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/signal.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
44 #include <asm/kmap_types.h>
45 #include <linux/uaccess.h>
46 #include <linux/nospec.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;
279 struct address_space *i_mapping;
282 truncate_setsize(file_inode(aio_ring_file), 0);
284 /* Prevent further access to the kioctx from migratepages */
285 i_mapping = aio_ring_file->f_mapping;
286 spin_lock(&i_mapping->private_lock);
287 i_mapping->private_data = NULL;
288 ctx->aio_ring_file = NULL;
289 spin_unlock(&i_mapping->private_lock);
295 static void aio_free_ring(struct kioctx *ctx)
299 /* Disconnect the kiotx from the ring file. This prevents future
300 * accesses to the kioctx from page migration.
302 put_aio_ring_file(ctx);
304 for (i = 0; i < ctx->nr_pages; i++) {
306 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
307 page_count(ctx->ring_pages[i]));
308 page = ctx->ring_pages[i];
311 ctx->ring_pages[i] = NULL;
315 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
316 kfree(ctx->ring_pages);
317 ctx->ring_pages = NULL;
321 static int aio_ring_mremap(struct vm_area_struct *vma)
323 struct file *file = vma->vm_file;
324 struct mm_struct *mm = vma->vm_mm;
325 struct kioctx_table *table;
326 int i, res = -EINVAL;
328 spin_lock(&mm->ioctx_lock);
330 table = rcu_dereference(mm->ioctx_table);
334 for (i = 0; i < table->nr; i++) {
337 ctx = rcu_dereference(table->table[i]);
338 if (ctx && ctx->aio_ring_file == file) {
339 if (!atomic_read(&ctx->dead)) {
340 ctx->user_id = ctx->mmap_base = vma->vm_start;
349 spin_unlock(&mm->ioctx_lock);
353 static const struct vm_operations_struct aio_ring_vm_ops = {
354 .mremap = aio_ring_mremap,
355 #if IS_ENABLED(CONFIG_MMU)
356 .fault = filemap_fault,
357 .map_pages = filemap_map_pages,
358 .page_mkwrite = filemap_page_mkwrite,
362 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
364 vma->vm_flags |= VM_DONTEXPAND;
365 vma->vm_ops = &aio_ring_vm_ops;
369 static const struct file_operations aio_ring_fops = {
370 .mmap = aio_ring_mmap,
373 #if IS_ENABLED(CONFIG_MIGRATION)
374 static int aio_migratepage(struct address_space *mapping, struct page *new,
375 struct page *old, enum migrate_mode mode)
383 * We cannot support the _NO_COPY case here, because copy needs to
384 * happen under the ctx->completion_lock. That does not work with the
385 * migration workflow of MIGRATE_SYNC_NO_COPY.
387 if (mode == MIGRATE_SYNC_NO_COPY)
392 /* mapping->private_lock here protects against the kioctx teardown. */
393 spin_lock(&mapping->private_lock);
394 ctx = mapping->private_data;
400 /* The ring_lock mutex. The prevents aio_read_events() from writing
401 * to the ring's head, and prevents page migration from mucking in
402 * a partially initialized kiotx.
404 if (!mutex_trylock(&ctx->ring_lock)) {
410 if (idx < (pgoff_t)ctx->nr_pages) {
411 /* Make sure the old page hasn't already been changed */
412 if (ctx->ring_pages[idx] != old)
420 /* Writeback must be complete */
421 BUG_ON(PageWriteback(old));
424 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
425 if (rc != MIGRATEPAGE_SUCCESS) {
430 /* Take completion_lock to prevent other writes to the ring buffer
431 * while the old page is copied to the new. This prevents new
432 * events from being lost.
