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.");
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 = rcu_dereference(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);
579 * free_ioctx() should be RCU delayed to synchronize against the RCU
580 * protected lookup_ioctx() and also needs process context to call
581 * aio_free_ring(), so the double bouncing through kioctx->free_rcu and
584 static void free_ioctx(struct work_struct *work)
586 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
588 pr_debug("freeing %p\n", ctx);
591 free_percpu(ctx->cpu);
592 percpu_ref_exit(&ctx->reqs);
593 percpu_ref_exit(&ctx->users);
594 kmem_cache_free(kioctx_cachep, ctx);
597 static void free_ioctx_rcufn(struct rcu_head *head)
599 struct kioctx *ctx = container_of(head, struct kioctx, free_rcu);
601 INIT_WORK(&ctx->free_work, free_ioctx);
602 schedule_work(&ctx->free_work);
605 static void free_ioctx_reqs(struct percpu_ref *ref)
607 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
609 /* At this point we know that there are no any in-flight requests */
610 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
611 complete(&ctx->rq_wait->comp);
613 /* Synchronize against RCU protected table->table[] dereferences */
614 call_rcu(&ctx->free_rcu, free_ioctx_rcufn);
618 * When this function runs, the kioctx has been removed from the "hash table"
619 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
620 * now it's safe to cancel any that need to be.
622 static void free_ioctx_users(struct percpu_ref *ref)
624 struct kioctx *ctx = container_of(ref, struct kioctx, users);
625 struct aio_kiocb *req;
627 spin_lock_irq(&ctx->ctx_lock);
629 while (!list_empty(&ctx->active_reqs)) {
630 req = list_first_entry(&ctx->active_reqs,
631 struct aio_kiocb, ki_list);
633 list_del_init(&req->ki_list);
636 spin_unlock_irq(&ctx->ctx_lock);
638 percpu_ref_kill(&ctx->reqs);
639 percpu_ref_put(&ctx->reqs);
642 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
645 struct kioctx_table *table, *old;
646 struct aio_ring *ring;
648 spin_lock(&mm->ioctx_lock);
649 table = rcu_dereference_raw(mm->ioctx_table);
653 for (i = 0; i < table->nr; i++)
654 if (!rcu_access_pointer(table->table[i])) {
656 rcu_assign_pointer(table->table[i], ctx);
657 spin_unlock(&mm->ioctx_lock);
659 /* While kioctx setup is in progress,
660 * we are protected from page migration
661 * changes ring_pages by ->ring_lock.
663 ring = kmap_atomic(ctx->ring_pages[0]);
669 new_nr = (table ? table->nr : 1) * 4;
670 spin_unlock(&mm->ioctx_lock);
672 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
679 spin_lock(&mm->ioctx_lock);
680 old = rcu_dereference_raw(mm->ioctx_table);
683 rcu_assign_pointer(mm->ioctx_table, table);
684 } else if (table->nr > old->nr) {
685 memcpy(table->table, old->table,
686 old->nr * sizeof(struct kioctx *));
688 rcu_assign_pointer(mm->ioctx_table, table);
697 static void aio_nr_sub(unsigned nr)
699 spin_lock(&aio_nr_lock);
700 if (WARN_ON(aio_nr - nr > aio_nr))
704 spin_unlock(&aio_nr_lock);
708 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
710 static struct kioctx *ioctx_alloc(unsigned nr_events)
712 struct mm_struct *mm = current->mm;
717 * We keep track of the number of available ringbuffer slots, to prevent
718 * overflow (reqs_available), and we also use percpu counters for this.
