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.
8 * Copyright 2018 Christoph Hellwig.
10 * See ../COPYING for licensing terms.
12 #define pr_fmt(fmt) "%s: " fmt, __func__
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
25 #include <linux/sched/signal.h>
27 #include <linux/file.h>
29 #include <linux/mman.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h>
46 #include <asm/kmap_types.h>
47 #include <linux/uaccess.h>
48 #include <linux/nospec.h>
54 #define AIO_RING_MAGIC 0xa10a10a1
55 #define AIO_RING_COMPAT_FEATURES 1
56 #define AIO_RING_INCOMPAT_FEATURES 0
58 unsigned id; /* kernel internal index number */
59 unsigned nr; /* number of io_events */
60 unsigned head; /* Written to by userland or under ring_lock
61 * mutex by aio_read_events_ring(). */
65 unsigned compat_features;
66 unsigned incompat_features;
67 unsigned header_length; /* size of aio_ring */
70 struct io_event io_events[];
71 }; /* 128 bytes + ring size */
74 * Plugging is meant to work with larger batches of IOs. If we don't
75 * have more than the below, then don't bother setting up a plug.
77 #define AIO_PLUG_THRESHOLD 2
79 #define AIO_RING_PAGES 8
84 struct kioctx __rcu *table[];
88 unsigned reqs_available;
92 struct completion comp;
97 struct percpu_ref users;
100 struct percpu_ref reqs;
102 unsigned long user_id;
104 struct __percpu kioctx_cpu *cpu;
107 * For percpu reqs_available, number of slots we move to/from global
112 * This is what userspace passed to io_setup(), it's not used for
113 * anything but counting against the global max_reqs quota.
115 * The real limit is nr_events - 1, which will be larger (see
120 /* Size of ringbuffer, in units of struct io_event */
123 unsigned long mmap_base;
124 unsigned long mmap_size;
126 struct page **ring_pages;
129 struct rcu_work free_rwork; /* see free_ioctx() */
132 * signals when all in-flight requests are done
134 struct ctx_rq_wait *rq_wait;
138 * This counts the number of available slots in the ringbuffer,
139 * so we avoid overflowing it: it's decremented (if positive)
140 * when allocating a kiocb and incremented when the resulting
141 * io_event is pulled off the ringbuffer.
143 * We batch accesses to it with a percpu version.
145 atomic_t reqs_available;
146 } ____cacheline_aligned_in_smp;
150 struct list_head active_reqs; /* used for cancellation */
151 } ____cacheline_aligned_in_smp;
154 struct mutex ring_lock;
155 wait_queue_head_t wait;
156 } ____cacheline_aligned_in_smp;
160 unsigned completed_events;
161 spinlock_t completion_lock;
162 } ____cacheline_aligned_in_smp;
164 struct page *internal_pages[AIO_RING_PAGES];
165 struct file *aio_ring_file;
171 * First field must be the file pointer in all the
172 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
176 struct work_struct work;
183 struct wait_queue_head *head;
187 bool work_need_resched;
188 struct wait_queue_entry wait;
189 struct work_struct work;
193 * NOTE! Each of the iocb union members has the file pointer
194 * as the first entry in their struct definition. So you can
195 * access the file pointer through any of the sub-structs,
196 * or directly as just 'ki_filp' in this struct.
200 struct file *ki_filp;
202 struct fsync_iocb fsync;
203 struct poll_iocb poll;
206 struct kioctx *ki_ctx;
207 kiocb_cancel_fn *ki_cancel;
209 struct io_event ki_res;
211 struct list_head ki_list; /* the aio core uses this
212 * for cancellation */
213 refcount_t ki_refcnt;
216 * If the aio_resfd field of the userspace iocb is not zero,
217 * this is the underlying eventfd context to deliver events to.
219 struct eventfd_ctx *ki_eventfd;
222 /*------ sysctl variables----*/
223 static DEFINE_SPINLOCK(aio_nr_lock);
224 unsigned long aio_nr; /* current system wide number of aio requests */
225 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
226 /*----end sysctl variables---*/
228 static struct kmem_cache *kiocb_cachep;
229 static struct kmem_cache *kioctx_cachep;
231 static struct vfsmount *aio_mnt;
233 static const struct file_operations aio_ring_fops;
234 static const struct address_space_operations aio_ctx_aops;
236 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
239 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
241 return ERR_CAST(inode);
243 inode->i_mapping->a_ops = &aio_ctx_aops;
244 inode->i_mapping->private_data = ctx;
245 inode->i_size = PAGE_SIZE * nr_pages;
247 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
248 O_RDWR, &aio_ring_fops);
254 static int aio_init_fs_context(struct fs_context *fc)
256 if (!init_pseudo(fc, AIO_RING_MAGIC))
258 fc->s_iflags |= SB_I_NOEXEC;
263 * Creates the slab caches used by the aio routines, panic on
264 * failure as this is done early during the boot sequence.
266 static int __init aio_setup(void)
268 static struct file_system_type aio_fs = {
270 .init_fs_context = aio_init_fs_context,
271 .kill_sb = kill_anon_super,
273 aio_mnt = kern_mount(&aio_fs);
275 panic("Failed to create aio fs mount.");
277 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
281 __initcall(aio_setup);
283 static void put_aio_ring_file(struct kioctx *ctx)
285 struct file *aio_ring_file = ctx->aio_ring_file;
286 struct address_space *i_mapping;
289 truncate_setsize(file_inode(aio_ring_file), 0);
291 /* Prevent further access to the kioctx from migratepages */
292 i_mapping = aio_ring_file->f_mapping;
293 spin_lock(&i_mapping->private_lock);
294 i_mapping->private_data = NULL;
295 ctx->aio_ring_file = NULL;
296 spin_unlock(&i_mapping->private_lock);
302 static void aio_free_ring(struct kioctx *ctx)
306 /* Disconnect the kiotx from the ring file. This prevents future
307 * accesses to the kioctx from page migration.