434 spin_lock_irqsave(&ctx->completion_lock, flags);
435 migrate_page_copy(new, old);
436 BUG_ON(ctx->ring_pages[idx] != old);
437 ctx->ring_pages[idx] = new;
438 spin_unlock_irqrestore(&ctx->completion_lock, flags);
440 /* The old page is no longer accessible. */
444 mutex_unlock(&ctx->ring_lock);
446 spin_unlock(&mapping->private_lock);
451 static const struct address_space_operations aio_ctx_aops = {
452 .set_page_dirty = __set_page_dirty_no_writeback,
453 #if IS_ENABLED(CONFIG_MIGRATION)
454 .migratepage = aio_migratepage,
458 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
460 struct aio_ring *ring;
461 struct mm_struct *mm = current->mm;
462 unsigned long size, unused;
467 /* Compensate for the ring buffer's head/tail overlap entry */
468 nr_events += 2; /* 1 is required, 2 for good luck */
470 size = sizeof(struct aio_ring);
471 size += sizeof(struct io_event) * nr_events;
473 nr_pages = PFN_UP(size);
477 file = aio_private_file(ctx, nr_pages);
479 ctx->aio_ring_file = NULL;
483 ctx->aio_ring_file = file;
484 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
485 / sizeof(struct io_event);
487 ctx->ring_pages = ctx->internal_pages;
488 if (nr_pages > AIO_RING_PAGES) {
489 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
491 if (!ctx->ring_pages) {
492 put_aio_ring_file(ctx);
497 for (i = 0; i < nr_pages; i++) {
499 page = find_or_create_page(file->f_mapping,
500 i, GFP_HIGHUSER | __GFP_ZERO);
503 pr_debug("pid(%d) page[%d]->count=%d\n",
504 current->pid, i, page_count(page));
505 SetPageUptodate(page);
508 ctx->ring_pages[i] = page;
512 if (unlikely(i != nr_pages)) {
517 ctx->mmap_size = nr_pages * PAGE_SIZE;
518 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
520 if (down_write_killable(&mm->mmap_sem)) {
526 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
527 PROT_READ | PROT_WRITE,
528 MAP_SHARED, 0, &unused, NULL);
529 up_write(&mm->mmap_sem);
530 if (IS_ERR((void *)ctx->mmap_base)) {
536 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
538 ctx->user_id = ctx->mmap_base;
539 ctx->nr_events = nr_events; /* trusted copy */
541 ring = kmap_atomic(ctx->ring_pages[0]);
542 ring->nr = nr_events; /* user copy */
544 ring->head = ring->tail = 0;
545 ring->magic = AIO_RING_MAGIC;
546 ring->compat_features = AIO_RING_COMPAT_FEATURES;
547 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
548 ring->header_length = sizeof(struct aio_ring);
550 flush_dcache_page(ctx->ring_pages[0]);
555 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
556 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
557 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
559 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
561 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
562 struct kioctx *ctx = req->ki_ctx;
565 spin_lock_irqsave(&ctx->ctx_lock, flags);
567 if (!req->ki_list.next)
568 list_add(&req->ki_list, &ctx->active_reqs);
570 req->ki_cancel = cancel;
572 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
574 EXPORT_SYMBOL(kiocb_set_cancel_fn);
576 static int kiocb_cancel(struct aio_kiocb *kiocb)
578 kiocb_cancel_fn *old, *cancel;
581 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
582 * actually has a cancel function, hence the cmpxchg()
585 cancel = ACCESS_ONCE(kiocb->ki_cancel);
587 if (!cancel || cancel == KIOCB_CANCELLED)
591 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
592 } while (cancel != old);
594 return cancel(&kiocb->common);
598 * free_ioctx() should be RCU delayed to synchronize against the RCU
599 * protected lookup_ioctx() and also needs process context to call
600 * aio_free_ring(), so the double bouncing through kioctx->free_rcu and
603 static void free_ioctx(struct work_struct *work)
605 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
607 pr_debug("freeing %p\n", ctx);
610 free_percpu(ctx->cpu);
611 percpu_ref_exit(&ctx->reqs);
612 percpu_ref_exit(&ctx->users);
613 kmem_cache_free(kioctx_cachep, ctx);
616 static void free_ioctx_rcufn(struct rcu_head *head)
618 struct kioctx *ctx = container_of(head, struct kioctx, free_rcu);
620 INIT_WORK(&ctx->free_work, free_ioctx);
621 schedule_work(&ctx->free_work);
624 static void free_ioctx_reqs(struct percpu_ref *ref)
626 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
628 /* At this point we know that there are no any in-flight requests */
629 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
630 complete(&ctx->rq_wait->comp);
632 /* Synchronize against RCU protected table->table[] dereferences */
633 call_rcu(&ctx->free_rcu, free_ioctx_rcufn);
637 * When this function runs, the kioctx has been removed from the "hash table"
638 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
639 * now it's safe to cancel any that need to be.
641 static void free_ioctx_users(struct percpu_ref *ref)
643 struct kioctx *ctx = container_of(ref, struct kioctx, users);
644 struct aio_kiocb *req;
646 spin_lock_irq(&ctx->ctx_lock);
648 while (!list_empty(&ctx->active_reqs)) {
649 req = list_first_entry(&ctx->active_reqs,
650 struct aio_kiocb, ki_list);
652 list_del_init(&req->ki_list);
655 spin_unlock_irq(&ctx->ctx_lock);
657 percpu_ref_kill(&ctx->reqs);
658 percpu_ref_put(&ctx->reqs);
661 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
664 struct kioctx_table *table, *old;
665 struct aio_ring *ring;
667 spin_lock(&mm->ioctx_lock);
668 table = rcu_dereference_raw(mm->ioctx_table);
672 for (i = 0; i < table->nr; i++)
673 if (!rcu_access_pointer(table->table[i])) {
675 rcu_assign_pointer(table->table[i], ctx);
676 spin_unlock(&mm->ioctx_lock);
678 /* While kioctx setup is in progress,
679 * we are protected from page migration
680 * changes ring_pages by ->ring_lock.
