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>
43 #include <linux/swork.h>
45 #include <asm/kmap_types.h>
46 #include <asm/uaccess.h>
50 #define AIO_RING_MAGIC 0xa10a10a1
51 #define AIO_RING_COMPAT_FEATURES 1
52 #define AIO_RING_INCOMPAT_FEATURES 0
54 unsigned id; /* kernel internal index number */
55 unsigned nr; /* number of io_events */
56 unsigned head; /* Written to by userland or under ring_lock
57 * mutex by aio_read_events_ring(). */
61 unsigned compat_features;
62 unsigned incompat_features;
63 unsigned header_length; /* size of aio_ring */
66 struct io_event io_events[0];
67 }; /* 128 bytes + ring size */
69 #define AIO_RING_PAGES 8
74 struct kioctx *table[];
78 unsigned reqs_available;
82 struct completion comp;
87 struct percpu_ref users;
90 struct percpu_ref reqs;
92 unsigned long user_id;
94 struct __percpu kioctx_cpu *cpu;
97 * For percpu reqs_available, number of slots we move to/from global
102 * This is what userspace passed to io_setup(), it's not used for
103 * anything but counting against the global max_reqs quota.
105 * The real limit is nr_events - 1, which will be larger (see
110 /* Size of ringbuffer, in units of struct io_event */
113 unsigned long mmap_base;
114 unsigned long mmap_size;
116 struct page **ring_pages;
119 struct swork_event free_work;
122 * signals when all in-flight requests are done
124 struct ctx_rq_wait *rq_wait;
128 * This counts the number of available slots in the ringbuffer,
129 * so we avoid overflowing it: it's decremented (if positive)
130 * when allocating a kiocb and incremented when the resulting
131 * io_event is pulled off the ringbuffer.
133 * We batch accesses to it with a percpu version.
135 atomic_t reqs_available;
136 } ____cacheline_aligned_in_smp;
140 struct list_head active_reqs; /* used for cancellation */
141 } ____cacheline_aligned_in_smp;
144 struct mutex ring_lock;
145 wait_queue_head_t wait;
146 } ____cacheline_aligned_in_smp;
150 unsigned completed_events;
151 spinlock_t completion_lock;
152 } ____cacheline_aligned_in_smp;
154 struct page *internal_pages[AIO_RING_PAGES];
155 struct file *aio_ring_file;
161 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
162 * cancelled or completed (this makes a certain amount of sense because
163 * successful cancellation - io_cancel() - does deliver the completion to
166 * And since most things don't implement kiocb cancellation and we'd really like
167 * kiocb completion to be lockless when possible, we use ki_cancel to
168 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
169 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
171 #define KIOCB_CANCELLED ((void *) (~0ULL))
176 struct kioctx *ki_ctx;
177 kiocb_cancel_fn *ki_cancel;
179 struct iocb __user *ki_user_iocb; /* user's aiocb */
180 __u64 ki_user_data; /* user's data for completion */
182 struct list_head ki_list; /* the aio core uses this
183 * for cancellation */
186 * If the aio_resfd field of the userspace iocb is not zero,
187 * this is the underlying eventfd context to deliver events to.
189 struct eventfd_ctx *ki_eventfd;
192 /*------ sysctl variables----*/
193 static DEFINE_SPINLOCK(aio_nr_lock);
194 unsigned long aio_nr; /* current system wide number of aio requests */
195 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
196 /*----end sysctl variables---*/
198 static struct kmem_cache *kiocb_cachep;
199 static struct kmem_cache *kioctx_cachep;
201 static struct vfsmount *aio_mnt;
203 static const struct file_operations aio_ring_fops;
204 static const struct address_space_operations aio_ctx_aops;
206 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
208 struct qstr this = QSTR_INIT("[aio]", 5);
211 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
213 return ERR_CAST(inode);
215 inode->i_mapping->a_ops = &aio_ctx_aops;
216 inode->i_mapping->private_data = ctx;
217 inode->i_size = PAGE_SIZE * nr_pages;
219 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
222 return ERR_PTR(-ENOMEM);
224 path.mnt = mntget(aio_mnt);
226 d_instantiate(path.dentry, inode);
227 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
233 file->f_flags = O_RDWR;
237 static struct dentry *aio_mount(struct file_system_type *fs_type,
238 int flags, const char *dev_name, void *data)
240 static const struct dentry_operations ops = {
241 .d_dname = simple_dname,
243 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
247 root->d_sb->s_iflags |= SB_I_NOEXEC;
252 * Creates the slab caches used by the aio routines, panic on
253 * failure as this is done early during the boot sequence.
