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 */
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
64 struct io_event io_events[0];
65 }; /* 128 bytes + ring size */
67 #define AIO_RING_PAGES 8
72 struct kioctx *table[];
76 unsigned reqs_available;
80 struct percpu_ref users;
83 struct percpu_ref reqs;
85 unsigned long user_id;
87 struct __percpu kioctx_cpu *cpu;
90 * For percpu reqs_available, number of slots we move to/from global
95 * This is what userspace passed to io_setup(), it's not used for
96 * anything but counting against the global max_reqs quota.
98 * The real limit is nr_events - 1, which will be larger (see
103 /* Size of ringbuffer, in units of struct io_event */
106 unsigned long mmap_base;
107 unsigned long mmap_size;
109 struct page **ring_pages;
112 struct work_struct free_work;
116 * This counts the number of available slots in the ringbuffer,
117 * so we avoid overflowing it: it's decremented (if positive)
118 * when allocating a kiocb and incremented when the resulting
119 * io_event is pulled off the ringbuffer.
121 * We batch accesses to it with a percpu version.
123 atomic_t reqs_available;
124 } ____cacheline_aligned_in_smp;
128 struct list_head active_reqs; /* used for cancellation */
129 } ____cacheline_aligned_in_smp;
132 struct mutex ring_lock;
133 wait_queue_head_t wait;
134 } ____cacheline_aligned_in_smp;
138 spinlock_t completion_lock;
139 } ____cacheline_aligned_in_smp;
141 struct page *internal_pages[AIO_RING_PAGES];
142 struct file *aio_ring_file;
147 /*------ sysctl variables----*/
148 static DEFINE_SPINLOCK(aio_nr_lock);
149 unsigned long aio_nr; /* current system wide number of aio requests */
150 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
151 /*----end sysctl variables---*/
153 static struct kmem_cache *kiocb_cachep;
154 static struct kmem_cache *kioctx_cachep;
156 static struct vfsmount *aio_mnt;
158 static const struct file_operations aio_ring_fops;
159 static const struct address_space_operations aio_ctx_aops;
161 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
163 struct qstr this = QSTR_INIT("[aio]", 5);
166 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
168 return ERR_CAST(inode);
170 inode->i_mapping->a_ops = &aio_ctx_aops;
171 inode->i_mapping->private_data = ctx;
172 inode->i_size = PAGE_SIZE * nr_pages;
174 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
177 return ERR_PTR(-ENOMEM);
179 path.mnt = mntget(aio_mnt);
181 d_instantiate(path.dentry, inode);
182 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
188 file->f_flags = O_RDWR;
189 file->private_data = ctx;
193 static struct dentry *aio_mount(struct file_system_type *fs_type,
194 int flags, const char *dev_name, void *data)
196 static const struct dentry_operations ops = {
197 .d_dname = simple_dname,
199 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
203 * Creates the slab caches used by the aio routines, panic on
204 * failure as this is done early during the boot sequence.