720 * So since up to half the slots might be on other cpu's percpu counters
721 * and unavailable, double nr_events so userspace sees what they
722 * expected: additionally, we move req_batch slots to/from percpu
723 * counters at a time, so make sure that isn't 0:
725 nr_events = max(nr_events, num_possible_cpus() * 4);
728 /* Prevent overflows */
729 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
730 pr_debug("ENOMEM: nr_events too high\n");
731 return ERR_PTR(-EINVAL);
734 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
735 return ERR_PTR(-EAGAIN);
737 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
739 return ERR_PTR(-ENOMEM);
741 ctx->max_reqs = nr_events;
743 spin_lock_init(&ctx->ctx_lock);
744 spin_lock_init(&ctx->completion_lock);
745 mutex_init(&ctx->ring_lock);
746 /* Protect against page migration throughout kiotx setup by keeping
747 * the ring_lock mutex held until setup is complete. */
748 mutex_lock(&ctx->ring_lock);
749 init_waitqueue_head(&ctx->wait);
751 INIT_LIST_HEAD(&ctx->active_reqs);
753 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
756 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
759 ctx->cpu = alloc_percpu(struct kioctx_cpu);
763 err = aio_setup_ring(ctx);
767 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
768 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
769 if (ctx->req_batch < 1)
772 /* limit the number of system wide aios */
773 spin_lock(&aio_nr_lock);
774 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
775 aio_nr + nr_events < aio_nr) {
776 spin_unlock(&aio_nr_lock);
780 aio_nr += ctx->max_reqs;
781 spin_unlock(&aio_nr_lock);
783 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
784 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
786 err = ioctx_add_table(ctx, mm);
790 /* Release the ring_lock mutex now that all setup is complete. */
791 mutex_unlock(&ctx->ring_lock);
793 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
794 ctx, ctx->user_id, mm, ctx->nr_events);
798 aio_nr_sub(ctx->max_reqs);
800 atomic_set(&ctx->dead, 1);
802 vm_munmap(ctx->mmap_base, ctx->mmap_size);
805 mutex_unlock(&ctx->ring_lock);
806 free_percpu(ctx->cpu);
807 percpu_ref_exit(&ctx->reqs);
808 percpu_ref_exit(&ctx->users);
809 kmem_cache_free(kioctx_cachep, ctx);
810 pr_debug("error allocating ioctx %d\n", err);
815 * Cancels all outstanding aio requests on an aio context. Used
816 * when the processes owning a context have all exited to encourage
817 * the rapid destruction of the kioctx.
819 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
820 struct ctx_rq_wait *wait)
822 struct kioctx_table *table;
824 spin_lock(&mm->ioctx_lock);
825 if (atomic_xchg(&ctx->dead, 1)) {
826 spin_unlock(&mm->ioctx_lock);
830 table = rcu_dereference_raw(mm->ioctx_table);
831 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
832 RCU_INIT_POINTER(table->table[ctx->id], NULL);
833 spin_unlock(&mm->ioctx_lock);
835 /* free_ioctx_reqs() will do the necessary RCU synchronization */
836 wake_up_all(&ctx->wait);
839 * It'd be more correct to do this in free_ioctx(), after all
840 * the outstanding kiocbs have finished - but by then io_destroy
841 * has already returned, so io_setup() could potentially return
842 * -EAGAIN with no ioctxs actually in use (as far as userspace
845 aio_nr_sub(ctx->max_reqs);
848 vm_munmap(ctx->mmap_base, ctx->mmap_size);
851 percpu_ref_kill(&ctx->users);
856 * exit_aio: called when the last user of mm goes away. At this point, there is
857 * no way for any new requests to be submited or any of the io_* syscalls to be
858 * called on the context.
860 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
863 void exit_aio(struct mm_struct *mm)
865 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
866 struct ctx_rq_wait wait;
872 atomic_set(&wait.count, table->nr);
873 init_completion(&wait.comp);
876 for (i = 0; i < table->nr; ++i) {
878 rcu_dereference_protected(table->table[i], true);
886 * We don't need to bother with munmap() here - exit_mmap(mm)
887 * is coming and it'll unmap everything. And we simply can't,
888 * this is not necessarily our ->mm.
889 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
890 * that it needs to unmap the area, just set it to 0.
