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 /* Backing dev info for aio fs.
169 * -no dirty page accounting or writeback happens
171 static struct backing_dev_info aio_fs_backing_dev_info = {
174 .capabilities = BDI_CAP_NO_ACCT_AND_WRITEBACK | BDI_CAP_MAP_COPY,
177 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
179 struct qstr this = QSTR_INIT("[aio]", 5);
182 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
184 return ERR_CAST(inode);
186 inode->i_mapping->a_ops = &aio_ctx_aops;
187 inode->i_mapping->private_data = ctx;
188 inode->i_mapping->backing_dev_info = &aio_fs_backing_dev_info;
189 inode->i_size = PAGE_SIZE * nr_pages;
191 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
194 return ERR_PTR(-ENOMEM);
196 path.mnt = mntget(aio_mnt);
198 d_instantiate(path.dentry, inode);
199 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
205 file->f_flags = O_RDWR;
206 file->private_data = ctx;
210 static struct dentry *aio_mount(struct file_system_type *fs_type,
211 int flags, const char *dev_name, void *data)
213 static const struct dentry_operations ops = {
214 .d_dname = simple_dname,
216 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
220 * Creates the slab caches used by the aio routines, panic on
221 * failure as this is done early during the boot sequence.
223 static int __init aio_setup(void)
225 static struct file_system_type aio_fs = {
228 .kill_sb = kill_anon_super,
230 aio_mnt = kern_mount(&aio_fs);
232 panic("Failed to create aio fs mount.");
234 if (bdi_init(&aio_fs_backing_dev_info))
235 panic("Failed to init aio fs backing dev info.");
237 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
238 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
240 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
244 __initcall(aio_setup);
246 static void put_aio_ring_file(struct kioctx *ctx)
248 struct file *aio_ring_file = ctx->aio_ring_file;
250 truncate_setsize(aio_ring_file->f_inode, 0);
252 /* Prevent further access to the kioctx from migratepages */
253 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
254 aio_ring_file->f_inode->i_mapping->private_data = NULL;
255 ctx->aio_ring_file = NULL;
256 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
262 static void aio_free_ring(struct kioctx *ctx)
266 /* Disconnect the kiotx from the ring file. This prevents future
267 * accesses to the kioctx from page migration.
269 put_aio_ring_file(ctx);
271 for (i = 0; i < ctx->nr_pages; i++) {
273 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
274 page_count(ctx->ring_pages[i]));
275 page = ctx->ring_pages[i];
278 ctx->ring_pages[i] = NULL;
282 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
283 kfree(ctx->ring_pages);
284 ctx->ring_pages = NULL;
288 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
290 vma->vm_ops = &generic_file_vm_ops;
294 static const struct file_operations aio_ring_fops = {
295 .mmap = aio_ring_mmap,
298 #if IS_ENABLED(CONFIG_MIGRATION)
299 static int aio_migratepage(struct address_space *mapping, struct page *new,
300 struct page *old, enum migrate_mode mode)
309 /* mapping->private_lock here protects against the kioctx teardown. */
310 spin_lock(&mapping->private_lock);
311 ctx = mapping->private_data;
317 /* The ring_lock mutex. The prevents aio_read_events() from writing
318 * to the ring's head, and prevents page migration from mucking in
319 * a partially initialized kiotx.
321 if (!mutex_trylock(&ctx->ring_lock)) {
327 if (idx < (pgoff_t)ctx->nr_pages) {
328 /* Make sure the old page hasn't already been changed */
329 if (ctx->ring_pages[idx] != old)
337 /* Writeback must be complete */
338 BUG_ON(PageWriteback(old));
341 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
342 if (rc != MIGRATEPAGE_SUCCESS) {
347 /* Take completion_lock to prevent other writes to the ring buffer
348 * while the old page is copied to the new. This prevents new
349 * events from being lost.