206 static int __init aio_setup(void)
208 static struct file_system_type aio_fs = {
211 .kill_sb = kill_anon_super,
213 aio_mnt = kern_mount(&aio_fs);
215 panic("Failed to create aio fs mount.");
217 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
218 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
220 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
224 __initcall(aio_setup);
226 static void put_aio_ring_file(struct kioctx *ctx)
228 struct file *aio_ring_file = ctx->aio_ring_file;
230 truncate_setsize(aio_ring_file->f_inode, 0);
232 /* Prevent further access to the kioctx from migratepages */
233 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
234 aio_ring_file->f_inode->i_mapping->private_data = NULL;
235 ctx->aio_ring_file = NULL;
236 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
242 static void aio_free_ring(struct kioctx *ctx)
246 for (i = 0; i < ctx->nr_pages; i++) {
248 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
249 page_count(ctx->ring_pages[i]));
250 page = ctx->ring_pages[i];
253 ctx->ring_pages[i] = NULL;
257 put_aio_ring_file(ctx);
259 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
260 kfree(ctx->ring_pages);
261 ctx->ring_pages = NULL;
265 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
267 vma->vm_ops = &generic_file_vm_ops;
271 static const struct file_operations aio_ring_fops = {
272 .mmap = aio_ring_mmap,
275 static int aio_set_page_dirty(struct page *page)
280 #if IS_ENABLED(CONFIG_MIGRATION)
281 static int aio_migratepage(struct address_space *mapping, struct page *new,
282 struct page *old, enum migrate_mode mode)
290 /* Make sure the old page hasn't already been changed */
291 spin_lock(&mapping->private_lock);
292 ctx = mapping->private_data;
295 spin_lock_irqsave(&ctx->completion_lock, flags);
297 if (idx < (pgoff_t)ctx->nr_pages) {
298 if (ctx->ring_pages[idx] != old)
302 spin_unlock_irqrestore(&ctx->completion_lock, flags);
305 spin_unlock(&mapping->private_lock);
310 /* Writeback must be complete */
311 BUG_ON(PageWriteback(old));
314 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
315 if (rc != MIGRATEPAGE_SUCCESS) {
320 /* We can potentially race against kioctx teardown here. Use the
321 * address_space's private data lock to protect the mapping's
324 spin_lock(&mapping->private_lock);
325 ctx = mapping->private_data;
328 spin_lock_irqsave(&ctx->completion_lock, flags);
329 migrate_page_copy(new, old);
331 if (idx < (pgoff_t)ctx->nr_pages) {
332 /* And only do the move if things haven't changed */
333 if (ctx->ring_pages[idx] == old)
334 ctx->ring_pages[idx] = new;
339 spin_unlock_irqrestore(&ctx->completion_lock, flags);
342 spin_unlock(&mapping->private_lock);
344 if (rc == MIGRATEPAGE_SUCCESS)
353 static const struct address_space_operations aio_ctx_aops = {
354 .set_page_dirty = aio_set_page_dirty,
355 #if IS_ENABLED(CONFIG_MIGRATION)
356 .migratepage = aio_migratepage,
360 static int aio_setup_ring(struct kioctx *ctx)
362 struct aio_ring *ring;
363 unsigned nr_events = ctx->max_reqs;
364 struct mm_struct *mm = current->mm;
365 unsigned long size, unused;
370 /* Compensate for the ring buffer's head/tail overlap entry */
371 nr_events += 2; /* 1 is required, 2 for good luck */
373 size = sizeof(struct aio_ring);
374 size += sizeof(struct io_event) * nr_events;
376 nr_pages = PFN_UP(size);
380 file = aio_private_file(ctx, nr_pages);
382 ctx->aio_ring_file = NULL;
386 ctx->aio_ring_file = file;
387 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
388 / sizeof(struct io_event);
390 ctx->ring_pages = ctx->internal_pages;
391 if (nr_pages > AIO_RING_PAGES) {
392 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
394 if (!ctx->ring_pages) {
395 put_aio_ring_file(ctx);
400 for (i = 0; i < nr_pages; i++) {
402 page = find_or_create_page(file->f_inode->i_mapping,
403 i, GFP_HIGHUSER | __GFP_ZERO);
406 pr_debug("pid(%d) page[%d]->count=%d\n",
407 current->pid, i, page_count(page));
408 SetPageUptodate(page);
412 ctx->ring_pages[i] = page;
416 if (unlikely(i != nr_pages)) {
