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 unsigned long user_id;
85 struct __percpu kioctx_cpu *cpu;
88 * For percpu reqs_available, number of slots we move to/from global
93 * This is what userspace passed to io_setup(), it's not used for
94 * anything but counting against the global max_reqs quota.
96 * The real limit is nr_events - 1, which will be larger (see
101 /* Size of ringbuffer, in units of struct io_event */
104 unsigned long mmap_base;
105 unsigned long mmap_size;
107 struct page **ring_pages;
110 struct rcu_head rcu_head;
111 struct work_struct free_work;
115 * This counts the number of available slots in the ringbuffer,
116 * so we avoid overflowing it: it's decremented (if positive)
117 * when allocating a kiocb and incremented when the resulting
118 * io_event is pulled off the ringbuffer.
120 * We batch accesses to it with a percpu version.
122 atomic_t reqs_available;
123 } ____cacheline_aligned_in_smp;
127 struct list_head active_reqs; /* used for cancellation */
128 } ____cacheline_aligned_in_smp;
131 struct mutex ring_lock;
132 wait_queue_head_t wait;
133 } ____cacheline_aligned_in_smp;
137 spinlock_t completion_lock;
138 } ____cacheline_aligned_in_smp;
140 struct page *internal_pages[AIO_RING_PAGES];
141 struct file *aio_ring_file;
146 /*------ sysctl variables----*/
147 static DEFINE_SPINLOCK(aio_nr_lock);
148 unsigned long aio_nr; /* current system wide number of aio requests */
149 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
150 /*----end sysctl variables---*/
152 static struct kmem_cache *kiocb_cachep;
153 static struct kmem_cache *kioctx_cachep;
155 static struct vfsmount *aio_mnt;
157 static const struct file_operations aio_ring_fops;
158 static const struct address_space_operations aio_ctx_aops;
160 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
162 struct qstr this = QSTR_INIT("[aio]", 5);
165 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
167 return ERR_CAST(inode);
169 inode->i_mapping->a_ops = &aio_ctx_aops;
170 inode->i_mapping->private_data = ctx;
171 inode->i_size = PAGE_SIZE * nr_pages;
173 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
176 return ERR_PTR(-ENOMEM);
178 path.mnt = mntget(aio_mnt);
180 d_instantiate(path.dentry, inode);
181 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
187 file->f_flags = O_RDWR;
188 file->private_data = ctx;
192 static struct dentry *aio_mount(struct file_system_type *fs_type,
193 int flags, const char *dev_name, void *data)
195 static const struct dentry_operations ops = {
196 .d_dname = simple_dname,
198 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
202 * Creates the slab caches used by the aio routines, panic on
203 * failure as this is done early during the boot sequence.
205 static int __init aio_setup(void)
207 static struct file_system_type aio_fs = {
210 .kill_sb = kill_anon_super,
212 aio_mnt = kern_mount(&aio_fs);
214 panic("Failed to create aio fs mount.");
216 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
217 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
219 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
223 __initcall(aio_setup);
225 static void put_aio_ring_file(struct kioctx *ctx)
227 struct file *aio_ring_file = ctx->aio_ring_file;
229 truncate_setsize(aio_ring_file->f_inode, 0);
231 /* Prevent further access to the kioctx from migratepages */
232 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
233 aio_ring_file->f_inode->i_mapping->private_data = NULL;
234 ctx->aio_ring_file = NULL;
235 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
241 static void aio_free_ring(struct kioctx *ctx)
245 for (i = 0; i < ctx->nr_pages; i++) {
246 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
247 page_count(ctx->ring_pages[i]));
248 put_page(ctx->ring_pages[i]);
251 put_aio_ring_file(ctx);
253 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages)
254 kfree(ctx->ring_pages);
257 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
259 vma->vm_ops = &generic_file_vm_ops;
263 static const struct file_operations aio_ring_fops = {
264 .mmap = aio_ring_mmap,
267 static int aio_set_page_dirty(struct page *page)
272 #if IS_ENABLED(CONFIG_MIGRATION)
273 static int aio_migratepage(struct address_space *mapping, struct page *new,
274 struct page *old, enum migrate_mode mode)
280 /* Writeback must be complete */
281 BUG_ON(PageWriteback(old));
284 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
