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 spinlock_t completion_lock;
145 } ____cacheline_aligned_in_smp;
147 struct page *internal_pages[AIO_RING_PAGES];
148 struct file *aio_ring_file;
153 /*------ sysctl variables----*/
154 static DEFINE_SPINLOCK(aio_nr_lock);
155 unsigned long aio_nr; /* current system wide number of aio requests */
156 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
157 /*----end sysctl variables---*/
159 static struct kmem_cache *kiocb_cachep;
160 static struct kmem_cache *kioctx_cachep;
162 static struct vfsmount *aio_mnt;
164 static const struct file_operations aio_ring_fops;
165 static const struct address_space_operations aio_ctx_aops;
167 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
169 struct qstr this = QSTR_INIT("[aio]", 5);
172 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
174 return ERR_CAST(inode);
176 inode->i_mapping->a_ops = &aio_ctx_aops;
177 inode->i_mapping->private_data = ctx;
178 inode->i_size = PAGE_SIZE * nr_pages;
180 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
183 return ERR_PTR(-ENOMEM);
185 path.mnt = mntget(aio_mnt);
187 d_instantiate(path.dentry, inode);
188 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
194 file->f_flags = O_RDWR;
198 static struct dentry *aio_mount(struct file_system_type *fs_type,
199 int flags, const char *dev_name, void *data)
201 static const struct dentry_operations ops = {
202 .d_dname = simple_dname,
204 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
208 * Creates the slab caches used by the aio routines, panic on
209 * failure as this is done early during the boot sequence.
211 static int __init aio_setup(void)
213 static struct file_system_type aio_fs = {
216 .kill_sb = kill_anon_super,
218 aio_mnt = kern_mount(&aio_fs);
220 panic("Failed to create aio fs mount.");
222 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
223 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
225 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
229 __initcall(aio_setup);
231 static void put_aio_ring_file(struct kioctx *ctx)
233 struct file *aio_ring_file = ctx->aio_ring_file;
235 truncate_setsize(aio_ring_file->f_inode, 0);
237 /* Prevent further access to the kioctx from migratepages */
238 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
239 aio_ring_file->f_inode->i_mapping->private_data = NULL;
240 ctx->aio_ring_file = NULL;
241 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
247 static void aio_free_ring(struct kioctx *ctx)
251 /* Disconnect the kiotx from the ring file. This prevents future
252 * accesses to the kioctx from page migration.
254 put_aio_ring_file(ctx);
256 for (i = 0; i < ctx->nr_pages; i++) {
258 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
259 page_count(ctx->ring_pages[i]));
260 page = ctx->ring_pages[i];
263 ctx->ring_pages[i] = NULL;
267 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
268 kfree(ctx->ring_pages);
269 ctx->ring_pages = NULL;
273 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
275 vma->vm_ops = &generic_file_vm_ops;
279 static const struct file_operations aio_ring_fops = {
280 .mmap = aio_ring_mmap,
283 static int aio_set_page_dirty(struct page *page)
288 #if IS_ENABLED(CONFIG_MIGRATION)
289 static int aio_migratepage(struct address_space *mapping, struct page *new,
290 struct page *old, enum migrate_mode mode)
299 /* mapping->private_lock here protects against the kioctx teardown. */
300 spin_lock(&mapping->private_lock);
301 ctx = mapping->private_data;
307 /* The ring_lock mutex. The prevents aio_read_events() from writing
308 * to the ring's head, and prevents page migration from mucking in
309 * a partially initialized kiotx.
311 if (!mutex_trylock(&ctx->ring_lock)) {
317 if (idx < (pgoff_t)ctx->nr_pages) {
318 /* Make sure the old page hasn't already been changed */
319 if (ctx->ring_pages[idx] != old)
327 /* Writeback must be complete */
328 BUG_ON(PageWriteback(old));
331 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
332 if (rc != MIGRATEPAGE_SUCCESS) {
337 /* Take completion_lock to prevent other writes to the ring buffer
338 * while the old page is copied to the new. This prevents new
339 * events from being lost.