421 ctx->mmap_size = nr_pages * PAGE_SIZE;
422 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
424 down_write(&mm->mmap_sem);
425 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
426 PROT_READ | PROT_WRITE,
427 MAP_SHARED, 0, &unused);
428 up_write(&mm->mmap_sem);
429 if (IS_ERR((void *)ctx->mmap_base)) {
435 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
437 ctx->user_id = ctx->mmap_base;
438 ctx->nr_events = nr_events; /* trusted copy */
440 ring = kmap_atomic(ctx->ring_pages[0]);
441 ring->nr = nr_events; /* user copy */
443 ring->head = ring->tail = 0;
444 ring->magic = AIO_RING_MAGIC;
445 ring->compat_features = AIO_RING_COMPAT_FEATURES;
446 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
447 ring->header_length = sizeof(struct aio_ring);
449 flush_dcache_page(ctx->ring_pages[0]);
454 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
455 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
456 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
458 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
460 struct kioctx *ctx = req->ki_ctx;
463 spin_lock_irqsave(&ctx->ctx_lock, flags);
465 if (!req->ki_list.next)
466 list_add(&req->ki_list, &ctx->active_reqs);
468 req->ki_cancel = cancel;
470 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
472 EXPORT_SYMBOL(kiocb_set_cancel_fn);
474 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
476 kiocb_cancel_fn *old, *cancel;
479 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
480 * actually has a cancel function, hence the cmpxchg()
483 cancel = ACCESS_ONCE(kiocb->ki_cancel);
485 if (!cancel || cancel == KIOCB_CANCELLED)
489 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
490 } while (cancel != old);
492 return cancel(kiocb);
495 static void free_ioctx(struct work_struct *work)
497 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
499 pr_debug("freeing %p\n", ctx);
502 free_percpu(ctx->cpu);
503 kmem_cache_free(kioctx_cachep, ctx);
506 static void free_ioctx_reqs(struct percpu_ref *ref)
508 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
510 INIT_WORK(&ctx->free_work, free_ioctx);
511 schedule_work(&ctx->free_work);
515 * When this function runs, the kioctx has been removed from the "hash table"
516 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
517 * now it's safe to cancel any that need to be.
519 static void free_ioctx_users(struct percpu_ref *ref)
521 struct kioctx *ctx = container_of(ref, struct kioctx, users);
524 spin_lock_irq(&ctx->ctx_lock);
526 while (!list_empty(&ctx->active_reqs)) {
527 req = list_first_entry(&ctx->active_reqs,
528 struct kiocb, ki_list);
530 list_del_init(&req->ki_list);
531 kiocb_cancel(ctx, req);
534 spin_unlock_irq(&ctx->ctx_lock);
536 percpu_ref_kill(&ctx->reqs);
537 percpu_ref_put(&ctx->reqs);
540 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
543 struct kioctx_table *table, *old;
544 struct aio_ring *ring;
546 spin_lock(&mm->ioctx_lock);
548 table = rcu_dereference(mm->ioctx_table);
552 for (i = 0; i < table->nr; i++)
553 if (!table->table[i]) {
555 table->table[i] = ctx;
557 spin_unlock(&mm->ioctx_lock);
559 ring = kmap_atomic(ctx->ring_pages[0]);
565 new_nr = (table ? table->nr : 1) * 4;
568 spin_unlock(&mm->ioctx_lock);
570 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
577 spin_lock(&mm->ioctx_lock);
579 old = rcu_dereference(mm->ioctx_table);
582 rcu_assign_pointer(mm->ioctx_table, table);
583 } else if (table->nr > old->nr) {
584 memcpy(table->table, old->table,
585 old->nr * sizeof(struct kioctx *));
587 rcu_assign_pointer(mm->ioctx_table, table);
596 static void aio_nr_sub(unsigned nr)
598 spin_lock(&aio_nr_lock);
599 if (WARN_ON(aio_nr - nr > aio_nr))
603 spin_unlock(&aio_nr_lock);
607 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
609 static struct kioctx *ioctx_alloc(unsigned nr_events)
611 struct mm_struct *mm = current->mm;
616 * We keep track of the number of available ringbuffer slots, to prevent
617 * overflow (reqs_available), and we also use percpu counters for this.
