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;
195 file->private_data = ctx;
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, 0xa10a10a1);
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 kioctx *ctx, 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 kmem_cache_free(kioctx_cachep, ctx);
512 static void free_ioctx_reqs(struct percpu_ref *ref)
514 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
516 /* At this point we know that there are no any in-flight requests */
517 if (ctx->requests_done)
518 complete(ctx->requests_done);
520 INIT_WORK(&ctx->free_work, free_ioctx);
521 schedule_work(&ctx->free_work);
525 * When this function runs, the kioctx has been removed from the "hash table"
526 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
527 * now it's safe to cancel any that need to be.
529 static void free_ioctx_users(struct percpu_ref *ref)
531 struct kioctx *ctx = container_of(ref, struct kioctx, users);
534 spin_lock_irq(&ctx->ctx_lock);
536 while (!list_empty(&ctx->active_reqs)) {
537 req = list_first_entry(&ctx->active_reqs,
538 struct kiocb, ki_list);
540 list_del_init(&req->ki_list);
541 kiocb_cancel(ctx, req);
544 spin_unlock_irq(&ctx->ctx_lock);
546 percpu_ref_kill(&ctx->reqs);
547 percpu_ref_put(&ctx->reqs);
550 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
553 struct kioctx_table *table, *old;
554 struct aio_ring *ring;
556 spin_lock(&mm->ioctx_lock);
558 table = rcu_dereference(mm->ioctx_table);
562 for (i = 0; i < table->nr; i++)
563 if (!table->table[i]) {
565 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;
582 spin_unlock(&mm->ioctx_lock);
584 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
591 spin_lock(&mm->ioctx_lock);
593 old = rcu_dereference(mm->ioctx_table);
596 rcu_assign_pointer(mm->ioctx_table, table);
597 } else if (table->nr > old->nr) {
598 memcpy(table->table, old->table,
599 old->nr * sizeof(struct kioctx *));
601 rcu_assign_pointer(mm->ioctx_table, table);
610 static void aio_nr_sub(unsigned nr)
612 spin_lock(&aio_nr_lock);
613 if (WARN_ON(aio_nr - nr > aio_nr))
617 spin_unlock(&aio_nr_lock);
621 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
623 static struct kioctx *ioctx_alloc(unsigned nr_events)
625 struct mm_struct *mm = current->mm;
630 * We keep track of the number of available ringbuffer slots, to prevent
631 * overflow (reqs_available), and we also use percpu counters for this.
633 * So since up to half the slots might be on other cpu's percpu counters
634 * and unavailable, double nr_events so userspace sees what they
635 * expected: additionally, we move req_batch slots to/from percpu
636 * counters at a time, so make sure that isn't 0:
638 nr_events = max(nr_events, num_possible_cpus() * 4);
641 /* Prevent overflows */
642 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
643 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
644 pr_debug("ENOMEM: nr_events too high\n");
645 return ERR_PTR(-EINVAL);
648 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
649 return ERR_PTR(-EAGAIN);
651 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
653 return ERR_PTR(-ENOMEM);
655 ctx->max_reqs = nr_events;
657 spin_lock_init(&ctx->ctx_lock);
658 spin_lock_init(&ctx->completion_lock);
659 mutex_init(&ctx->ring_lock);
660 /* Protect against page migration throughout kiotx setup by keeping
661 * the ring_lock mutex held until setup is complete. */
662 mutex_lock(&ctx->ring_lock);
663 init_waitqueue_head(&ctx->wait);
665 INIT_LIST_HEAD(&ctx->active_reqs);
667 if (percpu_ref_init(&ctx->users, free_ioctx_users))
670 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
673 ctx->cpu = alloc_percpu(struct kioctx_cpu);
677 err = aio_setup_ring(ctx);
681 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
682 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
683 if (ctx->req_batch < 1)
686 /* limit the number of system wide aios */
687 spin_lock(&aio_nr_lock);
688 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
689 aio_nr + nr_events < aio_nr) {
690 spin_unlock(&aio_nr_lock);
694 aio_nr += ctx->max_reqs;
695 spin_unlock(&aio_nr_lock);
697 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
698 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
700 err = ioctx_add_table(ctx, mm);
704 /* Release the ring_lock mutex now that all setup is complete. */
705 mutex_unlock(&ctx->ring_lock);
707 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
708 ctx, ctx->user_id, mm, ctx->nr_events);
712 aio_nr_sub(ctx->max_reqs);
716 mutex_unlock(&ctx->ring_lock);
717 free_percpu(ctx->cpu);
718 free_percpu(ctx->reqs.pcpu_count);
719 free_percpu(ctx->users.pcpu_count);
720 kmem_cache_free(kioctx_cachep, ctx);
721 pr_debug("error allocating ioctx %d\n", err);
726 * Cancels all outstanding aio requests on an aio context. Used
727 * when the processes owning a context have all exited to encourage
728 * the rapid destruction of the kioctx.