285 if (rc != MIGRATEPAGE_SUCCESS) {
292 /* We can potentially race against kioctx teardown here. Use the
293 * address_space's private data lock to protect the mapping's
296 spin_lock(&mapping->private_lock);
297 ctx = mapping->private_data;
300 spin_lock_irqsave(&ctx->completion_lock, flags);
301 migrate_page_copy(new, old);
303 if (idx < (pgoff_t)ctx->nr_pages)
304 ctx->ring_pages[idx] = new;
305 spin_unlock_irqrestore(&ctx->completion_lock, flags);
308 spin_unlock(&mapping->private_lock);
314 static const struct address_space_operations aio_ctx_aops = {
315 .set_page_dirty = aio_set_page_dirty,
316 #if IS_ENABLED(CONFIG_MIGRATION)
317 .migratepage = aio_migratepage,
321 static int aio_setup_ring(struct kioctx *ctx)
323 struct aio_ring *ring;
324 unsigned nr_events = ctx->max_reqs;
325 struct mm_struct *mm = current->mm;
326 unsigned long size, populate;
331 /* Compensate for the ring buffer's head/tail overlap entry */
332 nr_events += 2; /* 1 is required, 2 for good luck */
334 size = sizeof(struct aio_ring);
335 size += sizeof(struct io_event) * nr_events;
337 nr_pages = PFN_UP(size);
341 file = aio_private_file(ctx, nr_pages);
343 ctx->aio_ring_file = NULL;
347 for (i = 0; i < nr_pages; i++) {
349 page = find_or_create_page(file->f_inode->i_mapping,
350 i, GFP_HIGHUSER | __GFP_ZERO);
353 pr_debug("pid(%d) page[%d]->count=%d\n",
354 current->pid, i, page_count(page));
355 SetPageUptodate(page);
359 ctx->aio_ring_file = file;
360 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
361 / sizeof(struct io_event);
363 ctx->ring_pages = ctx->internal_pages;
364 if (nr_pages > AIO_RING_PAGES) {
365 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
367 if (!ctx->ring_pages)
371 ctx->mmap_size = nr_pages * PAGE_SIZE;
372 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
374 down_write(&mm->mmap_sem);
375 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
376 PROT_READ | PROT_WRITE,
377 MAP_SHARED | MAP_POPULATE, 0, &populate);
378 if (IS_ERR((void *)ctx->mmap_base)) {
379 up_write(&mm->mmap_sem);
385 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
387 /* We must do this while still holding mmap_sem for write, as we
388 * need to be protected against userspace attempting to mremap()
389 * or munmap() the ring buffer.
391 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
392 1, 0, ctx->ring_pages, NULL);
394 /* Dropping the reference here is safe as the page cache will hold
395 * onto the pages for us. It is also required so that page migration
396 * can unmap the pages and get the right reference count.
398 for (i = 0; i < ctx->nr_pages; i++)
399 put_page(ctx->ring_pages[i]);
401 up_write(&mm->mmap_sem);
403 if (unlikely(ctx->nr_pages != nr_pages)) {
408 ctx->user_id = ctx->mmap_base;
409 ctx->nr_events = nr_events; /* trusted copy */
411 ring = kmap_atomic(ctx->ring_pages[0]);
412 ring->nr = nr_events; /* user copy */
414 ring->head = ring->tail = 0;
415 ring->magic = AIO_RING_MAGIC;
416 ring->compat_features = AIO_RING_COMPAT_FEATURES;
417 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
418 ring->header_length = sizeof(struct aio_ring);
420 flush_dcache_page(ctx->ring_pages[0]);
425 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
426 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
427 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
429 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
431 struct kioctx *ctx = req->ki_ctx;
434 spin_lock_irqsave(&ctx->ctx_lock, flags);
436 if (!req->ki_list.next)
437 list_add(&req->ki_list, &ctx->active_reqs);
439 req->ki_cancel = cancel;
441 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
443 EXPORT_SYMBOL(kiocb_set_cancel_fn);
445 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
447 kiocb_cancel_fn *old, *cancel;
450 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
451 * actually has a cancel function, hence the cmpxchg()
454 cancel = ACCESS_ONCE(kiocb->ki_cancel);
456 if (!cancel || cancel == KIOCB_CANCELLED)
460 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
461 } while (cancel != old);
463 return cancel(kiocb);
466 static void free_ioctx_rcu(struct rcu_head *head)
468 struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
470 free_percpu(ctx->cpu);
471 kmem_cache_free(kioctx_cachep, ctx);
475 * When this function runs, the kioctx has been removed from the "hash table"
476 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
477 * now it's safe to cancel any that need to be.
