2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
9 * See ../COPYING for licensing terms.
11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
65 struct io_event io_events[0];
66 }; /* 128 bytes + ring size */
68 #define AIO_RING_PAGES 8
73 struct kioctx *table[];
77 unsigned reqs_available;
81 struct percpu_ref users;
84 struct percpu_ref reqs;
86 unsigned long user_id;
88 struct __percpu kioctx_cpu *cpu;
91 * For percpu reqs_available, number of slots we move to/from global
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
99 * The real limit is nr_events - 1, which will be larger (see
104 /* Size of ringbuffer, in units of struct io_event */
107 unsigned long mmap_base;
108 unsigned long mmap_size;
110 struct page **ring_pages;
113 struct work_struct free_work;
116 * signals when all in-flight requests are done
118 struct completion *requests_done;
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
127 * We batch accesses to it with a percpu version.
129 atomic_t reqs_available;
130 } ____cacheline_aligned_in_smp;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
138 struct mutex ring_lock;
139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
144 unsigned completed_events;
145 spinlock_t completion_lock;
146 } ____cacheline_aligned_in_smp;
148 struct page *internal_pages[AIO_RING_PAGES];
149 struct file *aio_ring_file;
154 /*------ sysctl variables----*/
155 static DEFINE_SPINLOCK(aio_nr_lock);
156 unsigned long aio_nr; /* current system wide number of aio requests */
157 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
158 /*----end sysctl variables---*/
160 static struct kmem_cache *kiocb_cachep;
161 static struct kmem_cache *kioctx_cachep;
163 static struct vfsmount *aio_mnt;
165 static const struct file_operations aio_ring_fops;
166 static const struct address_space_operations aio_ctx_aops;
168 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
170 struct qstr this = QSTR_INIT("[aio]", 5);
173 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
175 return ERR_CAST(inode);
177 inode->i_mapping->a_ops = &aio_ctx_aops;
178 inode->i_mapping->private_data = ctx;
179 inode->i_size = PAGE_SIZE * nr_pages;
181 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
184 return ERR_PTR(-ENOMEM);
186 path.mnt = mntget(aio_mnt);
188 d_instantiate(path.dentry, inode);
189 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
195 file->f_flags = O_RDWR;
196 file->private_data = ctx;
200 static struct dentry *aio_mount(struct file_system_type *fs_type,
201 int flags, const char *dev_name, void *data)
203 static const struct dentry_operations ops = {
204 .d_dname = simple_dname,
206 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
210 * Creates the slab caches used by the aio routines, panic on
211 * failure as this is done early during the boot sequence.
213 static int __init aio_setup(void)
215 static struct file_system_type aio_fs = {
218 .kill_sb = kill_anon_super,
220 aio_mnt = kern_mount(&aio_fs);
222 panic("Failed to create aio fs mount.");
224 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
225 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
227 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
231 __initcall(aio_setup);
233 static void put_aio_ring_file(struct kioctx *ctx)
235 struct file *aio_ring_file = ctx->aio_ring_file;
237 truncate_setsize(aio_ring_file->f_inode, 0);
239 /* Prevent further access to the kioctx from migratepages */
240 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
241 aio_ring_file->f_inode->i_mapping->private_data = NULL;
242 ctx->aio_ring_file = NULL;
243 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
249 static void aio_free_ring(struct kioctx *ctx)
253 /* Disconnect the kiotx from the ring file. This prevents future
254 * accesses to the kioctx from page migration.
256 put_aio_ring_file(ctx);
258 for (i = 0; i < ctx->nr_pages; i++) {
260 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
261 page_count(ctx->ring_pages[i]));
262 page = ctx->ring_pages[i];
265 ctx->ring_pages[i] = NULL;
269 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
270 kfree(ctx->ring_pages);
271 ctx->ring_pages = NULL;
275 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
277 vma->vm_ops = &generic_file_vm_ops;
281 static const struct file_operations aio_ring_fops = {
282 .mmap = aio_ring_mmap,
285 static int aio_set_page_dirty(struct page *page)
290 #if IS_ENABLED(CONFIG_MIGRATION)
291 static int aio_migratepage(struct address_space *mapping, struct page *new,
292 struct page *old, enum migrate_mode mode)
301 /* mapping->private_lock here protects against the kioctx teardown. */
302 spin_lock(&mapping->private_lock);
303 ctx = mapping->private_data;
309 /* The ring_lock mutex. The prevents aio_read_events() from writing
310 * to the ring's head, and prevents page migration from mucking in
311 * a partially initialized kiotx.