682 ring = kmap_atomic(ctx->ring_pages[0]);
688 new_nr = (table ? table->nr : 1) * 4;
689 spin_unlock(&mm->ioctx_lock);
691 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
698 spin_lock(&mm->ioctx_lock);
699 old = rcu_dereference_raw(mm->ioctx_table);
702 rcu_assign_pointer(mm->ioctx_table, table);
703 } else if (table->nr > old->nr) {
704 memcpy(table->table, old->table,
705 old->nr * sizeof(struct kioctx *));
707 rcu_assign_pointer(mm->ioctx_table, table);
716 static void aio_nr_sub(unsigned nr)
718 spin_lock(&aio_nr_lock);
719 if (WARN_ON(aio_nr - nr > aio_nr))
723 spin_unlock(&aio_nr_lock);
727 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
729 static struct kioctx *ioctx_alloc(unsigned nr_events)
731 struct mm_struct *mm = current->mm;
736 * Store the original nr_events -- what userspace passed to io_setup(),
737 * for counting against the global limit -- before it changes.
739 unsigned int max_reqs = nr_events;
742 * We keep track of the number of available ringbuffer slots, to prevent
743 * overflow (reqs_available), and we also use percpu counters for this.
745 * So since up to half the slots might be on other cpu's percpu counters
746 * and unavailable, double nr_events so userspace sees what they
747 * expected: additionally, we move req_batch slots to/from percpu
748 * counters at a time, so make sure that isn't 0:
750 nr_events = max(nr_events, num_possible_cpus() * 4);
753 /* Prevent overflows */
754 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
755 pr_debug("ENOMEM: nr_events too high\n");
756 return ERR_PTR(-EINVAL);
759 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
760 return ERR_PTR(-EAGAIN);
762 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
764 return ERR_PTR(-ENOMEM);
766 ctx->max_reqs = max_reqs;
768 spin_lock_init(&ctx->ctx_lock);
769 spin_lock_init(&ctx->completion_lock);
770 mutex_init(&ctx->ring_lock);
771 /* Protect against page migration throughout kiotx setup by keeping
772 * the ring_lock mutex held until setup is complete. */
773 mutex_lock(&ctx->ring_lock);
774 init_waitqueue_head(&ctx->wait);
776 INIT_LIST_HEAD(&ctx->active_reqs);
778 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
781 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
784 ctx->cpu = alloc_percpu(struct kioctx_cpu);
788 err = aio_setup_ring(ctx, nr_events);
792 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
793 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
794 if (ctx->req_batch < 1)
797 /* limit the number of system wide aios */
798 spin_lock(&aio_nr_lock);
799 if (aio_nr + ctx->max_reqs > aio_max_nr ||
800 aio_nr + ctx->max_reqs < aio_nr) {
801 spin_unlock(&aio_nr_lock);
805 aio_nr += ctx->max_reqs;
806 spin_unlock(&aio_nr_lock);
808 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
809 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
811 err = ioctx_add_table(ctx, mm);
815 /* Release the ring_lock mutex now that all setup is complete. */
816 mutex_unlock(&ctx->ring_lock);
818 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
819 ctx, ctx->user_id, mm, ctx->nr_events);
823 aio_nr_sub(ctx->max_reqs);
825 atomic_set(&ctx->dead, 1);
827 vm_munmap(ctx->mmap_base, ctx->mmap_size);
830 mutex_unlock(&ctx->ring_lock);
831 free_percpu(ctx->cpu);
832 percpu_ref_exit(&ctx->reqs);
833 percpu_ref_exit(&ctx->users);
834 kmem_cache_free(kioctx_cachep, ctx);
835 pr_debug("error allocating ioctx %d\n", err);
840 * Cancels all outstanding aio requests on an aio context. Used
841 * when the processes owning a context have all exited to encourage
842 * the rapid destruction of the kioctx.