479 static void free_ioctx(struct work_struct *work)
481 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
482 struct aio_ring *ring;
487 spin_lock_irq(&ctx->ctx_lock);
489 while (!list_empty(&ctx->active_reqs)) {
490 req = list_first_entry(&ctx->active_reqs,
491 struct kiocb, ki_list);
493 list_del_init(&req->ki_list);
494 kiocb_cancel(ctx, req);
497 spin_unlock_irq(&ctx->ctx_lock);
499 for_each_possible_cpu(cpu) {
500 struct kioctx_cpu *kcpu = per_cpu_ptr(ctx->cpu, cpu);
502 atomic_add(kcpu->reqs_available, &ctx->reqs_available);
503 kcpu->reqs_available = 0;
507 prepare_to_wait(&ctx->wait, &wait, TASK_UNINTERRUPTIBLE);
509 ring = kmap_atomic(ctx->ring_pages[0]);
510 avail = (ring->head <= ring->tail)
511 ? ring->tail - ring->head
512 : ctx->nr_events - ring->head + ring->tail;
514 atomic_add(avail, &ctx->reqs_available);
515 ring->head = ring->tail;
518 if (atomic_read(&ctx->reqs_available) >= ctx->nr_events - 1)
523 finish_wait(&ctx->wait, &wait);
525 WARN_ON(atomic_read(&ctx->reqs_available) > ctx->nr_events - 1);
529 pr_debug("freeing %p\n", ctx);
532 * Here the call_rcu() is between the wait_event() for reqs_active to
533 * hit 0, and freeing the ioctx.
535 * aio_complete() decrements reqs_active, but it has to touch the ioctx
536 * after to issue a wakeup so we use rcu.
538 call_rcu(&ctx->rcu_head, free_ioctx_rcu);
541 static void free_ioctx_ref(struct percpu_ref *ref)
543 struct kioctx *ctx = container_of(ref, struct kioctx, users);
545 INIT_WORK(&ctx->free_work, free_ioctx);
546 schedule_work(&ctx->free_work);
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);
557 table = rcu_dereference(mm->ioctx_table);
561 for (i = 0; i < table->nr; i++)
562 if (!table->table[i]) {
564 table->table[i] = ctx;
566 spin_unlock(&mm->ioctx_lock);
568 ring = kmap_atomic(ctx->ring_pages[0]);
574 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);
588 old = rcu_dereference(mm->ioctx_table);
591 rcu_assign_pointer(mm->ioctx_table, table);
592 } else if (table->nr > old->nr) {
593 memcpy(table->table, old->table,
594 old->nr * sizeof(struct kioctx *));
596 rcu_assign_pointer(mm->ioctx_table, table);
606 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
608 static struct kioctx *ioctx_alloc(unsigned nr_events)
610 struct mm_struct *mm = current->mm;
615 * We keep track of the number of available ringbuffer slots, to prevent
616 * overflow (reqs_available), and we also use percpu counters for this.
618 * So since up to half the slots might be on other cpu's percpu counters
619 * and unavailable, double nr_events so userspace sees what they
620 * expected: additionally, we move req_batch slots to/from percpu
621 * counters at a time, so make sure that isn't 0:
623 nr_events = max(nr_events, num_possible_cpus() * 4);
626 /* Prevent overflows */
627 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
628 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
629 pr_debug("ENOMEM: nr_events too high\n");
630 return ERR_PTR(-EINVAL);
633 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
634 return ERR_PTR(-EAGAIN);
636 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
638 return ERR_PTR(-ENOMEM);
640 ctx->max_reqs = nr_events;
642 if (percpu_ref_init(&ctx->users, free_ioctx_ref))
645 spin_lock_init(&ctx->ctx_lock);
646 spin_lock_init(&ctx->completion_lock);
647 mutex_init(&ctx->ring_lock);
648 init_waitqueue_head(&ctx->wait);
650 INIT_LIST_HEAD(&ctx->active_reqs);
652 ctx->cpu = alloc_percpu(struct kioctx_cpu);
656 if (aio_setup_ring(ctx) < 0)
659 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
660 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
661 if (ctx->req_batch < 1)
664 /* limit the number of system wide aios */
665 spin_lock(&aio_nr_lock);
666 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
667 aio_nr + nr_events < aio_nr) {
668 spin_unlock(&aio_nr_lock);
671 aio_nr += ctx->max_reqs;
672 spin_unlock(&aio_nr_lock);
674 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
676 err = ioctx_add_table(ctx, mm);
678 goto out_cleanup_put;
680 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
681 ctx, ctx->user_id, mm, ctx->nr_events);
685 percpu_ref_put(&ctx->users);
690 free_percpu(ctx->cpu);
692 free_percpu(ctx->users.pcpu_count);
694 put_aio_ring_file(ctx);
695 kmem_cache_free(kioctx_cachep, ctx);
696 pr_debug("error allocating ioctx %d\n", err);
701 * Cancels all outstanding aio requests on an aio context. Used
702 * when the processes owning a context have all exited to encourage
703 * the rapid destruction of the kioctx.