351 spin_lock_irqsave(&ctx->completion_lock, flags);
352 migrate_page_copy(new, old);
353 BUG_ON(ctx->ring_pages[idx] != old);
354 ctx->ring_pages[idx] = new;
355 spin_unlock_irqrestore(&ctx->completion_lock, flags);
357 /* The old page is no longer accessible. */
361 mutex_unlock(&ctx->ring_lock);
363 spin_unlock(&mapping->private_lock);
368 static const struct address_space_operations aio_ctx_aops = {
369 .set_page_dirty = __set_page_dirty_no_writeback,
370 #if IS_ENABLED(CONFIG_MIGRATION)
371 .migratepage = aio_migratepage,
375 static int aio_setup_ring(struct kioctx *ctx)
377 struct aio_ring *ring;
378 unsigned nr_events = ctx->max_reqs;
379 struct mm_struct *mm = current->mm;
380 unsigned long size, unused;
385 /* Compensate for the ring buffer's head/tail overlap entry */
386 nr_events += 2; /* 1 is required, 2 for good luck */
388 size = sizeof(struct aio_ring);
389 size += sizeof(struct io_event) * nr_events;
391 nr_pages = PFN_UP(size);
395 file = aio_private_file(ctx, nr_pages);
397 ctx->aio_ring_file = NULL;
401 ctx->aio_ring_file = file;
402 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
403 / sizeof(struct io_event);
405 ctx->ring_pages = ctx->internal_pages;
406 if (nr_pages > AIO_RING_PAGES) {
407 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
409 if (!ctx->ring_pages) {
410 put_aio_ring_file(ctx);
415 for (i = 0; i < nr_pages; i++) {
417 page = find_or_create_page(file->f_inode->i_mapping,
418 i, GFP_HIGHUSER | __GFP_ZERO);
421 pr_debug("pid(%d) page[%d]->count=%d\n",
422 current->pid, i, page_count(page));
423 SetPageUptodate(page);
426 ctx->ring_pages[i] = page;
430 if (unlikely(i != nr_pages)) {
435 ctx->mmap_size = nr_pages * PAGE_SIZE;
436 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
438 down_write(&mm->mmap_sem);
439 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
440 PROT_READ | PROT_WRITE,
441 MAP_SHARED, 0, &unused);
442 up_write(&mm->mmap_sem);
443 if (IS_ERR((void *)ctx->mmap_base)) {
449 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
451 ctx->user_id = ctx->mmap_base;
452 ctx->nr_events = nr_events; /* trusted copy */
454 ring = kmap_atomic(ctx->ring_pages[0]);
455 ring->nr = nr_events; /* user copy */
457 ring->head = ring->tail = 0;
458 ring->magic = AIO_RING_MAGIC;
459 ring->compat_features = AIO_RING_COMPAT_FEATURES;
460 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
461 ring->header_length = sizeof(struct aio_ring);
463 flush_dcache_page(ctx->ring_pages[0]);
468 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
469 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
470 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
472 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
474 struct kioctx *ctx = req->ki_ctx;
477 spin_lock_irqsave(&ctx->ctx_lock, flags);
479 if (!req->ki_list.next)
480 list_add(&req->ki_list, &ctx->active_reqs);
482 req->ki_cancel = cancel;
484 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
486 EXPORT_SYMBOL(kiocb_set_cancel_fn);
488 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
490 kiocb_cancel_fn *old, *cancel;
493 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
494 * actually has a cancel function, hence the cmpxchg()
497 cancel = ACCESS_ONCE(kiocb->ki_cancel);
499 if (!cancel || cancel == KIOCB_CANCELLED)
503 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
504 } while (cancel != old);
506 return cancel(kiocb);
509 static void free_ioctx(struct work_struct *work)
511 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
513 pr_debug("freeing %p\n", ctx);
516 free_percpu(ctx->cpu);
517 kmem_cache_free(kioctx_cachep, ctx);
520 static void free_ioctx_reqs(struct percpu_ref *ref)
522 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
524 /* At this point we know that there are no any in-flight requests */
525 if (ctx->requests_done)
526 complete(ctx->requests_done);
528 INIT_WORK(&ctx->free_work, free_ioctx);
529 schedule_work(&ctx->free_work);
533 * When this function runs, the kioctx has been removed from the "hash table"
534 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
535 * now it's safe to cancel any that need to be.