309 put_aio_ring_file(ctx);
311 for (i = 0; i < ctx->nr_pages; i++) {
313 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
314 page_count(ctx->ring_pages[i]));
315 page = ctx->ring_pages[i];
318 ctx->ring_pages[i] = NULL;
322 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
323 kfree(ctx->ring_pages);
324 ctx->ring_pages = NULL;
328 static int aio_ring_mremap(struct vm_area_struct *vma)
330 struct file *file = vma->vm_file;
331 struct mm_struct *mm = vma->vm_mm;
332 struct kioctx_table *table;
333 int i, res = -EINVAL;
335 spin_lock(&mm->ioctx_lock);
337 table = rcu_dereference(mm->ioctx_table);
338 for (i = 0; i < table->nr; i++) {
341 ctx = rcu_dereference(table->table[i]);
342 if (ctx && ctx->aio_ring_file == file) {
343 if (!atomic_read(&ctx->dead)) {
344 ctx->user_id = ctx->mmap_base = vma->vm_start;
352 spin_unlock(&mm->ioctx_lock);
356 static const struct vm_operations_struct aio_ring_vm_ops = {
357 .mremap = aio_ring_mremap,
358 #if IS_ENABLED(CONFIG_MMU)
359 .fault = filemap_fault,
360 .map_pages = filemap_map_pages,
361 .page_mkwrite = filemap_page_mkwrite,
365 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
367 vma->vm_flags |= VM_DONTEXPAND;
368 vma->vm_ops = &aio_ring_vm_ops;
372 static const struct file_operations aio_ring_fops = {
373 .mmap = aio_ring_mmap,
376 #if IS_ENABLED(CONFIG_MIGRATION)
377 static int aio_migratepage(struct address_space *mapping, struct page *new,
378 struct page *old, enum migrate_mode mode)
386 * We cannot support the _NO_COPY case here, because copy needs to
387 * happen under the ctx->completion_lock. That does not work with the
388 * migration workflow of MIGRATE_SYNC_NO_COPY.
390 if (mode == MIGRATE_SYNC_NO_COPY)
395 /* mapping->private_lock here protects against the kioctx teardown. */
396 spin_lock(&mapping->private_lock);
397 ctx = mapping->private_data;
403 /* The ring_lock mutex. The prevents aio_read_events() from writing
404 * to the ring's head, and prevents page migration from mucking in
405 * a partially initialized kiotx.
407 if (!mutex_trylock(&ctx->ring_lock)) {
413 if (idx < (pgoff_t)ctx->nr_pages) {
414 /* Make sure the old page hasn't already been changed */
415 if (ctx->ring_pages[idx] != old)
423 /* Writeback must be complete */
424 BUG_ON(PageWriteback(old));
427 rc = migrate_page_move_mapping(mapping, new, old, 1);
428 if (rc != MIGRATEPAGE_SUCCESS) {
433 /* Take completion_lock to prevent other writes to the ring buffer
434 * while the old page is copied to the new. This prevents new
435 * events from being lost.
437 spin_lock_irqsave(&ctx->completion_lock, flags);
438 migrate_page_copy(new, old);
439 BUG_ON(ctx->ring_pages[idx] != old);
440 ctx->ring_pages[idx] = new;
441 spin_unlock_irqrestore(&ctx->completion_lock, flags);
443 /* The old page is no longer accessible. */
447 mutex_unlock(&ctx->ring_lock);
449 spin_unlock(&mapping->private_lock);
454 static const struct address_space_operations aio_ctx_aops = {
455 .set_page_dirty = __set_page_dirty_no_writeback,
456 #if IS_ENABLED(CONFIG_MIGRATION)
457 .migratepage = aio_migratepage,
461 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
463 struct aio_ring *ring;
464 struct mm_struct *mm = current->mm;
465 unsigned long size, unused;
470 /* Compensate for the ring buffer's head/tail overlap entry */
471 nr_events += 2; /* 1 is required, 2 for good luck */
473 size = sizeof(struct aio_ring);
474 size += sizeof(struct io_event) * nr_events;
476 nr_pages = PFN_UP(size);
480 file = aio_private_file(ctx, nr_pages);
482 ctx->aio_ring_file = NULL;
486 ctx->aio_ring_file = file;
487 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
488 / sizeof(struct io_event);
490 ctx->ring_pages = ctx->internal_pages;
491 if (nr_pages > AIO_RING_PAGES) {
492 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
494 if (!ctx->ring_pages) {
495 put_aio_ring_file(ctx);
500 for (i = 0; i < nr_pages; i++) {
502 page = find_or_create_page(file->f_mapping,
503 i, GFP_HIGHUSER | __GFP_ZERO);
506 pr_debug("pid(%d) page[%d]->count=%d\n",
507 current->pid, i, page_count(page));
508 SetPageUptodate(page);
511 ctx->ring_pages[i] = page;
515 if (unlikely(i != nr_pages)) {
520 ctx->mmap_size = nr_pages * PAGE_SIZE;
521 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
523 if (mmap_write_lock_killable(mm)) {
529 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
530 PROT_READ | PROT_WRITE,
531 MAP_SHARED, 0, &unused, NULL);
532 mmap_write_unlock(mm);
533 if (IS_ERR((void *)ctx->mmap_base)) {
539 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
541 ctx->user_id = ctx->mmap_base;
542 ctx->nr_events = nr_events; /* trusted copy */
544 ring = kmap_atomic(ctx->ring_pages[0]);
545 ring->nr = nr_events; /* user copy */
547 ring->head = ring->tail = 0;
548 ring->magic = AIO_RING_MAGIC;
549 ring->compat_features = AIO_RING_COMPAT_FEATURES;
550 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
551 ring->header_length = sizeof(struct aio_ring);
553 flush_dcache_page(ctx->ring_pages[0]);
558 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
559 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
560 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
562 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
564 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
565 struct kioctx *ctx = req->ki_ctx;
568 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
571 spin_lock_irqsave(&ctx->ctx_lock, flags);
572 list_add_tail(&req->ki_list, &ctx->active_reqs);
573 req->ki_cancel = cancel;
574 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
576 EXPORT_SYMBOL(kiocb_set_cancel_fn);
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(). Use rcu_work.
583 static void free_ioctx(struct work_struct *work)
585 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
587 pr_debug("freeing %p\n", ctx);
590 free_percpu(ctx->cpu);
591 percpu_ref_exit(&ctx->reqs);
592 percpu_ref_exit(&ctx->users);
593 kmem_cache_free(kioctx_cachep, ctx);
596 static void free_ioctx_reqs(struct percpu_ref *ref)
598 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
600 /* At this point we know that there are no any in-flight requests */
601 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
602 complete(&ctx->rq_wait->comp);
604 /* Synchronize against RCU protected table->table[] dereferences */
605 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
606 queue_rcu_work(system_wq, &ctx->free_rwork);
610 * When this function runs, the kioctx has been removed from the "hash table"
611 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
612 * now it's safe to cancel any that need to be.