619 * So since up to half the slots might be on other cpu's percpu counters
620 * and unavailable, double nr_events so userspace sees what they
621 * expected: additionally, we move req_batch slots to/from percpu
622 * counters at a time, so make sure that isn't 0:
624 nr_events = max(nr_events, num_possible_cpus() * 4);
627 /* Prevent overflows */
628 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
629 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
630 pr_debug("ENOMEM: nr_events too high\n");
631 return ERR_PTR(-EINVAL);
634 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
635 return ERR_PTR(-EAGAIN);
637 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
639 return ERR_PTR(-ENOMEM);
641 ctx->max_reqs = nr_events;
643 if (percpu_ref_init(&ctx->users, free_ioctx_users))
646 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
649 spin_lock_init(&ctx->ctx_lock);
650 spin_lock_init(&ctx->completion_lock);
651 mutex_init(&ctx->ring_lock);
652 init_waitqueue_head(&ctx->wait);
654 INIT_LIST_HEAD(&ctx->active_reqs);
656 ctx->cpu = alloc_percpu(struct kioctx_cpu);
660 if (aio_setup_ring(ctx) < 0)
663 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
664 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
665 if (ctx->req_batch < 1)
668 /* limit the number of system wide aios */
669 spin_lock(&aio_nr_lock);
670 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
671 aio_nr + nr_events < aio_nr) {
672 spin_unlock(&aio_nr_lock);
676 aio_nr += ctx->max_reqs;
677 spin_unlock(&aio_nr_lock);
679 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
680 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
682 err = ioctx_add_table(ctx, mm);
686 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
687 ctx, ctx->user_id, mm, ctx->nr_events);
691 aio_nr_sub(ctx->max_reqs);
695 free_percpu(ctx->cpu);
696 free_percpu(ctx->reqs.pcpu_count);
697 free_percpu(ctx->users.pcpu_count);
698 kmem_cache_free(kioctx_cachep, ctx);
699 pr_debug("error allocating ioctx %d\n", err);
704 * Cancels all outstanding aio requests on an aio context. Used
705 * when the processes owning a context have all exited to encourage
706 * the rapid destruction of the kioctx.
708 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
710 if (!atomic_xchg(&ctx->dead, 1)) {
711 struct kioctx_table *table;
713 spin_lock(&mm->ioctx_lock);
715 table = rcu_dereference(mm->ioctx_table);
717 WARN_ON(ctx != table->table[ctx->id]);
718 table->table[ctx->id] = NULL;
720 spin_unlock(&mm->ioctx_lock);
722 /* percpu_ref_kill() will do the necessary call_rcu() */
723 wake_up_all(&ctx->wait);
726 * It'd be more correct to do this in free_ioctx(), after all
727 * the outstanding kiocbs have finished - but by then io_destroy
728 * has already returned, so io_setup() could potentially return
729 * -EAGAIN with no ioctxs actually in use (as far as userspace
732 aio_nr_sub(ctx->max_reqs);
735 vm_munmap(ctx->mmap_base, ctx->mmap_size);
737 percpu_ref_kill(&ctx->users);
741 /* wait_on_sync_kiocb:
742 * Waits on the given sync kiocb to complete.
744 ssize_t wait_on_sync_kiocb(struct kiocb *req)
746 while (!req->ki_ctx) {
747 set_current_state(TASK_UNINTERRUPTIBLE);
752 __set_current_state(TASK_RUNNING);
753 return req->ki_user_data;
755 EXPORT_SYMBOL(wait_on_sync_kiocb);
758 * exit_aio: called when the last user of mm goes away. At this point, there is
759 * no way for any new requests to be submited or any of the io_* syscalls to be
760 * called on the context.
762 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
765 void exit_aio(struct mm_struct *mm)
767 struct kioctx_table *table;
773 table = rcu_dereference(mm->ioctx_table);
776 if (!table || i >= table->nr) {
778 rcu_assign_pointer(mm->ioctx_table, NULL);
784 ctx = table->table[i++];
790 * We don't need to bother with munmap() here -
791 * exit_mmap(mm) is coming and it'll unmap everything.
792 * Since aio_free_ring() uses non-zero ->mmap_size
793 * as indicator that it needs to unmap the area,
794 * just set it to 0; aio_free_ring() is the only
795 * place that uses ->mmap_size, so it's safe.