537 static void free_ioctx_users(struct percpu_ref *ref)
539 struct kioctx *ctx = container_of(ref, struct kioctx, users);
542 spin_lock_irq(&ctx->ctx_lock);
544 while (!list_empty(&ctx->active_reqs)) {
545 req = list_first_entry(&ctx->active_reqs,
546 struct kiocb, ki_list);
548 list_del_init(&req->ki_list);
549 kiocb_cancel(ctx, req);
552 spin_unlock_irq(&ctx->ctx_lock);
554 percpu_ref_kill(&ctx->reqs);
555 percpu_ref_put(&ctx->reqs);
558 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
561 struct kioctx_table *table, *old;
562 struct aio_ring *ring;
564 spin_lock(&mm->ioctx_lock);
566 table = rcu_dereference(mm->ioctx_table);
570 for (i = 0; i < table->nr; i++)
571 if (!table->table[i]) {
573 table->table[i] = ctx;
575 spin_unlock(&mm->ioctx_lock);
577 /* While kioctx setup is in progress,
578 * we are protected from page migration
579 * changes ring_pages by ->ring_lock.
581 ring = kmap_atomic(ctx->ring_pages[0]);
587 new_nr = (table ? table->nr : 1) * 4;
590 spin_unlock(&mm->ioctx_lock);
592 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
599 spin_lock(&mm->ioctx_lock);
601 old = rcu_dereference(mm->ioctx_table);
604 rcu_assign_pointer(mm->ioctx_table, table);
605 } else if (table->nr > old->nr) {
606 memcpy(table->table, old->table,
607 old->nr * sizeof(struct kioctx *));
609 rcu_assign_pointer(mm->ioctx_table, table);
618 static void aio_nr_sub(unsigned nr)
620 spin_lock(&aio_nr_lock);
621 if (WARN_ON(aio_nr - nr > aio_nr))
625 spin_unlock(&aio_nr_lock);
629 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
631 static struct kioctx *ioctx_alloc(unsigned nr_events)
633 struct mm_struct *mm = current->mm;
638 * We keep track of the number of available ringbuffer slots, to prevent
639 * overflow (reqs_available), and we also use percpu counters for this.
641 * So since up to half the slots might be on other cpu's percpu counters
642 * and unavailable, double nr_events so userspace sees what they
643 * expected: additionally, we move req_batch slots to/from percpu
644 * counters at a time, so make sure that isn't 0:
646 nr_events = max(nr_events, num_possible_cpus() * 4);
649 /* Prevent overflows */
650 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
651 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
652 pr_debug("ENOMEM: nr_events too high\n");
653 return ERR_PTR(-EINVAL);
656 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
657 return ERR_PTR(-EAGAIN);
659 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
661 return ERR_PTR(-ENOMEM);
663 ctx->max_reqs = nr_events;
665 spin_lock_init(&ctx->ctx_lock);
666 spin_lock_init(&ctx->completion_lock);
667 mutex_init(&ctx->ring_lock);
668 /* Protect against page migration throughout kiotx setup by keeping
669 * the ring_lock mutex held until setup is complete. */
670 mutex_lock(&ctx->ring_lock);
671 init_waitqueue_head(&ctx->wait);
673 INIT_LIST_HEAD(&ctx->active_reqs);
675 if (percpu_ref_init(&ctx->users, free_ioctx_users))
678 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
681 ctx->cpu = alloc_percpu(struct kioctx_cpu);
685 err = aio_setup_ring(ctx);
689 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
690 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
691 if (ctx->req_batch < 1)
694 /* limit the number of system wide aios */
695 spin_lock(&aio_nr_lock);
696 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
697 aio_nr + nr_events < aio_nr) {
698 spin_unlock(&aio_nr_lock);
702 aio_nr += ctx->max_reqs;
703 spin_unlock(&aio_nr_lock);
705 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
706 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
708 err = ioctx_add_table(ctx, mm);
712 /* Release the ring_lock mutex now that all setup is complete. */
713 mutex_unlock(&ctx->ring_lock);
715 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
716 ctx, ctx->user_id, mm, ctx->nr_events);
720 aio_nr_sub(ctx->max_reqs);
724 mutex_unlock(&ctx->ring_lock);
725 free_percpu(ctx->cpu);
726 free_percpu(ctx->reqs.pcpu_count);
727 free_percpu(ctx->users.pcpu_count);
728 kmem_cache_free(kioctx_cachep, ctx);
729 pr_debug("error allocating ioctx %d\n", err);
734 * Cancels all outstanding aio requests on an aio context. Used
735 * when the processes owning a context have all exited to encourage
736 * the rapid destruction of the kioctx.
