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/nospec.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 __rcu *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 rcu_head free_rcu;
120 struct work_struct free_work; /* see free_ioctx() */
123 * signals when all in-flight requests are done
125 struct ctx_rq_wait *rq_wait;
129 * This counts the number of available slots in the ringbuffer,
130 * so we avoid overflowing it: it's decremented (if positive)
131 * when allocating a kiocb and incremented when the resulting
132 * io_event is pulled off the ringbuffer.
134 * We batch accesses to it with a percpu version.
136 atomic_t reqs_available;
137 } ____cacheline_aligned_in_smp;
141 struct list_head active_reqs; /* used for cancellation */
142 } ____cacheline_aligned_in_smp;
145 struct mutex ring_lock;
146 wait_queue_head_t wait;
147 } ____cacheline_aligned_in_smp;
151 unsigned completed_events;
152 spinlock_t completion_lock;
153 } ____cacheline_aligned_in_smp;
155 struct page *internal_pages[AIO_RING_PAGES];
156 struct file *aio_ring_file;
162 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
163 * cancelled or completed (this makes a certain amount of sense because
164 * successful cancellation - io_cancel() - does deliver the completion to
167 * And since most things don't implement kiocb cancellation and we'd really like
168 * kiocb completion to be lockless when possible, we use ki_cancel to
169 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
170 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
172 #define KIOCB_CANCELLED ((void *) (~0ULL))
177 struct kioctx *ki_ctx;
178 kiocb_cancel_fn *ki_cancel;
180 struct iocb __user *ki_user_iocb; /* user's aiocb */
181 __u64 ki_user_data; /* user's data for completion */
183 struct list_head ki_list; /* the aio core uses this
184 * for cancellation */
187 * If the aio_resfd field of the userspace iocb is not zero,
188 * this is the underlying eventfd context to deliver events to.
190 struct eventfd_ctx *ki_eventfd;
193 /*------ sysctl variables----*/
194 static DEFINE_SPINLOCK(aio_nr_lock);
195 unsigned long aio_nr; /* current system wide number of aio requests */
196 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
197 /*----end sysctl variables---*/
199 static struct kmem_cache *kiocb_cachep;
200 static struct kmem_cache *kioctx_cachep;
202 static struct vfsmount *aio_mnt;
204 static const struct file_operations aio_ring_fops;
205 static const struct address_space_operations aio_ctx_aops;
207 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
209 struct qstr this = QSTR_INIT("[aio]", 5);
212 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
214 return ERR_CAST(inode);
216 inode->i_mapping->a_ops = &aio_ctx_aops;
217 inode->i_mapping->private_data = ctx;
218 inode->i_size = PAGE_SIZE * nr_pages;
220 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
223 return ERR_PTR(-ENOMEM);
225 path.mnt = mntget(aio_mnt);
227 d_instantiate(path.dentry, inode);
228 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
234 file->f_flags = O_RDWR;
238 static struct dentry *aio_mount(struct file_system_type *fs_type,
239 int flags, const char *dev_name, void *data)
241 static const struct dentry_operations ops = {
242 .d_dname = simple_dname,
244 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, &ops,
248 root->d_sb->s_iflags |= SB_I_NOEXEC;
253 * Creates the slab caches used by the aio routines, panic on
254 * failure as this is done early during the boot sequence.
256 static int __init aio_setup(void)
258 static struct file_system_type aio_fs = {
261 .kill_sb = kill_anon_super,
263 aio_mnt = kern_mount(&aio_fs);
265 panic("Failed to create aio fs mount.");
266 aio_mnt->mnt_flags |= MNT_NOEXEC;
268 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
269 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
271 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
275 __initcall(aio_setup);
277 static void put_aio_ring_file(struct kioctx *ctx)
279 struct file *aio_ring_file = ctx->aio_ring_file;
281 truncate_setsize(aio_ring_file->f_inode, 0);
283 /* Prevent further access to the kioctx from migratepages */
284 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
285 aio_ring_file->f_inode->i_mapping->private_data = NULL;
286 ctx->aio_ring_file = NULL;
287 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
293 static void aio_free_ring(struct kioctx *ctx)
297 /* Disconnect the kiotx from the ring file. This prevents future
298 * accesses to the kioctx from page migration.