614 static void free_ioctx_users(struct percpu_ref *ref)
616 struct kioctx *ctx = container_of(ref, struct kioctx, users);
617 struct aio_kiocb *req;
619 spin_lock_irq(&ctx->ctx_lock);
621 while (!list_empty(&ctx->active_reqs)) {
622 req = list_first_entry(&ctx->active_reqs,
623 struct aio_kiocb, ki_list);
624 req->ki_cancel(&req->rw);
625 list_del_init(&req->ki_list);
628 spin_unlock_irq(&ctx->ctx_lock);
630 percpu_ref_kill(&ctx->reqs);
631 percpu_ref_put(&ctx->reqs);
634 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
637 struct kioctx_table *table, *old;
638 struct aio_ring *ring;
640 spin_lock(&mm->ioctx_lock);
641 table = rcu_dereference_raw(mm->ioctx_table);
645 for (i = 0; i < table->nr; i++)
646 if (!rcu_access_pointer(table->table[i])) {
648 rcu_assign_pointer(table->table[i], ctx);
649 spin_unlock(&mm->ioctx_lock);
651 /* While kioctx setup is in progress,
652 * we are protected from page migration
653 * changes ring_pages by ->ring_lock.
655 ring = kmap_atomic(ctx->ring_pages[0]);
661 new_nr = (table ? table->nr : 1) * 4;
662 spin_unlock(&mm->ioctx_lock);
664 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
671 spin_lock(&mm->ioctx_lock);
672 old = rcu_dereference_raw(mm->ioctx_table);
675 rcu_assign_pointer(mm->ioctx_table, table);
676 } else if (table->nr > old->nr) {
677 memcpy(table->table, old->table,
678 old->nr * sizeof(struct kioctx *));
680 rcu_assign_pointer(mm->ioctx_table, table);
689 static void aio_nr_sub(unsigned nr)
691 spin_lock(&aio_nr_lock);
692 if (WARN_ON(aio_nr - nr > aio_nr))
696 spin_unlock(&aio_nr_lock);
700 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
702 static struct kioctx *ioctx_alloc(unsigned nr_events)
704 struct mm_struct *mm = current->mm;
709 * Store the original nr_events -- what userspace passed to io_setup(),
710 * for counting against the global limit -- before it changes.
712 unsigned int max_reqs = nr_events;
715 * We keep track of the number of available ringbuffer slots, to prevent
716 * overflow (reqs_available), and we also use percpu counters for this.
718 * So since up to half the slots might be on other cpu's percpu counters
719 * and unavailable, double nr_events so userspace sees what they
720 * expected: additionally, we move req_batch slots to/from percpu
721 * counters at a time, so make sure that isn't 0:
723 nr_events = max(nr_events, num_possible_cpus() * 4);
726 /* Prevent overflows */
727 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
728 pr_debug("ENOMEM: nr_events too high\n");
729 return ERR_PTR(-EINVAL);
732 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
733 return ERR_PTR(-EAGAIN);
735 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
737 return ERR_PTR(-ENOMEM);
739 ctx->max_reqs = max_reqs;
741 spin_lock_init(&ctx->ctx_lock);
742 spin_lock_init(&ctx->completion_lock);
743 mutex_init(&ctx->ring_lock);
744 /* Protect against page migration throughout kiotx setup by keeping
745 * the ring_lock mutex held until setup is complete. */
746 mutex_lock(&ctx->ring_lock);
747 init_waitqueue_head(&ctx->wait);
749 INIT_LIST_HEAD(&ctx->active_reqs);
751 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
754 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
757 ctx->cpu = alloc_percpu(struct kioctx_cpu);
761 err = aio_setup_ring(ctx, nr_events);
765 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
766 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
767 if (ctx->req_batch < 1)
770 /* limit the number of system wide aios */
771 spin_lock(&aio_nr_lock);
772 if (aio_nr + ctx->max_reqs > aio_max_nr ||
773 aio_nr + ctx->max_reqs < aio_nr) {
774 spin_unlock(&aio_nr_lock);
778 aio_nr += ctx->max_reqs;
779 spin_unlock(&aio_nr_lock);
781 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
782 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
784 err = ioctx_add_table(ctx, mm);
788 /* Release the ring_lock mutex now that all setup is complete. */
789 mutex_unlock(&ctx->ring_lock);
791 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
792 ctx, ctx->user_id, mm, ctx->nr_events);
796 aio_nr_sub(ctx->max_reqs);
798 atomic_set(&ctx->dead, 1);
800 vm_munmap(ctx->mmap_base, ctx->mmap_size);
803 mutex_unlock(&ctx->ring_lock);
804 free_percpu(ctx->cpu);
805 percpu_ref_exit(&ctx->reqs);
806 percpu_ref_exit(&ctx->users);
807 kmem_cache_free(kioctx_cachep, ctx);
808 pr_debug("error allocating ioctx %d\n", err);
813 * Cancels all outstanding aio requests on an aio context. Used
814 * when the processes owning a context have all exited to encourage
815 * the rapid destruction of the kioctx.
817 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
818 struct ctx_rq_wait *wait)
820 struct kioctx_table *table;
822 spin_lock(&mm->ioctx_lock);
823 if (atomic_xchg(&ctx->dead, 1)) {
824 spin_unlock(&mm->ioctx_lock);
828 table = rcu_dereference_raw(mm->ioctx_table);
829 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
830 RCU_INIT_POINTER(table->table[ctx->id], NULL);
831 spin_unlock(&mm->ioctx_lock);
833 /* free_ioctx_reqs() will do the necessary RCU synchronization */
834 wake_up_all(&ctx->wait);
837 * It'd be more correct to do this in free_ioctx(), after all
838 * the outstanding kiocbs have finished - but by then io_destroy
839 * has already returned, so io_setup() could potentially return
840 * -EAGAIN with no ioctxs actually in use (as far as userspace
843 aio_nr_sub(ctx->max_reqs);
846 vm_munmap(ctx->mmap_base, ctx->mmap_size);
849 percpu_ref_kill(&ctx->users);
854 * exit_aio: called when the last user of mm goes away. At this point, there is
855 * no way for any new requests to be submited or any of the io_* syscalls to be
856 * called on the context.
858 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
861 void exit_aio(struct mm_struct *mm)
863 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
864 struct ctx_rq_wait wait;
870 atomic_set(&wait.count, table->nr);
871 init_completion(&wait.comp);
874 for (i = 0; i < table->nr; ++i) {
876 rcu_dereference_protected(table->table[i], true);
884 * We don't need to bother with munmap() here - exit_mmap(mm)
885 * is coming and it'll unmap everything. And we simply can't,
886 * this is not necessarily our ->mm.
887 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
888 * that it needs to unmap the area, just set it to 0.