255 static int __init aio_setup(void)
257 static struct file_system_type aio_fs = {
260 .kill_sb = kill_anon_super,
263 aio_mnt = kern_mount(&aio_fs);
265 panic("Failed to create aio fs mount.");
267 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
268 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
270 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
274 __initcall(aio_setup);
276 static void put_aio_ring_file(struct kioctx *ctx)
278 struct file *aio_ring_file = ctx->aio_ring_file;
280 truncate_setsize(aio_ring_file->f_inode, 0);
282 /* Prevent further access to the kioctx from migratepages */
283 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
284 aio_ring_file->f_inode->i_mapping->private_data = NULL;
285 ctx->aio_ring_file = NULL;
286 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
292 static void aio_free_ring(struct kioctx *ctx)
296 /* Disconnect the kiotx from the ring file. This prevents future
297 * accesses to the kioctx from page migration.
299 put_aio_ring_file(ctx);
301 for (i = 0; i < ctx->nr_pages; i++) {
303 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
304 page_count(ctx->ring_pages[i]));
305 page = ctx->ring_pages[i];
308 ctx->ring_pages[i] = NULL;
312 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
313 kfree(ctx->ring_pages);
314 ctx->ring_pages = NULL;
318 static int aio_ring_mremap(struct vm_area_struct *vma)
320 struct file *file = vma->vm_file;
321 struct mm_struct *mm = vma->vm_mm;
322 struct kioctx_table *table;
323 int i, res = -EINVAL;
325 spin_lock(&mm->ioctx_lock);
327 table = rcu_dereference(mm->ioctx_table);
328 for (i = 0; i < table->nr; i++) {
331 ctx = table->table[i];
332 if (ctx && ctx->aio_ring_file == file) {
333 if (!atomic_read(&ctx->dead)) {
334 ctx->user_id = ctx->mmap_base = vma->vm_start;
342 spin_unlock(&mm->ioctx_lock);
346 static const struct vm_operations_struct aio_ring_vm_ops = {
347 .mremap = aio_ring_mremap,
348 #if IS_ENABLED(CONFIG_MMU)
349 .fault = filemap_fault,
350 .map_pages = filemap_map_pages,
351 .page_mkwrite = filemap_page_mkwrite,
355 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
357 vma->vm_flags |= VM_DONTEXPAND;
358 vma->vm_ops = &aio_ring_vm_ops;
362 static const struct file_operations aio_ring_fops = {
363 .mmap = aio_ring_mmap,
366 #if IS_ENABLED(CONFIG_MIGRATION)
367 static int aio_migratepage(struct address_space *mapping, struct page *new,
368 struct page *old, enum migrate_mode mode)
377 /* mapping->private_lock here protects against the kioctx teardown. */
378 spin_lock(&mapping->private_lock);
379 ctx = mapping->private_data;
385 /* The ring_lock mutex. The prevents aio_read_events() from writing
386 * to the ring's head, and prevents page migration from mucking in
387 * a partially initialized kiotx.
389 if (!mutex_trylock(&ctx->ring_lock)) {
395 if (idx < (pgoff_t)ctx->nr_pages) {
396 /* Make sure the old page hasn't already been changed */
397 if (ctx->ring_pages[idx] != old)
405 /* Writeback must be complete */
406 BUG_ON(PageWriteback(old));
409 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
410 if (rc != MIGRATEPAGE_SUCCESS) {
415 /* Take completion_lock to prevent other writes to the ring buffer
416 * while the old page is copied to the new. This prevents new
417 * events from being lost.
419 spin_lock_irqsave(&ctx->completion_lock, flags);
420 migrate_page_copy(new, old);
421 BUG_ON(ctx->ring_pages[idx] != old);
422 ctx->ring_pages[idx] = new;
423 spin_unlock_irqrestore(&ctx->completion_lock, flags);
425 /* The old page is no longer accessible. */
429 mutex_unlock(&ctx->ring_lock);
431 spin_unlock(&mapping->private_lock);
436 static const struct address_space_operations aio_ctx_aops = {
437 .set_page_dirty = __set_page_dirty_no_writeback,
438 #if IS_ENABLED(CONFIG_MIGRATION)
439 .migratepage = aio_migratepage,
443 static int aio_setup_ring(struct kioctx *ctx)
445 struct aio_ring *ring;
446 unsigned nr_events = ctx->max_reqs;
447 struct mm_struct *mm = current->mm;
448 unsigned long size, unused;
453 /* Compensate for the ring buffer's head/tail overlap entry */
454 nr_events += 2; /* 1 is required, 2 for good luck */
456 size = sizeof(struct aio_ring);
457 size += sizeof(struct io_event) * nr_events;
459 nr_pages = PFN_UP(size);
463 file = aio_private_file(ctx, nr_pages);
465 ctx->aio_ring_file = NULL;
469 ctx->aio_ring_file = file;