705 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
707 if (!atomic_xchg(&ctx->dead, 1)) {
708 struct kioctx_table *table;
710 spin_lock(&mm->ioctx_lock);
712 table = rcu_dereference(mm->ioctx_table);
714 WARN_ON(ctx != table->table[ctx->id]);
715 table->table[ctx->id] = NULL;
717 spin_unlock(&mm->ioctx_lock);
719 /* percpu_ref_kill() will do the necessary call_rcu() */
720 wake_up_all(&ctx->wait);
723 * It'd be more correct to do this in free_ioctx(), after all
724 * the outstanding kiocbs have finished - but by then io_destroy
725 * has already returned, so io_setup() could potentially return
726 * -EAGAIN with no ioctxs actually in use (as far as userspace
729 spin_lock(&aio_nr_lock);
730 BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
731 aio_nr -= ctx->max_reqs;
732 spin_unlock(&aio_nr_lock);
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);
867 put_reqs_available(ctx, 1);
871 static void kiocb_free(struct kiocb *req)
875 if (req->ki_eventfd != NULL)
876 eventfd_ctx_put(req->ki_eventfd);
877 kmem_cache_free(kiocb_cachep, req);
880 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
882 struct aio_ring __user *ring = (void __user *)ctx_id;
883 struct mm_struct *mm = current->mm;
884 struct kioctx *ctx, *ret = NULL;
885 struct kioctx_table *table;
888 if (get_user(id, &ring->id))
892 table = rcu_dereference(mm->ioctx_table);
894 if (!table || id >= table->nr)
897 ctx = table->table[id];
898 if (ctx && ctx->user_id == ctx_id) {
899 percpu_ref_get(&ctx->users);
908 * Called when the io request on the given iocb is complete.
910 void aio_complete(struct kiocb *iocb, long res, long res2)
912 struct kioctx *ctx = iocb->ki_ctx;
913 struct aio_ring *ring;
914 struct io_event *ev_page, *event;
919 * Special case handling for sync iocbs:
920 * - events go directly into the iocb for fast handling
921 * - the sync task with the iocb in its stack holds the single iocb
922 * ref, no other paths have a way to get another ref
923 * - the sync task helpfully left a reference to itself in the iocb
925 if (is_sync_kiocb(iocb)) {
926 iocb->ki_user_data = res;
928 iocb->ki_ctx = ERR_PTR(-EXDEV);
929 wake_up_process(iocb->ki_obj.tsk);
934 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
935 * need to issue a wakeup after incrementing reqs_available.
939 if (iocb->ki_list.next) {
942 spin_lock_irqsave(&ctx->ctx_lock, flags);
943 list_del(&iocb->ki_list);
944 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
948 * Add a completion event to the ring buffer. Must be done holding
949 * ctx->completion_lock to prevent other code from messing with the tail
950 * pointer since we might be called from irq context.