300 put_aio_ring_file(ctx);
302 for (i = 0; i < ctx->nr_pages; i++) {
304 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
305 page_count(ctx->ring_pages[i]));
306 page = ctx->ring_pages[i];
309 ctx->ring_pages[i] = NULL;
313 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
314 kfree(ctx->ring_pages);
315 ctx->ring_pages = NULL;
319 static int aio_ring_mremap(struct vm_area_struct *vma)
321 struct file *file = vma->vm_file;
322 struct mm_struct *mm = vma->vm_mm;
323 struct kioctx_table *table;
324 int i, res = -EINVAL;
326 spin_lock(&mm->ioctx_lock);
328 table = rcu_dereference(mm->ioctx_table);
329 for (i = 0; i < table->nr; i++) {
332 ctx = rcu_dereference(table->table[i]);
333 if (ctx && ctx->aio_ring_file == file) {
334 if (!atomic_read(&ctx->dead)) {
335 ctx->user_id = ctx->mmap_base = vma->vm_start;
343 spin_unlock(&mm->ioctx_lock);
347 static const struct vm_operations_struct aio_ring_vm_ops = {
348 .mremap = aio_ring_mremap,
349 #if IS_ENABLED(CONFIG_MMU)
350 .fault = filemap_fault,
351 .map_pages = filemap_map_pages,
352 .page_mkwrite = filemap_page_mkwrite,
356 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
358 vma->vm_flags |= VM_DONTEXPAND;
359 vma->vm_ops = &aio_ring_vm_ops;
363 static const struct file_operations aio_ring_fops = {
364 .mmap = aio_ring_mmap,
367 #if IS_ENABLED(CONFIG_MIGRATION)
368 static int aio_migratepage(struct address_space *mapping, struct page *new,
369 struct page *old, enum migrate_mode mode)
378 /* mapping->private_lock here protects against the kioctx teardown. */
379 spin_lock(&mapping->private_lock);
380 ctx = mapping->private_data;
386 /* The ring_lock mutex. The prevents aio_read_events() from writing
387 * to the ring's head, and prevents page migration from mucking in
388 * a partially initialized kiotx.
390 if (!mutex_trylock(&ctx->ring_lock)) {
396 if (idx < (pgoff_t)ctx->nr_pages) {
397 /* Make sure the old page hasn't already been changed */
398 if (ctx->ring_pages[idx] != old)
406 /* Writeback must be complete */
407 BUG_ON(PageWriteback(old));
410 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
411 if (rc != MIGRATEPAGE_SUCCESS) {
416 /* Take completion_lock to prevent other writes to the ring buffer
417 * while the old page is copied to the new. This prevents new
418 * events from being lost.
420 spin_lock_irqsave(&ctx->completion_lock, flags);
421 migrate_page_copy(new, old);
422 BUG_ON(ctx->ring_pages[idx] != old);
423 ctx->ring_pages[idx] = new;
424 spin_unlock_irqrestore(&ctx->completion_lock, flags);
426 /* The old page is no longer accessible. */
430 mutex_unlock(&ctx->ring_lock);
432 spin_unlock(&mapping->private_lock);
437 static const struct address_space_operations aio_ctx_aops = {
438 .set_page_dirty = __set_page_dirty_no_writeback,
439 #if IS_ENABLED(CONFIG_MIGRATION)
440 .migratepage = aio_migratepage,
444 static int aio_setup_ring(struct kioctx *ctx)
446 struct aio_ring *ring;
447 unsigned nr_events = ctx->max_reqs;
448 struct mm_struct *mm = current->mm;
449 unsigned long size, unused;
454 /* Compensate for the ring buffer's head/tail overlap entry */
455 nr_events += 2; /* 1 is required, 2 for good luck */
457 size = sizeof(struct aio_ring);
458 size += sizeof(struct io_event) * nr_events;
460 nr_pages = PFN_UP(size);
464 file = aio_private_file(ctx, nr_pages);
466 ctx->aio_ring_file = NULL;
470 ctx->aio_ring_file = file;
471 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
472 / sizeof(struct io_event);
474 ctx->ring_pages = ctx->internal_pages;
475 if (nr_pages > AIO_RING_PAGES) {
476 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
478 if (!ctx->ring_pages) {
479 put_aio_ring_file(ctx);
484 for (i = 0; i < nr_pages; i++) {
486 page = find_or_create_page(file->f_inode->i_mapping,
487 i, GFP_HIGHUSER | __GFP_ZERO);
490 pr_debug("pid(%d) page[%d]->count=%d\n",
491 current->pid, i, page_count(page));
492 SetPageUptodate(page);
495 ctx->ring_pages[i] = page;
499 if (unlikely(i != nr_pages)) {
504 ctx->mmap_size = nr_pages * PAGE_SIZE;
505 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
507 down_write(&mm->mmap_sem);
508 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
509 PROT_READ | PROT_WRITE,
510 MAP_SHARED, 0, &unused);
511 up_write(&mm->mmap_sem);
512 if (IS_ERR((void *)ctx->mmap_base)) {
518 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
520 ctx->user_id = ctx->mmap_base;
521 ctx->nr_events = nr_events; /* trusted copy */
523 ring = kmap_atomic(ctx->ring_pages[0]);