470 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
471 / sizeof(struct io_event);
473 ctx->ring_pages = ctx->internal_pages;
474 if (nr_pages > AIO_RING_PAGES) {
475 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
477 if (!ctx->ring_pages) {
478 put_aio_ring_file(ctx);
483 for (i = 0; i < nr_pages; i++) {
485 page = find_or_create_page(file->f_inode->i_mapping,
486 i, GFP_HIGHUSER | __GFP_ZERO);
489 pr_debug("pid(%d) page[%d]->count=%d\n",
490 current->pid, i, page_count(page));
491 SetPageUptodate(page);
494 ctx->ring_pages[i] = page;
498 if (unlikely(i != nr_pages)) {
503 ctx->mmap_size = nr_pages * PAGE_SIZE;
504 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
506 down_write(&mm->mmap_sem);
507 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
508 PROT_READ | PROT_WRITE,
509 MAP_SHARED, 0, &unused);
510 up_write(&mm->mmap_sem);
511 if (IS_ERR((void *)ctx->mmap_base)) {
517 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
519 ctx->user_id = ctx->mmap_base;
520 ctx->nr_events = nr_events; /* trusted copy */
522 ring = kmap_atomic(ctx->ring_pages[0]);
523 ring->nr = nr_events; /* user copy */
525 ring->head = ring->tail = 0;
526 ring->magic = AIO_RING_MAGIC;
527 ring->compat_features = AIO_RING_COMPAT_FEATURES;
528 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
529 ring->header_length = sizeof(struct aio_ring);
531 flush_dcache_page(ctx->ring_pages[0]);
536 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
537 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
538 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
540 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
542 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
543 struct kioctx *ctx = req->ki_ctx;
546 spin_lock_irqsave(&ctx->ctx_lock, flags);
548 if (!req->ki_list.next)
549 list_add(&req->ki_list, &ctx->active_reqs);
551 req->ki_cancel = cancel;
553 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
555 EXPORT_SYMBOL(kiocb_set_cancel_fn);
557 static int kiocb_cancel(struct aio_kiocb *kiocb)
559 kiocb_cancel_fn *old, *cancel;
562 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
563 * actually has a cancel function, hence the cmpxchg()
566 cancel = ACCESS_ONCE(kiocb->ki_cancel);
568 if (!cancel || cancel == KIOCB_CANCELLED)
572 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
573 } while (cancel != old);
575 return cancel(&kiocb->common);
578 static void free_ioctx(struct swork_event *sev)
580 struct kioctx *ctx = container_of(sev, struct kioctx, free_work);
582 pr_debug("freeing %p\n", ctx);
585 free_percpu(ctx->cpu);
586 percpu_ref_exit(&ctx->reqs);
587 percpu_ref_exit(&ctx->users);
588 kmem_cache_free(kioctx_cachep, ctx);
591 static void free_ioctx_reqs(struct percpu_ref *ref)
593 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
595 /* At this point we know that there are no any in-flight requests */
596 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
597 complete(&ctx->rq_wait->comp);
599 INIT_SWORK(&ctx->free_work, free_ioctx);
600 swork_queue(&ctx->free_work);
604 * When this function runs, the kioctx has been removed from the "hash table"
605 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
606 * now it's safe to cancel any that need to be.
608 static void free_ioctx_users_work(struct swork_event *sev)
610 struct kioctx *ctx = container_of(sev, struct kioctx, free_work);
611 struct aio_kiocb *req;
613 spin_lock_irq(&ctx->ctx_lock);
615 while (!list_empty(&ctx->active_reqs)) {
616 req = list_first_entry(&ctx->active_reqs,
617 struct aio_kiocb, ki_list);
619 list_del_init(&req->ki_list);
623 spin_unlock_irq(&ctx->ctx_lock);
625 percpu_ref_kill(&ctx->reqs);
626 percpu_ref_put(&ctx->reqs);
629 static void free_ioctx_users(struct percpu_ref *ref)
631 struct kioctx *ctx = container_of(ref, struct kioctx, users);
633 INIT_SWORK(&ctx->free_work, free_ioctx_users_work);
634 swork_queue(&ctx->free_work);
637 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
640 struct kioctx_table *table, *old;
641 struct aio_ring *ring;
643 spin_lock(&mm->ioctx_lock);
644 table = rcu_dereference_raw(mm->ioctx_table);
648 for (i = 0; i < table->nr; i++)
649 if (!table->table[i]) {
651 table->table[i] = ctx;
652 spin_unlock(&mm->ioctx_lock);
654 /* While kioctx setup is in progress,
655 * we are protected from page migration
656 * changes ring_pages by ->ring_lock.