341 spin_lock_irqsave(&ctx->completion_lock, flags);
342 migrate_page_copy(new, old);
343 BUG_ON(ctx->ring_pages[idx] != old);
344 ctx->ring_pages[idx] = new;
345 spin_unlock_irqrestore(&ctx->completion_lock, flags);
347 /* The old page is no longer accessible. */
351 mutex_unlock(&ctx->ring_lock);
353 spin_unlock(&mapping->private_lock);
358 static const struct address_space_operations aio_ctx_aops = {
359 .set_page_dirty = aio_set_page_dirty,
360 #if IS_ENABLED(CONFIG_MIGRATION)
361 .migratepage = aio_migratepage,
365 static int aio_setup_ring(struct kioctx *ctx)
367 struct aio_ring *ring;
368 unsigned nr_events = ctx->max_reqs;
369 struct mm_struct *mm = current->mm;
370 unsigned long size, unused;
375 /* Compensate for the ring buffer's head/tail overlap entry */
376 nr_events += 2; /* 1 is required, 2 for good luck */
378 size = sizeof(struct aio_ring);
379 size += sizeof(struct io_event) * nr_events;
381 nr_pages = PFN_UP(size);
385 file = aio_private_file(ctx, nr_pages);
387 ctx->aio_ring_file = NULL;
391 ctx->aio_ring_file = file;
392 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
393 / sizeof(struct io_event);
395 ctx->ring_pages = ctx->internal_pages;
396 if (nr_pages > AIO_RING_PAGES) {
397 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
399 if (!ctx->ring_pages) {
400 put_aio_ring_file(ctx);
405 for (i = 0; i < nr_pages; i++) {
407 page = find_or_create_page(file->f_inode->i_mapping,
408 i, GFP_HIGHUSER | __GFP_ZERO);
411 pr_debug("pid(%d) page[%d]->count=%d\n",
412 current->pid, i, page_count(page));
413 SetPageUptodate(page);
417 ctx->ring_pages[i] = page;
421 if (unlikely(i != nr_pages)) {
426 ctx->mmap_size = nr_pages * PAGE_SIZE;
427 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
429 down_write(&mm->mmap_sem);
430 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
431 PROT_READ | PROT_WRITE,
432 MAP_SHARED, 0, &unused);
433 up_write(&mm->mmap_sem);
434 if (IS_ERR((void *)ctx->mmap_base)) {
440 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
442 ctx->user_id = ctx->mmap_base;
443 ctx->nr_events = nr_events; /* trusted copy */
445 ring = kmap_atomic(ctx->ring_pages[0]);
446 ring->nr = nr_events; /* user copy */
448 ring->head = ring->tail = 0;
449 ring->magic = AIO_RING_MAGIC;
450 ring->compat_features = AIO_RING_COMPAT_FEATURES;
451 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
452 ring->header_length = sizeof(struct aio_ring);
454 flush_dcache_page(ctx->ring_pages[0]);
459 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
460 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
461 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
463 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
465 struct kioctx *ctx = req->ki_ctx;
468 spin_lock_irqsave(&ctx->ctx_lock, flags);
470 if (!req->ki_list.next)
471 list_add(&req->ki_list, &ctx->active_reqs);
473 req->ki_cancel = cancel;
475 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
477 EXPORT_SYMBOL(kiocb_set_cancel_fn);
479 static int kiocb_cancel(struct kiocb *kiocb)
481 kiocb_cancel_fn *old, *cancel;
484 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
485 * actually has a cancel function, hence the cmpxchg()
488 cancel = ACCESS_ONCE(kiocb->ki_cancel);
490 if (!cancel || cancel == KIOCB_CANCELLED)
494 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
495 } while (cancel != old);
497 return cancel(kiocb);
500 static void free_ioctx(struct work_struct *work)
502 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
504 pr_debug("freeing %p\n", ctx);
507 free_percpu(ctx->cpu);
508 kmem_cache_free(kioctx_cachep, ctx);
511 static void free_ioctx_reqs(struct percpu_ref *ref)
513 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
515 /* At this point we know that there are no any in-flight requests */
516 if (ctx->requests_done)
517 complete(ctx->requests_done);
519 INIT_WORK(&ctx->free_work, free_ioctx);
520 schedule_work(&ctx->free_work);
524 * When this function runs, the kioctx has been removed from the "hash table"
525 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
526 * now it's safe to cancel any that need to be.