524 ring->nr = nr_events; /* user copy */
526 ring->head = ring->tail = 0;
527 ring->magic = AIO_RING_MAGIC;
528 ring->compat_features = AIO_RING_COMPAT_FEATURES;
529 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
530 ring->header_length = sizeof(struct aio_ring);
532 flush_dcache_page(ctx->ring_pages[0]);
537 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
538 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
539 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
541 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
543 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
544 struct kioctx *ctx = req->ki_ctx;
547 spin_lock_irqsave(&ctx->ctx_lock, flags);
549 if (!req->ki_list.next)
550 list_add(&req->ki_list, &ctx->active_reqs);
552 req->ki_cancel = cancel;
554 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
556 EXPORT_SYMBOL(kiocb_set_cancel_fn);
558 static int kiocb_cancel(struct aio_kiocb *kiocb)
560 kiocb_cancel_fn *old, *cancel;
563 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
564 * actually has a cancel function, hence the cmpxchg()
567 cancel = ACCESS_ONCE(kiocb->ki_cancel);
569 if (!cancel || cancel == KIOCB_CANCELLED)
573 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
574 } while (cancel != old);
576 return cancel(&kiocb->common);
580 * free_ioctx() should be RCU delayed to synchronize against the RCU
581 * protected lookup_ioctx() and also needs process context to call
582 * aio_free_ring(), so the double bouncing through kioctx->free_rcu and
585 static void free_ioctx(struct work_struct *work)
587 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
589 pr_debug("freeing %p\n", ctx);
592 free_percpu(ctx->cpu);
593 percpu_ref_exit(&ctx->reqs);
594 percpu_ref_exit(&ctx->users);
595 kmem_cache_free(kioctx_cachep, ctx);
598 static void free_ioctx_rcufn(struct rcu_head *head)
600 struct kioctx *ctx = container_of(head, struct kioctx, free_rcu);
602 INIT_WORK(&ctx->free_work, free_ioctx);
603 schedule_work(&ctx->free_work);
606 static void free_ioctx_reqs(struct percpu_ref *ref)
608 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
610 /* At this point we know that there are no any in-flight requests */
611 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
612 complete(&ctx->rq_wait->comp);
614 /* Synchronize against RCU protected table->table[] dereferences */
615 call_rcu(&ctx->free_rcu, free_ioctx_rcufn);
619 * When this function runs, the kioctx has been removed from the "hash table"
620 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
621 * now it's safe to cancel any that need to be.
623 static void free_ioctx_users(struct percpu_ref *ref)
625 struct kioctx *ctx = container_of(ref, struct kioctx, users);
626 struct aio_kiocb *req;
628 spin_lock_irq(&ctx->ctx_lock);
630 while (!list_empty(&ctx->active_reqs)) {
631 req = list_first_entry(&ctx->active_reqs,
632 struct aio_kiocb, ki_list);
634 list_del_init(&req->ki_list);
637 spin_unlock_irq(&ctx->ctx_lock);
639 percpu_ref_kill(&ctx->reqs);
640 percpu_ref_put(&ctx->reqs);
643 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
646 struct kioctx_table *table, *old;
647 struct aio_ring *ring;
649 spin_lock(&mm->ioctx_lock);
650 table = rcu_dereference_raw(mm->ioctx_table);
654 for (i = 0; i < table->nr; i++)
655 if (!rcu_access_pointer(table->table[i])) {
657 rcu_assign_pointer(table->table[i], ctx);
658 spin_unlock(&mm->ioctx_lock);
660 /* While kioctx setup is in progress,
661 * we are protected from page migration
662 * changes ring_pages by ->ring_lock.
664 ring = kmap_atomic(ctx->ring_pages[0]);
670 new_nr = (table ? table->nr : 1) * 4;
671 spin_unlock(&mm->ioctx_lock);
673 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
680 spin_lock(&mm->ioctx_lock);
681 old = rcu_dereference_raw(mm->ioctx_table);
684 rcu_assign_pointer(mm->ioctx_table, table);
685 } else if (table->nr > old->nr) {
686 memcpy(table->table, old->table,
687 old->nr * sizeof(struct kioctx *));
689 rcu_assign_pointer(mm->ioctx_table, table);
698 static void aio_nr_sub(unsigned nr)
700 spin_lock(&aio_nr_lock);
701 if (WARN_ON(aio_nr - nr > aio_nr))
705 spin_unlock(&aio_nr_lock);
709 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
711 static struct kioctx *ioctx_alloc(unsigned nr_events)
713 struct mm_struct *mm = current->mm;
718 * We keep track of the number of available ringbuffer slots, to prevent
719 * overflow (reqs_available), and we also use percpu counters for this.