730 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
731 struct completion *requests_done)
733 if (!atomic_xchg(&ctx->dead, 1)) {
734 struct kioctx_table *table;
736 spin_lock(&mm->ioctx_lock);
738 table = rcu_dereference(mm->ioctx_table);
740 WARN_ON(ctx != table->table[ctx->id]);
741 table->table[ctx->id] = NULL;
743 spin_unlock(&mm->ioctx_lock);
745 /* percpu_ref_kill() will do the necessary call_rcu() */
746 wake_up_all(&ctx->wait);
749 * It'd be more correct to do this in free_ioctx(), after all
750 * the outstanding kiocbs have finished - but by then io_destroy
751 * has already returned, so io_setup() could potentially return
752 * -EAGAIN with no ioctxs actually in use (as far as userspace
755 aio_nr_sub(ctx->max_reqs);
758 vm_munmap(ctx->mmap_base, ctx->mmap_size);
760 ctx->requests_done = requests_done;
761 percpu_ref_kill(&ctx->users);
764 complete(requests_done);
768 /* wait_on_sync_kiocb:
769 * Waits on the given sync kiocb to complete.
771 ssize_t wait_on_sync_kiocb(struct kiocb *req)
773 while (!req->ki_ctx) {
774 set_current_state(TASK_UNINTERRUPTIBLE);
779 __set_current_state(TASK_RUNNING);
780 return req->ki_user_data;
782 EXPORT_SYMBOL(wait_on_sync_kiocb);
785 * exit_aio: called when the last user of mm goes away. At this point, there is
786 * no way for any new requests to be submited or any of the io_* syscalls to be
787 * called on the context.
789 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
792 void exit_aio(struct mm_struct *mm)
794 struct kioctx_table *table;
800 table = rcu_dereference(mm->ioctx_table);
803 if (!table || i >= table->nr) {
805 rcu_assign_pointer(mm->ioctx_table, NULL);
811 ctx = table->table[i++];
817 * We don't need to bother with munmap() here -
818 * exit_mmap(mm) is coming and it'll unmap everything.
819 * Since aio_free_ring() uses non-zero ->mmap_size
820 * as indicator that it needs to unmap the area,
821 * just set it to 0; aio_free_ring() is the only
822 * place that uses ->mmap_size, so it's safe.
826 kill_ioctx(mm, ctx, NULL);
830 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
832 struct kioctx_cpu *kcpu;
835 kcpu = this_cpu_ptr(ctx->cpu);
837 kcpu->reqs_available += nr;
838 while (kcpu->reqs_available >= ctx->req_batch * 2) {
839 kcpu->reqs_available -= ctx->req_batch;
840 atomic_add(ctx->req_batch, &ctx->reqs_available);
846 static bool get_reqs_available(struct kioctx *ctx)
848 struct kioctx_cpu *kcpu;
852 kcpu = this_cpu_ptr(ctx->cpu);
854 if (!kcpu->reqs_available) {
855 int old, avail = atomic_read(&ctx->reqs_available);
858 if (avail < ctx->req_batch)
862 avail = atomic_cmpxchg(&ctx->reqs_available,
863 avail, avail - ctx->req_batch);
864 } while (avail != old);
866 kcpu->reqs_available += ctx->req_batch;
870 kcpu->reqs_available--;
877 * Allocate a slot for an aio request.