313 if (!mutex_trylock(&ctx->ring_lock)) {
319 if (idx < (pgoff_t)ctx->nr_pages) {
320 /* Make sure the old page hasn't already been changed */
321 if (ctx->ring_pages[idx] != old)
329 /* Writeback must be complete */
330 BUG_ON(PageWriteback(old));
333 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
334 if (rc != MIGRATEPAGE_SUCCESS) {
339 /* Take completion_lock to prevent other writes to the ring buffer
340 * while the old page is copied to the new. This prevents new
341 * events from being lost.
343 spin_lock_irqsave(&ctx->completion_lock, flags);
344 migrate_page_copy(new, old);
345 BUG_ON(ctx->ring_pages[idx] != old);
346 ctx->ring_pages[idx] = new;
347 spin_unlock_irqrestore(&ctx->completion_lock, flags);
349 /* The old page is no longer accessible. */
353 mutex_unlock(&ctx->ring_lock);
355 spin_unlock(&mapping->private_lock);
360 static const struct address_space_operations aio_ctx_aops = {
361 .set_page_dirty = aio_set_page_dirty,
362 #if IS_ENABLED(CONFIG_MIGRATION)
363 .migratepage = aio_migratepage,
367 static int aio_setup_ring(struct kioctx *ctx)
369 struct aio_ring *ring;
370 unsigned nr_events = ctx->max_reqs;
371 struct mm_struct *mm = current->mm;
372 unsigned long size, unused;
377 /* Compensate for the ring buffer's head/tail overlap entry */
378 nr_events += 2; /* 1 is required, 2 for good luck */
380 size = sizeof(struct aio_ring);
381 size += sizeof(struct io_event) * nr_events;
383 nr_pages = PFN_UP(size);
387 file = aio_private_file(ctx, nr_pages);
389 ctx->aio_ring_file = NULL;
393 ctx->aio_ring_file = file;
394 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
395 / sizeof(struct io_event);
397 ctx->ring_pages = ctx->internal_pages;
398 if (nr_pages > AIO_RING_PAGES) {
399 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
401 if (!ctx->ring_pages) {
402 put_aio_ring_file(ctx);
407 for (i = 0; i < nr_pages; i++) {
409 page = find_or_create_page(file->f_inode->i_mapping,
410 i, GFP_HIGHUSER | __GFP_ZERO);
413 pr_debug("pid(%d) page[%d]->count=%d\n",
414 current->pid, i, page_count(page));
415 SetPageUptodate(page);
419 ctx->ring_pages[i] = page;
423 if (unlikely(i != nr_pages)) {
428 ctx->mmap_size = nr_pages * PAGE_SIZE;
429 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
431 down_write(&mm->mmap_sem);
432 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
433 PROT_READ | PROT_WRITE,
434 MAP_SHARED, 0, &unused);
435 up_write(&mm->mmap_sem);
436 if (IS_ERR((void *)ctx->mmap_base)) {
442 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
444 ctx->user_id = ctx->mmap_base;
445 ctx->nr_events = nr_events; /* trusted copy */
447 ring = kmap_atomic(ctx->ring_pages[0]);
448 ring->nr = nr_events; /* user copy */
450 ring->head = ring->tail = 0;
451 ring->magic = AIO_RING_MAGIC;
452 ring->compat_features = AIO_RING_COMPAT_FEATURES;
453 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
454 ring->header_length = sizeof(struct aio_ring);
456 flush_dcache_page(ctx->ring_pages[0]);
461 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
462 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
463 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
465 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
467 struct kioctx *ctx = req->ki_ctx;
470 spin_lock_irqsave(&ctx->ctx_lock, flags);
472 if (!req->ki_list.next)
473 list_add(&req->ki_list, &ctx->active_reqs);
475 req->ki_cancel = cancel;
477 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
479 EXPORT_SYMBOL(kiocb_set_cancel_fn);
481 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
483 kiocb_cancel_fn *old, *cancel;
486 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
487 * actually has a cancel function, hence the cmpxchg()
490 cancel = ACCESS_ONCE(kiocb->ki_cancel);
492 if (!cancel || cancel == KIOCB_CANCELLED)
496 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
497 } while (cancel != old);
499 return cancel(kiocb);
502 static void free_ioctx(struct work_struct *work)
504 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
506 pr_debug("freeing %p\n", ctx);
509 free_percpu(ctx->cpu);
510 kmem_cache_free(kioctx_cachep, ctx);
513 static void free_ioctx_reqs(struct percpu_ref *ref)
515 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
517 /* At this point we know that there are no any in-flight requests */
518 if (ctx->requests_done)
519 complete(ctx->requests_done);
521 INIT_WORK(&ctx->free_work, free_ioctx);
522 schedule_work(&ctx->free_work);
526 * When this function runs, the kioctx has been removed from the "hash table"
527 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
528 * now it's safe to cancel any that need to be.