721 * So since up to half the slots might be on other cpu's percpu counters
722 * and unavailable, double nr_events so userspace sees what they
723 * expected: additionally, we move req_batch slots to/from percpu
724 * counters at a time, so make sure that isn't 0:
726 nr_events = max(nr_events, num_possible_cpus() * 4);
729 /* Prevent overflows */
730 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
731 pr_debug("ENOMEM: nr_events too high\n");
732 return ERR_PTR(-EINVAL);
735 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
736 return ERR_PTR(-EAGAIN);
738 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
740 return ERR_PTR(-ENOMEM);
742 ctx->max_reqs = nr_events;
744 spin_lock_init(&ctx->ctx_lock);
745 spin_lock_init(&ctx->completion_lock);
746 mutex_init(&ctx->ring_lock);
747 /* Protect against page migration throughout kiotx setup by keeping
748 * the ring_lock mutex held until setup is complete. */
749 mutex_lock(&ctx->ring_lock);
750 init_waitqueue_head(&ctx->wait);
752 INIT_LIST_HEAD(&ctx->active_reqs);
754 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
757 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
760 ctx->cpu = alloc_percpu(struct kioctx_cpu);
764 err = aio_setup_ring(ctx);
768 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
769 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
770 if (ctx->req_batch < 1)
773 /* limit the number of system wide aios */
774 spin_lock(&aio_nr_lock);
775 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
776 aio_nr + nr_events < aio_nr) {
777 spin_unlock(&aio_nr_lock);
781 aio_nr += ctx->max_reqs;
782 spin_unlock(&aio_nr_lock);
784 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
785 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
787 err = ioctx_add_table(ctx, mm);
791 /* Release the ring_lock mutex now that all setup is complete. */
792 mutex_unlock(&ctx->ring_lock);
794 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
795 ctx, ctx->user_id, mm, ctx->nr_events);
799 aio_nr_sub(ctx->max_reqs);
801 atomic_set(&ctx->dead, 1);
803 vm_munmap(ctx->mmap_base, ctx->mmap_size);
806 mutex_unlock(&ctx->ring_lock);
807 free_percpu(ctx->cpu);
808 percpu_ref_exit(&ctx->reqs);
809 percpu_ref_exit(&ctx->users);
810 kmem_cache_free(kioctx_cachep, ctx);
811 pr_debug("error allocating ioctx %d\n", err);
816 * Cancels all outstanding aio requests on an aio context. Used
817 * when the processes owning a context have all exited to encourage
818 * the rapid destruction of the kioctx.
820 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
821 struct ctx_rq_wait *wait)
823 struct kioctx_table *table;
825 spin_lock(&mm->ioctx_lock);
826 if (atomic_xchg(&ctx->dead, 1)) {
827 spin_unlock(&mm->ioctx_lock);
831 table = rcu_dereference_raw(mm->ioctx_table);
832 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
833 RCU_INIT_POINTER(table->table[ctx->id], NULL);
834 spin_unlock(&mm->ioctx_lock);
836 /* free_ioctx_reqs() will do the necessary RCU synchronization */
837 wake_up_all(&ctx->wait);
840 * It'd be more correct to do this in free_ioctx(), after all
841 * the outstanding kiocbs have finished - but by then io_destroy
842 * has already returned, so io_setup() could potentially return
843 * -EAGAIN with no ioctxs actually in use (as far as userspace
846 aio_nr_sub(ctx->max_reqs);
849 vm_munmap(ctx->mmap_base, ctx->mmap_size);
852 percpu_ref_kill(&ctx->users);
857 * exit_aio: called when the last user of mm goes away. At this point, there is
858 * no way for any new requests to be submited or any of the io_* syscalls to be
859 * called on the context.
861 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
864 void exit_aio(struct mm_struct *mm)
866 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
867 struct ctx_rq_wait wait;
873 atomic_set(&wait.count, table->nr);
874 init_completion(&wait.comp);
877 for (i = 0; i < table->nr; ++i) {
879 rcu_dereference_protected(table->table[i], true);
887 * We don't need to bother with munmap() here - exit_mmap(mm)
888 * is coming and it'll unmap everything. And we simply can't,
889 * this is not necessarily our ->mm.
890 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
891 * that it needs to unmap the area, just set it to 0.