878 * Returns NULL if no requests are free.
880 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
884 if (!get_reqs_available(ctx))
887 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
891 percpu_ref_get(&ctx->reqs);
896 put_reqs_available(ctx, 1);
900 static void kiocb_free(struct kiocb *req)
904 if (req->ki_eventfd != NULL)
905 eventfd_ctx_put(req->ki_eventfd);
906 kmem_cache_free(kiocb_cachep, req);
909 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
911 struct aio_ring __user *ring = (void __user *)ctx_id;
912 struct mm_struct *mm = current->mm;
913 struct kioctx *ctx, *ret = NULL;
914 struct kioctx_table *table;
917 if (get_user(id, &ring->id))
921 table = rcu_dereference(mm->ioctx_table);
923 if (!table || id >= table->nr)
926 ctx = table->table[id];
927 if (ctx && ctx->user_id == ctx_id) {
928 percpu_ref_get(&ctx->users);
937 * Called when the io request on the given iocb is complete.
939 void aio_complete(struct kiocb *iocb, long res, long res2)
941 struct kioctx *ctx = iocb->ki_ctx;
942 struct aio_ring *ring;
943 struct io_event *ev_page, *event;
948 * Special case handling for sync iocbs:
949 * - events go directly into the iocb for fast handling
950 * - the sync task with the iocb in its stack holds the single iocb
951 * ref, no other paths have a way to get another ref
952 * - the sync task helpfully left a reference to itself in the iocb
954 if (is_sync_kiocb(iocb)) {
955 iocb->ki_user_data = res;
957 iocb->ki_ctx = ERR_PTR(-EXDEV);
958 wake_up_process(iocb->ki_obj.tsk);
962 if (iocb->ki_list.next) {
965 spin_lock_irqsave(&ctx->ctx_lock, flags);
966 list_del(&iocb->ki_list);
967 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
971 * Add a completion event to the ring buffer. Must be done holding
972 * ctx->completion_lock to prevent other code from messing with the tail
973 * pointer since we might be called from irq context.
975 spin_lock_irqsave(&ctx->completion_lock, flags);
978 pos = tail + AIO_EVENTS_OFFSET;
980 if (++tail >= ctx->nr_events)
983 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
984 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
986 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
987 event->data = iocb->ki_user_data;
991 kunmap_atomic(ev_page);
992 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
994 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
995 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
998 /* after flagging the request as done, we
999 * must never even look at it again
1001 smp_wmb(); /* make event visible before updating tail */
1005 ring = kmap_atomic(ctx->ring_pages[0]);
1007 kunmap_atomic(ring);
1008 flush_dcache_page(ctx->ring_pages[0]);
1010 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1012 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1015 * Check if the user asked us to deliver the result through an
1016 * eventfd. The eventfd_signal() function is safe to be called
1019 if (iocb->ki_eventfd != NULL)
1020 eventfd_signal(iocb->ki_eventfd, 1);
1022 /* everything turned out well, dispose of the aiocb. */
1024 put_reqs_available(ctx, 1);
1027 * We have to order our ring_info tail store above and test
1028 * of the wait list below outside the wait lock. This is
1029 * like in wake_up_bit() where clearing a bit has to be
1030 * ordered with the unlocked test.