528 static void free_ioctx_users(struct percpu_ref *ref)
530 struct kioctx *ctx = container_of(ref, struct kioctx, users);
533 spin_lock_irq(&ctx->ctx_lock);
535 while (!list_empty(&ctx->active_reqs)) {
536 req = list_first_entry(&ctx->active_reqs,
537 struct kiocb, ki_list);
539 list_del_init(&req->ki_list);
543 spin_unlock_irq(&ctx->ctx_lock);
545 percpu_ref_kill(&ctx->reqs);
546 percpu_ref_put(&ctx->reqs);
549 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
552 struct kioctx_table *table, *old;
553 struct aio_ring *ring;
555 spin_lock(&mm->ioctx_lock);
556 table = rcu_dereference_raw(mm->ioctx_table);
560 for (i = 0; i < table->nr; i++)
561 if (!table->table[i]) {
563 table->table[i] = ctx;
564 spin_unlock(&mm->ioctx_lock);
566 /* While kioctx setup is in progress,
567 * we are protected from page migration
568 * changes ring_pages by ->ring_lock.
570 ring = kmap_atomic(ctx->ring_pages[0]);
576 new_nr = (table ? table->nr : 1) * 4;
577 spin_unlock(&mm->ioctx_lock);
579 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
586 spin_lock(&mm->ioctx_lock);
587 old = rcu_dereference_raw(mm->ioctx_table);
590 rcu_assign_pointer(mm->ioctx_table, table);
591 } else if (table->nr > old->nr) {
592 memcpy(table->table, old->table,
593 old->nr * sizeof(struct kioctx *));
595 rcu_assign_pointer(mm->ioctx_table, table);
604 static void aio_nr_sub(unsigned nr)
606 spin_lock(&aio_nr_lock);
607 if (WARN_ON(aio_nr - nr > aio_nr))
611 spin_unlock(&aio_nr_lock);
615 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
617 static struct kioctx *ioctx_alloc(unsigned nr_events)
619 struct mm_struct *mm = current->mm;
624 * We keep track of the number of available ringbuffer slots, to prevent
625 * overflow (reqs_available), and we also use percpu counters for this.
627 * So since up to half the slots might be on other cpu's percpu counters
628 * and unavailable, double nr_events so userspace sees what they
629 * expected: additionally, we move req_batch slots to/from percpu
630 * counters at a time, so make sure that isn't 0:
632 nr_events = max(nr_events, num_possible_cpus() * 4);
635 /* Prevent overflows */
636 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
637 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
638 pr_debug("ENOMEM: nr_events too high\n");
639 return ERR_PTR(-EINVAL);
642 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
643 return ERR_PTR(-EAGAIN);
645 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
647 return ERR_PTR(-ENOMEM);
649 ctx->max_reqs = nr_events;
651 spin_lock_init(&ctx->ctx_lock);
652 spin_lock_init(&ctx->completion_lock);
653 mutex_init(&ctx->ring_lock);
654 /* Protect against page migration throughout kiotx setup by keeping
655 * the ring_lock mutex held until setup is complete. */
656 mutex_lock(&ctx->ring_lock);
657 init_waitqueue_head(&ctx->wait);
659 INIT_LIST_HEAD(&ctx->active_reqs);
661 if (percpu_ref_init(&ctx->users, free_ioctx_users))
664 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
667 ctx->cpu = alloc_percpu(struct kioctx_cpu);
671 err = aio_setup_ring(ctx);
675 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
676 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
677 if (ctx->req_batch < 1)
680 /* limit the number of system wide aios */
681 spin_lock(&aio_nr_lock);
682 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
683 aio_nr + nr_events < aio_nr) {
684 spin_unlock(&aio_nr_lock);
688 aio_nr += ctx->max_reqs;
689 spin_unlock(&aio_nr_lock);
691 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
692 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
694 err = ioctx_add_table(ctx, mm);
698 /* Release the ring_lock mutex now that all setup is complete. */
699 mutex_unlock(&ctx->ring_lock);
701 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
702 ctx, ctx->user_id, mm, ctx->nr_events);
706 aio_nr_sub(ctx->max_reqs);
710 mutex_unlock(&ctx->ring_lock);
711 free_percpu(ctx->cpu);
712 free_percpu(ctx->reqs.pcpu_count);
713 free_percpu(ctx->users.pcpu_count);
714 kmem_cache_free(kioctx_cachep, ctx);
715 pr_debug("error allocating ioctx %d\n", err);
720 * Cancels all outstanding aio requests on an aio context. Used
721 * when the processes owning a context have all exited to encourage
722 * the rapid destruction of the kioctx.