844 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
845 struct ctx_rq_wait *wait)
847 struct kioctx_table *table;
849 spin_lock(&mm->ioctx_lock);
850 if (atomic_xchg(&ctx->dead, 1)) {
851 spin_unlock(&mm->ioctx_lock);
855 table = rcu_dereference_raw(mm->ioctx_table);
856 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
857 RCU_INIT_POINTER(table->table[ctx->id], NULL);
858 spin_unlock(&mm->ioctx_lock);
860 /* free_ioctx_reqs() will do the necessary RCU synchronization */
861 wake_up_all(&ctx->wait);
864 * It'd be more correct to do this in free_ioctx(), after all
865 * the outstanding kiocbs have finished - but by then io_destroy
866 * has already returned, so io_setup() could potentially return
867 * -EAGAIN with no ioctxs actually in use (as far as userspace
870 aio_nr_sub(ctx->max_reqs);
873 vm_munmap(ctx->mmap_base, ctx->mmap_size);
876 percpu_ref_kill(&ctx->users);
881 * exit_aio: called when the last user of mm goes away. At this point, there is
882 * no way for any new requests to be submited or any of the io_* syscalls to be
883 * called on the context.
885 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
888 void exit_aio(struct mm_struct *mm)
890 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
891 struct ctx_rq_wait wait;
897 atomic_set(&wait.count, table->nr);
898 init_completion(&wait.comp);
901 for (i = 0; i < table->nr; ++i) {
903 rcu_dereference_protected(table->table[i], true);
911 * We don't need to bother with munmap() here - exit_mmap(mm)
912 * is coming and it'll unmap everything. And we simply can't,
913 * this is not necessarily our ->mm.
914 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
915 * that it needs to unmap the area, just set it to 0.
918 kill_ioctx(mm, ctx, &wait);
921 if (!atomic_sub_and_test(skipped, &wait.count)) {
922 /* Wait until all IO for the context are done. */
923 wait_for_completion(&wait.comp);
926 RCU_INIT_POINTER(mm->ioctx_table, NULL);
930 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
932 struct kioctx_cpu *kcpu;
935 local_irq_save(flags);
936 kcpu = this_cpu_ptr(ctx->cpu);
937 kcpu->reqs_available += nr;
939 while (kcpu->reqs_available >= ctx->req_batch * 2) {
940 kcpu->reqs_available -= ctx->req_batch;
941 atomic_add(ctx->req_batch, &ctx->reqs_available);
944 local_irq_restore(flags);
947 static bool get_reqs_available(struct kioctx *ctx)
949 struct kioctx_cpu *kcpu;
953 local_irq_save(flags);
954 kcpu = this_cpu_ptr(ctx->cpu);
955 if (!kcpu->reqs_available) {
956 int old, avail = atomic_read(&ctx->reqs_available);
959 if (avail < ctx->req_batch)
963 avail = atomic_cmpxchg(&ctx->reqs_available,
964 avail, avail - ctx->req_batch);
965 } while (avail != old);
967 kcpu->reqs_available += ctx->req_batch;
971 kcpu->reqs_available--;
973 local_irq_restore(flags);
977 /* refill_reqs_available
978 * Updates the reqs_available reference counts used for tracking the
979 * number of free slots in the completion ring. This can be called
980 * from aio_complete() (to optimistically update reqs_available) or
981 * from aio_get_req() (the we're out of events case). It must be
982 * called holding ctx->completion_lock.
984 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
987 unsigned events_in_ring, completed;
989 /* Clamp head since userland can write to it. */
990 head %= ctx->nr_events;
992 events_in_ring = tail - head;
994 events_in_ring = ctx->nr_events - (head - tail);
996 completed = ctx->completed_events;
997 if (events_in_ring < completed)
998 completed -= events_in_ring;
1005 ctx->completed_events -= completed;
1006 put_reqs_available(ctx, completed);
1009 /* user_refill_reqs_available
1010 * Called to refill reqs_available when aio_get_req() encounters an
1011 * out of space in the completion ring.
1013 static void user_refill_reqs_available(struct kioctx *ctx)
1015 spin_lock_irq(&ctx->completion_lock);
1016 if (ctx->completed_events) {
1017 struct aio_ring *ring;
1020 /* Access of ring->head may race with aio_read_events_ring()
1021 * here, but that's okay since whether we read the old version
1022 * or the new version, and either will be valid. The important
1023 * part is that head cannot pass tail since we prevent
1024 * aio_complete() from updating tail by holding
1025 * ctx->completion_lock. Even if head is invalid, the check
1026 * against ctx->completed_events below will make sure we do the
1029 ring = kmap_atomic(ctx->ring_pages[0]);
1031 kunmap_atomic(ring);
1033 refill_reqs_available(ctx, head, ctx->tail);
1036 spin_unlock_irq(&ctx->completion_lock);
1040 * Allocate a slot for an aio request.
1041 * Returns NULL if no requests are free.