894 kill_ioctx(mm, ctx, &wait);
897 if (!atomic_sub_and_test(skipped, &wait.count)) {
898 /* Wait until all IO for the context are done. */
899 wait_for_completion(&wait.comp);
902 RCU_INIT_POINTER(mm->ioctx_table, NULL);
906 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
908 struct kioctx_cpu *kcpu;
911 local_irq_save(flags);
912 kcpu = this_cpu_ptr(ctx->cpu);
913 kcpu->reqs_available += nr;
915 while (kcpu->reqs_available >= ctx->req_batch * 2) {
916 kcpu->reqs_available -= ctx->req_batch;
917 atomic_add(ctx->req_batch, &ctx->reqs_available);
920 local_irq_restore(flags);
923 static bool get_reqs_available(struct kioctx *ctx)
925 struct kioctx_cpu *kcpu;
929 local_irq_save(flags);
930 kcpu = this_cpu_ptr(ctx->cpu);
931 if (!kcpu->reqs_available) {
932 int old, avail = atomic_read(&ctx->reqs_available);
935 if (avail < ctx->req_batch)
939 avail = atomic_cmpxchg(&ctx->reqs_available,
940 avail, avail - ctx->req_batch);
941 } while (avail != old);
943 kcpu->reqs_available += ctx->req_batch;
947 kcpu->reqs_available--;
949 local_irq_restore(flags);
953 /* refill_reqs_available
954 * Updates the reqs_available reference counts used for tracking the
955 * number of free slots in the completion ring. This can be called
956 * from aio_complete() (to optimistically update reqs_available) or
957 * from aio_get_req() (the we're out of events case). It must be
958 * called holding ctx->completion_lock.
960 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
963 unsigned events_in_ring, completed;
965 /* Clamp head since userland can write to it. */
966 head %= ctx->nr_events;
968 events_in_ring = tail - head;
970 events_in_ring = ctx->nr_events - (head - tail);
972 completed = ctx->completed_events;
973 if (events_in_ring < completed)
974 completed -= events_in_ring;
981 ctx->completed_events -= completed;
982 put_reqs_available(ctx, completed);
985 /* user_refill_reqs_available
986 * Called to refill reqs_available when aio_get_req() encounters an
987 * out of space in the completion ring.
989 static void user_refill_reqs_available(struct kioctx *ctx)
991 spin_lock_irq(&ctx->completion_lock);
992 if (ctx->completed_events) {
993 struct aio_ring *ring;
996 /* Access of ring->head may race with aio_read_events_ring()
997 * here, but that's okay since whether we read the old version
998 * or the new version, and either will be valid. The important
999 * part is that head cannot pass tail since we prevent
1000 * aio_complete() from updating tail by holding
1001 * ctx->completion_lock. Even if head is invalid, the check
1002 * against ctx->completed_events below will make sure we do the
1005 ring = kmap_atomic(ctx->ring_pages[0]);
1007 kunmap_atomic(ring);
1009 refill_reqs_available(ctx, head, ctx->tail);
1012 spin_unlock_irq(&ctx->completion_lock);
1016 * Allocate a slot for an aio request.
1017 * Returns NULL if no requests are free.
1019 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1021 struct aio_kiocb *req;
1023 if (!get_reqs_available(ctx)) {
1024 user_refill_reqs_available(ctx);
1025 if (!get_reqs_available(ctx))
1029 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1033 percpu_ref_get(&ctx->reqs);
1038 put_reqs_available(ctx, 1);
1042 static void kiocb_free(struct aio_kiocb *req)
1044 if (req->common.ki_filp)
1045 fput(req->common.ki_filp);
1046 if (req->ki_eventfd != NULL)
1047 eventfd_ctx_put(req->ki_eventfd);
1048 kmem_cache_free(kiocb_cachep, req);
1051 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1053 struct aio_ring __user *ring = (void __user *)ctx_id;
1054 struct mm_struct *mm = current->mm;
1055 struct kioctx *ctx, *ret = NULL;
1056 struct kioctx_table *table;
1059 if (get_user(id, &ring->id))
1063 table = rcu_dereference(mm->ioctx_table);
1065 if (!table || id >= table->nr)
1068 id = array_index_nospec(id, table->nr);
1069 ctx = rcu_dereference(table->table[id]);
1070 if (ctx && ctx->user_id == ctx_id) {
1071 if (percpu_ref_tryget_live(&ctx->users))
1080 * Called when the io request on the given iocb is complete.
1082 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1084 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1085 struct kioctx *ctx = iocb->ki_ctx;
1086 struct aio_ring *ring;
1087 struct io_event *ev_page, *event;
1088 unsigned tail, pos, head;
1089 unsigned long flags;
1092 * Special case handling for sync iocbs:
1093 * - events go directly into the iocb for fast handling
1094 * - the sync task with the iocb in its stack holds the single iocb
1095 * ref, no other paths have a way to get another ref
1096 * - the sync task helpfully left a reference to itself in the iocb
1098 BUG_ON(is_sync_kiocb(kiocb));
1100 if (iocb->ki_list.next) {
1101 unsigned long flags;
1103 spin_lock_irqsave(&ctx->ctx_lock, flags);
1104 list_del(&iocb->ki_list);
1105 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1109 * Add a completion event to the ring buffer. Must be done holding
1110 * ctx->completion_lock to prevent other code from messing with the tail
1111 * pointer since we might be called from irq context.