530 static void free_ioctx_users(struct percpu_ref *ref)
532 struct kioctx *ctx = container_of(ref, struct kioctx, users);
535 spin_lock_irq(&ctx->ctx_lock);
537 while (!list_empty(&ctx->active_reqs)) {
538 req = list_first_entry(&ctx->active_reqs,
539 struct kiocb, ki_list);
541 list_del_init(&req->ki_list);
542 kiocb_cancel(ctx, req);
545 spin_unlock_irq(&ctx->ctx_lock);
547 percpu_ref_kill(&ctx->reqs);
548 percpu_ref_put(&ctx->reqs);
551 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
554 struct kioctx_table *table, *old;
555 struct aio_ring *ring;
557 spin_lock(&mm->ioctx_lock);
559 table = rcu_dereference(mm->ioctx_table);
563 for (i = 0; i < table->nr; i++)
564 if (!table->table[i]) {
566 table->table[i] = ctx;
568 spin_unlock(&mm->ioctx_lock);
570 /* While kioctx setup is in progress,
571 * we are protected from page migration
572 * changes ring_pages by ->ring_lock.
574 ring = kmap_atomic(ctx->ring_pages[0]);
580 new_nr = (table ? table->nr : 1) * 4;
583 spin_unlock(&mm->ioctx_lock);
585 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
592 spin_lock(&mm->ioctx_lock);
594 old = rcu_dereference(mm->ioctx_table);
597 rcu_assign_pointer(mm->ioctx_table, table);
598 } else if (table->nr > old->nr) {
599 memcpy(table->table, old->table,
600 old->nr * sizeof(struct kioctx *));
602 rcu_assign_pointer(mm->ioctx_table, table);
611 static void aio_nr_sub(unsigned nr)
613 spin_lock(&aio_nr_lock);
614 if (WARN_ON(aio_nr - nr > aio_nr))
618 spin_unlock(&aio_nr_lock);
622 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
624 static struct kioctx *ioctx_alloc(unsigned nr_events)
626 struct mm_struct *mm = current->mm;
631 * We keep track of the number of available ringbuffer slots, to prevent
632 * overflow (reqs_available), and we also use percpu counters for this.
634 * So since up to half the slots might be on other cpu's percpu counters
635 * and unavailable, double nr_events so userspace sees what they
636 * expected: additionally, we move req_batch slots to/from percpu
637 * counters at a time, so make sure that isn't 0:
639 nr_events = max(nr_events, num_possible_cpus() * 4);
642 /* Prevent overflows */
643 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
644 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
645 pr_debug("ENOMEM: nr_events too high\n");
646 return ERR_PTR(-EINVAL);
649 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
650 return ERR_PTR(-EAGAIN);
652 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
654 return ERR_PTR(-ENOMEM);
656 ctx->max_reqs = nr_events;
658 spin_lock_init(&ctx->ctx_lock);
659 spin_lock_init(&ctx->completion_lock);
660 mutex_init(&ctx->ring_lock);
661 /* Protect against page migration throughout kiotx setup by keeping
662 * the ring_lock mutex held until setup is complete. */
663 mutex_lock(&ctx->ring_lock);
664 init_waitqueue_head(&ctx->wait);
666 INIT_LIST_HEAD(&ctx->active_reqs);
668 if (percpu_ref_init(&ctx->users, free_ioctx_users))
671 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
674 ctx->cpu = alloc_percpu(struct kioctx_cpu);
678 err = aio_setup_ring(ctx);
682 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
683 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
684 if (ctx->req_batch < 1)
687 /* limit the number of system wide aios */
688 spin_lock(&aio_nr_lock);
689 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
690 aio_nr + nr_events < aio_nr) {
691 spin_unlock(&aio_nr_lock);
695 aio_nr += ctx->max_reqs;
696 spin_unlock(&aio_nr_lock);
698 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
699 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
701 err = ioctx_add_table(ctx, mm);
705 /* Release the ring_lock mutex now that all setup is complete. */
706 mutex_unlock(&ctx->ring_lock);
708 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
709 ctx, ctx->user_id, mm, ctx->nr_events);
713 aio_nr_sub(ctx->max_reqs);
717 mutex_unlock(&ctx->ring_lock);
718 free_percpu(ctx->cpu);
719 free_percpu(ctx->reqs.pcpu_count);
720 free_percpu(ctx->users.pcpu_count);
721 kmem_cache_free(kioctx_cachep, ctx);
722 pr_debug("error allocating ioctx %d\n", err);
727 * Cancels all outstanding aio requests on an aio context. Used
728 * when the processes owning a context have all exited to encourage
729 * the rapid destruction of the kioctx.