1043 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1045 struct aio_kiocb *req;
1047 if (!get_reqs_available(ctx)) {
1048 user_refill_reqs_available(ctx);
1049 if (!get_reqs_available(ctx))
1053 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1057 percpu_ref_get(&ctx->reqs);
1062 put_reqs_available(ctx, 1);
1066 static void kiocb_free(struct aio_kiocb *req)
1068 if (req->common.ki_filp)
1069 fput(req->common.ki_filp);
1070 if (req->ki_eventfd != NULL)
1071 eventfd_ctx_put(req->ki_eventfd);
1072 kmem_cache_free(kiocb_cachep, req);
1075 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1077 struct aio_ring __user *ring = (void __user *)ctx_id;
1078 struct mm_struct *mm = current->mm;
1079 struct kioctx *ctx, *ret = NULL;
1080 struct kioctx_table *table;
1083 if (get_user(id, &ring->id))
1087 table = rcu_dereference(mm->ioctx_table);
1089 if (!table || id >= table->nr)
1092 id = array_index_nospec(id, table->nr);
1093 ctx = rcu_dereference(table->table[id]);
1094 if (ctx && ctx->user_id == ctx_id) {
1095 if (percpu_ref_tryget_live(&ctx->users))
1104 * Called when the io request on the given iocb is complete.
1106 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1108 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1109 struct kioctx *ctx = iocb->ki_ctx;
1110 struct aio_ring *ring;
1111 struct io_event *ev_page, *event;
1112 unsigned tail, pos, head;
1113 unsigned long flags;
1115 if (kiocb->ki_flags & IOCB_WRITE) {
1116 struct file *file = kiocb->ki_filp;
1119 * Tell lockdep we inherited freeze protection from submission
1122 if (S_ISREG(file_inode(file)->i_mode))
1123 __sb_writers_acquired(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1124 file_end_write(file);
1128 * Special case handling for sync iocbs:
1129 * - events go directly into the iocb for fast handling
1130 * - the sync task with the iocb in its stack holds the single iocb
1131 * ref, no other paths have a way to get another ref
1132 * - the sync task helpfully left a reference to itself in the iocb
1134 BUG_ON(is_sync_kiocb(kiocb));
1136 if (iocb->ki_list.next) {
1137 unsigned long flags;
1139 spin_lock_irqsave(&ctx->ctx_lock, flags);
1140 list_del(&iocb->ki_list);
1141 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1145 * Add a completion event to the ring buffer. Must be done holding
1146 * ctx->completion_lock to prevent other code from messing with the tail
1147 * pointer since we might be called from irq context.
1149 spin_lock_irqsave(&ctx->completion_lock, flags);
1152 pos = tail + AIO_EVENTS_OFFSET;
1154 if (++tail >= ctx->nr_events)
1157 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1158 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1160 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1161 event->data = iocb->ki_user_data;
1165 kunmap_atomic(ev_page);
1166 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1168 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1169 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1172 /* after flagging the request as done, we
1173 * must never even look at it again
1175 smp_wmb(); /* make event visible before updating tail */
1179 ring = kmap_atomic(ctx->ring_pages[0]);
1182 kunmap_atomic(ring);
1183 flush_dcache_page(ctx->ring_pages[0]);
1185 ctx->completed_events++;
1186 if (ctx->completed_events > 1)
1187 refill_reqs_available(ctx, head, tail);
1188 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1190 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1193 * Check if the user asked us to deliver the result through an
1194 * eventfd. The eventfd_signal() function is safe to be called
1197 if (iocb->ki_eventfd != NULL)
1198 eventfd_signal(iocb->ki_eventfd, 1);
1200 /* everything turned out well, dispose of the aiocb. */
1204 * We have to order our ring_info tail store above and test
1205 * of the wait list below outside the wait lock. This is
1206 * like in wake_up_bit() where clearing a bit has to be
1207 * ordered with the unlocked test.
1211 if (waitqueue_active(&ctx->wait))
1212 wake_up(&ctx->wait);
1214 percpu_ref_put(&ctx->reqs);
1217 /* aio_read_events_ring
1218 * Pull an event off of the ioctx's event ring. Returns the number of
1221 static long aio_read_events_ring(struct kioctx *ctx,
1222 struct io_event __user *event, long nr)
1224 struct aio_ring *ring;
1225 unsigned head, tail, pos;
1230 * The mutex can block and wake us up and that will cause
1231 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1232 * and repeat. This should be rare enough that it doesn't cause
1233 * peformance issues. See the comment in read_events() for more detail.
1235 sched_annotate_sleep();
1236 mutex_lock(&ctx->ring_lock);
1238 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1239 ring = kmap_atomic(ctx->ring_pages[0]);
1242 kunmap_atomic(ring);
1245 * Ensure that once we've read the current tail pointer, that
1246 * we also see the events that were stored up to the tail.