738 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
739 struct completion *requests_done)
741 if (!atomic_xchg(&ctx->dead, 1)) {
742 struct kioctx_table *table;
744 spin_lock(&mm->ioctx_lock);
746 table = rcu_dereference(mm->ioctx_table);
748 WARN_ON(ctx != table->table[ctx->id]);
749 table->table[ctx->id] = NULL;
751 spin_unlock(&mm->ioctx_lock);
753 /* percpu_ref_kill() will do the necessary call_rcu() */
754 wake_up_all(&ctx->wait);
757 * It'd be more correct to do this in free_ioctx(), after all
758 * the outstanding kiocbs have finished - but by then io_destroy
759 * has already returned, so io_setup() could potentially return
760 * -EAGAIN with no ioctxs actually in use (as far as userspace
763 aio_nr_sub(ctx->max_reqs);
766 vm_munmap(ctx->mmap_base, ctx->mmap_size);
768 ctx->requests_done = requests_done;
769 percpu_ref_kill(&ctx->users);
772 complete(requests_done);
776 /* wait_on_sync_kiocb:
777 * Waits on the given sync kiocb to complete.
779 ssize_t wait_on_sync_kiocb(struct kiocb *req)
781 while (!req->ki_ctx) {
782 set_current_state(TASK_UNINTERRUPTIBLE);
787 __set_current_state(TASK_RUNNING);
788 return req->ki_user_data;
790 EXPORT_SYMBOL(wait_on_sync_kiocb);
793 * exit_aio: called when the last user of mm goes away. At this point, there is
794 * no way for any new requests to be submited or any of the io_* syscalls to be
795 * called on the context.
797 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
800 void exit_aio(struct mm_struct *mm)
802 struct kioctx_table *table;
807 struct completion requests_done =
808 COMPLETION_INITIALIZER_ONSTACK(requests_done);
811 table = rcu_dereference(mm->ioctx_table);
814 if (!table || i >= table->nr) {
816 rcu_assign_pointer(mm->ioctx_table, NULL);
822 ctx = table->table[i++];
828 * We don't need to bother with munmap() here -
829 * exit_mmap(mm) is coming and it'll unmap everything.
830 * Since aio_free_ring() uses non-zero ->mmap_size
831 * as indicator that it needs to unmap the area,
832 * just set it to 0; aio_free_ring() is the only
833 * place that uses ->mmap_size, so it's safe.
837 kill_ioctx(mm, ctx, &requests_done);
839 /* Wait until all IO for the context are done. */
840 wait_for_completion(&requests_done);
844 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
846 struct kioctx_cpu *kcpu;
850 kcpu = this_cpu_ptr(ctx->cpu);
852 local_irq_save(flags);
853 kcpu->reqs_available += nr;
855 while (kcpu->reqs_available >= ctx->req_batch * 2) {
856 kcpu->reqs_available -= ctx->req_batch;
857 atomic_add(ctx->req_batch, &ctx->reqs_available);
860 local_irq_restore(flags);
864 static bool get_reqs_available(struct kioctx *ctx)
866 struct kioctx_cpu *kcpu;
871 kcpu = this_cpu_ptr(ctx->cpu);
873 local_irq_save(flags);
874 if (!kcpu->reqs_available) {
875 int old, avail = atomic_read(&ctx->reqs_available);
878 if (avail < ctx->req_batch)
882 avail = atomic_cmpxchg(&ctx->reqs_available,
883 avail, avail - ctx->req_batch);
884 } while (avail != old);
886 kcpu->reqs_available += ctx->req_batch;
890 kcpu->reqs_available--;
892 local_irq_restore(flags);
897 /* refill_reqs_available
898 * Updates the reqs_available reference counts used for tracking the
899 * number of free slots in the completion ring. This can be called
900 * from aio_complete() (to optimistically update reqs_available) or
901 * from aio_get_req() (the we're out of events case). It must be
902 * called holding ctx->completion_lock.
904 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
907 unsigned events_in_ring, completed;
909 /* Clamp head since userland can write to it. */
910 head %= ctx->nr_events;
912 events_in_ring = tail - head;
914 events_in_ring = ctx->nr_events - (head - tail);
916 completed = ctx->completed_events;
917 if (events_in_ring < completed)
918 completed -= events_in_ring;
925 ctx->completed_events -= completed;
926 put_reqs_available(ctx, completed);
929 /* user_refill_reqs_available
930 * Called to refill reqs_available when aio_get_req() encounters an
931 * out of space in the completion ring.