658 ring = kmap_atomic(ctx->ring_pages[0]);
664 new_nr = (table ? table->nr : 1) * 4;
665 spin_unlock(&mm->ioctx_lock);
667 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
674 spin_lock(&mm->ioctx_lock);
675 old = rcu_dereference_raw(mm->ioctx_table);
678 rcu_assign_pointer(mm->ioctx_table, table);
679 } else if (table->nr > old->nr) {
680 memcpy(table->table, old->table,
681 old->nr * sizeof(struct kioctx *));
683 rcu_assign_pointer(mm->ioctx_table, table);
692 static void aio_nr_sub(unsigned nr)
694 spin_lock(&aio_nr_lock);
695 if (WARN_ON(aio_nr - nr > aio_nr))
699 spin_unlock(&aio_nr_lock);
703 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
705 static struct kioctx *ioctx_alloc(unsigned nr_events)
707 struct mm_struct *mm = current->mm;
712 * We keep track of the number of available ringbuffer slots, to prevent
713 * overflow (reqs_available), and we also use percpu counters for this.
715 * So since up to half the slots might be on other cpu's percpu counters
716 * and unavailable, double nr_events so userspace sees what they
717 * expected: additionally, we move req_batch slots to/from percpu
718 * counters at a time, so make sure that isn't 0:
720 nr_events = max(nr_events, num_possible_cpus() * 4);
723 /* Prevent overflows */
724 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
725 pr_debug("ENOMEM: nr_events too high\n");
726 return ERR_PTR(-EINVAL);
729 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
730 return ERR_PTR(-EAGAIN);
732 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
734 return ERR_PTR(-ENOMEM);
736 ctx->max_reqs = nr_events;
738 spin_lock_init(&ctx->ctx_lock);
739 spin_lock_init(&ctx->completion_lock);
740 mutex_init(&ctx->ring_lock);
741 /* Protect against page migration throughout kiotx setup by keeping
742 * the ring_lock mutex held until setup is complete. */
743 mutex_lock(&ctx->ring_lock);
744 init_waitqueue_head(&ctx->wait);
746 INIT_LIST_HEAD(&ctx->active_reqs);
748 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754 ctx->cpu = alloc_percpu(struct kioctx_cpu);
758 err = aio_setup_ring(ctx);
762 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
763 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
764 if (ctx->req_batch < 1)
767 /* limit the number of system wide aios */
768 spin_lock(&aio_nr_lock);
769 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
770 aio_nr + nr_events < aio_nr) {
771 spin_unlock(&aio_nr_lock);
775 aio_nr += ctx->max_reqs;
776 spin_unlock(&aio_nr_lock);
778 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
779 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
781 err = ioctx_add_table(ctx, mm);
785 /* Release the ring_lock mutex now that all setup is complete. */
786 mutex_unlock(&ctx->ring_lock);
788 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
789 ctx, ctx->user_id, mm, ctx->nr_events);
793 aio_nr_sub(ctx->max_reqs);
795 atomic_set(&ctx->dead, 1);
797 vm_munmap(ctx->mmap_base, ctx->mmap_size);
800 mutex_unlock(&ctx->ring_lock);
801 free_percpu(ctx->cpu);
802 percpu_ref_exit(&ctx->reqs);
803 percpu_ref_exit(&ctx->users);
804 kmem_cache_free(kioctx_cachep, ctx);
805 pr_debug("error allocating ioctx %d\n", err);
810 * Cancels all outstanding aio requests on an aio context. Used
811 * when the processes owning a context have all exited to encourage
812 * the rapid destruction of the kioctx.
814 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
815 struct ctx_rq_wait *wait)
817 struct kioctx_table *table;
819 spin_lock(&mm->ioctx_lock);
820 if (atomic_xchg(&ctx->dead, 1)) {
821 spin_unlock(&mm->ioctx_lock);
825 table = rcu_dereference_raw(mm->ioctx_table);
826 WARN_ON(ctx != table->table[ctx->id]);
827 table->table[ctx->id] = NULL;
828 spin_unlock(&mm->ioctx_lock);
830 /* percpu_ref_kill() will do the necessary call_rcu() */
831 wake_up_all(&ctx->wait);
834 * It'd be more correct to do this in free_ioctx(), after all
835 * the outstanding kiocbs have finished - but by then io_destroy
836 * has already returned, so io_setup() could potentially return
837 * -EAGAIN with no ioctxs actually in use (as far as userspace
840 aio_nr_sub(ctx->max_reqs);
843 vm_munmap(ctx->mmap_base, ctx->mmap_size);
846 percpu_ref_kill(&ctx->users);
851 * exit_aio: called when the last user of mm goes away. At this point, there is
852 * no way for any new requests to be submited or any of the io_* syscalls to be
853 * called on the context.
855 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
858 void exit_aio(struct mm_struct *mm)
860 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
861 struct ctx_rq_wait wait;
867 atomic_set(&wait.count, table->nr);
868 init_completion(&wait.comp);
871 for (i = 0; i < table->nr; ++i) {
872 struct kioctx *ctx = table->table[i];
880 * We don't need to bother with munmap() here - exit_mmap(mm)
881 * is coming and it'll unmap everything. And we simply can't,
882 * this is not necessarily our ->mm.
883 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
884 * that it needs to unmap the area, just set it to 0.