803 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
805 struct kioctx_cpu *kcpu;
808 kcpu = this_cpu_ptr(ctx->cpu);
810 kcpu->reqs_available += nr;
811 while (kcpu->reqs_available >= ctx->req_batch * 2) {
812 kcpu->reqs_available -= ctx->req_batch;
813 atomic_add(ctx->req_batch, &ctx->reqs_available);
819 static bool get_reqs_available(struct kioctx *ctx)
821 struct kioctx_cpu *kcpu;
825 kcpu = this_cpu_ptr(ctx->cpu);
827 if (!kcpu->reqs_available) {
828 int old, avail = atomic_read(&ctx->reqs_available);
831 if (avail < ctx->req_batch)
835 avail = atomic_cmpxchg(&ctx->reqs_available,
836 avail, avail - ctx->req_batch);
837 } while (avail != old);
839 kcpu->reqs_available += ctx->req_batch;
843 kcpu->reqs_available--;
850 * Allocate a slot for an aio request.
851 * Returns NULL if no requests are free.
853 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
857 if (!get_reqs_available(ctx))
860 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
864 percpu_ref_get(&ctx->reqs);
869 put_reqs_available(ctx, 1);
873 static void kiocb_free(struct kiocb *req)
877 if (req->ki_eventfd != NULL)
878 eventfd_ctx_put(req->ki_eventfd);
879 kmem_cache_free(kiocb_cachep, req);
882 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
884 struct aio_ring __user *ring = (void __user *)ctx_id;
885 struct mm_struct *mm = current->mm;
886 struct kioctx *ctx, *ret = NULL;
887 struct kioctx_table *table;
890 if (get_user(id, &ring->id))
894 table = rcu_dereference(mm->ioctx_table);
896 if (!table || id >= table->nr)
899 ctx = table->table[id];
900 if (ctx && ctx->user_id == ctx_id) {
901 percpu_ref_get(&ctx->users);
910 * Called when the io request on the given iocb is complete.
912 void aio_complete(struct kiocb *iocb, long res, long res2)
914 struct kioctx *ctx = iocb->ki_ctx;
915 struct aio_ring *ring;
916 struct io_event *ev_page, *event;
921 * Special case handling for sync iocbs:
922 * - events go directly into the iocb for fast handling
923 * - the sync task with the iocb in its stack holds the single iocb
924 * ref, no other paths have a way to get another ref
925 * - the sync task helpfully left a reference to itself in the iocb
927 if (is_sync_kiocb(iocb)) {
928 iocb->ki_user_data = res;
930 iocb->ki_ctx = ERR_PTR(-EXDEV);
931 wake_up_process(iocb->ki_obj.tsk);
935 if (iocb->ki_list.next) {
938 spin_lock_irqsave(&ctx->ctx_lock, flags);
939 list_del(&iocb->ki_list);
940 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
944 * Add a completion event to the ring buffer. Must be done holding
945 * ctx->completion_lock to prevent other code from messing with the tail
946 * pointer since we might be called from irq context.
948 spin_lock_irqsave(&ctx->completion_lock, flags);
951 pos = tail + AIO_EVENTS_OFFSET;
953 if (++tail >= ctx->nr_events)
956 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
957 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
959 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
960 event->data = iocb->ki_user_data;
964 kunmap_atomic(ev_page);
965 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
967 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
968 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
971 /* after flagging the request as done, we
972 * must never even look at it again
974 smp_wmb(); /* make event visible before updating tail */
978 ring = kmap_atomic(ctx->ring_pages[0]);
981 flush_dcache_page(ctx->ring_pages[0]);
983 spin_unlock_irqrestore(&ctx->completion_lock, flags);
985 pr_debug("added to ring %p at [%u]\n", iocb, tail);
988 * Check if the user asked us to deliver the result through an
989 * eventfd. The eventfd_signal() function is safe to be called
992 if (iocb->ki_eventfd != NULL)
993 eventfd_signal(iocb->ki_eventfd, 1);
995 /* everything turned out well, dispose of the aiocb. */
999 * We have to order our ring_info tail store above and test
1000 * of the wait list below outside the wait lock. This is
1001 * like in wake_up_bit() where clearing a bit has to be
1002 * ordered with the unlocked test.