933 static void user_refill_reqs_available(struct kioctx *ctx)
935 spin_lock_irq(&ctx->completion_lock);
936 if (ctx->completed_events) {
937 struct aio_ring *ring;
940 /* Access of ring->head may race with aio_read_events_ring()
941 * here, but that's okay since whether we read the old version
942 * or the new version, and either will be valid. The important
943 * part is that head cannot pass tail since we prevent
944 * aio_complete() from updating tail by holding
945 * ctx->completion_lock. Even if head is invalid, the check
946 * against ctx->completed_events below will make sure we do the
949 ring = kmap_atomic(ctx->ring_pages[0]);
953 refill_reqs_available(ctx, head, ctx->tail);
956 spin_unlock_irq(&ctx->completion_lock);
960 * Allocate a slot for an aio request.
961 * Returns NULL if no requests are free.
963 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
967 if (!get_reqs_available(ctx)) {
968 user_refill_reqs_available(ctx);
969 if (!get_reqs_available(ctx))
973 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
977 percpu_ref_get(&ctx->reqs);
982 put_reqs_available(ctx, 1);
986 static void kiocb_free(struct kiocb *req)
990 if (req->ki_eventfd != NULL)
991 eventfd_ctx_put(req->ki_eventfd);
992 kmem_cache_free(kiocb_cachep, req);
995 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
997 struct aio_ring __user *ring = (void __user *)ctx_id;
998 struct mm_struct *mm = current->mm;
999 struct kioctx *ctx, *ret = NULL;
1000 struct kioctx_table *table;
1003 if (get_user(id, &ring->id))
1007 table = rcu_dereference(mm->ioctx_table);
1009 if (!table || id >= table->nr)
1012 ctx = table->table[id];
1013 if (ctx && ctx->user_id == ctx_id) {
1014 percpu_ref_get(&ctx->users);
1023 * Called when the io request on the given iocb is complete.
1025 void aio_complete(struct kiocb *iocb, long res, long res2)
1027 struct kioctx *ctx = iocb->ki_ctx;
1028 struct aio_ring *ring;
1029 struct io_event *ev_page, *event;
1030 unsigned tail, pos, head;
1031 unsigned long flags;
1034 * Special case handling for sync iocbs:
1035 * - events go directly into the iocb for fast handling
1036 * - the sync task with the iocb in its stack holds the single iocb
1037 * ref, no other paths have a way to get another ref
1038 * - the sync task helpfully left a reference to itself in the iocb
1040 if (is_sync_kiocb(iocb)) {
1041 iocb->ki_user_data = res;
1043 iocb->ki_ctx = ERR_PTR(-EXDEV);
1044 wake_up_process(iocb->ki_obj.tsk);
1048 if (iocb->ki_list.next) {
1049 unsigned long flags;
1051 spin_lock_irqsave(&ctx->ctx_lock, flags);
1052 list_del(&iocb->ki_list);
1053 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1057 * Add a completion event to the ring buffer. Must be done holding
1058 * ctx->completion_lock to prevent other code from messing with the tail
1059 * pointer since we might be called from irq context.
1061 spin_lock_irqsave(&ctx->completion_lock, flags);
1064 pos = tail + AIO_EVENTS_OFFSET;
1066 if (++tail >= ctx->nr_events)
1069 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1070 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1072 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
1073 event->data = iocb->ki_user_data;
1077 kunmap_atomic(ev_page);
1078 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1080 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1081 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
1084 /* after flagging the request as done, we
1085 * must never even look at it again
1087 smp_wmb(); /* make event visible before updating tail */
1091 ring = kmap_atomic(ctx->ring_pages[0]);
1094 kunmap_atomic(ring);
1095 flush_dcache_page(ctx->ring_pages[0]);
1097 ctx->completed_events++;
1098 if (ctx->completed_events > 1)
1099 refill_reqs_available(ctx, head, tail);
1100 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1102 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1105 * Check if the user asked us to deliver the result through an
1106 * eventfd. The eventfd_signal() function is safe to be called
1109 if (iocb->ki_eventfd != NULL)
1110 eventfd_signal(iocb->ki_eventfd, 1);
1112 /* everything turned out well, dispose of the aiocb. */
1116 * We have to order our ring_info tail store above and test
1117 * of the wait list below outside the wait lock. This is
1118 * like in wake_up_bit() where clearing a bit has to be
1119 * ordered with the unlocked test.