891 kill_ioctx(mm, ctx, &wait);
894 if (!atomic_sub_and_test(skipped, &wait.count)) {
895 /* Wait until all IO for the context are done. */
896 wait_for_completion(&wait.comp);
899 RCU_INIT_POINTER(mm->ioctx_table, NULL);
903 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
905 struct kioctx_cpu *kcpu;
908 local_irq_save(flags);
909 kcpu = this_cpu_ptr(ctx->cpu);
910 kcpu->reqs_available += nr;
912 while (kcpu->reqs_available >= ctx->req_batch * 2) {
913 kcpu->reqs_available -= ctx->req_batch;
914 atomic_add(ctx->req_batch, &ctx->reqs_available);
917 local_irq_restore(flags);
920 static bool __get_reqs_available(struct kioctx *ctx)
922 struct kioctx_cpu *kcpu;
926 local_irq_save(flags);
927 kcpu = this_cpu_ptr(ctx->cpu);
928 if (!kcpu->reqs_available) {
929 int old, avail = atomic_read(&ctx->reqs_available);
932 if (avail < ctx->req_batch)
936 avail = atomic_cmpxchg(&ctx->reqs_available,
937 avail, avail - ctx->req_batch);
938 } while (avail != old);
940 kcpu->reqs_available += ctx->req_batch;
944 kcpu->reqs_available--;
946 local_irq_restore(flags);
950 /* refill_reqs_available
951 * Updates the reqs_available reference counts used for tracking the
952 * number of free slots in the completion ring. This can be called
953 * from aio_complete() (to optimistically update reqs_available) or
954 * from aio_get_req() (the we're out of events case). It must be
955 * called holding ctx->completion_lock.
957 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
960 unsigned events_in_ring, completed;
962 /* Clamp head since userland can write to it. */
963 head %= ctx->nr_events;
965 events_in_ring = tail - head;
967 events_in_ring = ctx->nr_events - (head - tail);
969 completed = ctx->completed_events;
970 if (events_in_ring < completed)
971 completed -= events_in_ring;
978 ctx->completed_events -= completed;
979 put_reqs_available(ctx, completed);
982 /* user_refill_reqs_available
983 * Called to refill reqs_available when aio_get_req() encounters an
984 * out of space in the completion ring.
986 static void user_refill_reqs_available(struct kioctx *ctx)
988 spin_lock_irq(&ctx->completion_lock);
989 if (ctx->completed_events) {
990 struct aio_ring *ring;
993 /* Access of ring->head may race with aio_read_events_ring()
994 * here, but that's okay since whether we read the old version
995 * or the new version, and either will be valid. The important
996 * part is that head cannot pass tail since we prevent
997 * aio_complete() from updating tail by holding
998 * ctx->completion_lock. Even if head is invalid, the check
999 * against ctx->completed_events below will make sure we do the
1002 ring = kmap_atomic(ctx->ring_pages[0]);
1004 kunmap_atomic(ring);
1006 refill_reqs_available(ctx, head, ctx->tail);
1009 spin_unlock_irq(&ctx->completion_lock);
1012 static bool get_reqs_available(struct kioctx *ctx)
1014 if (__get_reqs_available(ctx))
1016 user_refill_reqs_available(ctx);
1017 return __get_reqs_available(ctx);
1021 * Allocate a slot for an aio request.
1022 * Returns NULL if no requests are free.
1024 * The refcount is initialized to 2 - one for the async op completion,
1025 * one for the synchronous code that does this.
1027 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1029 struct aio_kiocb *req;
1031 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1035 if (unlikely(!get_reqs_available(ctx))) {
1036 kmem_cache_free(kiocb_cachep, req);
1040 percpu_ref_get(&ctx->reqs);
1042 INIT_LIST_HEAD(&req->ki_list);
1043 refcount_set(&req->ki_refcnt, 2);
1044 req->ki_eventfd = NULL;
1048 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1050 struct aio_ring __user *ring = (void __user *)ctx_id;
1051 struct mm_struct *mm = current->mm;
1052 struct kioctx *ctx, *ret = NULL;
1053 struct kioctx_table *table;
1056 if (get_user(id, &ring->id))
1060 table = rcu_dereference(mm->ioctx_table);
1062 if (!table || id >= table->nr)
1065 id = array_index_nospec(id, table->nr);
1066 ctx = rcu_dereference(table->table[id]);
1067 if (ctx && ctx->user_id == ctx_id) {
1068 if (percpu_ref_tryget_live(&ctx->users))
1076 static inline void iocb_destroy(struct aio_kiocb *iocb)
1078 if (iocb->ki_eventfd)
1079 eventfd_ctx_put(iocb->ki_eventfd);
1081 fput(iocb->ki_filp);
1082 percpu_ref_put(&iocb->ki_ctx->reqs);
1083 kmem_cache_free(kiocb_cachep, iocb);
1087 * Called when the io request on the given iocb is complete.
1089 static void aio_complete(struct aio_kiocb *iocb)
1091 struct kioctx *ctx = iocb->ki_ctx;
1092 struct aio_ring *ring;
1093 struct io_event *ev_page, *event;
1094 unsigned tail, pos, head;
1095 unsigned long flags;
1098 * Add a completion event to the ring buffer. Must be done holding
1099 * ctx->completion_lock to prevent other code from messing with the tail
1100 * pointer since we might be called from irq context.
1102 spin_lock_irqsave(&ctx->completion_lock, flags);
1105 pos = tail + AIO_EVENTS_OFFSET;
1107 if (++tail >= ctx->nr_events)
1110 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1111 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1113 *event = iocb->ki_res;
1115 kunmap_atomic(ev_page);
1116 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1118 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1119 (void __user *)(unsigned long)iocb->ki_res.obj,
1120 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1122 /* after flagging the request as done, we
1123 * must never even look at it again
1125 smp_wmb(); /* make event visible before updating tail */
1129 ring = kmap_atomic(ctx->ring_pages[0]);
1132 kunmap_atomic(ring);
1133 flush_dcache_page(ctx->ring_pages[0]);
1135 ctx->completed_events++;
1136 if (ctx->completed_events > 1)
1137 refill_reqs_available(ctx, head, tail);
1138 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1140 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1143 * Check if the user asked us to deliver the result through an
1144 * eventfd. The eventfd_signal() function is safe to be called
1147 if (iocb->ki_eventfd)
1148 eventfd_signal(iocb->ki_eventfd, 1);
1151 * We have to order our ring_info tail store above and test
1152 * of the wait list below outside the wait lock. This is
1153 * like in wake_up_bit() where clearing a bit has to be
1154 * ordered with the unlocked test.
1158 if (waitqueue_active(&ctx->wait))
1159 wake_up(&ctx->wait);
1162 static inline void iocb_put(struct aio_kiocb *iocb)
1164 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1170 /* aio_read_events_ring
1171 * Pull an event off of the ioctx's event ring. Returns the number of
1174 static long aio_read_events_ring(struct kioctx *ctx,
1175 struct io_event __user *event, long nr)
1177 struct aio_ring *ring;
1178 unsigned head, tail, pos;
1183 * The mutex can block and wake us up and that will cause
1184 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1185 * and repeat. This should be rare enough that it doesn't cause
1186 * peformance issues. See the comment in read_events() for more detail.