731 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
732 struct completion *requests_done)
734 if (!atomic_xchg(&ctx->dead, 1)) {
735 struct kioctx_table *table;
737 spin_lock(&mm->ioctx_lock);
739 table = rcu_dereference(mm->ioctx_table);
741 WARN_ON(ctx != table->table[ctx->id]);
742 table->table[ctx->id] = NULL;
744 spin_unlock(&mm->ioctx_lock);
746 /* percpu_ref_kill() will do the necessary call_rcu() */
747 wake_up_all(&ctx->wait);
750 * It'd be more correct to do this in free_ioctx(), after all
751 * the outstanding kiocbs have finished - but by then io_destroy
752 * has already returned, so io_setup() could potentially return
753 * -EAGAIN with no ioctxs actually in use (as far as userspace
756 aio_nr_sub(ctx->max_reqs);
759 vm_munmap(ctx->mmap_base, ctx->mmap_size);
761 ctx->requests_done = requests_done;
762 percpu_ref_kill(&ctx->users);
765 complete(requests_done);
769 /* wait_on_sync_kiocb:
770 * Waits on the given sync kiocb to complete.
772 ssize_t wait_on_sync_kiocb(struct kiocb *req)
774 while (!req->ki_ctx) {
775 set_current_state(TASK_UNINTERRUPTIBLE);
780 __set_current_state(TASK_RUNNING);
781 return req->ki_user_data;
783 EXPORT_SYMBOL(wait_on_sync_kiocb);
786 * exit_aio: called when the last user of mm goes away. At this point, there is
787 * no way for any new requests to be submited or any of the io_* syscalls to be
788 * called on the context.
790 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
793 void exit_aio(struct mm_struct *mm)
795 struct kioctx_table *table;
801 table = rcu_dereference(mm->ioctx_table);
804 if (!table || i >= table->nr) {
806 rcu_assign_pointer(mm->ioctx_table, NULL);
812 ctx = table->table[i++];
818 * We don't need to bother with munmap() here -
819 * exit_mmap(mm) is coming and it'll unmap everything.
820 * Since aio_free_ring() uses non-zero ->mmap_size
821 * as indicator that it needs to unmap the area,
822 * just set it to 0; aio_free_ring() is the only
823 * place that uses ->mmap_size, so it's safe.
827 kill_ioctx(mm, ctx, NULL);
831 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
833 struct kioctx_cpu *kcpu;
837 kcpu = this_cpu_ptr(ctx->cpu);
839 local_irq_save(flags);
840 kcpu->reqs_available += nr;
842 while (kcpu->reqs_available >= ctx->req_batch * 2) {
843 kcpu->reqs_available -= ctx->req_batch;
844 atomic_add(ctx->req_batch, &ctx->reqs_available);
847 local_irq_restore(flags);
851 static bool get_reqs_available(struct kioctx *ctx)
853 struct kioctx_cpu *kcpu;
858 kcpu = this_cpu_ptr(ctx->cpu);
860 local_irq_save(flags);
861 if (!kcpu->reqs_available) {
862 int old, avail = atomic_read(&ctx->reqs_available);
865 if (avail < ctx->req_batch)
869 avail = atomic_cmpxchg(&ctx->reqs_available,
870 avail, avail - ctx->req_batch);
871 } while (avail != old);
873 kcpu->reqs_available += ctx->req_batch;
877 kcpu->reqs_available--;
879 local_irq_restore(flags);
884 /* refill_reqs_available
885 * Updates the reqs_available reference counts used for tracking the
886 * number of free slots in the completion ring. This can be called
887 * from aio_complete() (to optimistically update reqs_available) or
888 * from aio_get_req() (the we're out of events case). It must be
889 * called holding ctx->completion_lock.