887 kill_ioctx(mm, ctx, &wait);
890 if (!atomic_sub_and_test(skipped, &wait.count)) {
891 /* Wait until all IO for the context are done. */
892 wait_for_completion(&wait.comp);
895 RCU_INIT_POINTER(mm->ioctx_table, NULL);
899 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
901 struct kioctx_cpu *kcpu;
904 local_irq_save(flags);
905 kcpu = this_cpu_ptr(ctx->cpu);
906 kcpu->reqs_available += nr;
908 while (kcpu->reqs_available >= ctx->req_batch * 2) {
909 kcpu->reqs_available -= ctx->req_batch;
910 atomic_add(ctx->req_batch, &ctx->reqs_available);
913 local_irq_restore(flags);
916 static bool get_reqs_available(struct kioctx *ctx)
918 struct kioctx_cpu *kcpu;
922 local_irq_save(flags);
923 kcpu = this_cpu_ptr(ctx->cpu);
924 if (!kcpu->reqs_available) {
925 int old, avail = atomic_read(&ctx->reqs_available);
928 if (avail < ctx->req_batch)
932 avail = atomic_cmpxchg(&ctx->reqs_available,
933 avail, avail - ctx->req_batch);
934 } while (avail != old);
936 kcpu->reqs_available += ctx->req_batch;
940 kcpu->reqs_available--;
942 local_irq_restore(flags);
946 /* refill_reqs_available
947 * Updates the reqs_available reference counts used for tracking the
948 * number of free slots in the completion ring. This can be called
949 * from aio_complete() (to optimistically update reqs_available) or
950 * from aio_get_req() (the we're out of events case). It must be
951 * called holding ctx->completion_lock.
953 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
956 unsigned events_in_ring, completed;
958 /* Clamp head since userland can write to it. */
959 head %= ctx->nr_events;
961 events_in_ring = tail - head;
963 events_in_ring = ctx->nr_events - (head - tail);
965 completed = ctx->completed_events;
966 if (events_in_ring < completed)
967 completed -= events_in_ring;
974 ctx->completed_events -= completed;
975 put_reqs_available(ctx, completed);
978 /* user_refill_reqs_available
979 * Called to refill reqs_available when aio_get_req() encounters an
980 * out of space in the completion ring.
982 static void user_refill_reqs_available(struct kioctx *ctx)
984 spin_lock_irq(&ctx->completion_lock);
985 if (ctx->completed_events) {
986 struct aio_ring *ring;
989 /* Access of ring->head may race with aio_read_events_ring()
990 * here, but that's okay since whether we read the old version
991 * or the new version, and either will be valid. The important
992 * part is that head cannot pass tail since we prevent
993 * aio_complete() from updating tail by holding
994 * ctx->completion_lock. Even if head is invalid, the check
995 * against ctx->completed_events below will make sure we do the
998 ring = kmap_atomic(ctx->ring_pages[0]);
1000 kunmap_atomic(ring);
1002 refill_reqs_available(ctx, head, ctx->tail);
1005 spin_unlock_irq(&ctx->completion_lock);
1009 * Allocate a slot for an aio request.
1010 * Returns NULL if no requests are free.
1012 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1014 struct aio_kiocb *req;
1016 if (!get_reqs_available(ctx)) {
1017 user_refill_reqs_available(ctx);
1018 if (!get_reqs_available(ctx))
1022 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1026 percpu_ref_get(&ctx->reqs);
1031 put_reqs_available(ctx, 1);
1035 static void kiocb_free(struct aio_kiocb *req)
1037 if (req->common.ki_filp)
1038 fput(req->common.ki_filp);
1039 if (req->ki_eventfd != NULL)
1040 eventfd_ctx_put(req->ki_eventfd);
1041 kmem_cache_free(kiocb_cachep, req);
1044 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1046 struct aio_ring __user *ring = (void __user *)ctx_id;
1047 struct mm_struct *mm = current->mm;
1048 struct kioctx *ctx, *ret = NULL;
1049 struct kioctx_table *table;
1052 if (get_user(id, &ring->id))
1056 table = rcu_dereference(mm->ioctx_table);
1058 if (!table || id >= table->nr)
1061 ctx = table->table[id];
1062 if (ctx && ctx->user_id == ctx_id) {
1063 percpu_ref_get(&ctx->users);
1072 * Called when the io request on the given iocb is complete.
1074 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1076 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1077 struct kioctx *ctx = iocb->ki_ctx;
1078 struct aio_ring *ring;
1079 struct io_event *ev_page, *event;
1080 unsigned tail, pos, head;
1081 unsigned long flags;
1084 * Special case handling for sync iocbs:
1085 * - events go directly into the iocb for fast handling
1086 * - the sync task with the iocb in its stack holds the single iocb
1087 * ref, no other paths have a way to get another ref
1088 * - the sync task helpfully left a reference to itself in the iocb
1090 BUG_ON(is_sync_kiocb(kiocb));
1092 if (iocb->ki_list.next) {
1093 unsigned long flags;
1095 spin_lock_irqsave(&ctx->ctx_lock, flags);
1096 list_del(&iocb->ki_list);
1097 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1101 * Add a completion event to the ring buffer. Must be done holding
1102 * ctx->completion_lock to prevent other code from messing with the tail
1103 * pointer since we might be called from irq context.