1188 sched_annotate_sleep();
1189 mutex_lock(&ctx->ring_lock);
1191 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1192 ring = kmap_atomic(ctx->ring_pages[0]);
1195 kunmap_atomic(ring);
1198 * Ensure that once we've read the current tail pointer, that
1199 * we also see the events that were stored up to the tail.
1203 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1208 head %= ctx->nr_events;
1209 tail %= ctx->nr_events;
1213 struct io_event *ev;
1216 avail = (head <= tail ? tail : ctx->nr_events) - head;
1220 pos = head + AIO_EVENTS_OFFSET;
1221 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1222 pos %= AIO_EVENTS_PER_PAGE;
1224 avail = min(avail, nr - ret);
1225 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1228 copy_ret = copy_to_user(event + ret, ev + pos,
1229 sizeof(*ev) * avail);
1232 if (unlikely(copy_ret)) {
1239 head %= ctx->nr_events;
1242 ring = kmap_atomic(ctx->ring_pages[0]);
1244 kunmap_atomic(ring);
1245 flush_dcache_page(ctx->ring_pages[0]);
1247 pr_debug("%li h%u t%u\n", ret, head, tail);
1249 mutex_unlock(&ctx->ring_lock);
1254 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1255 struct io_event __user *event, long *i)
1257 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1262 if (unlikely(atomic_read(&ctx->dead)))
1268 return ret < 0 || *i >= min_nr;
1271 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1272 struct io_event __user *event,
1278 * Note that aio_read_events() is being called as the conditional - i.e.
1279 * we're calling it after prepare_to_wait() has set task state to
1280 * TASK_INTERRUPTIBLE.
1282 * But aio_read_events() can block, and if it blocks it's going to flip
1283 * the task state back to TASK_RUNNING.
1285 * This should be ok, provided it doesn't flip the state back to
1286 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1287 * will only happen if the mutex_lock() call blocks, and we then find
1288 * the ringbuffer empty. So in practice we should be ok, but it's
1289 * something to be aware of when touching this code.
1292 aio_read_events(ctx, min_nr, nr, event, &ret);
1294 wait_event_interruptible_hrtimeout(ctx->wait,
1295 aio_read_events(ctx, min_nr, nr, event, &ret),
1301 * Create an aio_context capable of receiving at least nr_events.
1302 * ctxp must not point to an aio_context that already exists, and
1303 * must be initialized to 0 prior to the call. On successful
1304 * creation of the aio_context, *ctxp is filled in with the resulting
1305 * handle. May fail with -EINVAL if *ctxp is not initialized,
1306 * if the specified nr_events exceeds internal limits. May fail
1307 * with -EAGAIN if the specified nr_events exceeds the user's limit
1308 * of available events. May fail with -ENOMEM if insufficient kernel
1309 * resources are available. May fail with -EFAULT if an invalid
1310 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1313 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1315 struct kioctx *ioctx = NULL;
1319 ret = get_user(ctx, ctxp);
1324 if (unlikely(ctx || nr_events == 0)) {
1325 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1330 ioctx = ioctx_alloc(nr_events);
1331 ret = PTR_ERR(ioctx);
1332 if (!IS_ERR(ioctx)) {
1333 ret = put_user(ioctx->user_id, ctxp);
1335 kill_ioctx(current->mm, ioctx, NULL);
1336 percpu_ref_put(&ioctx->users);
1343 #ifdef CONFIG_COMPAT
1344 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1346 struct kioctx *ioctx = NULL;
1350 ret = get_user(ctx, ctx32p);
1355 if (unlikely(ctx || nr_events == 0)) {
1356 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1361 ioctx = ioctx_alloc(nr_events);
1362 ret = PTR_ERR(ioctx);
1363 if (!IS_ERR(ioctx)) {
1364 /* truncating is ok because it's a user address */
1365 ret = put_user((u32)ioctx->user_id, ctx32p);
1367 kill_ioctx(current->mm, ioctx, NULL);
1368 percpu_ref_put(&ioctx->users);
1377 * Destroy the aio_context specified. May cancel any outstanding
1378 * AIOs and block on completion. Will fail with -ENOSYS if not
1379 * implemented. May fail with -EINVAL if the context pointed to
1382 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1384 struct kioctx *ioctx = lookup_ioctx(ctx);
1385 if (likely(NULL != ioctx)) {
1386 struct ctx_rq_wait wait;
1389 init_completion(&wait.comp);
1390 atomic_set(&wait.count, 1);
1392 /* Pass requests_done to kill_ioctx() where it can be set
1393 * in a thread-safe way. If we try to set it here then we have
1394 * a race condition if two io_destroy() called simultaneously.
1396 ret = kill_ioctx(current->mm, ioctx, &wait);
1397 percpu_ref_put(&ioctx->users);
1399 /* Wait until all IO for the context are done. Otherwise kernel
1400 * keep using user-space buffers even if user thinks the context
1404 wait_for_completion(&wait.comp);
1408 pr_debug("EINVAL: invalid context id\n");
1412 static void aio_remove_iocb(struct aio_kiocb *iocb)
1414 struct kioctx *ctx = iocb->ki_ctx;
1415 unsigned long flags;
1417 spin_lock_irqsave(&ctx->ctx_lock, flags);
1418 list_del(&iocb->ki_list);
1419 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1422 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1424 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1426 if (!list_empty_careful(&iocb->ki_list))
1427 aio_remove_iocb(iocb);
1429 if (kiocb->ki_flags & IOCB_WRITE) {
1430 struct inode *inode = file_inode(kiocb->ki_filp);
1433 * Tell lockdep we inherited freeze protection from submission
1436 if (S_ISREG(inode->i_mode))
1437 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1438 file_end_write(kiocb->ki_filp);
1441 iocb->ki_res.res = res;
1442 iocb->ki_res.res2 = res2;
1446 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1450 req->ki_complete = aio_complete_rw;
1451 req->private = NULL;
1452 req->ki_pos = iocb->aio_offset;
1453 req->ki_flags = iocb_flags(req->ki_filp);
1454 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1455 req->ki_flags |= IOCB_EVENTFD;
1456 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1457 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1459 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1460 * aio_reqprio is interpreted as an I/O scheduling
1461 * class and priority.
1463 ret = ioprio_check_cap(iocb->aio_reqprio);
1465 pr_debug("aio ioprio check cap error: %d\n", ret);
1469 req->ki_ioprio = iocb->aio_reqprio;
1471 req->ki_ioprio = get_current_ioprio();
1473 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1477 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1481 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1482 struct iovec **iovec, bool vectored, bool compat,
1483 struct iov_iter *iter)
1485 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1486 size_t len = iocb->aio_nbytes;
1489 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1494 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1497 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1503 case -ERESTARTNOINTR:
1504 case -ERESTARTNOHAND:
1505 case -ERESTART_RESTARTBLOCK:
1507 * There's no easy way to restart the syscall since other AIO's
1508 * may be already running. Just fail this IO with EINTR.