893 kill_ioctx(mm, ctx, &wait);
896 if (!atomic_sub_and_test(skipped, &wait.count)) {
897 /* Wait until all IO for the context are done. */
898 wait_for_completion(&wait.comp);
901 RCU_INIT_POINTER(mm->ioctx_table, NULL);
905 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
907 struct kioctx_cpu *kcpu;
910 local_irq_save(flags);
911 kcpu = this_cpu_ptr(ctx->cpu);
912 kcpu->reqs_available += nr;
914 while (kcpu->reqs_available >= ctx->req_batch * 2) {
915 kcpu->reqs_available -= ctx->req_batch;
916 atomic_add(ctx->req_batch, &ctx->reqs_available);
919 local_irq_restore(flags);
922 static bool get_reqs_available(struct kioctx *ctx)
924 struct kioctx_cpu *kcpu;
928 local_irq_save(flags);
929 kcpu = this_cpu_ptr(ctx->cpu);
930 if (!kcpu->reqs_available) {
931 int old, avail = atomic_read(&ctx->reqs_available);
934 if (avail < ctx->req_batch)
938 avail = atomic_cmpxchg(&ctx->reqs_available,
939 avail, avail - ctx->req_batch);
940 } while (avail != old);
942 kcpu->reqs_available += ctx->req_batch;
946 kcpu->reqs_available--;
948 local_irq_restore(flags);
952 /* refill_reqs_available
953 * Updates the reqs_available reference counts used for tracking the
954 * number of free slots in the completion ring. This can be called
955 * from aio_complete() (to optimistically update reqs_available) or
956 * from aio_get_req() (the we're out of events case). It must be
957 * called holding ctx->completion_lock.
959 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
962 unsigned events_in_ring, completed;
964 /* Clamp head since userland can write to it. */
965 head %= ctx->nr_events;
967 events_in_ring = tail - head;
969 events_in_ring = ctx->nr_events - (head - tail);
971 completed = ctx->completed_events;
972 if (events_in_ring < completed)
973 completed -= events_in_ring;
980 ctx->completed_events -= completed;
981 put_reqs_available(ctx, completed);
984 /* user_refill_reqs_available
985 * Called to refill reqs_available when aio_get_req() encounters an
986 * out of space in the completion ring.
988 static void user_refill_reqs_available(struct kioctx *ctx)
990 spin_lock_irq(&ctx->completion_lock);
991 if (ctx->completed_events) {
992 struct aio_ring *ring;
995 /* Access of ring->head may race with aio_read_events_ring()
996 * here, but that's okay since whether we read the old version
997 * or the new version, and either will be valid. The important
998 * part is that head cannot pass tail since we prevent
999 * aio_complete() from updating tail by holding
1000 * ctx->completion_lock. Even if head is invalid, the check
1001 * against ctx->completed_events below will make sure we do the
1004 ring = kmap_atomic(ctx->ring_pages[0]);
1006 kunmap_atomic(ring);
1008 refill_reqs_available(ctx, head, ctx->tail);
1011 spin_unlock_irq(&ctx->completion_lock);
1015 * Allocate a slot for an aio request.
1016 * Returns NULL if no requests are free.
1018 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1020 struct aio_kiocb *req;
1022 if (!get_reqs_available(ctx)) {
1023 user_refill_reqs_available(ctx);
1024 if (!get_reqs_available(ctx))
1028 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1032 percpu_ref_get(&ctx->reqs);
1037 put_reqs_available(ctx, 1);
1041 static void kiocb_free(struct aio_kiocb *req)
1043 if (req->common.ki_filp)
1044 fput(req->common.ki_filp);
1045 if (req->ki_eventfd != NULL)
1046 eventfd_ctx_put(req->ki_eventfd);
1047 kmem_cache_free(kiocb_cachep, req);
1050 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1052 struct aio_ring __user *ring = (void __user *)ctx_id;
1053 struct mm_struct *mm = current->mm;
1054 struct kioctx *ctx, *ret = NULL;
1055 struct kioctx_table *table;
1058 if (get_user(id, &ring->id))
1062 table = rcu_dereference(mm->ioctx_table);
1064 if (!table || id >= table->nr)
1067 id = array_index_nospec(id, table->nr);
1068 ctx = rcu_dereference(table->table[id]);
1069 if (ctx && ctx->user_id == ctx_id) {
1070 if (percpu_ref_tryget_live(&ctx->users))
1079 * Called when the io request on the given iocb is complete.