724 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
725 struct completion *requests_done)
727 struct kioctx_table *table;
729 if (atomic_xchg(&ctx->dead, 1))
733 spin_lock(&mm->ioctx_lock);
734 table = rcu_dereference_raw(mm->ioctx_table);
735 WARN_ON(ctx != table->table[ctx->id]);
736 table->table[ctx->id] = NULL;
737 spin_unlock(&mm->ioctx_lock);
739 /* percpu_ref_kill() will do the necessary call_rcu() */
740 wake_up_all(&ctx->wait);
743 * It'd be more correct to do this in free_ioctx(), after all
744 * the outstanding kiocbs have finished - but by then io_destroy
745 * has already returned, so io_setup() could potentially return
746 * -EAGAIN with no ioctxs actually in use (as far as userspace
749 aio_nr_sub(ctx->max_reqs);
752 vm_munmap(ctx->mmap_base, ctx->mmap_size);
754 ctx->requests_done = requests_done;
755 percpu_ref_kill(&ctx->users);
759 /* wait_on_sync_kiocb:
760 * Waits on the given sync kiocb to complete.
762 ssize_t wait_on_sync_kiocb(struct kiocb *req)
764 while (!req->ki_ctx) {
765 set_current_state(TASK_UNINTERRUPTIBLE);
770 __set_current_state(TASK_RUNNING);
771 return req->ki_user_data;
773 EXPORT_SYMBOL(wait_on_sync_kiocb);
776 * exit_aio: called when the last user of mm goes away. At this point, there is
777 * no way for any new requests to be submited or any of the io_* syscalls to be
778 * called on the context.
780 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
783 void exit_aio(struct mm_struct *mm)
785 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
791 for (i = 0; i < table->nr; ++i) {
792 struct kioctx *ctx = table->table[i];
797 * We don't need to bother with munmap() here - exit_mmap(mm)
798 * is coming and it'll unmap everything. And we simply can't,
799 * this is not necessarily our ->mm.
800 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
801 * that it needs to unmap the area, just set it to 0.
804 kill_ioctx(mm, ctx, NULL);
807 RCU_INIT_POINTER(mm->ioctx_table, NULL);
811 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
813 struct kioctx_cpu *kcpu;
816 local_irq_save(flags);
817 kcpu = this_cpu_ptr(ctx->cpu);
818 kcpu->reqs_available += nr;
820 while (kcpu->reqs_available >= ctx->req_batch * 2) {
821 kcpu->reqs_available -= ctx->req_batch;
822 atomic_add(ctx->req_batch, &ctx->reqs_available);
825 local_irq_restore(flags);
828 static bool get_reqs_available(struct kioctx *ctx)
830 struct kioctx_cpu *kcpu;
834 local_irq_save(flags);
835 kcpu = this_cpu_ptr(ctx->cpu);
836 if (!kcpu->reqs_available) {
837 int old, avail = atomic_read(&ctx->reqs_available);
840 if (avail < ctx->req_batch)
844 avail = atomic_cmpxchg(&ctx->reqs_available,
845 avail, avail - ctx->req_batch);
846 } while (avail != old);
848 kcpu->reqs_available += ctx->req_batch;
852 kcpu->reqs_available--;
854 local_irq_restore(flags);
859 * Allocate a slot for an aio request.
860 * Returns NULL if no requests are free.