1113 spin_lock_irqsave(&ctx->completion_lock, flags);
1116 pos = tail + AIO_EVENTS_OFFSET;
1118 if (++tail >= ctx->nr_events)
1121 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1122 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1124 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1125 event->data = iocb->ki_user_data;
1129 kunmap_atomic(ev_page);
1130 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1132 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1133 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1136 /* after flagging the request as done, we
1137 * must never even look at it again
1139 smp_wmb(); /* make event visible before updating tail */
1143 ring = kmap_atomic(ctx->ring_pages[0]);
1146 kunmap_atomic(ring);
1147 flush_dcache_page(ctx->ring_pages[0]);
1149 ctx->completed_events++;
1150 if (ctx->completed_events > 1)
1151 refill_reqs_available(ctx, head, tail);
1152 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1154 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1157 * Check if the user asked us to deliver the result through an
1158 * eventfd. The eventfd_signal() function is safe to be called
1161 if (iocb->ki_eventfd != NULL)
1162 eventfd_signal(iocb->ki_eventfd, 1);
1164 /* everything turned out well, dispose of the aiocb. */
1168 * We have to order our ring_info tail store above and test
1169 * of the wait list below outside the wait lock. This is
1170 * like in wake_up_bit() where clearing a bit has to be
1171 * ordered with the unlocked test.
1175 if (waitqueue_active(&ctx->wait))
1176 wake_up(&ctx->wait);
1178 percpu_ref_put(&ctx->reqs);
1181 /* aio_read_events_ring
1182 * Pull an event off of the ioctx's event ring. Returns the number of
1185 static long aio_read_events_ring(struct kioctx *ctx,
1186 struct io_event __user *event, long nr)
1188 struct aio_ring *ring;
1189 unsigned head, tail, pos;
1194 * The mutex can block and wake us up and that will cause
1195 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1196 * and repeat. This should be rare enough that it doesn't cause
1197 * peformance issues. See the comment in read_events() for more detail.
1199 sched_annotate_sleep();
1200 mutex_lock(&ctx->ring_lock);
1202 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1203 ring = kmap_atomic(ctx->ring_pages[0]);
1206 kunmap_atomic(ring);
1209 * Ensure that once we've read the current tail pointer, that
1210 * we also see the events that were stored up to the tail.
1214 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1219 head %= ctx->nr_events;
1220 tail %= ctx->nr_events;
1224 struct io_event *ev;
1227 avail = (head <= tail ? tail : ctx->nr_events) - head;
1231 avail = min(avail, nr - ret);
1232 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1233 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1235 pos = head + AIO_EVENTS_OFFSET;
1236 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1237 pos %= AIO_EVENTS_PER_PAGE;
1240 copy_ret = copy_to_user(event + ret, ev + pos,
1241 sizeof(*ev) * avail);
1244 if (unlikely(copy_ret)) {
1251 head %= ctx->nr_events;
1254 ring = kmap_atomic(ctx->ring_pages[0]);
1256 kunmap_atomic(ring);
1257 flush_dcache_page(ctx->ring_pages[0]);
1259 pr_debug("%li h%u t%u\n", ret, head, tail);
1261 mutex_unlock(&ctx->ring_lock);
1266 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1267 struct io_event __user *event, long *i)
1269 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1274 if (unlikely(atomic_read(&ctx->dead)))
1280 return ret < 0 || *i >= min_nr;
1283 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1284 struct io_event __user *event,
1285 struct timespec __user *timeout)
1287 ktime_t until = { .tv64 = KTIME_MAX };
1293 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1296 until = timespec_to_ktime(ts);
1300 * Note that aio_read_events() is being called as the conditional - i.e.
1301 * we're calling it after prepare_to_wait() has set task state to
1302 * TASK_INTERRUPTIBLE.
1304 * But aio_read_events() can block, and if it blocks it's going to flip
1305 * the task state back to TASK_RUNNING.
1307 * This should be ok, provided it doesn't flip the state back to
1308 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1309 * will only happen if the mutex_lock() call blocks, and we then find
1310 * the ringbuffer empty. So in practice we should be ok, but it's
1311 * something to be aware of when touching this code.
1313 if (until.tv64 == 0)
1314 aio_read_events(ctx, min_nr, nr, event, &ret);
1316 wait_event_interruptible_hrtimeout(ctx->wait,
1317 aio_read_events(ctx, min_nr, nr, event, &ret),
1320 if (!ret && signal_pending(current))
1327 * Create an aio_context capable of receiving at least nr_events.