1513 req->ki_complete(req, ret, 0);
1517 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1518 bool vectored, bool compat)
1520 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1521 struct iov_iter iter;
1525 ret = aio_prep_rw(req, iocb);
1528 file = req->ki_filp;
1529 if (unlikely(!(file->f_mode & FMODE_READ)))
1532 if (unlikely(!file->f_op->read_iter))
1535 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1538 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1540 aio_rw_done(req, call_read_iter(file, req, &iter));
1545 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1546 bool vectored, bool compat)
1548 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1549 struct iov_iter iter;
1553 ret = aio_prep_rw(req, iocb);
1556 file = req->ki_filp;
1558 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1560 if (unlikely(!file->f_op->write_iter))
1563 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1566 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1569 * Open-code file_start_write here to grab freeze protection,
1570 * which will be released by another thread in
1571 * aio_complete_rw(). Fool lockdep by telling it the lock got
1572 * released so that it doesn't complain about the held lock when
1573 * we return to userspace.
1575 if (S_ISREG(file_inode(file)->i_mode)) {
1576 sb_start_write(file_inode(file)->i_sb);
1577 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1579 req->ki_flags |= IOCB_WRITE;
1580 aio_rw_done(req, call_write_iter(file, req, &iter));
1586 static void aio_fsync_work(struct work_struct *work)
1588 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1589 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1591 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1592 revert_creds(old_cred);
1593 put_cred(iocb->fsync.creds);
1597 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1600 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1601 iocb->aio_rw_flags))
1604 if (unlikely(!req->file->f_op->fsync))
1607 req->creds = prepare_creds();
1611 req->datasync = datasync;
1612 INIT_WORK(&req->work, aio_fsync_work);
1613 schedule_work(&req->work);
1617 static void aio_poll_put_work(struct work_struct *work)
1619 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1620 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1626 * Safely lock the waitqueue which the request is on, synchronizing with the
1627 * case where the ->poll() provider decides to free its waitqueue early.
1629 * Returns true on success, meaning that req->head->lock was locked, req->wait
1630 * is on req->head, and an RCU read lock was taken. Returns false if the
1631 * request was already removed from its waitqueue (which might no longer exist).
1633 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1635 wait_queue_head_t *head;
1638 * While we hold the waitqueue lock and the waitqueue is nonempty,
1639 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1640 * lock in the first place can race with the waitqueue being freed.
1642 * We solve this as eventpoll does: by taking advantage of the fact that
1643 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1644 * we enter rcu_read_lock() and see that the pointer to the queue is
1645 * non-NULL, we can then lock it without the memory being freed out from
1646 * under us, then check whether the request is still on the queue.
1648 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1649 * case the caller deletes the entry from the queue, leaving it empty.
1650 * In that case, only RCU prevents the queue memory from being freed.
1653 head = smp_load_acquire(&req->head);
1655 spin_lock(&head->lock);
1656 if (!list_empty(&req->wait.entry))
1658 spin_unlock(&head->lock);
1664 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1666 spin_unlock(&req->head->lock);
1670 static void aio_poll_complete_work(struct work_struct *work)
1672 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1673 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1674 struct poll_table_struct pt = { ._key = req->events };
1675 struct kioctx *ctx = iocb->ki_ctx;
1678 if (!READ_ONCE(req->cancelled))
1679 mask = vfs_poll(req->file, &pt) & req->events;
1682 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1683 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1684 * synchronize with them. In the cancellation case the list_del_init
1685 * itself is not actually needed, but harmless so we keep it in to
1686 * avoid further branches in the fast path.
1688 spin_lock_irq(&ctx->ctx_lock);
1689 if (poll_iocb_lock_wq(req)) {
1690 if (!mask && !READ_ONCE(req->cancelled)) {
1692 * The request isn't actually ready to be completed yet.
1693 * Reschedule completion if another wakeup came in.
1695 if (req->work_need_resched) {
1696 schedule_work(&req->work);
1697 req->work_need_resched = false;
1699 req->work_scheduled = false;
1701 poll_iocb_unlock_wq(req);
1702 spin_unlock_irq(&ctx->ctx_lock);
1705 list_del_init(&req->wait.entry);
1706 poll_iocb_unlock_wq(req);
1707 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1708 list_del_init(&iocb->ki_list);
1709 iocb->ki_res.res = mangle_poll(mask);
1710 spin_unlock_irq(&ctx->ctx_lock);
1715 /* assumes we are called with irqs disabled */
1716 static int aio_poll_cancel(struct kiocb *iocb)
1718 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1719 struct poll_iocb *req = &aiocb->poll;
1721 if (poll_iocb_lock_wq(req)) {
1722 WRITE_ONCE(req->cancelled, true);
1723 if (!req->work_scheduled) {
1724 schedule_work(&aiocb->poll.work);
1725 req->work_scheduled = true;
1727 poll_iocb_unlock_wq(req);
1728 } /* else, the request was force-cancelled by POLLFREE already */
1733 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1736 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1737 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1738 __poll_t mask = key_to_poll(key);
1739 unsigned long flags;
1741 /* for instances that support it check for an event match first: */
1742 if (mask && !(mask & req->events))
1746 * Complete the request inline if possible. This requires that three
1747 * conditions be met:
1748 * 1. An event mask must have been passed. If a plain wakeup was done
1749 * instead, then mask == 0 and we have to call vfs_poll() to get
1750 * the events, so inline completion isn't possible.
1751 * 2. The completion work must not have already been scheduled.
1752 * 3. ctx_lock must not be busy. We have to use trylock because we
1753 * already hold the waitqueue lock, so this inverts the normal
1754 * locking order. Use irqsave/irqrestore because not all
1755 * filesystems (e.g. fuse) call this function with IRQs disabled,
1756 * yet IRQs have to be disabled before ctx_lock is obtained.
1758 if (mask && !req->work_scheduled &&
1759 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1760 struct kioctx *ctx = iocb->ki_ctx;
1762 list_del_init(&req->wait.entry);
1763 list_del(&iocb->ki_list);
1764 iocb->ki_res.res = mangle_poll(mask);
1765 if (iocb->ki_eventfd && eventfd_signal_count()) {
1767 INIT_WORK(&req->work, aio_poll_put_work);
1768 schedule_work(&req->work);
1770 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1775 * Schedule the completion work if needed. If it was already
1776 * scheduled, record that another wakeup came in.
1778 * Don't remove the request from the waitqueue here, as it might
1779 * not actually be complete yet (we won't know until vfs_poll()
1780 * is called), and we must not miss any wakeups. POLLFREE is an
1781 * exception to this; see below.