862 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
866 if (!get_reqs_available(ctx))
869 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
873 percpu_ref_get(&ctx->reqs);
878 put_reqs_available(ctx, 1);
882 static void kiocb_free(struct kiocb *req)
886 if (req->ki_eventfd != NULL)
887 eventfd_ctx_put(req->ki_eventfd);
888 kmem_cache_free(kiocb_cachep, req);
891 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
893 struct aio_ring __user *ring = (void __user *)ctx_id;
894 struct mm_struct *mm = current->mm;
895 struct kioctx *ctx, *ret = NULL;
896 struct kioctx_table *table;
899 if (get_user(id, &ring->id))
903 table = rcu_dereference(mm->ioctx_table);
905 if (!table || id >= table->nr)
908 ctx = table->table[id];
909 if (ctx && ctx->user_id == ctx_id) {
910 percpu_ref_get(&ctx->users);
919 * Called when the io request on the given iocb is complete.
921 void aio_complete(struct kiocb *iocb, long res, long res2)
923 struct kioctx *ctx = iocb->ki_ctx;
924 struct aio_ring *ring;
925 struct io_event *ev_page, *event;
930 * Special case handling for sync iocbs:
931 * - events go directly into the iocb for fast handling
932 * - the sync task with the iocb in its stack holds the single iocb
933 * ref, no other paths have a way to get another ref
934 * - the sync task helpfully left a reference to itself in the iocb
936 if (is_sync_kiocb(iocb)) {
937 iocb->ki_user_data = res;
939 iocb->ki_ctx = ERR_PTR(-EXDEV);
940 wake_up_process(iocb->ki_obj.tsk);
944 if (iocb->ki_list.next) {
947 spin_lock_irqsave(&ctx->ctx_lock, flags);
948 list_del(&iocb->ki_list);
949 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
953 * Add a completion event to the ring buffer. Must be done holding
954 * ctx->completion_lock to prevent other code from messing with the tail
955 * pointer since we might be called from irq context.
957 spin_lock_irqsave(&ctx->completion_lock, flags);
960 pos = tail + AIO_EVENTS_OFFSET;
962 if (++tail >= ctx->nr_events)
965 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
966 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
968 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
969 event->data = iocb->ki_user_data;
973 kunmap_atomic(ev_page);
974 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
976 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
977 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
980 /* after flagging the request as done, we
981 * must never even look at it again
983 smp_wmb(); /* make event visible before updating tail */
987 ring = kmap_atomic(ctx->ring_pages[0]);
990 flush_dcache_page(ctx->ring_pages[0]);
992 spin_unlock_irqrestore(&ctx->completion_lock, flags);
994 pr_debug("added to ring %p at [%u]\n", iocb, tail);
997 * Check if the user asked us to deliver the result through an
998 * eventfd. The eventfd_signal() function is safe to be called
1001 if (iocb->ki_eventfd != NULL)
1002 eventfd_signal(iocb->ki_eventfd, 1);
1004 /* everything turned out well, dispose of the aiocb. */
1006 put_reqs_available(ctx, 1);
1009 * We have to order our ring_info tail store above and test
1010 * of the wait list below outside the wait lock. This is
1011 * like in wake_up_bit() where clearing a bit has to be
1012 * ordered with the unlocked test.