1034 if (waitqueue_active(&ctx->wait))
1035 wake_up(&ctx->wait);
1037 percpu_ref_put(&ctx->reqs);
1039 EXPORT_SYMBOL(aio_complete);
1042 * Pull an event off of the ioctx's event ring. Returns the number of
1045 static long aio_read_events_ring(struct kioctx *ctx,
1046 struct io_event __user *event, long nr)
1048 struct aio_ring *ring;
1049 unsigned head, tail, pos;
1053 mutex_lock(&ctx->ring_lock);
1055 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1056 ring = kmap_atomic(ctx->ring_pages[0]);
1059 kunmap_atomic(ring);
1061 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1066 head %= ctx->nr_events;
1067 tail %= ctx->nr_events;
1071 struct io_event *ev;
1074 avail = (head <= tail ? tail : ctx->nr_events) - head;
1078 avail = min(avail, nr - ret);
1079 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1080 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1082 pos = head + AIO_EVENTS_OFFSET;
1083 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1084 pos %= AIO_EVENTS_PER_PAGE;
1087 copy_ret = copy_to_user(event + ret, ev + pos,
1088 sizeof(*ev) * avail);
1091 if (unlikely(copy_ret)) {
1098 head %= ctx->nr_events;
1101 ring = kmap_atomic(ctx->ring_pages[0]);
1103 kunmap_atomic(ring);
1104 flush_dcache_page(ctx->ring_pages[0]);
1106 pr_debug("%li h%u t%u\n", ret, head, tail);
1108 mutex_unlock(&ctx->ring_lock);
1113 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1114 struct io_event __user *event, long *i)
1116 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1121 if (unlikely(atomic_read(&ctx->dead)))
1127 return ret < 0 || *i >= min_nr;
1130 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1131 struct io_event __user *event,
1132 struct timespec __user *timeout)
1134 ktime_t until = { .tv64 = KTIME_MAX };
1140 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1143 until = timespec_to_ktime(ts);
1147 * Note that aio_read_events() is being called as the conditional - i.e.
1148 * we're calling it after prepare_to_wait() has set task state to
1149 * TASK_INTERRUPTIBLE.
1151 * But aio_read_events() can block, and if it blocks it's going to flip
1152 * the task state back to TASK_RUNNING.
1154 * This should be ok, provided it doesn't flip the state back to
1155 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1156 * will only happen if the mutex_lock() call blocks, and we then find
1157 * the ringbuffer empty. So in practice we should be ok, but it's
1158 * something to be aware of when touching this code.
1160 wait_event_interruptible_hrtimeout(ctx->wait,
1161 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1163 if (!ret && signal_pending(current))
1170 * Create an aio_context capable of receiving at least nr_events.
1171 * ctxp must not point to an aio_context that already exists, and
1172 * must be initialized to 0 prior to the call. On successful
1173 * creation of the aio_context, *ctxp is filled in with the resulting
1174 * handle. May fail with -EINVAL if *ctxp is not initialized,
1175 * if the specified nr_events exceeds internal limits. May fail
1176 * with -EAGAIN if the specified nr_events exceeds the user's limit
1177 * of available events. May fail with -ENOMEM if insufficient kernel
1178 * resources are available. May fail with -EFAULT if an invalid
1179 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1182 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1184 struct kioctx *ioctx = NULL;
1188 ret = get_user(ctx, ctxp);
1193 if (unlikely(ctx || nr_events == 0)) {
1194 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1199 ioctx = ioctx_alloc(nr_events);
1200 ret = PTR_ERR(ioctx);
1201 if (!IS_ERR(ioctx)) {
1202 ret = put_user(ioctx->user_id, ctxp);
1204 kill_ioctx(current->mm, ioctx, NULL);
1205 percpu_ref_put(&ioctx->users);
1213 * Destroy the aio_context specified. May cancel any outstanding
1214 * AIOs and block on completion. Will fail with -ENOSYS if not
1215 * implemented. May fail with -EINVAL if the context pointed to
1218 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1220 struct kioctx *ioctx = lookup_ioctx(ctx);
1221 if (likely(NULL != ioctx)) {
1222 struct completion requests_done =
1223 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1225 /* Pass requests_done to kill_ioctx() where it can be set
1226 * in a thread-safe way. If we try to set it here then we have
1227 * a race condition if two io_destroy() called simultaneously.