1250 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1255 head %= ctx->nr_events;
1256 tail %= ctx->nr_events;
1260 struct io_event *ev;
1263 avail = (head <= tail ? tail : ctx->nr_events) - head;
1267 avail = min(avail, nr - ret);
1268 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1269 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1271 pos = head + AIO_EVENTS_OFFSET;
1272 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1273 pos %= AIO_EVENTS_PER_PAGE;
1276 copy_ret = copy_to_user(event + ret, ev + pos,
1277 sizeof(*ev) * avail);
1280 if (unlikely(copy_ret)) {
1287 head %= ctx->nr_events;
1290 ring = kmap_atomic(ctx->ring_pages[0]);
1292 kunmap_atomic(ring);
1293 flush_dcache_page(ctx->ring_pages[0]);
1295 pr_debug("%li h%u t%u\n", ret, head, tail);
1297 mutex_unlock(&ctx->ring_lock);
1302 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1303 struct io_event __user *event, long *i)
1305 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1310 if (unlikely(atomic_read(&ctx->dead)))
1316 return ret < 0 || *i >= min_nr;
1319 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1320 struct io_event __user *event,
1321 struct timespec __user *timeout)
1323 ktime_t until = KTIME_MAX;
1329 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1332 until = timespec_to_ktime(ts);
1336 * Note that aio_read_events() is being called as the conditional - i.e.
1337 * we're calling it after prepare_to_wait() has set task state to
1338 * TASK_INTERRUPTIBLE.
1340 * But aio_read_events() can block, and if it blocks it's going to flip
1341 * the task state back to TASK_RUNNING.
1343 * This should be ok, provided it doesn't flip the state back to
1344 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1345 * will only happen if the mutex_lock() call blocks, and we then find
1346 * the ringbuffer empty. So in practice we should be ok, but it's
1347 * something to be aware of when touching this code.
1350 aio_read_events(ctx, min_nr, nr, event, &ret);
1352 wait_event_interruptible_hrtimeout(ctx->wait,
1353 aio_read_events(ctx, min_nr, nr, event, &ret),
1356 if (!ret && signal_pending(current))
1363 * Create an aio_context capable of receiving at least nr_events.
1364 * ctxp must not point to an aio_context that already exists, and
1365 * must be initialized to 0 prior to the call. On successful
1366 * creation of the aio_context, *ctxp is filled in with the resulting
1367 * handle. May fail with -EINVAL if *ctxp is not initialized,
1368 * if the specified nr_events exceeds internal limits. May fail
1369 * with -EAGAIN if the specified nr_events exceeds the user's limit
1370 * of available events. May fail with -ENOMEM if insufficient kernel
1371 * resources are available. May fail with -EFAULT if an invalid
1372 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1375 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1377 struct kioctx *ioctx = NULL;
1381 ret = get_user(ctx, ctxp);
1386 if (unlikely(ctx || nr_events == 0)) {
1387 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1392 ioctx = ioctx_alloc(nr_events);
1393 ret = PTR_ERR(ioctx);
1394 if (!IS_ERR(ioctx)) {
1395 ret = put_user(ioctx->user_id, ctxp);
1397 kill_ioctx(current->mm, ioctx, NULL);
1398 percpu_ref_put(&ioctx->users);
1405 #ifdef CONFIG_COMPAT
1406 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1408 struct kioctx *ioctx = NULL;
1412 ret = get_user(ctx, ctx32p);
1417 if (unlikely(ctx || nr_events == 0)) {
1418 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1423 ioctx = ioctx_alloc(nr_events);
1424 ret = PTR_ERR(ioctx);
1425 if (!IS_ERR(ioctx)) {
1426 /* truncating is ok because it's a user address */
1427 ret = put_user((u32)ioctx->user_id, ctx32p);
1429 kill_ioctx(current->mm, ioctx, NULL);
1430 percpu_ref_put(&ioctx->users);
1439 * Destroy the aio_context specified. May cancel any outstanding
1440 * AIOs and block on completion. Will fail with -ENOSYS if not
1441 * implemented. May fail with -EINVAL if the context pointed to
1444 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1446 struct kioctx *ioctx = lookup_ioctx(ctx);
1447 if (likely(NULL != ioctx)) {
1448 struct ctx_rq_wait wait;
1451 init_completion(&wait.comp);
1452 atomic_set(&wait.count, 1);
1454 /* Pass requests_done to kill_ioctx() where it can be set
1455 * in a thread-safe way. If we try to set it here then we have
1456 * a race condition if two io_destroy() called simultaneously.