891 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
894 unsigned events_in_ring, completed;
896 /* Clamp head since userland can write to it. */
897 head %= ctx->nr_events;
899 events_in_ring = tail - head;
901 events_in_ring = ctx->nr_events - (head - tail);
903 completed = ctx->completed_events;
904 if (events_in_ring < completed)
905 completed -= events_in_ring;
912 ctx->completed_events -= completed;
913 put_reqs_available(ctx, completed);
916 /* user_refill_reqs_available
917 * Called to refill reqs_available when aio_get_req() encounters an
918 * out of space in the completion ring.
920 static void user_refill_reqs_available(struct kioctx *ctx)
922 spin_lock_irq(&ctx->completion_lock);
923 if (ctx->completed_events) {
924 struct aio_ring *ring;
927 /* Access of ring->head may race with aio_read_events_ring()
928 * here, but that's okay since whether we read the old version
929 * or the new version, and either will be valid. The important
930 * part is that head cannot pass tail since we prevent
931 * aio_complete() from updating tail by holding
932 * ctx->completion_lock. Even if head is invalid, the check
933 * against ctx->completed_events below will make sure we do the
936 ring = kmap_atomic(ctx->ring_pages[0]);
940 refill_reqs_available(ctx, head, ctx->tail);
943 spin_unlock_irq(&ctx->completion_lock);
947 * Allocate a slot for an aio request.
948 * Returns NULL if no requests are free.
950 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
954 if (!get_reqs_available(ctx)) {
955 user_refill_reqs_available(ctx);
956 if (!get_reqs_available(ctx))
960 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
964 percpu_ref_get(&ctx->reqs);
969 put_reqs_available(ctx, 1);
973 static void kiocb_free(struct kiocb *req)
977 if (req->ki_eventfd != NULL)
978 eventfd_ctx_put(req->ki_eventfd);
979 kmem_cache_free(kiocb_cachep, req);
982 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
984 struct aio_ring __user *ring = (void __user *)ctx_id;
985 struct mm_struct *mm = current->mm;
986 struct kioctx *ctx, *ret = NULL;
987 struct kioctx_table *table;
990 if (get_user(id, &ring->id))
994 table = rcu_dereference(mm->ioctx_table);
996 if (!table || id >= table->nr)
999 ctx = table->table[id];
1000 if (ctx && ctx->user_id == ctx_id) {
1001 percpu_ref_get(&ctx->users);
1010 * Called when the io request on the given iocb is complete.
1012 void aio_complete(struct kiocb *iocb, long res, long res2)
1014 struct kioctx *ctx = iocb->ki_ctx;
1015 struct aio_ring *ring;
1016 struct io_event *ev_page, *event;
1017 unsigned tail, pos, head;
1018 unsigned long flags;
1021 * Special case handling for sync iocbs:
1022 * - events go directly into the iocb for fast handling
1023 * - the sync task with the iocb in its stack holds the single iocb
1024 * ref, no other paths have a way to get another ref
1025 * - the sync task helpfully left a reference to itself in the iocb
1027 if (is_sync_kiocb(iocb)) {
1028 iocb->ki_user_data = res;
1030 iocb->ki_ctx = ERR_PTR(-EXDEV);
1031 wake_up_process(iocb->ki_obj.tsk);
1035 if (iocb->ki_list.next) {
1036 unsigned long flags;
1038 spin_lock_irqsave(&ctx->ctx_lock, flags);
1039 list_del(&iocb->ki_list);
1040 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1044 * Add a completion event to the ring buffer. Must be done holding
1045 * ctx->completion_lock to prevent other code from messing with the tail
1046 * pointer since we might be called from irq context.
1048 spin_lock_irqsave(&ctx->completion_lock, flags);
1051 pos = tail + AIO_EVENTS_OFFSET;
1053 if (++tail >= ctx->nr_events)
1056 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1057 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1059 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1060 event->data = iocb->ki_user_data;
1064 kunmap_atomic(ev_page);
1065 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1067 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1068 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1071 /* after flagging the request as done, we
1072 * must never even look at it again
1074 smp_wmb(); /* make event visible before updating tail */
1078 ring = kmap_atomic(ctx->ring_pages[0]);
1081 kunmap_atomic(ring);
1082 flush_dcache_page(ctx->ring_pages[0]);
1084 ctx->completed_events++;
1085 if (ctx->completed_events > 1)
1086 refill_reqs_available(ctx, head, tail);
1087 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1089 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1092 * Check if the user asked us to deliver the result through an
1093 * eventfd. The eventfd_signal() function is safe to be called
1096 if (iocb->ki_eventfd != NULL)
1097 eventfd_signal(iocb->ki_eventfd, 1);
1099 /* everything turned out well, dispose of the aiocb. */
1103 * We have to order our ring_info tail store above and test
1104 * of the wait list below outside the wait lock. This is
1105 * like in wake_up_bit() where clearing a bit has to be
1106 * ordered with the unlocked test.