1016 if (waitqueue_active(&ctx->wait))
1017 wake_up(&ctx->wait);
1019 percpu_ref_put(&ctx->reqs);
1021 EXPORT_SYMBOL(aio_complete);
1023 /* aio_read_events_ring
1024 * Pull an event off of the ioctx's event ring. Returns the number of
1027 static long aio_read_events_ring(struct kioctx *ctx,
1028 struct io_event __user *event, long nr)
1030 struct aio_ring *ring;
1031 unsigned head, tail, pos;
1035 mutex_lock(&ctx->ring_lock);
1037 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1038 ring = kmap_atomic(ctx->ring_pages[0]);
1041 kunmap_atomic(ring);
1043 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1048 head %= ctx->nr_events;
1049 tail %= ctx->nr_events;
1053 struct io_event *ev;
1056 avail = (head <= tail ? tail : ctx->nr_events) - head;
1060 avail = min(avail, nr - ret);
1061 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1062 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1064 pos = head + AIO_EVENTS_OFFSET;
1065 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1066 pos %= AIO_EVENTS_PER_PAGE;
1069 copy_ret = copy_to_user(event + ret, ev + pos,
1070 sizeof(*ev) * avail);
1073 if (unlikely(copy_ret)) {
1080 head %= ctx->nr_events;
1083 ring = kmap_atomic(ctx->ring_pages[0]);
1085 kunmap_atomic(ring);
1086 flush_dcache_page(ctx->ring_pages[0]);
1088 pr_debug("%li h%u t%u\n", ret, head, tail);
1090 mutex_unlock(&ctx->ring_lock);
1095 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1096 struct io_event __user *event, long *i)
1098 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1103 if (unlikely(atomic_read(&ctx->dead)))
1109 return ret < 0 || *i >= min_nr;
1112 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1113 struct io_event __user *event,
1114 struct timespec __user *timeout)
1116 ktime_t until = { .tv64 = KTIME_MAX };
1122 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1125 until = timespec_to_ktime(ts);
1129 * Note that aio_read_events() is being called as the conditional - i.e.
1130 * we're calling it after prepare_to_wait() has set task state to
1131 * TASK_INTERRUPTIBLE.
1133 * But aio_read_events() can block, and if it blocks it's going to flip
1134 * the task state back to TASK_RUNNING.
1136 * This should be ok, provided it doesn't flip the state back to
1137 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1138 * will only happen if the mutex_lock() call blocks, and we then find
1139 * the ringbuffer empty. So in practice we should be ok, but it's
1140 * something to be aware of when touching this code.
1142 wait_event_interruptible_hrtimeout(ctx->wait,
1143 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1145 if (!ret && signal_pending(current))
1152 * Create an aio_context capable of receiving at least nr_events.
1153 * ctxp must not point to an aio_context that already exists, and
1154 * must be initialized to 0 prior to the call. On successful
1155 * creation of the aio_context, *ctxp is filled in with the resulting
1156 * handle. May fail with -EINVAL if *ctxp is not initialized,
1157 * if the specified nr_events exceeds internal limits. May fail
1158 * with -EAGAIN if the specified nr_events exceeds the user's limit
1159 * of available events. May fail with -ENOMEM if insufficient kernel
1160 * resources are available. May fail with -EFAULT if an invalid
1161 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1164 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1166 struct kioctx *ioctx = NULL;
1170 ret = get_user(ctx, ctxp);
1175 if (unlikely(ctx || nr_events == 0)) {
1176 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1181 ioctx = ioctx_alloc(nr_events);
1182 ret = PTR_ERR(ioctx);
1183 if (!IS_ERR(ioctx)) {
1184 ret = put_user(ioctx->user_id, ctxp);
1186 kill_ioctx(current->mm, ioctx, NULL);
1187 percpu_ref_put(&ioctx->users);
1195 * Destroy the aio_context specified. May cancel any outstanding
1196 * AIOs and block on completion. Will fail with -ENOSYS if not
1197 * implemented. May fail with -EINVAL if the context pointed to
1200 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1202 struct kioctx *ioctx = lookup_ioctx(ctx);
1203 if (likely(NULL != ioctx)) {
1204 struct completion requests_done =
1205 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1208 /* Pass requests_done to kill_ioctx() where it can be set
1209 * in a thread-safe way. If we try to set it here then we have
1210 * a race condition if two io_destroy() called simultaneously.
