx->nr_events;
1135 struct io_event *ev;
1138 avail = (head <= tail ? tail : ctx->nr_events) - head;
1142 avail = min(avail, nr - ret);
1143 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1144 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1146 pos = head + AIO_EVENTS_OFFSET;
1147 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1148 pos %= AIO_EVENTS_PER_PAGE;
1151 copy_ret = copy_to_user(event + ret, ev + pos,
1152 sizeof(*ev) * avail);
1155 if (unlikely(copy_ret)) {
1162 head %= ctx->nr_events;
1165 ring = kmap_atomic(ctx->ring_pages[0]);
1167 kunmap_atomic(ring);
1168 flush_dcache_page(ctx->ring_pages[0]);
1170 pr_debug("%li h%u t%u\n", ret, head, tail);
1172 mutex_unlock(&ctx->ring_lock);
1177 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1178 struct io_event __user *event, long *i)
1180 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1185 if (unlikely(atomic_read(&ctx->dead)))
1191 return ret < 0 || *i >= min_nr;
1194 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1195 struct io_event __user *event,
1196 struct timespec __user *timeout)
1198 ktime_t until = { .tv64 = KTIME_MAX };
1204 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1207 until = timespec_to_ktime(ts);
1211 * Note that aio_read_events() is being called as the conditional - i.e.
1212 * we're calling it after prepare_to_wait() has set task state to
1213 * TASK_INTERRUPTIBLE.
1215 * But aio_read_events() can block, and if it blocks it's going to flip
1216 * the task state back to TASK_RUNNING.
1218 * This should be ok, provided it doesn't flip the state back to
1219 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1220 * will only happen if the mutex_lock() call blocks, and we then find
1221 * the ringbuffer empty. So in practice we should be ok, but it's
1222 * something to be aware of when touching this code.
1224 wait_event_interruptible_hrtimeout(ctx->wait,
1225 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1227 if (!ret && signal_pending(current))
1234 * Create an aio_context capable of receiving at least nr_events.
1235 * ctxp must not point to an aio_context that already exists, and
1236 * must be initialized to 0 prior to the call. On successful
1237 * creation of the aio_context, *ctxp is filled in with the resulting
1238 * handle. May fail with -EINVAL if *ctxp is not initialized,
1239 * if the specified nr_events exceeds internal limits. May fail
1240 * with -EAGAIN if the specified nr_events exceeds the user's limit
1241 * of available events. May fail with -ENOMEM if insufficient kernel
1242 * resources are available. May fail with -EFAULT if an invalid
1243 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1246 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1248 struct kioctx *ioctx = NULL;
1252 ret = get_user(ctx, ctxp);
1257 if (unlikely(ctx || nr_events == 0)) {
1258 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1263 ioctx = ioctx_alloc(nr_events);
1264 ret = PTR_ERR(ioctx);
1265 if (!IS_ERR(ioctx)) {
1266 ret = put_user(ioctx->user_id, ctxp);
1268 kill_ioctx(current->mm, ioctx, NULL);
1269 percpu_ref_put(&ioctx->users);
1277 * Destroy the aio_context specified. May cancel any outstanding
1278 * AIOs and block on completion. Will fail with -ENOSYS if not
1279 * implemented. May fail with -EINVAL if the context pointed to
1282 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1284 struct kioctx *ioctx = lookup_ioctx(ctx);
1285 if (likely(NULL != ioctx)) {
1286 struct completion requests_done =
1287 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1290 /* Pass requests_done to kill_ioctx() where it can be set
1291 * in a thread-safe way. If we try to set it here then we have
1292 * a race condition if two io_destroy() called simultaneously.