1081 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1083 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1084 struct kioctx *ctx = iocb->ki_ctx;
1085 struct aio_ring *ring;
1086 struct io_event *ev_page, *event;
1087 unsigned tail, pos, head;
1088 unsigned long flags;
1091 * Special case handling for sync iocbs:
1092 * - events go directly into the iocb for fast handling
1093 * - the sync task with the iocb in its stack holds the single iocb
1094 * ref, no other paths have a way to get another ref
1095 * - the sync task helpfully left a reference to itself in the iocb
1097 BUG_ON(is_sync_kiocb(kiocb));
1099 if (iocb->ki_list.next) {
1100 unsigned long flags;
1102 spin_lock_irqsave(&ctx->ctx_lock, flags);
1103 list_del(&iocb->ki_list);
1104 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1108 * Add a completion event to the ring buffer. Must be done holding
1109 * ctx->completion_lock to prevent other code from messing with the tail
1110 * pointer since we might be called from irq context.
1112 spin_lock_irqsave(&ctx->completion_lock, flags);
1115 pos = tail + AIO_EVENTS_OFFSET;
1117 if (++tail >= ctx->nr_events)
1120 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1121 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1123 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1124 event->data = iocb->ki_user_data;
1128 kunmap_atomic(ev_page);
1129 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1131 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1132 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1135 /* after flagging the request as done, we
1136 * must never even look at it again
1138 smp_wmb(); /* make event visible before updating tail */
1142 ring = kmap_atomic(ctx->ring_pages[0]);
1145 kunmap_atomic(ring);
1146 flush_dcache_page(ctx->ring_pages[0]);
1148 ctx->completed_events++;
1149 if (ctx->completed_events > 1)
1150 refill_reqs_available(ctx, head, tail);
1151 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1153 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1156 * Check if the user asked us to deliver the result through an
1157 * eventfd. The eventfd_signal() function is safe to be called
1160 if (iocb->ki_eventfd != NULL)
1161 eventfd_signal(iocb->ki_eventfd, 1);
1163 /* everything turned out well, dispose of the aiocb. */
1167 * We have to order our ring_info tail store above and test
1168 * of the wait list below outside the wait lock. This is
1169 * like in wake_up_bit() where clearing a bit has to be
1170 * ordered with the unlocked test.
1174 if (waitqueue_active(&ctx->wait))
1175 wake_up(&ctx->wait);
1177 percpu_ref_put(&ctx->reqs);
1180 /* aio_read_events_ring
1181 * Pull an event off of the ioctx's event ring. Returns the number of
1184 static long aio_read_events_ring(struct kioctx *ctx,
1185 struct io_event __user *event, long nr)
1187 struct aio_ring *ring;
1188 unsigned head, tail, pos;
1193 * The mutex can block and wake us up and that will cause
1194 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1195 * and repeat. This should be rare enough that it doesn't cause
1196 * peformance issues. See the comment in read_events() for more detail.
1198 sched_annotate_sleep();
1199 mutex_lock(&ctx->ring_lock);
1201 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1202 ring = kmap_atomic(ctx->ring_pages[0]);
1205 kunmap_atomic(ring);
1208 * Ensure that once we've read the current tail pointer, that
1209 * we also see the events that were stored up to the tail.