1229 kill_ioctx(current->mm, ioctx, &requests_done);
1230 percpu_ref_put(&ioctx->users);
1232 /* Wait until all IO for the context are done. Otherwise kernel
1233 * keep using user-space buffers even if user thinks the context
1236 wait_for_completion(&requests_done);
1240 pr_debug("EINVAL: io_destroy: invalid context id\n");
1244 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1245 unsigned long, loff_t);
1247 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1248 int rw, char __user *buf,
1249 unsigned long *nr_segs,
1250 struct iovec **iovec,
1255 *nr_segs = kiocb->ki_nbytes;
1257 #ifdef CONFIG_COMPAT
1259 ret = compat_rw_copy_check_uvector(rw,
1260 (struct compat_iovec __user *)buf,
1261 *nr_segs, 1, *iovec, iovec);
1264 ret = rw_copy_check_uvector(rw,
1265 (struct iovec __user *)buf,
1266 *nr_segs, 1, *iovec, iovec);
1270 /* ki_nbytes now reflect bytes instead of segs */
1271 kiocb->ki_nbytes = ret;
1275 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1276 int rw, char __user *buf,
1277 unsigned long *nr_segs,
1278 struct iovec *iovec)
1280 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1283 iovec->iov_base = buf;
1284 iovec->iov_len = kiocb->ki_nbytes;
1291 * Performs the initial checks and aio retry method
1292 * setup for the kiocb at the time of io submission.
1294 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1295 char __user *buf, bool compat)
1297 struct file *file = req->ki_filp;
1299 unsigned long nr_segs;
1303 struct iovec inline_vec, *iovec = &inline_vec;
1306 case IOCB_CMD_PREAD:
1307 case IOCB_CMD_PREADV:
1310 rw_op = file->f_op->aio_read;
1313 case IOCB_CMD_PWRITE:
1314 case IOCB_CMD_PWRITEV:
1317 rw_op = file->f_op->aio_write;
1320 if (unlikely(!(file->f_mode & mode)))
1326 ret = (opcode == IOCB_CMD_PREADV ||
1327 opcode == IOCB_CMD_PWRITEV)
1328 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1330 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1333 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1335 if (iovec != &inline_vec)
1340 req->ki_nbytes = ret;
1342 /* XXX: move/kill - rw_verify_area()? */
1343 /* This matches the pread()/pwrite() logic */
1344 if (req->ki_pos < 0) {
1350 file_start_write(file);
1352 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1355 file_end_write(file);
1358 case IOCB_CMD_FDSYNC:
1359 if (!file->f_op->aio_fsync)
1362 ret = file->f_op->aio_fsync(req, 1);
1365 case IOCB_CMD_FSYNC:
1366 if (!file->f_op->aio_fsync)
1369 ret = file->f_op->aio_fsync(req, 0);
1373 pr_debug("EINVAL: no operation provided\n");
1377 if (iovec != &inline_vec)
1380 if (ret != -EIOCBQUEUED) {
1382 * There's no easy way to restart the syscall since other AIO's
1383 * may be already running. Just fail this IO with EINTR.
1385 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1386 ret == -ERESTARTNOHAND ||
1387 ret == -ERESTART_RESTARTBLOCK))
1389 aio_complete(req, ret, 0);
1395 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1396 struct iocb *iocb, bool compat)
1401 /* enforce forwards compatibility on users */
1402 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1403 pr_debug("EINVAL: reserve field set\n");
1407 /* prevent overflows */
1409 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1410 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1411 ((ssize_t)iocb->aio_nbytes < 0)
1413 pr_debug("EINVAL: io_submit: overflow check\n");
1417 req = aio_get_req(ctx);
1421 req->ki_filp = fget(iocb->aio_fildes);
1422 if (unlikely(!req->ki_filp)) {
1427 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1429 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1430 * instance of the file* now. The file descriptor must be
1431 * an eventfd() fd, and will be signaled for each completed
1432 * event using the eventfd_signal() function.