952 spin_lock_irqsave(&ctx->completion_lock, flags);
955 pos = tail + AIO_EVENTS_OFFSET;
957 if (++tail >= ctx->nr_events)
960 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
961 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
963 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
964 event->data = iocb->ki_user_data;
968 kunmap_atomic(ev_page);
969 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
971 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
972 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
975 /* after flagging the request as done, we
976 * must never even look at it again
978 smp_wmb(); /* make event visible before updating tail */
982 ring = kmap_atomic(ctx->ring_pages[0]);
985 flush_dcache_page(ctx->ring_pages[0]);
987 spin_unlock_irqrestore(&ctx->completion_lock, flags);
989 pr_debug("added to ring %p at [%u]\n", iocb, tail);
992 * Check if the user asked us to deliver the result through an
993 * eventfd. The eventfd_signal() function is safe to be called
996 if (iocb->ki_eventfd != NULL)
997 eventfd_signal(iocb->ki_eventfd, 1);
999 /* everything turned out well, dispose of the aiocb. */
1003 * We have to order our ring_info tail store above and test
1004 * of the wait list below outside the wait lock. This is
1005 * like in wake_up_bit() where clearing a bit has to be
1006 * ordered with the unlocked test.
1010 if (waitqueue_active(&ctx->wait))
1011 wake_up(&ctx->wait);
1015 EXPORT_SYMBOL(aio_complete);
1018 * Pull an event off of the ioctx's event ring. Returns the number of
1021 static long aio_read_events_ring(struct kioctx *ctx,
1022 struct io_event __user *event, long nr)
1024 struct aio_ring *ring;
1025 unsigned head, tail, pos;
1029 mutex_lock(&ctx->ring_lock);
1031 ring = kmap_atomic(ctx->ring_pages[0]);
1034 kunmap_atomic(ring);
1036 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1043 struct io_event *ev;
1046 avail = (head <= tail ? tail : ctx->nr_events) - head;
1050 avail = min(avail, nr - ret);
1051 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1052 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1054 pos = head + AIO_EVENTS_OFFSET;
1055 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1056 pos %= AIO_EVENTS_PER_PAGE;
1059 copy_ret = copy_to_user(event + ret, ev + pos,
1060 sizeof(*ev) * avail);
1063 if (unlikely(copy_ret)) {
1070 head %= ctx->nr_events;
1073 ring = kmap_atomic(ctx->ring_pages[0]);
1075 kunmap_atomic(ring);
1076 flush_dcache_page(ctx->ring_pages[0]);
1078 pr_debug("%li h%u t%u\n", ret, head, tail);
1080 put_reqs_available(ctx, ret);
1082 mutex_unlock(&ctx->ring_lock);
1087 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1088 struct io_event __user *event, long *i)
1090 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1095 if (unlikely(atomic_read(&ctx->dead)))
1101 return ret < 0 || *i >= min_nr;
1104 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1105 struct io_event __user *event,
1106 struct timespec __user *timeout)
1108 ktime_t until = { .tv64 = KTIME_MAX };
1114 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1117 until = timespec_to_ktime(ts);
1121 * Note that aio_read_events() is being called as the conditional - i.e.
1122 * we're calling it after prepare_to_wait() has set task state to
1123 * TASK_INTERRUPTIBLE.
1125 * But aio_read_events() can block, and if it blocks it's going to flip
1126 * the task state back to TASK_RUNNING.
1128 * This should be ok, provided it doesn't flip the state back to
1129 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1130 * will only happen if the mutex_lock() call blocks, and we then find
1131 * the ringbuffer empty. So in practice we should be ok, but it's
1132 * something to be aware of when touching this code.
1134 wait_event_interruptible_hrtimeout(ctx->wait,
1135 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1137 if (!ret && signal_pending(current))
1144 * Create an aio_context capable of receiving at least nr_events.