1783 if (req->work_scheduled) {
1784 req->work_need_resched = true;
1786 schedule_work(&req->work);
1787 req->work_scheduled = true;
1791 * If the waitqueue is being freed early but we can't complete
1792 * the request inline, we have to tear down the request as best
1793 * we can. That means immediately removing the request from its
1794 * waitqueue and preventing all further accesses to the
1795 * waitqueue via the request. We also need to schedule the
1796 * completion work (done above). Also mark the request as
1797 * cancelled, to potentially skip an unneeded call to ->poll().
1799 if (mask & POLLFREE) {
1800 WRITE_ONCE(req->cancelled, true);
1801 list_del_init(&req->wait.entry);
1804 * Careful: this *must* be the last step, since as soon
1805 * as req->head is NULL'ed out, the request can be
1806 * completed and freed, since aio_poll_complete_work()
1807 * will no longer need to take the waitqueue lock.
1809 smp_store_release(&req->head, NULL);
1815 struct aio_poll_table {
1816 struct poll_table_struct pt;
1817 struct aio_kiocb *iocb;
1823 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1824 struct poll_table_struct *p)
1826 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1828 /* multiple wait queues per file are not supported */
1829 if (unlikely(pt->queued)) {
1830 pt->error = -EINVAL;
1836 pt->iocb->poll.head = head;
1837 add_wait_queue(head, &pt->iocb->poll.wait);
1840 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1842 struct kioctx *ctx = aiocb->ki_ctx;
1843 struct poll_iocb *req = &aiocb->poll;
1844 struct aio_poll_table apt;
1845 bool cancel = false;
1848 /* reject any unknown events outside the normal event mask. */
1849 if ((u16)iocb->aio_buf != iocb->aio_buf)
1851 /* reject fields that are not defined for poll */
1852 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1855 INIT_WORK(&req->work, aio_poll_complete_work);
1856 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1859 req->cancelled = false;
1860 req->work_scheduled = false;
1861 req->work_need_resched = false;
1863 apt.pt._qproc = aio_poll_queue_proc;
1864 apt.pt._key = req->events;
1867 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1869 /* initialized the list so that we can do list_empty checks */
1870 INIT_LIST_HEAD(&req->wait.entry);
1871 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1873 mask = vfs_poll(req->file, &apt.pt) & req->events;
1874 spin_lock_irq(&ctx->ctx_lock);
1875 if (likely(apt.queued)) {
1876 bool on_queue = poll_iocb_lock_wq(req);
1878 if (!on_queue || req->work_scheduled) {
1880 * aio_poll_wake() already either scheduled the async
1881 * completion work, or completed the request inline.
1883 if (apt.error) /* unsupported case: multiple queues */
1888 if (mask || apt.error) {
1889 /* Steal to complete synchronously. */
1890 list_del_init(&req->wait.entry);
1891 } else if (cancel) {
1892 /* Cancel if possible (may be too late though). */
1893 WRITE_ONCE(req->cancelled, true);
1894 } else if (on_queue) {
1896 * Actually waiting for an event, so add the request to
1897 * active_reqs so that it can be cancelled if needed.
1899 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1900 aiocb->ki_cancel = aio_poll_cancel;
1903 poll_iocb_unlock_wq(req);
1905 if (mask) { /* no async, we'd stolen it */
1906 aiocb->ki_res.res = mangle_poll(mask);
1909 spin_unlock_irq(&ctx->ctx_lock);
1915 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1916 struct iocb __user *user_iocb, struct aio_kiocb *req,
1919 req->ki_filp = fget(iocb->aio_fildes);
1920 if (unlikely(!req->ki_filp))
1923 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1924 struct eventfd_ctx *eventfd;
1926 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1927 * instance of the file* now. The file descriptor must be
1928 * an eventfd() fd, and will be signaled for each completed
1929 * event using the eventfd_signal() function.
1931 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1932 if (IS_ERR(eventfd))
1933 return PTR_ERR(eventfd);
1935 req->ki_eventfd = eventfd;
1938 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1939 pr_debug("EFAULT: aio_key\n");
1943 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1944 req->ki_res.data = iocb->aio_data;
1945 req->ki_res.res = 0;
1946 req->ki_res.res2 = 0;
1948 switch (iocb->aio_lio_opcode) {
1949 case IOCB_CMD_PREAD:
1950 return aio_read(&req->rw, iocb, false, compat);
1951 case IOCB_CMD_PWRITE:
1952 return aio_write(&req->rw, iocb, false, compat);
1953 case IOCB_CMD_PREADV:
1954 return aio_read(&req->rw, iocb, true, compat);
1955 case IOCB_CMD_PWRITEV:
1956 return aio_write(&req->rw, iocb, true, compat);
1957 case IOCB_CMD_FSYNC:
1958 return aio_fsync(&req->fsync, iocb, false);
1959 case IOCB_CMD_FDSYNC:
1960 return aio_fsync(&req->fsync, iocb, true);
1962 return aio_poll(req, iocb);
1964 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1969 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1972 struct aio_kiocb *req;
1976 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1979 /* enforce forwards compatibility on users */
1980 if (unlikely(iocb.aio_reserved2)) {
1981 pr_debug("EINVAL: reserve field set\n");
1985 /* prevent overflows */
1987 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1988 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1989 ((ssize_t)iocb.aio_nbytes < 0)
1991 pr_debug("EINVAL: overflow check\n");
1995 req = aio_get_req(ctx);
1999 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2001 /* Done with the synchronous reference */
2005 * If err is 0, we'd either done aio_complete() ourselves or have
2006 * arranged for that to be done asynchronously. Anything non-zero
2007 * means that we need to destroy req ourselves.
2009 if (unlikely(err)) {
2011 put_reqs_available(ctx, 1);
2017 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2018 * the number of iocbs queued. May return -EINVAL if the aio_context
2019 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2020 * *iocbpp[0] is not properly initialized, if the operation specified
2021 * is invalid for the file descriptor in the iocb. May fail with
2022 * -EFAULT if any of the data structures point to invalid data. May
2023 * fail with -EBADF if the file descriptor specified in the first
2024 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2025 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2026 * fail with -ENOSYS if not implemented.