1110 if (waitqueue_active(&ctx->wait))
1111 wake_up(&ctx->wait);
1113 percpu_ref_put(&ctx->reqs);
1115 EXPORT_SYMBOL(aio_complete);
1118 * Pull an event off of the ioctx's event ring. Returns the number of
1121 static long aio_read_events_ring(struct kioctx *ctx,
1122 struct io_event __user *event, long nr)
1124 struct aio_ring *ring;
1125 unsigned head, tail, pos;
1129 mutex_lock(&ctx->ring_lock);
1131 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1132 ring = kmap_atomic(ctx->ring_pages[0]);
1135 kunmap_atomic(ring);
1137 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1142 head %= ctx->nr_events;
1143 tail %= ctx->nr_events;
1147 struct io_event *ev;
1150 avail = (head <= tail ? tail : ctx->nr_events) - head;
1154 avail = min(avail, nr - ret);
1155 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1156 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1158 pos = head + AIO_EVENTS_OFFSET;
1159 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1160 pos %= AIO_EVENTS_PER_PAGE;
1163 copy_ret = copy_to_user(event + ret, ev + pos,
1164 sizeof(*ev) * avail);
1167 if (unlikely(copy_ret)) {
1174 head %= ctx->nr_events;
1177 ring = kmap_atomic(ctx->ring_pages[0]);
1179 kunmap_atomic(ring);
1180 flush_dcache_page(ctx->ring_pages[0]);
1182 pr_debug("%li h%u t%u\n", ret, head, tail);
1184 mutex_unlock(&ctx->ring_lock);
1189 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1190 struct io_event __user *event, long *i)
1192 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1197 if (unlikely(atomic_read(&ctx->dead)))
1203 return ret < 0 || *i >= min_nr;
1206 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1207 struct io_event __user *event,
1208 struct timespec __user *timeout)
1210 ktime_t until = { .tv64 = KTIME_MAX };
1216 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1219 until = timespec_to_ktime(ts);
1223 * Note that aio_read_events() is being called as the conditional - i.e.
1224 * we're calling it after prepare_to_wait() has set task state to
1225 * TASK_INTERRUPTIBLE.
1227 * But aio_read_events() can block, and if it blocks it's going to flip
1228 * the task state back to TASK_RUNNING.
1230 * This should be ok, provided it doesn't flip the state back to
1231 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1232 * will only happen if the mutex_lock() call blocks, and we then find
1233 * the ringbuffer empty. So in practice we should be ok, but it's
1234 * something to be aware of when touching this code.
1236 wait_event_interruptible_hrtimeout(ctx->wait,
1237 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1239 if (!ret && signal_pending(current))
1246 * Create an aio_context capable of receiving at least nr_events.
1247 * ctxp must not point to an aio_context that already exists, and
1248 * must be initialized to 0 prior to the call. On successful
1249 * creation of the aio_context, *ctxp is filled in with the resulting
1250 * handle. May fail with -EINVAL if *ctxp is not initialized,
1251 * if the specified nr_events exceeds internal limits. May fail
1252 * with -EAGAIN if the specified nr_events exceeds the user's limit
1253 * of available events. May fail with -ENOMEM if insufficient kernel
1254 * resources are available. May fail with -EFAULT if an invalid
1255 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1258 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1260 struct kioctx *ioctx = NULL;
1264 ret = get_user(ctx, ctxp);
1269 if (unlikely(ctx || nr_events == 0)) {
1270 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1275 ioctx = ioctx_alloc(nr_events);
1276 ret = PTR_ERR(ioctx);
1277 if (!IS_ERR(ioctx)) {
1278 ret = put_user(ioctx->user_id, ctxp);
1280 kill_ioctx(current->mm, ioctx, NULL);
1281 percpu_ref_put(&ioctx->users);
1289 * Destroy the aio_context specified. May cancel any outstanding
1290 * AIOs and block on completion. Will fail with -ENOSYS if not
1291 * implemented. May fail with -EINVAL if the context pointed to
1294 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1296 struct kioctx *ioctx = lookup_ioctx(ctx);
1297 if (likely(NULL != ioctx)) {
1298 struct completion requests_done =
1299 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1301 /* Pass requests_done to kill_ioctx() where it can be set
1302 * in a thread-safe way. If we try to set it here then we have
1303 * a race condition if two io_destroy() called simultaneously.