1145 * ctxp must not point to an aio_context that already exists, and
1146 * must be initialized to 0 prior to the call. On successful
1147 * creation of the aio_context, *ctxp is filled in with the resulting
1148 * handle. May fail with -EINVAL if *ctxp is not initialized,
1149 * if the specified nr_events exceeds internal limits. May fail
1150 * with -EAGAIN if the specified nr_events exceeds the user's limit
1151 * of available events. May fail with -ENOMEM if insufficient kernel
1152 * resources are available. May fail with -EFAULT if an invalid
1153 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1156 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1158 struct kioctx *ioctx = NULL;
1162 ret = get_user(ctx, ctxp);
1167 if (unlikely(ctx || nr_events == 0)) {
1168 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1173 ioctx = ioctx_alloc(nr_events);
1174 ret = PTR_ERR(ioctx);
1175 if (!IS_ERR(ioctx)) {
1176 ret = put_user(ioctx->user_id, ctxp);
1178 kill_ioctx(current->mm, ioctx);
1179 percpu_ref_put(&ioctx->users);
1187 * Destroy the aio_context specified. May cancel any outstanding
1188 * AIOs and block on completion. Will fail with -ENOSYS if not
1189 * implemented. May fail with -EINVAL if the context pointed to
1192 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1194 struct kioctx *ioctx = lookup_ioctx(ctx);
1195 if (likely(NULL != ioctx)) {
1196 kill_ioctx(current->mm, ioctx);
1197 percpu_ref_put(&ioctx->users);
1200 pr_debug("EINVAL: io_destroy: invalid context id\n");
1204 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1205 unsigned long, loff_t);
1207 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1208 int rw, char __user *buf,
1209 unsigned long *nr_segs,
1210 struct iovec **iovec,
1215 *nr_segs = kiocb->ki_nbytes;
1217 #ifdef CONFIG_COMPAT
1219 ret = compat_rw_copy_check_uvector(rw,
1220 (struct compat_iovec __user *)buf,
1221 *nr_segs, 1, *iovec, iovec);
1224 ret = rw_copy_check_uvector(rw,
1225 (struct iovec __user *)buf,
1226 *nr_segs, 1, *iovec, iovec);
1230 /* ki_nbytes now reflect bytes instead of segs */
1231 kiocb->ki_nbytes = ret;
1235 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1236 int rw, char __user *buf,
1237 unsigned long *nr_segs,
1238 struct iovec *iovec)
1240 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1243 iovec->iov_base = buf;
1244 iovec->iov_len = kiocb->ki_nbytes;
1251 * Performs the initial checks and aio retry method
1252 * setup for the kiocb at the time of io submission.
1254 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1255 char __user *buf, bool compat)
1257 struct file *file = req->ki_filp;
1259 unsigned long nr_segs;
1263 struct iovec inline_vec, *iovec = &inline_vec;
1266 case IOCB_CMD_PREAD:
1267 case IOCB_CMD_PREADV:
1270 rw_op = file->f_op->aio_read;
1273 case IOCB_CMD_PWRITE:
1274 case IOCB_CMD_PWRITEV:
1277 rw_op = file->f_op->aio_write;
1280 if (unlikely(!(file->f_mode & mode)))
1286 ret = (opcode == IOCB_CMD_PREADV ||
1287 opcode == IOCB_CMD_PWRITEV)
1288 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1290 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1295 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1297 if (iovec != &inline_vec)
1302 req->ki_nbytes = ret;
1304 /* XXX: move/kill - rw_verify_area()? */
1305 /* This matches the pread()/pwrite() logic */
1306 if (req->ki_pos < 0) {
1312 file_start_write(file);
1314 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1317 file_end_write(file);
1320 case IOCB_CMD_FDSYNC:
1321 if (!file->f_op->aio_fsync)
1324 ret = file->f_op->aio_fsync(req, 1);
1327 case IOCB_CMD_FSYNC:
1328 if (!file->f_op->aio_fsync)
1331 ret = file->f_op->aio_fsync(req, 0);
1335 pr_debug("EINVAL: no operation provided\n");
1339 if (iovec != &inline_vec)
1342 if (ret != -EIOCBQUEUED) {
1344 * There's no easy way to restart the syscall since other AIO's
1345 * may be already running. Just fail this IO with EINTR.
1347 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1348 ret == -ERESTARTNOHAND ||
1349 ret == -ERESTART_RESTARTBLOCK))
1351 aio_complete(req, ret, 0);
1357 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1358 struct iocb *iocb, bool compat)
1363 /* enforce forwards compatibility on users */
1364 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1365 pr_debug("EINVAL: reserve field set\n");
1369 /* prevent overflows */
1371 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1372 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1373 ((ssize_t)iocb->aio_nbytes < 0)
1375 pr_debug("EINVAL: io_submit: overflow check\n");
1379 req = aio_get_req(ctx);
1383 req->ki_filp = fget(iocb->aio_fildes);
1384 if (unlikely(!req->ki_filp)) {
1389 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1391 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1392 * instance of the file* now. The file descriptor must be
1393 * an eventfd() fd, and will be signaled for each completed
1394 * event using the eventfd_signal() function.