1458 ret = kill_ioctx(current->mm, ioctx, &wait);
1459 percpu_ref_put(&ioctx->users);
1461 /* Wait until all IO for the context are done. Otherwise kernel
1462 * keep using user-space buffers even if user thinks the context
1466 wait_for_completion(&wait.comp);
1470 pr_debug("EINVAL: invalid context id\n");
1474 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1475 bool vectored, bool compat, struct iov_iter *iter)
1477 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1478 size_t len = iocb->aio_nbytes;
1481 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1485 #ifdef CONFIG_COMPAT
1487 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1490 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1493 static inline ssize_t aio_ret(struct kiocb *req, ssize_t ret)
1499 case -ERESTARTNOINTR:
1500 case -ERESTARTNOHAND:
1501 case -ERESTART_RESTARTBLOCK:
1503 * There's no easy way to restart the syscall since other AIO's
1504 * may be already running. Just fail this IO with EINTR.
1509 aio_complete(req, ret, 0);
1514 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1517 struct file *file = req->ki_filp;
1518 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1519 struct iov_iter iter;
1522 if (unlikely(!(file->f_mode & FMODE_READ)))
1524 if (unlikely(!file->f_op->read_iter))
1527 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1530 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1532 ret = aio_ret(req, call_read_iter(file, req, &iter));
1537 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1540 struct file *file = req->ki_filp;
1541 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1542 struct iov_iter iter;
1545 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1547 if (unlikely(!file->f_op->write_iter))
1550 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1553 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1555 req->ki_flags |= IOCB_WRITE;
1556 file_start_write(file);
1557 ret = aio_ret(req, call_write_iter(file, req, &iter));
1559 * We release freeze protection in aio_complete(). Fool lockdep
1560 * by telling it the lock got released so that it doesn't
1561 * complain about held lock when we return to userspace.
1563 if (S_ISREG(file_inode(file)->i_mode))
1564 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1570 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1571 struct iocb *iocb, bool compat)
1573 struct aio_kiocb *req;
1577 /* enforce forwards compatibility on users */
1578 if (unlikely(iocb->aio_reserved2)) {
1579 pr_debug("EINVAL: reserve field set\n");
1583 /* prevent overflows */
1585 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1586 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1587 ((ssize_t)iocb->aio_nbytes < 0)
1589 pr_debug("EINVAL: overflow check\n");
1593 req = aio_get_req(ctx);
1597 req->common.ki_filp = file = fget(iocb->aio_fildes);
1598 if (unlikely(!req->common.ki_filp)) {
1602 req->common.ki_pos = iocb->aio_offset;
1603 req->common.ki_complete = aio_complete;
1604 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1605 req->common.ki_hint = file_write_hint(file);
1607 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1609 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1610 * instance of the file* now. The file descriptor must be
1611 * an eventfd() fd, and will be signaled for each completed
1612 * event using the eventfd_signal() function.
1614 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1615 if (IS_ERR(req->ki_eventfd)) {
1616 ret = PTR_ERR(req->ki_eventfd);
1617 req->ki_eventfd = NULL;
1621 req->common.ki_flags |= IOCB_EVENTFD;
1624 ret = kiocb_set_rw_flags(&req->common, iocb->aio_rw_flags);
1625 if (unlikely(ret)) {
1626 pr_debug("EINVAL: aio_rw_flags\n");
1630 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1631 if (unlikely(ret)) {
1632 pr_debug("EFAULT: aio_key\n");
1636 req->ki_user_iocb = user_iocb;
1637 req->ki_user_data = iocb->aio_data;
1640 switch (iocb->aio_lio_opcode) {
1641 case IOCB_CMD_PREAD:
1642 ret = aio_read(&req->common, iocb, false, compat);
1644 case IOCB_CMD_PWRITE:
1645 ret = aio_write(&req->common, iocb, false, compat);
1647 case IOCB_CMD_PREADV:
1648 ret = aio_read(&req->common, iocb, true, compat);
1650 case IOCB_CMD_PWRITEV:
1651 ret = aio_write(&req->common, iocb, true, compat);
1654 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1660 if (ret && ret != -EIOCBQUEUED)
1664 put_reqs_available(ctx, 1);
1665 percpu_ref_put(&ctx->reqs);
1670 static long do_io_submit(aio_context_t ctx_id, long nr,
1671 struct iocb __user *__user *iocbpp, bool compat)
1676 struct blk_plug plug;
1678 if (unlikely(nr < 0))
1681 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1682 nr = LONG_MAX/sizeof(*iocbpp);
1684 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1687 ctx = lookup_ioctx(ctx_id);
1688 if (unlikely(!ctx)) {
1689 pr_debug("EINVAL: invalid context id\n");
1693 blk_start_plug(&plug);
1696 * AKPM: should this return a partial result if some of the IOs were
1697 * successfully submitted?