1328 * ctxp must not point to an aio_context that already exists, and
1329 * must be initialized to 0 prior to the call. On successful
1330 * creation of the aio_context, *ctxp is filled in with the resulting
1331 * handle. May fail with -EINVAL if *ctxp is not initialized,
1332 * if the specified nr_events exceeds internal limits. May fail
1333 * with -EAGAIN if the specified nr_events exceeds the user's limit
1334 * of available events. May fail with -ENOMEM if insufficient kernel
1335 * resources are available. May fail with -EFAULT if an invalid
1336 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1339 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1341 struct kioctx *ioctx = NULL;
1342 unsigned long ctx = 0;
1345 ret = get_user(ctx, ctxp);
1350 if (unlikely(ctx || nr_events == 0)) {
1351 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1356 ioctx = ioctx_alloc(nr_events);
1357 ret = PTR_ERR(ioctx);
1358 if (!IS_ERR(ioctx)) {
1359 ret = put_user(ioctx->user_id, ctxp);
1361 kill_ioctx(current->mm, ioctx, NULL);
1362 percpu_ref_put(&ioctx->users);
1370 * Destroy the aio_context specified. May cancel any outstanding
1371 * AIOs and block on completion. Will fail with -ENOSYS if not
1372 * implemented. May fail with -EINVAL if the context pointed to
1375 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1377 struct kioctx *ioctx = lookup_ioctx(ctx);
1378 if (likely(NULL != ioctx)) {
1379 struct ctx_rq_wait wait;
1382 init_completion(&wait.comp);
1383 atomic_set(&wait.count, 1);
1385 /* Pass requests_done to kill_ioctx() where it can be set
1386 * in a thread-safe way. If we try to set it here then we have
1387 * a race condition if two io_destroy() called simultaneously.
1389 ret = kill_ioctx(current->mm, ioctx, &wait);
1390 percpu_ref_put(&ioctx->users);
1392 /* Wait until all IO for the context are done. Otherwise kernel
1393 * keep using user-space buffers even if user thinks the context
1397 wait_for_completion(&wait.comp);
1401 pr_debug("EINVAL: invalid context id\n");
1405 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1407 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1408 struct iovec **iovec,
1410 struct iov_iter *iter)
1412 #ifdef CONFIG_COMPAT
1414 return compat_import_iovec(rw,
1415 (struct compat_iovec __user *)buf,
1416 len, UIO_FASTIOV, iovec, iter);
1418 return import_iovec(rw, (struct iovec __user *)buf,
1419 len, UIO_FASTIOV, iovec, iter);
1424 * Performs the initial checks and io submission.
1426 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1427 char __user *buf, size_t len, bool compat)
1429 struct file *file = req->ki_filp;
1433 rw_iter_op *iter_op;
1434 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1435 struct iov_iter iter;
1438 case IOCB_CMD_PREAD:
1439 case IOCB_CMD_PREADV:
1442 iter_op = file->f_op->read_iter;
1445 case IOCB_CMD_PWRITE:
1446 case IOCB_CMD_PWRITEV:
1449 iter_op = file->f_op->write_iter;
1452 if (unlikely(!(file->f_mode & mode)))
1458 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1459 ret = aio_setup_vectored_rw(rw, buf, len,
1460 &iovec, compat, &iter);
1462 ret = import_single_range(rw, buf, len, iovec, &iter);
1466 ret = rw_verify_area(rw, file, &req->ki_pos,
1467 iov_iter_count(&iter));
1477 file_start_write(file);
1479 ret = iter_op(req, &iter);
1482 file_end_write(file);
1487 case IOCB_CMD_FDSYNC:
1488 if (!file->f_op->aio_fsync)
1491 ret = file->f_op->aio_fsync(req, 1);
1494 case IOCB_CMD_FSYNC:
1495 if (!file->f_op->aio_fsync)
1498 ret = file->f_op->aio_fsync(req, 0);
1502 pr_debug("EINVAL: no operation provided\n");
1506 if (ret != -EIOCBQUEUED) {
1508 * There's no easy way to restart the syscall since other AIO's
1509 * may be already running. Just fail this IO with EINTR.
1511 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1512 ret == -ERESTARTNOHAND ||
1513 ret == -ERESTART_RESTARTBLOCK))
1515 aio_complete(req, ret, 0);
1521 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1522 struct iocb *iocb, bool compat)
1524 struct aio_kiocb *req;
1527 /* enforce forwards compatibility on users */
1528 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1529 pr_debug("EINVAL: reserve field set\n");
1533 /* prevent overflows */
1535 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1536 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1537 ((ssize_t)iocb->aio_nbytes < 0)
1539 pr_debug("EINVAL: overflow check\n");
1543 req = aio_get_req(ctx);
1547 req->common.ki_filp = fget(iocb->aio_fildes);
1548 if (unlikely(!req->common.ki_filp)) {
1552 req->common.ki_pos = iocb->aio_offset;
1553 req->common.ki_complete = aio_complete;
1554 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1556 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1558 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1559 * instance of the file* now. The file descriptor must be
1560 * an eventfd() fd, and will be signaled for each completed
1561 * event using the eventfd_signal() function.