1305 kill_ioctx(current->mm, ioctx, &requests_done);
1306 percpu_ref_put(&ioctx->users);
1308 /* Wait until all IO for the context are done. Otherwise kernel
1309 * keep using user-space buffers even if user thinks the context
1312 wait_for_completion(&requests_done);
1316 pr_debug("EINVAL: io_destroy: invalid context id\n");
1320 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1321 unsigned long, loff_t);
1323 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1324 int rw, char __user *buf,
1325 unsigned long *nr_segs,
1326 struct iovec **iovec,
1331 *nr_segs = kiocb->ki_nbytes;
1333 #ifdef CONFIG_COMPAT
1335 ret = compat_rw_copy_check_uvector(rw,
1336 (struct compat_iovec __user *)buf,
1337 *nr_segs, 1, *iovec, iovec);
1340 ret = rw_copy_check_uvector(rw,
1341 (struct iovec __user *)buf,
1342 *nr_segs, 1, *iovec, iovec);
1346 /* ki_nbytes now reflect bytes instead of segs */
1347 kiocb->ki_nbytes = ret;
1351 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1352 int rw, char __user *buf,
1353 unsigned long *nr_segs,
1354 struct iovec *iovec)
1356 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1359 iovec->iov_base = buf;
1360 iovec->iov_len = kiocb->ki_nbytes;
1367 * Performs the initial checks and aio retry method
1368 * setup for the kiocb at the time of io submission.
1370 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1371 char __user *buf, bool compat)
1373 struct file *file = req->ki_filp;
1375 unsigned long nr_segs;
1379 struct iovec inline_vec, *iovec = &inline_vec;
1382 case IOCB_CMD_PREAD:
1383 case IOCB_CMD_PREADV:
1386 rw_op = file->f_op->aio_read;
1389 case IOCB_CMD_PWRITE:
1390 case IOCB_CMD_PWRITEV:
1393 rw_op = file->f_op->aio_write;
1396 if (unlikely(!(file->f_mode & mode)))
1402 ret = (opcode == IOCB_CMD_PREADV ||
1403 opcode == IOCB_CMD_PWRITEV)
1404 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1406 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1409 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1411 if (iovec != &inline_vec)
1416 req->ki_nbytes = ret;
1418 /* XXX: move/kill - rw_verify_area()? */
1419 /* This matches the pread()/pwrite() logic */
1420 if (req->ki_pos < 0) {
1426 file_start_write(file);
1428 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1431 file_end_write(file);
1434 case IOCB_CMD_FDSYNC:
1435 if (!file->f_op->aio_fsync)
1438 ret = file->f_op->aio_fsync(req, 1);
1441 case IOCB_CMD_FSYNC:
1442 if (!file->f_op->aio_fsync)
1445 ret = file->f_op->aio_fsync(req, 0);
1449 pr_debug("EINVAL: no operation provided\n");
1453 if (iovec != &inline_vec)
1456 if (ret != -EIOCBQUEUED) {
1458 * There's no easy way to restart the syscall since other AIO's
1459 * may be already running. Just fail this IO with EINTR.
1461 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1462 ret == -ERESTARTNOHAND ||
1463 ret == -ERESTART_RESTARTBLOCK))
1465 aio_complete(req, ret, 0);
1471 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1472 struct iocb *iocb, bool compat)
1477 /* enforce forwards compatibility on users */
1478 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1479 pr_debug("EINVAL: reserve field set\n");
1483 /* prevent overflows */
1485 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1486 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1487 ((ssize_t)iocb->aio_nbytes < 0)
1489 pr_debug("EINVAL: io_submit: overflow check\n");
1493 req = aio_get_req(ctx);
1497 req->ki_filp = fget(iocb->aio_fildes);
1498 if (unlikely(!req->ki_filp)) {
1503 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1505 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1506 * instance of the file* now. The file descriptor must be
1507 * an eventfd() fd, and will be signaled for each completed
1508 * event using the eventfd_signal() function.