1294 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1295 percpu_ref_put(&ioctx->users);
1297 /* Wait until all IO for the context are done. Otherwise kernel
1298 * keep using user-space buffers even if user thinks the context
1302 wait_for_completion(&requests_done);
1306 pr_debug("EINVAL: io_destroy: invalid context id\n");
1310 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1311 unsigned long, loff_t);
1312 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1314 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1315 int rw, char __user *buf,
1316 unsigned long *nr_segs,
1317 struct iovec **iovec,
1322 *nr_segs = kiocb->ki_nbytes;
1324 #ifdef CONFIG_COMPAT
1326 ret = compat_rw_copy_check_uvector(rw,
1327 (struct compat_iovec __user *)buf,
1328 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1331 ret = rw_copy_check_uvector(rw,
1332 (struct iovec __user *)buf,
1333 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1337 /* ki_nbytes now reflect bytes instead of segs */
1338 kiocb->ki_nbytes = ret;
1342 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1343 int rw, char __user *buf,
1344 unsigned long *nr_segs,
1345 struct iovec *iovec)
1347 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1350 iovec->iov_base = buf;
1351 iovec->iov_len = kiocb->ki_nbytes;
1358 * Performs the initial checks and io submission.
1360 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1361 char __user *buf, bool compat)
1363 struct file *file = req->ki_filp;
1365 unsigned long nr_segs;
1369 rw_iter_op *iter_op;
1370 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1371 struct iov_iter iter;
1374 case IOCB_CMD_PREAD:
1375 case IOCB_CMD_PREADV:
1378 rw_op = file->f_op->aio_read;
1379 iter_op = file->f_op->read_iter;
1382 case IOCB_CMD_PWRITE:
1383 case IOCB_CMD_PWRITEV:
1386 rw_op = file->f_op->aio_write;
1387 iter_op = file->f_op->write_iter;
1390 if (unlikely(!(file->f_mode & mode)))
1393 if (!rw_op && !iter_op)
1396 ret = (opcode == IOCB_CMD_PREADV ||
1397 opcode == IOCB_CMD_PWRITEV)
1398 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1400 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1403 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1405 if (iovec != inline_vecs)
1410 req->ki_nbytes = ret;
1412 /* XXX: move/kill - rw_verify_area()? */
1413 /* This matches the pread()/pwrite() logic */
1414 if (req->ki_pos < 0) {
1420 file_start_write(file);
1423 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1424 ret = iter_op(req, &iter);
1426 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1430 file_end_write(file);
1433 case IOCB_CMD_FDSYNC:
1434 if (!file->f_op->aio_fsync)
1437 ret = file->f_op->aio_fsync(req, 1);
1440 case IOCB_CMD_FSYNC:
1441 if (!file->f_op->aio_fsync)
1444 ret = file->f_op->aio_fsync(req, 0);
1448 pr_debug("EINVAL: no operation provided\n");
1452 if (iovec != inline_vecs)
1455 if (ret != -EIOCBQUEUED) {
1457 * There's no easy way to restart the syscall since other AIO's
1458 * may be already running. Just fail this IO with EINTR.
1460 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1461 ret == -ERESTARTNOHAND ||
1462 ret == -ERESTART_RESTARTBLOCK))
1464 aio_complete(req, ret, 0);
1470 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1471 struct iocb *iocb, bool compat)
1476 /* enforce forwards compatibility on users */
1477 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1478 pr_debug("EINVAL: reserve field set\n");
1482 /* prevent overflows */
1484 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1485 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1486 ((ssize_t)iocb->aio_nbytes < 0)
1488 pr_debug("EINVAL: io_submit: overflow check\n");
1492 req = aio_get_req(ctx);
1496 req->ki_filp = fget(iocb->aio_fildes);
1497 if (unlikely(!req->ki_filp)) {
1502 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1504 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1505 * instance of the file* now. The file descriptor must be
1506 * an eventfd() fd, and will be signaled for each completed
1507 * event using the eventfd_signal() function.