1213 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1218 head %= ctx->nr_events;
1219 tail %= ctx->nr_events;
1223 struct io_event *ev;
1226 avail = (head <= tail ? tail : ctx->nr_events) - head;
1230 avail = min(avail, nr - ret);
1231 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1232 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1234 pos = head + AIO_EVENTS_OFFSET;
1235 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1236 pos %= AIO_EVENTS_PER_PAGE;
1239 copy_ret = copy_to_user(event + ret, ev + pos,
1240 sizeof(*ev) * avail);
1243 if (unlikely(copy_ret)) {
1250 head %= ctx->nr_events;
1253 ring = kmap_atomic(ctx->ring_pages[0]);
1255 kunmap_atomic(ring);
1256 flush_dcache_page(ctx->ring_pages[0]);
1258 pr_debug("%li h%u t%u\n", ret, head, tail);
1260 mutex_unlock(&ctx->ring_lock);
1265 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1266 struct io_event __user *event, long *i)
1268 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1273 if (unlikely(atomic_read(&ctx->dead)))
1279 return ret < 0 || *i >= min_nr;
1282 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1283 struct io_event __user *event,
1284 struct timespec __user *timeout)
1286 ktime_t until = { .tv64 = KTIME_MAX };
1292 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1295 until = timespec_to_ktime(ts);
1299 * Note that aio_read_events() is being called as the conditional - i.e.
1300 * we're calling it after prepare_to_wait() has set task state to
1301 * TASK_INTERRUPTIBLE.
1303 * But aio_read_events() can block, and if it blocks it's going to flip
1304 * the task state back to TASK_RUNNING.
1306 * This should be ok, provided it doesn't flip the state back to
1307 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1308 * will only happen if the mutex_lock() call blocks, and we then find
1309 * the ringbuffer empty. So in practice we should be ok, but it's
1310 * something to be aware of when touching this code.
1312 if (until.tv64 == 0)
1313 aio_read_events(ctx, min_nr, nr, event, &ret);
1315 wait_event_interruptible_hrtimeout(ctx->wait,
1316 aio_read_events(ctx, min_nr, nr, event, &ret),
1319 if (!ret && signal_pending(current))
1326 * Create an aio_context capable of receiving at least nr_events.
1327 * ctxp must not point to an aio_context that already exists, and
1328 * must be initialized to 0 prior to the call. On successful
1329 * creation of the aio_context, *ctxp is filled in with the resulting
1330 * handle. May fail with -EINVAL if *ctxp is not initialized,
1331 * if the specified nr_events exceeds internal limits. May fail
1332 * with -EAGAIN if the specified nr_events exceeds the user's limit
1333 * of available events. May fail with -ENOMEM if insufficient kernel
1334 * resources are available. May fail with -EFAULT if an invalid
1335 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1338 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1340 struct kioctx *ioctx = NULL;
1344 ret = get_user(ctx, ctxp);
1349 if (unlikely(ctx || nr_events == 0)) {
1350 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1355 ioctx = ioctx_alloc(nr_events);
1356 ret = PTR_ERR(ioctx);
1357 if (!IS_ERR(ioctx)) {
1358 ret = put_user(ioctx->user_id, ctxp);
1360 kill_ioctx(current->mm, ioctx, NULL);
1361 percpu_ref_put(&ioctx->users);
1369 * Destroy the aio_context specified. May cancel any outstanding
1370 * AIOs and block on completion. Will fail with -ENOSYS if not
1371 * implemented. May fail with -EINVAL if the context pointed to
1374 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1376 struct kioctx *ioctx = lookup_ioctx(ctx);
1377 if (likely(NULL != ioctx)) {
1378 struct ctx_rq_wait wait;
1381 init_completion(&wait.comp);
1382 atomic_set(&wait.count, 1);
1384 /* Pass requests_done to kill_ioctx() where it can be set
1385 * in a thread-safe way. If we try to set it here then we have
1386 * a race condition if two io_destroy() called simultaneously.