1212 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1213 percpu_ref_put(&ioctx->users);
1215 /* Wait until all IO for the context are done. Otherwise kernel
1216 * keep using user-space buffers even if user thinks the context
1220 wait_for_completion(&requests_done);
1224 pr_debug("EINVAL: io_destroy: invalid context id\n");
1228 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1229 unsigned long, loff_t);
1230 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1232 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1233 int rw, char __user *buf,
1234 unsigned long *nr_segs,
1235 struct iovec **iovec,
1240 *nr_segs = kiocb->ki_nbytes;
1242 #ifdef CONFIG_COMPAT
1244 ret = compat_rw_copy_check_uvector(rw,
1245 (struct compat_iovec __user *)buf,
1246 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1249 ret = rw_copy_check_uvector(rw,
1250 (struct iovec __user *)buf,
1251 *nr_segs, UIO_FASTIOV, *iovec, iovec);
1255 /* ki_nbytes now reflect bytes instead of segs */
1256 kiocb->ki_nbytes = ret;
1260 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1261 int rw, char __user *buf,
1262 unsigned long *nr_segs,
1263 struct iovec *iovec)
1265 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1268 iovec->iov_base = buf;
1269 iovec->iov_len = kiocb->ki_nbytes;
1276 * Performs the initial checks and io submission.
1278 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1279 char __user *buf, bool compat)
1281 struct file *file = req->ki_filp;
1283 unsigned long nr_segs;
1287 rw_iter_op *iter_op;
1288 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1289 struct iov_iter iter;
1292 case IOCB_CMD_PREAD:
1293 case IOCB_CMD_PREADV:
1296 rw_op = file->f_op->aio_read;
1297 iter_op = file->f_op->read_iter;
1300 case IOCB_CMD_PWRITE:
1301 case IOCB_CMD_PWRITEV:
1304 rw_op = file->f_op->aio_write;
1305 iter_op = file->f_op->write_iter;
1308 if (unlikely(!(file->f_mode & mode)))
1311 if (!rw_op && !iter_op)
1314 ret = (opcode == IOCB_CMD_PREADV ||
1315 opcode == IOCB_CMD_PWRITEV)
1316 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1318 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1321 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1323 if (iovec != inline_vecs)
1328 req->ki_nbytes = ret;
1330 /* XXX: move/kill - rw_verify_area()? */
1331 /* This matches the pread()/pwrite() logic */
1332 if (req->ki_pos < 0) {
1338 file_start_write(file);
1341 iov_iter_init(&iter, rw, iovec, nr_segs, req->ki_nbytes);
1342 ret = iter_op(req, &iter);
1344 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1348 file_end_write(file);
1351 case IOCB_CMD_FDSYNC:
1352 if (!file->f_op->aio_fsync)
1355 ret = file->f_op->aio_fsync(req, 1);
1358 case IOCB_CMD_FSYNC:
1359 if (!file->f_op->aio_fsync)
1362 ret = file->f_op->aio_fsync(req, 0);
1366 pr_debug("EINVAL: no operation provided\n");
1370 if (iovec != inline_vecs)
1373 if (ret != -EIOCBQUEUED) {
1375 * There's no easy way to restart the syscall since other AIO's
1376 * may be already running. Just fail this IO with EINTR.
1378 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1379 ret == -ERESTARTNOHAND ||
1380 ret == -ERESTART_RESTARTBLOCK))
1382 aio_complete(req, ret, 0);
1388 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1389 struct iocb *iocb, bool compat)
1394 /* enforce forwards compatibility on users */
1395 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1396 pr_debug("EINVAL: reserve field set\n");
1400 /* prevent overflows */
1402 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1403 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1404 ((ssize_t)iocb->aio_nbytes < 0)
1406 pr_debug("EINVAL: io_submit: overflow check\n");
1410 req = aio_get_req(ctx);
1414 req->ki_filp = fget(iocb->aio_fildes);
1415 if (unlikely(!req->ki_filp)) {
1420 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1422 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1423 * instance of the file* now. The file descriptor must be
1424 * an eventfd() fd, and will be signaled for each completed
1425 * event using the eventfd_signal() function.