1434 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1435 if (IS_ERR(req->ki_eventfd)) {
1436 ret = PTR_ERR(req->ki_eventfd);
1437 req->ki_eventfd = NULL;
1442 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1443 if (unlikely(ret)) {
1444 pr_debug("EFAULT: aio_key\n");
1448 req->ki_obj.user = user_iocb;
1449 req->ki_user_data = iocb->aio_data;
1450 req->ki_pos = iocb->aio_offset;
1451 req->ki_nbytes = iocb->aio_nbytes;
1453 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1454 (char __user *)(unsigned long)iocb->aio_buf,
1461 put_reqs_available(ctx, 1);
1462 percpu_ref_put(&ctx->reqs);
1467 long do_io_submit(aio_context_t ctx_id, long nr,
1468 struct iocb __user *__user *iocbpp, bool compat)
1473 struct blk_plug plug;
1475 if (unlikely(nr < 0))
1478 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1479 nr = LONG_MAX/sizeof(*iocbpp);
1481 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1484 ctx = lookup_ioctx(ctx_id);
1485 if (unlikely(!ctx)) {
1486 pr_debug("EINVAL: invalid context id\n");
1490 blk_start_plug(&plug);
1493 * AKPM: should this return a partial result if some of the IOs were
1494 * successfully submitted?
1496 for (i=0; i<nr; i++) {
1497 struct iocb __user *user_iocb;
1500 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1505 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1510 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1514 blk_finish_plug(&plug);
1516 percpu_ref_put(&ctx->users);
1521 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1522 * the number of iocbs queued. May return -EINVAL if the aio_context
1523 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1524 * *iocbpp[0] is not properly initialized, if the operation specified
1525 * is invalid for the file descriptor in the iocb. May fail with
1526 * -EFAULT if any of the data structures point to invalid data. May
1527 * fail with -EBADF if the file descriptor specified in the first
1528 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1529 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1530 * fail with -ENOSYS if not implemented.
1532 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1533 struct iocb __user * __user *, iocbpp)
1535 return do_io_submit(ctx_id, nr, iocbpp, 0);
1539 * Finds a given iocb for cancellation.
1541 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1544 struct list_head *pos;
1546 assert_spin_locked(&ctx->ctx_lock);
1548 if (key != KIOCB_KEY)
1551 /* TODO: use a hash or array, this sucks. */
1552 list_for_each(pos, &ctx->active_reqs) {
1553 struct kiocb *kiocb = list_kiocb(pos);
1554 if (kiocb->ki_obj.user == iocb)
1561 * Attempts to cancel an iocb previously passed to io_submit. If
1562 * the operation is successfully cancelled, the resulting event is
1563 * copied into the memory pointed to by result without being placed
1564 * into the completion queue and 0 is returned. May fail with
1565 * -EFAULT if any of the data structures pointed to are invalid.
1566 * May fail with -EINVAL if aio_context specified by ctx_id is
1567 * invalid. May fail with -EAGAIN if the iocb specified was not
1568 * cancelled. Will fail with -ENOSYS if not implemented.
1570 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1571 struct io_event __user *, result)
1574 struct kiocb *kiocb;
1578 ret = get_user(key, &iocb->aio_key);
1582 ctx = lookup_ioctx(ctx_id);
1586 spin_lock_irq(&ctx->ctx_lock);
1588 kiocb = lookup_kiocb(ctx, iocb, key);
1590 ret = kiocb_cancel(ctx, kiocb);
1594 spin_unlock_irq(&ctx->ctx_lock);
1598 * The result argument is no longer used - the io_event is
1599 * always delivered via the ring buffer. -EINPROGRESS indicates
1600 * cancellation is progress:
1605 percpu_ref_put(&ctx->users);
1611 * Attempts to read at least min_nr events and up to nr events from
1612 * the completion queue for the aio_context specified by ctx_id. If
1613 * it succeeds, the number of read events is returned. May fail with
1614 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1615 * out of range, if timeout is out of range. May fail with -EFAULT
1616 * if any of the memory specified is invalid. May return 0 or
1617 * < min_nr if the timeout specified by timeout has elapsed
1618 * before sufficient events are available, where timeout == NULL
1619 * specifies an infinite timeout. Note that the timeout pointed to by
1620 * timeout is relative. Will fail with -ENOSYS if not implemented.
1622 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1625 struct io_event __user *, events,
1626 struct timespec __user *, timeout)
1628 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1631 if (likely(ioctx)) {
1632 if (likely(min_nr <= nr && min_nr >= 0))
1633 ret = read_events(ioctx, min_nr, nr, events, timeout);
1634 percpu_ref_put(&ioctx->users);