1699 for (i=0; i<nr; i++) {
1700 struct iocb __user *user_iocb;
1703 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1708 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1713 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1717 blk_finish_plug(&plug);
1719 percpu_ref_put(&ctx->users);
1724 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1725 * the number of iocbs queued. May return -EINVAL if the aio_context
1726 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1727 * *iocbpp[0] is not properly initialized, if the operation specified
1728 * is invalid for the file descriptor in the iocb. May fail with
1729 * -EFAULT if any of the data structures point to invalid data. May
1730 * fail with -EBADF if the file descriptor specified in the first
1731 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1732 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1733 * fail with -ENOSYS if not implemented.
1735 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1736 struct iocb __user * __user *, iocbpp)
1738 return do_io_submit(ctx_id, nr, iocbpp, 0);
1741 #ifdef CONFIG_COMPAT
1743 copy_iocb(long nr, u32 __user *ptr32, struct iocb __user * __user *ptr64)
1748 for (i = 0; i < nr; ++i) {
1749 if (get_user(uptr, ptr32 + i))
1751 if (put_user(compat_ptr(uptr), ptr64 + i))
1757 #define MAX_AIO_SUBMITS (PAGE_SIZE/sizeof(struct iocb *))
1759 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1760 int, nr, u32 __user *, iocb)
1762 struct iocb __user * __user *iocb64;
1765 if (unlikely(nr < 0))
1768 if (nr > MAX_AIO_SUBMITS)
1769 nr = MAX_AIO_SUBMITS;
1771 iocb64 = compat_alloc_user_space(nr * sizeof(*iocb64));
1772 ret = copy_iocb(nr, iocb, iocb64);
1774 ret = do_io_submit(ctx_id, nr, iocb64, 1);
1780 * Finds a given iocb for cancellation.
1782 static struct aio_kiocb *
1783 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1785 struct aio_kiocb *kiocb;
1787 assert_spin_locked(&ctx->ctx_lock);
1789 if (key != KIOCB_KEY)
1792 /* TODO: use a hash or array, this sucks. */
1793 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1794 if (kiocb->ki_user_iocb == iocb)
1801 * Attempts to cancel an iocb previously passed to io_submit. If
1802 * the operation is successfully cancelled, the resulting event is
1803 * copied into the memory pointed to by result without being placed
1804 * into the completion queue and 0 is returned. May fail with
1805 * -EFAULT if any of the data structures pointed to are invalid.
1806 * May fail with -EINVAL if aio_context specified by ctx_id is
1807 * invalid. May fail with -EAGAIN if the iocb specified was not
1808 * cancelled. Will fail with -ENOSYS if not implemented.
1810 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1811 struct io_event __user *, result)
1814 struct aio_kiocb *kiocb;
1818 ret = get_user(key, &iocb->aio_key);
1822 ctx = lookup_ioctx(ctx_id);
1826 spin_lock_irq(&ctx->ctx_lock);
1828 kiocb = lookup_kiocb(ctx, iocb, key);
1830 ret = kiocb_cancel(kiocb);
1834 spin_unlock_irq(&ctx->ctx_lock);
1838 * The result argument is no longer used - the io_event is
1839 * always delivered via the ring buffer. -EINPROGRESS indicates
1840 * cancellation is progress:
1845 percpu_ref_put(&ctx->users);
1851 * Attempts to read at least min_nr events and up to nr events from
1852 * the completion queue for the aio_context specified by ctx_id. If
1853 * it succeeds, the number of read events is returned. May fail with
1854 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1855 * out of range, if timeout is out of range. May fail with -EFAULT
1856 * if any of the memory specified is invalid. May return 0 or
1857 * < min_nr if the timeout specified by timeout has elapsed
1858 * before sufficient events are available, where timeout == NULL
1859 * specifies an infinite timeout. Note that the timeout pointed to by
1860 * timeout is relative. Will fail with -ENOSYS if not implemented.
1862 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1865 struct io_event __user *, events,
1866 struct timespec __user *, timeout)
1868 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1871 if (likely(ioctx)) {
1872 if (likely(min_nr <= nr && min_nr >= 0))
1873 ret = read_events(ioctx, min_nr, nr, events, timeout);
1874 percpu_ref_put(&ioctx->users);
1879 #ifdef CONFIG_COMPAT
1880 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
1881 compat_long_t, min_nr,
1883 struct io_event __user *, events,
1884 struct compat_timespec __user *, timeout)
1887 struct timespec __user *ut = NULL;
1890 if (compat_get_timespec(&t, timeout))
1893 ut = compat_alloc_user_space(sizeof(*ut));
1894 if (copy_to_user(ut, &t, sizeof(t)))
1897 return sys_io_getevents(ctx_id, min_nr, nr, events, ut);