1563 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1564 if (IS_ERR(req->ki_eventfd)) {
1565 ret = PTR_ERR(req->ki_eventfd);
1566 req->ki_eventfd = NULL;
1570 req->common.ki_flags |= IOCB_EVENTFD;
1573 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1574 if (unlikely(ret)) {
1575 pr_debug("EFAULT: aio_key\n");
1579 req->ki_user_iocb = user_iocb;
1580 req->ki_user_data = iocb->aio_data;
1582 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1583 (char __user *)(unsigned long)iocb->aio_buf,
1591 put_reqs_available(ctx, 1);
1592 percpu_ref_put(&ctx->reqs);
1597 long do_io_submit(aio_context_t ctx_id, long nr,
1598 struct iocb __user *__user *iocbpp, bool compat)
1603 struct blk_plug plug;
1605 if (unlikely(nr < 0))
1608 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1609 nr = LONG_MAX/sizeof(*iocbpp);
1611 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1614 ctx = lookup_ioctx(ctx_id);
1615 if (unlikely(!ctx)) {
1616 pr_debug("EINVAL: invalid context id\n");
1620 blk_start_plug(&plug);
1623 * AKPM: should this return a partial result if some of the IOs were
1624 * successfully submitted?
1626 for (i=0; i<nr; i++) {
1627 struct iocb __user *user_iocb;
1630 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1635 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1640 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1644 blk_finish_plug(&plug);
1646 percpu_ref_put(&ctx->users);
1651 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1652 * the number of iocbs queued. May return -EINVAL if the aio_context
1653 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1654 * *iocbpp[0] is not properly initialized, if the operation specified
1655 * is invalid for the file descriptor in the iocb. May fail with
1656 * -EFAULT if any of the data structures point to invalid data. May
1657 * fail with -EBADF if the file descriptor specified in the first
1658 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1659 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1660 * fail with -ENOSYS if not implemented.
1662 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1663 struct iocb __user * __user *, iocbpp)
1665 return do_io_submit(ctx_id, nr, iocbpp, 0);
1669 * Finds a given iocb for cancellation.
1671 static struct aio_kiocb *
1672 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1674 struct aio_kiocb *kiocb;
1676 assert_spin_locked(&ctx->ctx_lock);
1678 if (key != KIOCB_KEY)
1681 /* TODO: use a hash or array, this sucks. */
1682 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1683 if (kiocb->ki_user_iocb == iocb)
1690 * Attempts to cancel an iocb previously passed to io_submit. If
1691 * the operation is successfully cancelled, the resulting event is
1692 * copied into the memory pointed to by result without being placed
1693 * into the completion queue and 0 is returned. May fail with
1694 * -EFAULT if any of the data structures pointed to are invalid.
1695 * May fail with -EINVAL if aio_context specified by ctx_id is
1696 * invalid. May fail with -EAGAIN if the iocb specified was not
1697 * cancelled. Will fail with -ENOSYS if not implemented.
1699 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1700 struct io_event __user *, result)
1703 struct aio_kiocb *kiocb;
1707 ret = get_user(key, &iocb->aio_key);
1711 ctx = lookup_ioctx(ctx_id);
1715 spin_lock_irq(&ctx->ctx_lock);
1717 kiocb = lookup_kiocb(ctx, iocb, key);
1719 ret = kiocb_cancel(kiocb);
1723 spin_unlock_irq(&ctx->ctx_lock);
1727 * The result argument is no longer used - the io_event is
1728 * always delivered via the ring buffer. -EINPROGRESS indicates
1729 * cancellation is progress:
1734 percpu_ref_put(&ctx->users);
1740 * Attempts to read at least min_nr events and up to nr events from
1741 * the completion queue for the aio_context specified by ctx_id. If
1742 * it succeeds, the number of read events is returned. May fail with
1743 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1744 * out of range, if timeout is out of range. May fail with -EFAULT
1745 * if any of the memory specified is invalid. May return 0 or
1746 * < min_nr if the timeout specified by timeout has elapsed
1747 * before sufficient events are available, where timeout == NULL
1748 * specifies an infinite timeout. Note that the timeout pointed to by
1749 * timeout is relative. Will fail with -ENOSYS if not implemented.
1751 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1754 struct io_event __user *, events,
1755 struct timespec __user *, timeout)
1757 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1760 if (likely(ioctx)) {
1761 if (likely(min_nr <= nr && min_nr >= 0))
1762 ret = read_events(ioctx, min_nr, nr, events, timeout);
1763 percpu_ref_put(&ioctx->users);