1510 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1511 if (IS_ERR(req->ki_eventfd)) {
1512 ret = PTR_ERR(req->ki_eventfd);
1513 req->ki_eventfd = NULL;
1518 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1519 if (unlikely(ret)) {
1520 pr_debug("EFAULT: aio_key\n");
1524 req->ki_obj.user = user_iocb;
1525 req->ki_user_data = iocb->aio_data;
1526 req->ki_pos = iocb->aio_offset;
1527 req->ki_nbytes = iocb->aio_nbytes;
1529 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1530 (char __user *)(unsigned long)iocb->aio_buf,
1537 put_reqs_available(ctx, 1);
1538 percpu_ref_put(&ctx->reqs);
1543 long do_io_submit(aio_context_t ctx_id, long nr,
1544 struct iocb __user *__user *iocbpp, bool compat)
1549 struct blk_plug plug;
1551 if (unlikely(nr < 0))
1554 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1555 nr = LONG_MAX/sizeof(*iocbpp);
1557 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1560 ctx = lookup_ioctx(ctx_id);
1561 if (unlikely(!ctx)) {
1562 pr_debug("EINVAL: invalid context id\n");
1566 blk_start_plug(&plug);
1569 * AKPM: should this return a partial result if some of the IOs were
1570 * successfully submitted?
1572 for (i=0; i<nr; i++) {
1573 struct iocb __user *user_iocb;
1576 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1581 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1586 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1590 blk_finish_plug(&plug);
1592 percpu_ref_put(&ctx->users);
1597 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1598 * the number of iocbs queued. May return -EINVAL if the aio_context
1599 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1600 * *iocbpp[0] is not properly initialized, if the operation specified
1601 * is invalid for the file descriptor in the iocb. May fail with
1602 * -EFAULT if any of the data structures point to invalid data. May
1603 * fail with -EBADF if the file descriptor specified in the first
1604 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1605 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1606 * fail with -ENOSYS if not implemented.
1608 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1609 struct iocb __user * __user *, iocbpp)
1611 return do_io_submit(ctx_id, nr, iocbpp, 0);
1615 * Finds a given iocb for cancellation.
1617 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1620 struct list_head *pos;
1622 assert_spin_locked(&ctx->ctx_lock);
1624 if (key != KIOCB_KEY)
1627 /* TODO: use a hash or array, this sucks. */
1628 list_for_each(pos, &ctx->active_reqs) {
1629 struct kiocb *kiocb = list_kiocb(pos);
1630 if (kiocb->ki_obj.user == iocb)
1637 * Attempts to cancel an iocb previously passed to io_submit. If
1638 * the operation is successfully cancelled, the resulting event is
1639 * copied into the memory pointed to by result without being placed
1640 * into the completion queue and 0 is returned. May fail with
1641 * -EFAULT if any of the data structures pointed to are invalid.
1642 * May fail with -EINVAL if aio_context specified by ctx_id is
1643 * invalid. May fail with -EAGAIN if the iocb specified was not
1644 * cancelled. Will fail with -ENOSYS if not implemented.
1646 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1647 struct io_event __user *, result)
1650 struct kiocb *kiocb;
1654 ret = get_user(key, &iocb->aio_key);
1658 ctx = lookup_ioctx(ctx_id);
1662 spin_lock_irq(&ctx->ctx_lock);
1664 kiocb = lookup_kiocb(ctx, iocb, key);
1666 ret = kiocb_cancel(ctx, kiocb);
1670 spin_unlock_irq(&ctx->ctx_lock);
1674 * The result argument is no longer used - the io_event is
1675 * always delivered via the ring buffer. -EINPROGRESS indicates
1676 * cancellation is progress:
1681 percpu_ref_put(&ctx->users);
1687 * Attempts to read at least min_nr events and up to nr events from
1688 * the completion queue for the aio_context specified by ctx_id. If
1689 * it succeeds, the number of read events is returned. May fail with
1690 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1691 * out of range, if timeout is out of range. May fail with -EFAULT
1692 * if any of the memory specified is invalid. May return 0 or
1693 * < min_nr if the timeout specified by timeout has elapsed
1694 * before sufficient events are available, where timeout == NULL
1695 * specifies an infinite timeout. Note that the timeout pointed to by
1696 * timeout is relative. Will fail with -ENOSYS if not implemented.
1698 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1701 struct io_event __user *, events,
1702 struct timespec __user *, timeout)
1704 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1707 if (likely(ioctx)) {
1708 if (likely(min_nr <= nr && min_nr >= 0))
1709 ret = read_events(ioctx, min_nr, nr, events, timeout);
1710 percpu_ref_put(&ioctx->users);