2028 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2029 struct iocb __user * __user *, iocbpp)
2034 struct blk_plug plug;
2036 if (unlikely(nr < 0))
2039 ctx = lookup_ioctx(ctx_id);
2040 if (unlikely(!ctx)) {
2041 pr_debug("EINVAL: invalid context id\n");
2045 if (nr > ctx->nr_events)
2046 nr = ctx->nr_events;
2048 if (nr > AIO_PLUG_THRESHOLD)
2049 blk_start_plug(&plug);
2050 for (i = 0; i < nr; i++) {
2051 struct iocb __user *user_iocb;
2053 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2058 ret = io_submit_one(ctx, user_iocb, false);
2062 if (nr > AIO_PLUG_THRESHOLD)
2063 blk_finish_plug(&plug);
2065 percpu_ref_put(&ctx->users);
2069 #ifdef CONFIG_COMPAT
2070 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2071 int, nr, compat_uptr_t __user *, iocbpp)
2076 struct blk_plug plug;
2078 if (unlikely(nr < 0))
2081 ctx = lookup_ioctx(ctx_id);
2082 if (unlikely(!ctx)) {
2083 pr_debug("EINVAL: invalid context id\n");
2087 if (nr > ctx->nr_events)
2088 nr = ctx->nr_events;
2090 if (nr > AIO_PLUG_THRESHOLD)
2091 blk_start_plug(&plug);
2092 for (i = 0; i < nr; i++) {
2093 compat_uptr_t user_iocb;
2095 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2100 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2104 if (nr > AIO_PLUG_THRESHOLD)
2105 blk_finish_plug(&plug);
2107 percpu_ref_put(&ctx->users);
2113 * Attempts to cancel an iocb previously passed to io_submit. If
2114 * the operation is successfully cancelled, the resulting event is
2115 * copied into the memory pointed to by result without being placed
2116 * into the completion queue and 0 is returned. May fail with
2117 * -EFAULT if any of the data structures pointed to are invalid.
2118 * May fail with -EINVAL if aio_context specified by ctx_id is
2119 * invalid. May fail with -EAGAIN if the iocb specified was not
2120 * cancelled. Will fail with -ENOSYS if not implemented.
2122 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2123 struct io_event __user *, result)
2126 struct aio_kiocb *kiocb;
2129 u64 obj = (u64)(unsigned long)iocb;
2131 if (unlikely(get_user(key, &iocb->aio_key)))
2133 if (unlikely(key != KIOCB_KEY))
2136 ctx = lookup_ioctx(ctx_id);
2140 spin_lock_irq(&ctx->ctx_lock);
2141 /* TODO: use a hash or array, this sucks. */
2142 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2143 if (kiocb->ki_res.obj == obj) {
2144 ret = kiocb->ki_cancel(&kiocb->rw);
2145 list_del_init(&kiocb->ki_list);
2149 spin_unlock_irq(&ctx->ctx_lock);
2153 * The result argument is no longer used - the io_event is
2154 * always delivered via the ring buffer. -EINPROGRESS indicates
2155 * cancellation is progress:
2160 percpu_ref_put(&ctx->users);
2165 static long do_io_getevents(aio_context_t ctx_id,
2168 struct io_event __user *events,
2169 struct timespec64 *ts)
2171 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2172 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2175 if (likely(ioctx)) {
2176 if (likely(min_nr <= nr && min_nr >= 0))
2177 ret = read_events(ioctx, min_nr, nr, events, until);
2178 percpu_ref_put(&ioctx->users);
2185 * Attempts to read at least min_nr events and up to nr events from
2186 * the completion queue for the aio_context specified by ctx_id. If
2187 * it succeeds, the number of read events is returned. May fail with
2188 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2189 * out of range, if timeout is out of range. May fail with -EFAULT
2190 * if any of the memory specified is invalid. May return 0 or
2191 * < min_nr if the timeout specified by timeout has elapsed
2192 * before sufficient events are available, where timeout == NULL
2193 * specifies an infinite timeout. Note that the timeout pointed to by
2194 * timeout is relative. Will fail with -ENOSYS if not implemented.
2198 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2201 struct io_event __user *, events,
2202 struct __kernel_timespec __user *, timeout)
2204 struct timespec64 ts;
2207 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2210 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2211 if (!ret && signal_pending(current))
2218 struct __aio_sigset {
2219 const sigset_t __user *sigmask;
2223 SYSCALL_DEFINE6(io_pgetevents,
2224 aio_context_t, ctx_id,
2227 struct io_event __user *, events,
2228 struct __kernel_timespec __user *, timeout,
2229 const struct __aio_sigset __user *, usig)
2231 struct __aio_sigset ksig = { NULL, };
2232 struct timespec64 ts;
2236 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2239 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2242 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2246 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2248 interrupted = signal_pending(current);
2249 restore_saved_sigmask_unless(interrupted);
2250 if (interrupted && !ret)
2251 ret = -ERESTARTNOHAND;
2256 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2258 SYSCALL_DEFINE6(io_pgetevents_time32,
2259 aio_context_t, ctx_id,
2262 struct io_event __user *, events,
2263 struct old_timespec32 __user *, timeout,
2264 const struct __aio_sigset __user *, usig)
2266 struct __aio_sigset ksig = { NULL, };
2267 struct timespec64 ts;
2271 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2274 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2278 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2282 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2284 interrupted = signal_pending(current);
2285 restore_saved_sigmask_unless(interrupted);
2286 if (interrupted && !ret)
2287 ret = -ERESTARTNOHAND;
2294 #if defined(CONFIG_COMPAT_32BIT_TIME)
2296 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2299 struct io_event __user *, events,
2300 struct old_timespec32 __user *, timeout)
2302 struct timespec64 t;
2305 if (timeout && get_old_timespec32(&t, timeout))
2308 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2309 if (!ret && signal_pending(current))
2316 #ifdef CONFIG_COMPAT
2318 struct __compat_aio_sigset {
2319 compat_uptr_t sigmask;
2320 compat_size_t sigsetsize;
2323 #if defined(CONFIG_COMPAT_32BIT_TIME)
2325 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2326 compat_aio_context_t, ctx_id,
2327 compat_long_t, min_nr,
2329 struct io_event __user *, events,
2330 struct old_timespec32 __user *, timeout,
2331 const struct __compat_aio_sigset __user *, usig)
2333 struct __compat_aio_sigset ksig = { 0, };
2334 struct timespec64 t;
2338 if (timeout && get_old_timespec32(&t, timeout))
2341 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2344 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2348 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2350 interrupted = signal_pending(current);
2351 restore_saved_sigmask_unless(interrupted);
2352 if (interrupted && !ret)
2353 ret = -ERESTARTNOHAND;
2360 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2361 compat_aio_context_t, ctx_id,
2362 compat_long_t, min_nr,
2364 struct io_event __user *, events,
2365 struct __kernel_timespec __user *, timeout,
2366 const struct __compat_aio_sigset __user *, usig)
2368 struct __compat_aio_sigset ksig = { 0, };
2369 struct timespec64 t;
2373 if (timeout && get_timespec64(&t, timeout))
2376 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2379 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2383 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2385 interrupted = signal_pending(current);
2386 restore_saved_sigmask_unless(interrupted);
2387 if (interrupted && !ret)
2388 ret = -ERESTARTNOHAND;