1105 spin_lock_irqsave(&ctx->completion_lock, flags);
1108 pos = tail + AIO_EVENTS_OFFSET;
1110 if (++tail >= ctx->nr_events)
1113 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1114 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1116 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1117 event->data = iocb->ki_user_data;
1121 kunmap_atomic(ev_page);
1122 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1124 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1125 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1128 /* after flagging the request as done, we
1129 * must never even look at it again
1131 smp_wmb(); /* make event visible before updating tail */
1135 ring = kmap_atomic(ctx->ring_pages[0]);
1138 kunmap_atomic(ring);
1139 flush_dcache_page(ctx->ring_pages[0]);
1141 ctx->completed_events++;
1142 if (ctx->completed_events > 1)
1143 refill_reqs_available(ctx, head, tail);
1144 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1146 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1149 * Check if the user asked us to deliver the result through an
1150 * eventfd. The eventfd_signal() function is safe to be called
1153 if (iocb->ki_eventfd != NULL)
1154 eventfd_signal(iocb->ki_eventfd, 1);
1156 /* everything turned out well, dispose of the aiocb. */
1160 * We have to order our ring_info tail store above and test
1161 * of the wait list below outside the wait lock. This is
1162 * like in wake_up_bit() where clearing a bit has to be
1163 * ordered with the unlocked test.
1167 if (waitqueue_active(&ctx->wait))
1168 wake_up(&ctx->wait);
1170 percpu_ref_put(&ctx->reqs);
1173 /* aio_read_events_ring
1174 * Pull an event off of the ioctx's event ring. Returns the number of
1177 static long aio_read_events_ring(struct kioctx *ctx,
1178 struct io_event __user *event, long nr)
1180 struct aio_ring *ring;
1181 unsigned head, tail, pos;
1186 * The mutex can block and wake us up and that will cause
1187 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1188 * and repeat. This should be rare enough that it doesn't cause
1189 * peformance issues. See the comment in read_events() for more detail.
1191 sched_annotate_sleep();
1192 mutex_lock(&ctx->ring_lock);
1194 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1195 ring = kmap_atomic(ctx->ring_pages[0]);
1198 kunmap_atomic(ring);
1201 * Ensure that once we've read the current tail pointer, that
1202 * we also see the events that were stored up to the tail.
1206 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1211 head %= ctx->nr_events;
1212 tail %= ctx->nr_events;
1216 struct io_event *ev;
1219 avail = (head <= tail ? tail : ctx->nr_events) - head;
1223 avail = min(avail, nr - ret);
1224 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1225 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1227 pos = head + AIO_EVENTS_OFFSET;
1228 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1229 pos %= AIO_EVENTS_PER_PAGE;
1232 copy_ret = copy_to_user(event + ret, ev + pos,
1233 sizeof(*ev) * avail);
1236 if (unlikely(copy_ret)) {
1243 head %= ctx->nr_events;
1246 ring = kmap_atomic(ctx->ring_pages[0]);
1248 kunmap_atomic(ring);
1249 flush_dcache_page(ctx->ring_pages[0]);
1251 pr_debug("%li h%u t%u\n", ret, head, tail);
1253 mutex_unlock(&ctx->ring_lock);
1258 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1259 struct io_event __user *event, long *i)
1261 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1266 if (unlikely(atomic_read(&ctx->dead)))
1272 return ret < 0 || *i >= min_nr;
1275 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1276 struct io_event __user *event,
1277 struct timespec __user *timeout)
1279 ktime_t until = { .tv64 = KTIME_MAX };
1285 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1288 until = timespec_to_ktime(ts);
1292 * Note that aio_read_events() is being called as the conditional - i.e.
1293 * we're calling it after prepare_to_wait() has set task state to
1294 * TASK_INTERRUPTIBLE.
1296 * But aio_read_events() can block, and if it blocks it's going to flip
1297 * the task state back to TASK_RUNNING.
1299 * This should be ok, provided it doesn't flip the state back to
1300 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1301 * will only happen if the mutex_lock() call blocks, and we then find
1302 * the ringbuffer empty. So in practice we should be ok, but it's
1303 * something to be aware of when touching this code.
1305 if (until.tv64 == 0)
1306 aio_read_events(ctx, min_nr, nr, event, &ret);
1308 wait_event_interruptible_hrtimeout(ctx->wait,
1309 aio_read_events(ctx, min_nr, nr, event, &ret),
1312 if (!ret && signal_pending(current))
1319 * Create an aio_context capable of receiving at least nr_events.