1123 if (waitqueue_active(&ctx->wait))
1124 wake_up(&ctx->wait);
1126 percpu_ref_put(&ctx->reqs);
1128 EXPORT_SYMBOL(aio_complete);
1131 * Pull an event off of the ioctx's event ring. Returns the number of
1134 static long aio_read_events_ring(struct kioctx *ctx,
1135 struct io_event __user *event, long nr)
1137 struct aio_ring *ring;
1138 unsigned head, tail, pos;
1142 mutex_lock(&ctx->ring_lock);
1144 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1145 ring = kmap_atomic(ctx->ring_pages[0]);
1148 kunmap_atomic(ring);
1151 * Ensure that once we've read the current tail pointer, that
1152 * we also see the events that were stored up to the tail.
1156 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1161 head %= ctx->nr_events;
1162 tail %= ctx->nr_events;
1166 struct io_event *ev;
1169 avail = (head <= tail ? tail : ctx->nr_events) - head;
1173 avail = min(avail, nr - ret);
1174 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1175 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1177 pos = head + AIO_EVENTS_OFFSET;
1178 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1179 pos %= AIO_EVENTS_PER_PAGE;
1182 copy_ret = copy_to_user(event + ret, ev + pos,
1183 sizeof(*ev) * avail);
1186 if (unlikely(copy_ret)) {
1193 head %= ctx->nr_events;
1196 ring = kmap_atomic(ctx->ring_pages[0]);
1198 kunmap_atomic(ring);
1199 flush_dcache_page(ctx->ring_pages[0]);
1201 pr_debug("%li h%u t%u\n", ret, head, tail);
1203 mutex_unlock(&ctx->ring_lock);
1208 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1209 struct io_event __user *event, long *i)
1211 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1216 if (unlikely(atomic_read(&ctx->dead)))
1222 return ret < 0 || *i >= min_nr;
1225 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1226 struct io_event __user *event,
1227 struct timespec __user *timeout)
1229 ktime_t until = { .tv64 = KTIME_MAX };
1235 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1238 until = timespec_to_ktime(ts);
1242 * Note that aio_read_events() is being called as the conditional - i.e.
1243 * we're calling it after prepare_to_wait() has set task state to
1244 * TASK_INTERRUPTIBLE.
1246 * But aio_read_events() can block, and if it blocks it's going to flip
1247 * the task state back to TASK_RUNNING.
1249 * This should be ok, provided it doesn't flip the state back to
1250 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1251 * will only happen if the mutex_lock() call blocks, and we then find
1252 * the ringbuffer empty. So in practice we should be ok, but it's
1253 * something to be aware of when touching this code.
1255 wait_event_interruptible_hrtimeout(ctx->wait,
1256 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1258 if (!ret && signal_pending(current))
1265 * Create an aio_context capable of receiving at least nr_events.
1266 * ctxp must not point to an aio_context that already exists, and
1267 * must be initialized to 0 prior to the call. On successful
1268 * creation of the aio_context, *ctxp is filled in with the resulting
1269 * handle. May fail with -EINVAL if *ctxp is not initialized,
1270 * if the specified nr_events exceeds internal limits. May fail
1271 * with -EAGAIN if the specified nr_events exceeds the user's limit
1272 * of available events. May fail with -ENOMEM if insufficient kernel
1273 * resources are available. May fail with -EFAULT if an invalid
1274 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1277 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1279 struct kioctx *ioctx = NULL;
1283 ret = get_user(ctx, ctxp);
1288 if (unlikely(ctx || nr_events == 0)) {
1289 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1294 ioctx = ioctx_alloc(nr_events);
1295 ret = PTR_ERR(ioctx);
1296 if (!IS_ERR(ioctx)) {
1297 ret = put_user(ioctx->user_id, ctxp);
1299 kill_ioctx(current->mm, ioctx, NULL);
1300 percpu_ref_put(&ioctx->users);
1308 * Destroy the aio_context specified. May cancel any outstanding
1309 * AIOs and block on completion. Will fail with -ENOSYS if not
1310 * implemented. May fail with -EINVAL if the context pointed to
1313 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1315 struct kioctx *ioctx = lookup_ioctx(ctx);
1316 if (likely(NULL != ioctx)) {
1317 struct completion requests_done =
1318 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1320 /* Pass requests_done to kill_ioctx() where it can be set
1321 * in a thread-safe way. If we try to set it here then we have
1322 * a race condition if two io_destroy() called simultaneously.