1006 if (waitqueue_active(&ctx->wait))
1007 wake_up(&ctx->wait);
1009 percpu_ref_put(&ctx->reqs);
1011 EXPORT_SYMBOL(aio_complete);
1014 * Pull an event off of the ioctx's event ring. Returns the number of
1017 static long aio_read_events_ring(struct kioctx *ctx,
1018 struct io_event __user *event, long nr)
1020 struct aio_ring *ring;
1021 unsigned head, tail, pos;
1025 mutex_lock(&ctx->ring_lock);
1027 ring = kmap_atomic(ctx->ring_pages[0]);
1030 kunmap_atomic(ring);
1032 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1039 struct io_event *ev;
1042 avail = (head <= tail ? tail : ctx->nr_events) - head;
1046 avail = min(avail, nr - ret);
1047 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1048 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1050 pos = head + AIO_EVENTS_OFFSET;
1051 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1052 pos %= AIO_EVENTS_PER_PAGE;
1055 copy_ret = copy_to_user(event + ret, ev + pos,
1056 sizeof(*ev) * avail);
1059 if (unlikely(copy_ret)) {
1066 head %= ctx->nr_events;
1069 ring = kmap_atomic(ctx->ring_pages[0]);
1071 kunmap_atomic(ring);
1072 flush_dcache_page(ctx->ring_pages[0]);
1074 pr_debug("%li h%u t%u\n", ret, head, tail);
1076 put_reqs_available(ctx, ret);
1078 mutex_unlock(&ctx->ring_lock);
1083 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1084 struct io_event __user *event, long *i)
1086 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1091 if (unlikely(atomic_read(&ctx->dead)))
1097 return ret < 0 || *i >= min_nr;
1100 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1101 struct io_event __user *event,
1102 struct timespec __user *timeout)
1104 ktime_t until = { .tv64 = KTIME_MAX };
1110 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1113 until = timespec_to_ktime(ts);
1117 * Note that aio_read_events() is being called as the conditional - i.e.
1118 * we're calling it after prepare_to_wait() has set task state to
1119 * TASK_INTERRUPTIBLE.
1121 * But aio_read_events() can block, and if it blocks it's going to flip
1122 * the task state back to TASK_RUNNING.
1124 * This should be ok, provided it doesn't flip the state back to
1125 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1126 * will only happen if the mutex_lock() call blocks, and we then find
1127 * the ringbuffer empty. So in practice we should be ok, but it's
1128 * something to be aware of when touching this code.
1130 wait_event_interruptible_hrtimeout(ctx->wait,
1131 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1133 if (!ret && signal_pending(current))
1140 * Create an aio_context capable of receiving at least nr_events.
1141 * ctxp must not point to an aio_context that already exists, and
1142 * must be initialized to 0 prior to the call. On successful
1143 * creation of the aio_context, *ctxp is filled in with the resulting
1144 * handle. May fail with -EINVAL if *ctxp is not initialized,
1145 * if the specified nr_events exceeds internal limits. May fail
1146 * with -EAGAIN if the specified nr_events exceeds the user's limit
1147 * of available events. May fail with -ENOMEM if insufficient kernel
1148 * resources are available. May fail with -EFAULT if an invalid
1149 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1152 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1154 struct kioctx *ioctx = NULL;
1158 ret = get_user(ctx, ctxp);
1163 if (unlikely(ctx || nr_events == 0)) {
1164 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1169 ioctx = ioctx_alloc(nr_events);
1170 ret = PTR_ERR(ioctx);
1171 if (!IS_ERR(ioctx)) {
1172 ret = put_user(ioctx->user_id, ctxp);
1174 kill_ioctx(current->mm, ioctx);
1175 percpu_ref_put(&ioctx->users);
1183 * Destroy the aio_context specified. May cancel any outstanding
1184 * AIOs and block on completion. Will fail with -ENOSYS if not
1185 * implemented. May fail with -EINVAL if the context pointed to
1188 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1190 struct kioctx *ioctx = lookup_ioctx(ctx);
1191 if (likely(NULL != ioctx)) {
1192 kill_ioctx(current->mm, ioctx);
1193 percpu_ref_put(&ioctx->users);
1196 pr_debug("EINVAL: io_destroy: invalid context id\n");
1200 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1201 unsigned long, loff_t);
1203 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1204 int rw, char __user *buf,
1205 unsigned long *nr_segs,
1206 struct iovec **iovec,
1211 *nr_segs = kiocb->ki_nbytes;
1213 #ifdef CONFIG_COMPAT
1215 ret = compat_rw_copy_check_uvector(rw,
1216 (struct compat_iovec __user *)buf,
1217 *nr_segs, 1, *iovec, iovec);
1220 ret = rw_copy_check_uvector(rw,
1221 (struct iovec __user *)buf,
1222 *nr_segs, 1, *iovec, iovec);
1226 /* ki_nbytes now reflect bytes instead of segs */
1227 kiocb->ki_nbytes = ret;
1231 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1232 int rw, char __user *buf,
1233 unsigned long *nr_segs,
1234 struct iovec *iovec)
1236 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1239 iovec->iov_base = buf;
1240 iovec->iov_len = kiocb->ki_nbytes;
1247 * Performs the initial checks and aio retry method
1248 * setup for the kiocb at the time of io submission.