1427 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1428 if (IS_ERR(req->ki_eventfd)) {
1429 ret = PTR_ERR(req->ki_eventfd);
1430 req->ki_eventfd = NULL;
1435 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1436 if (unlikely(ret)) {
1437 pr_debug("EFAULT: aio_key\n");
1441 req->ki_obj.user = user_iocb;
1442 req->ki_user_data = iocb->aio_data;
1443 req->ki_pos = iocb->aio_offset;
1444 req->ki_nbytes = iocb->aio_nbytes;
1446 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1447 (char __user *)(unsigned long)iocb->aio_buf,
1454 put_reqs_available(ctx, 1);
1455 percpu_ref_put(&ctx->reqs);
1460 long do_io_submit(aio_context_t ctx_id, long nr,
1461 struct iocb __user *__user *iocbpp, bool compat)
1466 struct blk_plug plug;
1468 if (unlikely(nr < 0))
1471 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1472 nr = LONG_MAX/sizeof(*iocbpp);
1474 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1477 ctx = lookup_ioctx(ctx_id);
1478 if (unlikely(!ctx)) {
1479 pr_debug("EINVAL: invalid context id\n");
1483 blk_start_plug(&plug);
1486 * AKPM: should this return a partial result if some of the IOs were
1487 * successfully submitted?
1489 for (i=0; i<nr; i++) {
1490 struct iocb __user *user_iocb;
1493 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1498 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1503 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1507 blk_finish_plug(&plug);
1509 percpu_ref_put(&ctx->users);
1514 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1515 * the number of iocbs queued. May return -EINVAL if the aio_context
1516 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1517 * *iocbpp[0] is not properly initialized, if the operation specified
1518 * is invalid for the file descriptor in the iocb. May fail with
1519 * -EFAULT if any of the data structures point to invalid data. May
1520 * fail with -EBADF if the file descriptor specified in the first
1521 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1522 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1523 * fail with -ENOSYS if not implemented.
1525 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1526 struct iocb __user * __user *, iocbpp)
1528 return do_io_submit(ctx_id, nr, iocbpp, 0);
1532 * Finds a given iocb for cancellation.
1534 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1537 struct list_head *pos;
1539 assert_spin_locked(&ctx->ctx_lock);
1541 if (key != KIOCB_KEY)
1544 /* TODO: use a hash or array, this sucks. */
1545 list_for_each(pos, &ctx->active_reqs) {
1546 struct kiocb *kiocb = list_kiocb(pos);
1547 if (kiocb->ki_obj.user == iocb)
1554 * Attempts to cancel an iocb previously passed to io_submit. If
1555 * the operation is successfully cancelled, the resulting event is
1556 * copied into the memory pointed to by result without being placed
1557 * into the completion queue and 0 is returned. May fail with
1558 * -EFAULT if any of the data structures pointed to are invalid.
1559 * May fail with -EINVAL if aio_context specified by ctx_id is
1560 * invalid. May fail with -EAGAIN if the iocb specified was not
1561 * cancelled. Will fail with -ENOSYS if not implemented.
1563 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1564 struct io_event __user *, result)
1567 struct kiocb *kiocb;
1571 ret = get_user(key, &iocb->aio_key);
1575 ctx = lookup_ioctx(ctx_id);
1579 spin_lock_irq(&ctx->ctx_lock);
1581 kiocb = lookup_kiocb(ctx, iocb, key);
1583 ret = kiocb_cancel(kiocb);
1587 spin_unlock_irq(&ctx->ctx_lock);
1591 * The result argument is no longer used - the io_event is
1592 * always delivered via the ring buffer. -EINPROGRESS indicates
1593 * cancellation is progress:
1598 percpu_ref_put(&ctx->users);
1604 * Attempts to read at least min_nr events and up to nr events from
1605 * the completion queue for the aio_context specified by ctx_id. If
1606 * it succeeds, the number of read events is returned. May fail with
1607 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1608 * out of range, if timeout is out of range. May fail with -EFAULT
1609 * if any of the memory specified is invalid. May return 0 or
1610 * < min_nr if the timeout specified by timeout has elapsed
1611 * before sufficient events are available, where timeout == NULL
1612 * specifies an infinite timeout. Note that the timeout pointed to by
1613 * timeout is relative. Will fail with -ENOSYS if not implemented.
1615 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1618 struct io_event __user *, events,
1619 struct timespec __user *, timeout)
1621 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1624 if (likely(ioctx)) {
1625 if (likely(min_nr <= nr && min_nr >= 0))
1626 ret = read_events(ioctx, min_nr, nr, events, timeout);
1627 percpu_ref_put(&ioctx->users);