1324 kill_ioctx(current->mm, ioctx, &requests_done);
1325 percpu_ref_put(&ioctx->users);
1327 /* Wait until all IO for the context are done. Otherwise kernel
1328 * keep using user-space buffers even if user thinks the context
1331 wait_for_completion(&requests_done);
1335 pr_debug("EINVAL: io_destroy: invalid context id\n");
1339 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1340 unsigned long, loff_t);
1342 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1343 int rw, char __user *buf,
1344 unsigned long *nr_segs,
1345 struct iovec **iovec,
1350 *nr_segs = kiocb->ki_nbytes;
1352 #ifdef CONFIG_COMPAT
1354 ret = compat_rw_copy_check_uvector(rw,
1355 (struct compat_iovec __user *)buf,
1356 *nr_segs, 1, *iovec, iovec);
1359 ret = rw_copy_check_uvector(rw,
1360 (struct iovec __user *)buf,
1361 *nr_segs, 1, *iovec, iovec);
1365 /* ki_nbytes now reflect bytes instead of segs */
1366 kiocb->ki_nbytes = ret;
1370 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1371 int rw, char __user *buf,
1372 unsigned long *nr_segs,
1373 struct iovec *iovec)
1375 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1378 iovec->iov_base = buf;
1379 iovec->iov_len = kiocb->ki_nbytes;
1386 * Performs the initial checks and aio retry method
1387 * setup for the kiocb at the time of io submission.
1389 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1390 char __user *buf, bool compat)
1392 struct file *file = req->ki_filp;
1394 unsigned long nr_segs;
1398 struct iovec inline_vec, *iovec = &inline_vec;
1401 case IOCB_CMD_PREAD:
1402 case IOCB_CMD_PREADV:
1405 rw_op = file->f_op->aio_read;
1408 case IOCB_CMD_PWRITE:
1409 case IOCB_CMD_PWRITEV:
1412 rw_op = file->f_op->aio_write;
1415 if (unlikely(!(file->f_mode & mode)))
1421 ret = (opcode == IOCB_CMD_PREADV ||
1422 opcode == IOCB_CMD_PWRITEV)
1423 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1425 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1428 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1430 if (iovec != &inline_vec)
1435 req->ki_nbytes = ret;
1437 /* XXX: move/kill - rw_verify_area()? */
1438 /* This matches the pread()/pwrite() logic */
1439 if (req->ki_pos < 0) {
1445 file_start_write(file);
1447 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1450 file_end_write(file);
1453 case IOCB_CMD_FDSYNC:
1454 if (!file->f_op->aio_fsync)
1457 ret = file->f_op->aio_fsync(req, 1);
1460 case IOCB_CMD_FSYNC:
1461 if (!file->f_op->aio_fsync)
1464 ret = file->f_op->aio_fsync(req, 0);
1468 pr_debug("EINVAL: no operation provided\n");
1472 if (iovec != &inline_vec)
1475 if (ret != -EIOCBQUEUED) {
1477 * There's no easy way to restart the syscall since other AIO's
1478 * may be already running. Just fail this IO with EINTR.
1480 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1481 ret == -ERESTARTNOHAND ||
1482 ret == -ERESTART_RESTARTBLOCK))
1484 aio_complete(req, ret, 0);
1490 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1491 struct iocb *iocb, bool compat)
1496 /* enforce forwards compatibility on users */
1497 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1498 pr_debug("EINVAL: reserve field set\n");
1502 /* prevent overflows */
1504 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1505 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1506 ((ssize_t)iocb->aio_nbytes < 0)
1508 pr_debug("EINVAL: io_submit: overflow check\n");
1512 req = aio_get_req(ctx);
1516 req->ki_filp = fget(iocb->aio_fildes);
1517 if (unlikely(!req->ki_filp)) {
1522 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1524 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1525 * instance of the file* now. The file descriptor must be
1526 * an eventfd() fd, and will be signaled for each completed
1527 * event using the eventfd_signal() function.