1250 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1251 char __user *buf, bool compat)
1253 struct file *file = req->ki_filp;
1255 unsigned long nr_segs;
1259 struct iovec inline_vec, *iovec = &inline_vec;
1262 case IOCB_CMD_PREAD:
1263 case IOCB_CMD_PREADV:
1266 rw_op = file->f_op->aio_read;
1269 case IOCB_CMD_PWRITE:
1270 case IOCB_CMD_PWRITEV:
1273 rw_op = file->f_op->aio_write;
1276 if (unlikely(!(file->f_mode & mode)))
1282 ret = (opcode == IOCB_CMD_PREADV ||
1283 opcode == IOCB_CMD_PWRITEV)
1284 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1286 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1291 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1293 if (iovec != &inline_vec)
1298 req->ki_nbytes = ret;
1300 /* XXX: move/kill - rw_verify_area()? */
1301 /* This matches the pread()/pwrite() logic */
1302 if (req->ki_pos < 0) {
1308 file_start_write(file);
1310 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1313 file_end_write(file);
1316 case IOCB_CMD_FDSYNC:
1317 if (!file->f_op->aio_fsync)
1320 ret = file->f_op->aio_fsync(req, 1);
1323 case IOCB_CMD_FSYNC:
1324 if (!file->f_op->aio_fsync)
1327 ret = file->f_op->aio_fsync(req, 0);
1331 pr_debug("EINVAL: no operation provided\n");
1335 if (iovec != &inline_vec)
1338 if (ret != -EIOCBQUEUED) {
1340 * There's no easy way to restart the syscall since other AIO's
1341 * may be already running. Just fail this IO with EINTR.
1343 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1344 ret == -ERESTARTNOHAND ||
1345 ret == -ERESTART_RESTARTBLOCK))
1347 aio_complete(req, ret, 0);
1353 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1354 struct iocb *iocb, bool compat)
1359 /* enforce forwards compatibility on users */
1360 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1361 pr_debug("EINVAL: reserve field set\n");
1365 /* prevent overflows */
1367 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1368 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1369 ((ssize_t)iocb->aio_nbytes < 0)
1371 pr_debug("EINVAL: io_submit: overflow check\n");
1375 req = aio_get_req(ctx);
1379 req->ki_filp = fget(iocb->aio_fildes);
1380 if (unlikely(!req->ki_filp)) {
1385 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1387 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1388 * instance of the file* now. The file descriptor must be
1389 * an eventfd() fd, and will be signaled for each completed
1390 * event using the eventfd_signal() function.