1320 * ctxp must not point to an aio_context that already exists, and
1321 * must be initialized to 0 prior to the call. On successful
1322 * creation of the aio_context, *ctxp is filled in with the resulting
1323 * handle. May fail with -EINVAL if *ctxp is not initialized,
1324 * if the specified nr_events exceeds internal limits. May fail
1325 * with -EAGAIN if the specified nr_events exceeds the user's limit
1326 * of available events. May fail with -ENOMEM if insufficient kernel
1327 * resources are available. May fail with -EFAULT if an invalid
1328 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1331 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1333 struct kioctx *ioctx = NULL;
1337 ret = get_user(ctx, ctxp);
1342 if (unlikely(ctx || nr_events == 0)) {
1343 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1348 ioctx = ioctx_alloc(nr_events);
1349 ret = PTR_ERR(ioctx);
1350 if (!IS_ERR(ioctx)) {
1351 ret = put_user(ioctx->user_id, ctxp);
1353 kill_ioctx(current->mm, ioctx, NULL);
1354 percpu_ref_put(&ioctx->users);
1362 * Destroy the aio_context specified. May cancel any outstanding
1363 * AIOs and block on completion. Will fail with -ENOSYS if not
1364 * implemented. May fail with -EINVAL if the context pointed to
1367 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1369 struct kioctx *ioctx = lookup_ioctx(ctx);
1370 if (likely(NULL != ioctx)) {
1371 struct ctx_rq_wait wait;
1374 init_completion(&wait.comp);
1375 atomic_set(&wait.count, 1);
1377 /* Pass requests_done to kill_ioctx() where it can be set
1378 * in a thread-safe way. If we try to set it here then we have
1379 * a race condition if two io_destroy() called simultaneously.
1381 ret = kill_ioctx(current->mm, ioctx, &wait);
1382 percpu_ref_put(&ioctx->users);
1384 /* Wait until all IO for the context are done. Otherwise kernel
1385 * keep using user-space buffers even if user thinks the context
1389 wait_for_completion(&wait.comp);
1393 pr_debug("EINVAL: invalid context id\n");
1397 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1399 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1400 struct iovec **iovec,
1402 struct iov_iter *iter)
1404 #ifdef CONFIG_COMPAT
1406 return compat_import_iovec(rw,
1407 (struct compat_iovec __user *)buf,
1408 len, UIO_FASTIOV, iovec, iter);
1410 return import_iovec(rw, (struct iovec __user *)buf,
1411 len, UIO_FASTIOV, iovec, iter);
1416 * Performs the initial checks and io submission.
1418 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1419 char __user *buf, size_t len, bool compat)
1421 struct file *file = req->ki_filp;
1425 rw_iter_op *iter_op;
1426 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1427 struct iov_iter iter;
1430 case IOCB_CMD_PREAD:
1431 case IOCB_CMD_PREADV:
1434 iter_op = file->f_op->read_iter;
1437 case IOCB_CMD_PWRITE:
1438 case IOCB_CMD_PWRITEV:
1441 iter_op = file->f_op->write_iter;
1444 if (unlikely(!(file->f_mode & mode)))
1450 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1451 ret = aio_setup_vectored_rw(rw, buf, len,
1452 &iovec, compat, &iter);
1454 ret = import_single_range(rw, buf, len, iovec, &iter);
1458 ret = rw_verify_area(rw, file, &req->ki_pos,
1459 iov_iter_count(&iter));
1468 file_start_write(file);
1470 ret = iter_op(req, &iter);
1473 file_end_write(file);
1477 case IOCB_CMD_FDSYNC:
1478 if (!file->f_op->aio_fsync)
1481 ret = file->f_op->aio_fsync(req, 1);
1484 case IOCB_CMD_FSYNC:
1485 if (!file->f_op->aio_fsync)
1488 ret = file->f_op->aio_fsync(req, 0);
1492 pr_debug("EINVAL: no operation provided\n");
1496 if (ret != -EIOCBQUEUED) {
1498 * There's no easy way to restart the syscall since other AIO's
1499 * may be already running. Just fail this IO with EINTR.
1501 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1502 ret == -ERESTARTNOHAND ||
1503 ret == -ERESTART_RESTARTBLOCK))
1505 aio_complete(req, ret, 0);
1511 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1512 struct iocb *iocb, bool compat)
1514 struct aio_kiocb *req;
1517 /* enforce forwards compatibility on users */
1518 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1519 pr_debug("EINVAL: reserve field set\n");
1523 /* prevent overflows */
1525 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1526 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1527 ((ssize_t)iocb->aio_nbytes < 0)
1529 pr_debug("EINVAL: overflow check\n");
1533 req = aio_get_req(ctx);
1537 req->common.ki_filp = fget(iocb->aio_fildes);
1538 if (unlikely(!req->common.ki_filp)) {
1542 req->common.ki_pos = iocb->aio_offset;
1543 req->common.ki_complete = aio_complete;
1544 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1546 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1548 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1549 * instance of the file* now. The file descriptor must be
1550 * an eventfd() fd, and will be signaled for each completed
1551 * event using the eventfd_signal() function.