1509 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1510 if (IS_ERR(req->ki_eventfd)) {
1511 ret = PTR_ERR(req->ki_eventfd);
1512 req->ki_eventfd = NULL;
1517 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1518 if (unlikely(ret)) {
1519 pr_debug("EFAULT: aio_key\n");
1523 req->ki_obj.user = user_iocb;
1524 req->ki_user_data = iocb->aio_data;
1525 req->ki_pos = iocb->aio_offset;
1526 req->ki_nbytes = iocb->aio_nbytes;
1528 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1529 (char __user *)(unsigned long)iocb->aio_buf,
1536 put_reqs_available(ctx, 1);
1537 percpu_ref_put(&ctx->reqs);
1542 long do_io_submit(aio_context_t ctx_id, long nr,
1543 struct iocb __user *__user *iocbpp, bool compat)
1548 struct blk_plug plug;
1550 if (unlikely(nr < 0))
1553 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1554 nr = LONG_MAX/sizeof(*iocbpp);
1556 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1559 ctx = lookup_ioctx(ctx_id);
1560 if (unlikely(!ctx)) {
1561 pr_debug("EINVAL: invalid context id\n");
1565 blk_start_plug(&plug);
1568 * AKPM: should this return a partial result if some of the IOs were
1569 * successfully submitted?
1571 for (i=0; i<nr; i++) {
1572 struct iocb __user *user_iocb;
1575 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1580 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1585 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1589 blk_finish_plug(&plug);
1591 percpu_ref_put(&ctx->users);
1596 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1597 * the number of iocbs queued. May return -EINVAL if the aio_context
1598 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1599 * *iocbpp[0] is not properly initialized, if the operation specified
1600 * is invalid for the file descriptor in the iocb. May fail with
1601 * -EFAULT if any of the data structures point to invalid data. May
1602 * fail with -EBADF if the file descriptor specified in the first
1603 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1604 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1605 * fail with -ENOSYS if not implemented.
1607 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1608 struct iocb __user * __user *, iocbpp)
1610 return do_io_submit(ctx_id, nr, iocbpp, 0);
1614 * Finds a given iocb for cancellation.
1616 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1619 struct list_head *pos;
1621 assert_spin_locked(&ctx->ctx_lock);
1623 if (key != KIOCB_KEY)
1626 /* TODO: use a hash or array, this sucks. */
1627 list_for_each(pos, &ctx->active_reqs) {
1628 struct kiocb *kiocb = list_kiocb(pos);
1629 if (kiocb->ki_obj.user == iocb)
1636 * Attempts to cancel an iocb previously passed to io_submit. If
1637 * the operation is successfully cancelled, the resulting event is
1638 * copied into the memory pointed to by result without being placed
1639 * into the completion queue and 0 is returned. May fail with
1640 * -EFAULT if any of the data structures pointed to are invalid.
1641 * May fail with -EINVAL if aio_context specified by ctx_id is
1642 * invalid. May fail with -EAGAIN if the iocb specified was not
1643 * cancelled. Will fail with -ENOSYS if not implemented.
1645 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1646 struct io_event __user *, result)
1649 struct kiocb *kiocb;
1653 ret = get_user(key, &iocb->aio_key);
1657 ctx = lookup_ioctx(ctx_id);
1661 spin_lock_irq(&ctx->ctx_lock);
1663 kiocb = lookup_kiocb(ctx, iocb, key);
1665 ret = kiocb_cancel(kiocb);
1669 spin_unlock_irq(&ctx->ctx_lock);
1673 * The result argument is no longer used - the io_event is
1674 * always delivered via the ring buffer. -EINPROGRESS indicates
1675 * cancellation is progress:
1680 percpu_ref_put(&ctx->users);
1686 * Attempts to read at least min_nr events and up to nr events from
1687 * the completion queue for the aio_context specified by ctx_id. If
1688 * it succeeds, the number of read events is returned. May fail with
1689 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1690 * out of range, if timeout is out of range. May fail with -EFAULT
1691 * if any of the memory specified is invalid. May return 0 or
1692 * < min_nr if the timeout specified by timeout has elapsed
1693 * before sufficient events are available, where timeout == NULL
1694 * specifies an infinite timeout. Note that the timeout pointed to by
1695 * timeout is relative. Will fail with -ENOSYS if not implemented.
1697 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1700 struct io_event __user *, events,
1701 struct timespec __user *, timeout)
1703 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1706 if (likely(ioctx)) {
1707 if (likely(min_nr <= nr && min_nr >= 0))
1708 ret = read_events(ioctx, min_nr, nr, events, timeout);
1709 percpu_ref_put(&ioctx->users);