1529 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1530 if (IS_ERR(req->ki_eventfd)) {
1531 ret = PTR_ERR(req->ki_eventfd);
1532 req->ki_eventfd = NULL;
1537 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1538 if (unlikely(ret)) {
1539 pr_debug("EFAULT: aio_key\n");
1543 req->ki_obj.user = user_iocb;
1544 req->ki_user_data = iocb->aio_data;
1545 req->ki_pos = iocb->aio_offset;
1546 req->ki_nbytes = iocb->aio_nbytes;
1548 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1549 (char __user *)(unsigned long)iocb->aio_buf,
1556 put_reqs_available(ctx, 1);
1557 percpu_ref_put(&ctx->reqs);
1562 long do_io_submit(aio_context_t ctx_id, long nr,
1563 struct iocb __user *__user *iocbpp, bool compat)
1568 struct blk_plug plug;
1570 if (unlikely(nr < 0))
1573 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1574 nr = LONG_MAX/sizeof(*iocbpp);
1576 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1579 ctx = lookup_ioctx(ctx_id);
1580 if (unlikely(!ctx)) {
1581 pr_debug("EINVAL: invalid context id\n");
1585 blk_start_plug(&plug);
1588 * AKPM: should this return a partial result if some of the IOs were
1589 * successfully submitted?
1591 for (i=0; i<nr; i++) {
1592 struct iocb __user *user_iocb;
1595 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1600 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1605 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1609 blk_finish_plug(&plug);
1611 percpu_ref_put(&ctx->users);
1616 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1617 * the number of iocbs queued. May return -EINVAL if the aio_context
1618 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1619 * *iocbpp[0] is not properly initialized, if the operation specified
1620 * is invalid for the file descriptor in the iocb. May fail with
1621 * -EFAULT if any of the data structures point to invalid data. May
1622 * fail with -EBADF if the file descriptor specified in the first
1623 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1624 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1625 * fail with -ENOSYS if not implemented.
1627 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1628 struct iocb __user * __user *, iocbpp)
1630 return do_io_submit(ctx_id, nr, iocbpp, 0);
1634 * Finds a given iocb for cancellation.
1636 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1639 struct list_head *pos;
1641 assert_spin_locked(&ctx->ctx_lock);
1643 if (key != KIOCB_KEY)
1646 /* TODO: use a hash or array, this sucks. */
1647 list_for_each(pos, &ctx->active_reqs) {
1648 struct kiocb *kiocb = list_kiocb(pos);
1649 if (kiocb->ki_obj.user == iocb)
1656 * Attempts to cancel an iocb previously passed to io_submit. If
1657 * the operation is successfully cancelled, the resulting event is
1658 * copied into the memory pointed to by result without being placed
1659 * into the completion queue and 0 is returned. May fail with
1660 * -EFAULT if any of the data structures pointed to are invalid.
1661 * May fail with -EINVAL if aio_context specified by ctx_id is
1662 * invalid. May fail with -EAGAIN if the iocb specified was not
1663 * cancelled. Will fail with -ENOSYS if not implemented.
1665 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1666 struct io_event __user *, result)
1669 struct kiocb *kiocb;
1673 ret = get_user(key, &iocb->aio_key);
1677 ctx = lookup_ioctx(ctx_id);
1681 spin_lock_irq(&ctx->ctx_lock);
1683 kiocb = lookup_kiocb(ctx, iocb, key);
1685 ret = kiocb_cancel(ctx, kiocb);
1689 spin_unlock_irq(&ctx->ctx_lock);
1693 * The result argument is no longer used - the io_event is
1694 * always delivered via the ring buffer. -EINPROGRESS indicates
1695 * cancellation is progress:
1700 percpu_ref_put(&ctx->users);
1706 * Attempts to read at least min_nr events and up to nr events from
1707 * the completion queue for the aio_context specified by ctx_id. If
1708 * it succeeds, the number of read events is returned. May fail with
1709 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1710 * out of range, if timeout is out of range. May fail with -EFAULT
1711 * if any of the memory specified is invalid. May return 0 or
1712 * < min_nr if the timeout specified by timeout has elapsed
1713 * before sufficient events are available, where timeout == NULL
1714 * specifies an infinite timeout. Note that the timeout pointed to by
1715 * timeout is relative. Will fail with -ENOSYS if not implemented.
1717 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1720 struct io_event __user *, events,
1721 struct timespec __user *, timeout)
1723 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1726 if (likely(ioctx)) {
1727 if (likely(min_nr <= nr && min_nr >= 0))
1728 ret = read_events(ioctx, min_nr, nr, events, timeout);
1729 percpu_ref_put(&ioctx->users);