1388 ret = kill_ioctx(current->mm, ioctx, &wait);
1389 percpu_ref_put(&ioctx->users);
1391 /* Wait until all IO for the context are done. Otherwise kernel
1392 * keep using user-space buffers even if user thinks the context
1396 wait_for_completion(&wait.comp);
1400 pr_debug("EINVAL: invalid context id\n");
1404 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1406 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1407 struct iovec **iovec,
1409 struct iov_iter *iter)
1411 #ifdef CONFIG_COMPAT
1413 return compat_import_iovec(rw,
1414 (struct compat_iovec __user *)buf,
1415 len, UIO_FASTIOV, iovec, iter);
1417 return import_iovec(rw, (struct iovec __user *)buf,
1418 len, UIO_FASTIOV, iovec, iter);
1423 * Performs the initial checks and io submission.
1425 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1426 char __user *buf, size_t len, bool compat)
1428 struct file *file = req->ki_filp;
1432 rw_iter_op *iter_op;
1433 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1434 struct iov_iter iter;
1437 case IOCB_CMD_PREAD:
1438 case IOCB_CMD_PREADV:
1441 iter_op = file->f_op->read_iter;
1444 case IOCB_CMD_PWRITE:
1445 case IOCB_CMD_PWRITEV:
1448 iter_op = file->f_op->write_iter;
1451 if (unlikely(!(file->f_mode & mode)))
1457 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1458 ret = aio_setup_vectored_rw(rw, buf, len,
1459 &iovec, compat, &iter);
1461 ret = import_single_range(rw, buf, len, iovec, &iter);
1465 ret = rw_verify_area(rw, file, &req->ki_pos,
1466 iov_iter_count(&iter));
1475 file_start_write(file);
1477 ret = iter_op(req, &iter);
1480 file_end_write(file);
1484 case IOCB_CMD_FDSYNC:
1485 if (!file->f_op->aio_fsync)
1488 ret = file->f_op->aio_fsync(req, 1);
1491 case IOCB_CMD_FSYNC:
1492 if (!file->f_op->aio_fsync)
1495 ret = file->f_op->aio_fsync(req, 0);
1499 pr_debug("EINVAL: no operation provided\n");
1503 if (ret != -EIOCBQUEUED) {
1505 * There's no easy way to restart the syscall since other AIO's
1506 * may be already running. Just fail this IO with EINTR.
1508 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1509 ret == -ERESTARTNOHAND ||
1510 ret == -ERESTART_RESTARTBLOCK))
1512 aio_complete(req, ret, 0);
1518 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1519 struct iocb *iocb, bool compat)
1521 struct aio_kiocb *req;
1524 /* enforce forwards compatibility on users */
1525 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1526 pr_debug("EINVAL: reserve field set\n");
1530 /* prevent overflows */
1532 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1533 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1534 ((ssize_t)iocb->aio_nbytes < 0)
1536 pr_debug("EINVAL: overflow check\n");
1540 req = aio_get_req(ctx);
1544 req->common.ki_filp = fget(iocb->aio_fildes);
1545 if (unlikely(!req->common.ki_filp)) {
1549 req->common.ki_pos = iocb->aio_offset;
1550 req->common.ki_complete = aio_complete;
1551 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1553 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1555 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1556 * instance of the file* now. The file descriptor must be
1557 * an eventfd() fd, and will be signaled for each completed
1558 * event using the eventfd_signal() function.