1392 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1393 if (IS_ERR(req->ki_eventfd)) {
1394 ret = PTR_ERR(req->ki_eventfd);
1395 req->ki_eventfd = NULL;
1400 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1401 if (unlikely(ret)) {
1402 pr_debug("EFAULT: aio_key\n");
1406 req->ki_obj.user = user_iocb;
1407 req->ki_user_data = iocb->aio_data;
1408 req->ki_pos = iocb->aio_offset;
1409 req->ki_nbytes = iocb->aio_nbytes;
1411 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1412 (char __user *)(unsigned long)iocb->aio_buf,
1419 put_reqs_available(ctx, 1);
1420 percpu_ref_put(&ctx->reqs);
1425 long do_io_submit(aio_context_t ctx_id, long nr,
1426 struct iocb __user *__user *iocbpp, bool compat)
1431 struct blk_plug plug;
1433 if (unlikely(nr < 0))
1436 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1437 nr = LONG_MAX/sizeof(*iocbpp);
1439 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1442 ctx = lookup_ioctx(ctx_id);
1443 if (unlikely(!ctx)) {
1444 pr_debug("EINVAL: invalid context id\n");
1448 blk_start_plug(&plug);
1451 * AKPM: should this return a partial result if some of the IOs were
1452 * successfully submitted?
1454 for (i=0; i<nr; i++) {
1455 struct iocb __user *user_iocb;
1458 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1463 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1468 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1472 blk_finish_plug(&plug);
1474 percpu_ref_put(&ctx->users);
1479 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1480 * the number of iocbs queued. May return -EINVAL if the aio_context
1481 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1482 * *iocbpp[0] is not properly initialized, if the operation specified
1483 * is invalid for the file descriptor in the iocb. May fail with
1484 * -EFAULT if any of the data structures point to invalid data. May
1485 * fail with -EBADF if the file descriptor specified in the first
1486 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1487 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1488 * fail with -ENOSYS if not implemented.
1490 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1491 struct iocb __user * __user *, iocbpp)
1493 return do_io_submit(ctx_id, nr, iocbpp, 0);
1497 * Finds a given iocb for cancellation.
1499 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1502 struct list_head *pos;
1504 assert_spin_locked(&ctx->ctx_lock);
1506 if (key != KIOCB_KEY)
1509 /* TODO: use a hash or array, this sucks. */
1510 list_for_each(pos, &ctx->active_reqs) {
1511 struct kiocb *kiocb = list_kiocb(pos);
1512 if (kiocb->ki_obj.user == iocb)
1519 * Attempts to cancel an iocb previously passed to io_submit. If
1520 * the operation is successfully cancelled, the resulting event is
1521 * copied into the memory pointed to by result without being placed
1522 * into the completion queue and 0 is returned. May fail with
1523 * -EFAULT if any of the data structures pointed to are invalid.
1524 * May fail with -EINVAL if aio_context specified by ctx_id is
1525 * invalid. May fail with -EAGAIN if the iocb specified was not
1526 * cancelled. Will fail with -ENOSYS if not implemented.
1528 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1529 struct io_event __user *, result)
1532 struct kiocb *kiocb;
1536 ret = get_user(key, &iocb->aio_key);
1540 ctx = lookup_ioctx(ctx_id);
1544 spin_lock_irq(&ctx->ctx_lock);
1546 kiocb = lookup_kiocb(ctx, iocb, key);
1548 ret = kiocb_cancel(ctx, kiocb);
1552 spin_unlock_irq(&ctx->ctx_lock);
1556 * The result argument is no longer used - the io_event is
1557 * always delivered via the ring buffer. -EINPROGRESS indicates
1558 * cancellation is progress:
1563 percpu_ref_put(&ctx->users);
1569 * Attempts to read at least min_nr events and up to nr events from
1570 * the completion queue for the aio_context specified by ctx_id. If
1571 * it succeeds, the number of read events is returned. May fail with
1572 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1573 * out of range, if timeout is out of range. May fail with -EFAULT
1574 * if any of the memory specified is invalid. May return 0 or
1575 * < min_nr if the timeout specified by timeout has elapsed
1576 * before sufficient events are available, where timeout == NULL
1577 * specifies an infinite timeout. Note that the timeout pointed to by
1578 * timeout is relative. Will fail with -ENOSYS if not implemented.
1580 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1583 struct io_event __user *, events,
1584 struct timespec __user *, timeout)
1586 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1589 if (likely(ioctx)) {
1590 if (likely(min_nr <= nr && min_nr >= 0))
1591 ret = read_events(ioctx, min_nr, nr, events, timeout);
1592 percpu_ref_put(&ioctx->users);