1396 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1397 if (IS_ERR(req->ki_eventfd)) {
1398 ret = PTR_ERR(req->ki_eventfd);
1399 req->ki_eventfd = NULL;
1404 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1405 if (unlikely(ret)) {
1406 pr_debug("EFAULT: aio_key\n");
1410 req->ki_obj.user = user_iocb;
1411 req->ki_user_data = iocb->aio_data;
1412 req->ki_pos = iocb->aio_offset;
1413 req->ki_nbytes = iocb->aio_nbytes;
1415 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1416 (char __user *)(unsigned long)iocb->aio_buf,
1423 put_reqs_available(ctx, 1);
1428 long do_io_submit(aio_context_t ctx_id, long nr,
1429 struct iocb __user *__user *iocbpp, bool compat)
1434 struct blk_plug plug;
1436 if (unlikely(nr < 0))
1439 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1440 nr = LONG_MAX/sizeof(*iocbpp);
1442 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1445 ctx = lookup_ioctx(ctx_id);
1446 if (unlikely(!ctx)) {
1447 pr_debug("EINVAL: invalid context id\n");
1451 blk_start_plug(&plug);
1454 * AKPM: should this return a partial result if some of the IOs were
1455 * successfully submitted?
1457 for (i=0; i<nr; i++) {
1458 struct iocb __user *user_iocb;
1461 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1466 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1471 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1475 blk_finish_plug(&plug);
1477 percpu_ref_put(&ctx->users);
1482 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1483 * the number of iocbs queued. May return -EINVAL if the aio_context
1484 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1485 * *iocbpp[0] is not properly initialized, if the operation specified
1486 * is invalid for the file descriptor in the iocb. May fail with
1487 * -EFAULT if any of the data structures point to invalid data. May
1488 * fail with -EBADF if the file descriptor specified in the first
1489 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1490 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1491 * fail with -ENOSYS if not implemented.
1493 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1494 struct iocb __user * __user *, iocbpp)
1496 return do_io_submit(ctx_id, nr, iocbpp, 0);
1500 * Finds a given iocb for cancellation.
1502 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1505 struct list_head *pos;
1507 assert_spin_locked(&ctx->ctx_lock);
1509 if (key != KIOCB_KEY)
1512 /* TODO: use a hash or array, this sucks. */
1513 list_for_each(pos, &ctx->active_reqs) {
1514 struct kiocb *kiocb = list_kiocb(pos);
1515 if (kiocb->ki_obj.user == iocb)
1522 * Attempts to cancel an iocb previously passed to io_submit. If
1523 * the operation is successfully cancelled, the resulting event is
1524 * copied into the memory pointed to by result without being placed
1525 * into the completion queue and 0 is returned. May fail with
1526 * -EFAULT if any of the data structures pointed to are invalid.
1527 * May fail with -EINVAL if aio_context specified by ctx_id is
1528 * invalid. May fail with -EAGAIN if the iocb specified was not
1529 * cancelled. Will fail with -ENOSYS if not implemented.
1531 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1532 struct io_event __user *, result)
1535 struct kiocb *kiocb;
1539 ret = get_user(key, &iocb->aio_key);
1543 ctx = lookup_ioctx(ctx_id);
1547 spin_lock_irq(&ctx->ctx_lock);
1549 kiocb = lookup_kiocb(ctx, iocb, key);
1551 ret = kiocb_cancel(ctx, kiocb);
1555 spin_unlock_irq(&ctx->ctx_lock);
1559 * The result argument is no longer used - the io_event is
1560 * always delivered via the ring buffer. -EINPROGRESS indicates
1561 * cancellation is progress:
1566 percpu_ref_put(&ctx->users);
1572 * Attempts to read at least min_nr events and up to nr events from
1573 * the completion queue for the aio_context specified by ctx_id. If
1574 * it succeeds, the number of read events is returned. May fail with
1575 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1576 * out of range, if timeout is out of range. May fail with -EFAULT
1577 * if any of the memory specified is invalid. May return 0 or
1578 * < min_nr if the timeout specified by timeout has elapsed
1579 * before sufficient events are available, where timeout == NULL
1580 * specifies an infinite timeout. Note that the timeout pointed to by
1581 * timeout is relative. Will fail with -ENOSYS if not implemented.
1583 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1586 struct io_event __user *, events,
1587 struct timespec __user *, timeout)
1589 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1592 if (likely(ioctx)) {
1593 if (likely(min_nr <= nr && min_nr >= 0))
1594 ret = read_events(ioctx, min_nr, nr, events, timeout);
1595 percpu_ref_put(&ioctx->users);