1553 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1554 if (IS_ERR(req->ki_eventfd)) {
1555 ret = PTR_ERR(req->ki_eventfd);
1556 req->ki_eventfd = NULL;
1560 req->common.ki_flags |= IOCB_EVENTFD;
1563 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1564 if (unlikely(ret)) {
1565 pr_debug("EFAULT: aio_key\n");
1569 req->ki_user_iocb = user_iocb;
1570 req->ki_user_data = iocb->aio_data;
1572 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1573 (char __user *)(unsigned long)iocb->aio_buf,
1581 put_reqs_available(ctx, 1);
1582 percpu_ref_put(&ctx->reqs);
1587 long do_io_submit(aio_context_t ctx_id, long nr,
1588 struct iocb __user *__user *iocbpp, bool compat)
1593 struct blk_plug plug;
1595 if (unlikely(nr < 0))
1598 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1599 nr = LONG_MAX/sizeof(*iocbpp);
1601 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1604 ctx = lookup_ioctx(ctx_id);
1605 if (unlikely(!ctx)) {
1606 pr_debug("EINVAL: invalid context id\n");
1610 blk_start_plug(&plug);
1613 * AKPM: should this return a partial result if some of the IOs were
1614 * successfully submitted?
1616 for (i=0; i<nr; i++) {
1617 struct iocb __user *user_iocb;
1620 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1625 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1630 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1634 blk_finish_plug(&plug);
1636 percpu_ref_put(&ctx->users);
1641 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1642 * the number of iocbs queued. May return -EINVAL if the aio_context
1643 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1644 * *iocbpp[0] is not properly initialized, if the operation specified
1645 * is invalid for the file descriptor in the iocb. May fail with
1646 * -EFAULT if any of the data structures point to invalid data. May
1647 * fail with -EBADF if the file descriptor specified in the first
1648 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1649 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1650 * fail with -ENOSYS if not implemented.
1652 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1653 struct iocb __user * __user *, iocbpp)
1655 return do_io_submit(ctx_id, nr, iocbpp, 0);
1659 * Finds a given iocb for cancellation.
1661 static struct aio_kiocb *
1662 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1664 struct aio_kiocb *kiocb;
1666 assert_spin_locked(&ctx->ctx_lock);
1668 if (key != KIOCB_KEY)
1671 /* TODO: use a hash or array, this sucks. */
1672 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1673 if (kiocb->ki_user_iocb == iocb)
1680 * Attempts to cancel an iocb previously passed to io_submit. If
1681 * the operation is successfully cancelled, the resulting event is
1682 * copied into the memory pointed to by result without being placed
1683 * into the completion queue and 0 is returned. May fail with
1684 * -EFAULT if any of the data structures pointed to are invalid.
1685 * May fail with -EINVAL if aio_context specified by ctx_id is
1686 * invalid. May fail with -EAGAIN if the iocb specified was not
1687 * cancelled. Will fail with -ENOSYS if not implemented.
1689 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1690 struct io_event __user *, result)
1693 struct aio_kiocb *kiocb;
1697 ret = get_user(key, &iocb->aio_key);
1701 ctx = lookup_ioctx(ctx_id);
1705 spin_lock_irq(&ctx->ctx_lock);
1707 kiocb = lookup_kiocb(ctx, iocb, key);
1709 ret = kiocb_cancel(kiocb);
1713 spin_unlock_irq(&ctx->ctx_lock);
1717 * The result argument is no longer used - the io_event is
1718 * always delivered via the ring buffer. -EINPROGRESS indicates
1719 * cancellation is progress:
1724 percpu_ref_put(&ctx->users);
1730 * Attempts to read at least min_nr events and up to nr events from
1731 * the completion queue for the aio_context specified by ctx_id. If
1732 * it succeeds, the number of read events is returned. May fail with
1733 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1734 * out of range, if timeout is out of range. May fail with -EFAULT
1735 * if any of the memory specified is invalid. May return 0 or
1736 * < min_nr if the timeout specified by timeout has elapsed
1737 * before sufficient events are available, where timeout == NULL
1738 * specifies an infinite timeout. Note that the timeout pointed to by
1739 * timeout is relative. Will fail with -ENOSYS if not implemented.
1741 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1744 struct io_event __user *, events,
1745 struct timespec __user *, timeout)
1747 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1750 if (likely(ioctx)) {
1751 if (likely(min_nr <= nr && min_nr >= 0))
1752 ret = read_events(ioctx, min_nr, nr, events, timeout);
1753 percpu_ref_put(&ioctx->users);