1560 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1561 if (IS_ERR(req->ki_eventfd)) {
1562 ret = PTR_ERR(req->ki_eventfd);
1563 req->ki_eventfd = NULL;
1567 req->common.ki_flags |= IOCB_EVENTFD;
1570 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1571 if (unlikely(ret)) {
1572 pr_debug("EFAULT: aio_key\n");
1576 req->ki_user_iocb = user_iocb;
1577 req->ki_user_data = iocb->aio_data;
1579 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1580 (char __user *)(unsigned long)iocb->aio_buf,
1588 put_reqs_available(ctx, 1);
1589 percpu_ref_put(&ctx->reqs);
1594 long do_io_submit(aio_context_t ctx_id, long nr,
1595 struct iocb __user *__user *iocbpp, bool compat)
1600 struct blk_plug plug;
1602 if (unlikely(nr < 0))
1605 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1606 nr = LONG_MAX/sizeof(*iocbpp);
1608 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1611 ctx = lookup_ioctx(ctx_id);
1612 if (unlikely(!ctx)) {
1613 pr_debug("EINVAL: invalid context id\n");
1617 blk_start_plug(&plug);
1620 * AKPM: should this return a partial result if some of the IOs were
1621 * successfully submitted?
1623 for (i=0; i<nr; i++) {
1624 struct iocb __user *user_iocb;
1627 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1632 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1637 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1641 blk_finish_plug(&plug);
1643 percpu_ref_put(&ctx->users);
1648 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1649 * the number of iocbs queued. May return -EINVAL if the aio_context
1650 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1651 * *iocbpp[0] is not properly initialized, if the operation specified
1652 * is invalid for the file descriptor in the iocb. May fail with
1653 * -EFAULT if any of the data structures point to invalid data. May
1654 * fail with -EBADF if the file descriptor specified in the first
1655 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1656 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1657 * fail with -ENOSYS if not implemented.
1659 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1660 struct iocb __user * __user *, iocbpp)
1662 return do_io_submit(ctx_id, nr, iocbpp, 0);
1666 * Finds a given iocb for cancellation.
1668 static struct aio_kiocb *
1669 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1671 struct aio_kiocb *kiocb;
1673 assert_spin_locked(&ctx->ctx_lock);
1675 if (key != KIOCB_KEY)
1678 /* TODO: use a hash or array, this sucks. */
1679 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1680 if (kiocb->ki_user_iocb == iocb)
1687 * Attempts to cancel an iocb previously passed to io_submit. If
1688 * the operation is successfully cancelled, the resulting event is
1689 * copied into the memory pointed to by result without being placed
1690 * into the completion queue and 0 is returned. May fail with
1691 * -EFAULT if any of the data structures pointed to are invalid.
1692 * May fail with -EINVAL if aio_context specified by ctx_id is
1693 * invalid. May fail with -EAGAIN if the iocb specified was not
1694 * cancelled. Will fail with -ENOSYS if not implemented.
1696 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1697 struct io_event __user *, result)
1700 struct aio_kiocb *kiocb;
1704 ret = get_user(key, &iocb->aio_key);
1708 ctx = lookup_ioctx(ctx_id);
1712 spin_lock_irq(&ctx->ctx_lock);
1714 kiocb = lookup_kiocb(ctx, iocb, key);
1716 ret = kiocb_cancel(kiocb);
1720 spin_unlock_irq(&ctx->ctx_lock);
1724 * The result argument is no longer used - the io_event is
1725 * always delivered via the ring buffer. -EINPROGRESS indicates
1726 * cancellation is progress:
1731 percpu_ref_put(&ctx->users);
1737 * Attempts to read at least min_nr events and up to nr events from
1738 * the completion queue for the aio_context specified by ctx_id. If
1739 * it succeeds, the number of read events is returned. May fail with
1740 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1741 * out of range, if timeout is out of range. May fail with -EFAULT
1742 * if any of the memory specified is invalid. May return 0 or
1743 * < min_nr if the timeout specified by timeout has elapsed
1744 * before sufficient events are available, where timeout == NULL
1745 * specifies an infinite timeout. Note that the timeout pointed to by
1746 * timeout is relative. Will fail with -ENOSYS if not implemented.
1748 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1751 struct io_event __user *, events,
1752 struct timespec __user *, timeout)
1754 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1757 if (likely(ioctx)) {
1758 if (likely(min_nr <= nr && min_nr >= 0))
1759 ret = read_events(ioctx, min_nr, nr, events, timeout);
1760 percpu_ref_put(&ioctx->users);