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);
341 for (i = 0; i < table->nr; i++) {
344 ctx = rcu_dereference(table->table[i]);
345 if (ctx && ctx->aio_ring_file == file) {
346 if (!atomic_read(&ctx->dead)) {
347 ctx->user_id = ctx->mmap_base = vma->vm_start;
356 spin_unlock(&mm->ioctx_lock);
360 static const struct vm_operations_struct aio_ring_vm_ops = {
361 .mremap = aio_ring_mremap,
362 #if IS_ENABLED(CONFIG_MMU)
363 .fault = filemap_fault,
364 .map_pages = filemap_map_pages,
365 .page_mkwrite = filemap_page_mkwrite,
369 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
371 vma->vm_flags |= VM_DONTEXPAND;
372 vma->vm_ops = &aio_ring_vm_ops;
376 static const struct file_operations aio_ring_fops = {
377 .mmap = aio_ring_mmap,
380 #if IS_ENABLED(CONFIG_MIGRATION)
381 static int aio_migratepage(struct address_space *mapping, struct page *new,
382 struct page *old, enum migrate_mode mode)
390 * We cannot support the _NO_COPY case here, because copy needs to
391 * happen under the ctx->completion_lock. That does not work with the
392 * migration workflow of MIGRATE_SYNC_NO_COPY.
394 if (mode == MIGRATE_SYNC_NO_COPY)
399 /* mapping->private_lock here protects against the kioctx teardown. */
400 spin_lock(&mapping->private_lock);
401 ctx = mapping->private_data;
407 /* The ring_lock mutex. The prevents aio_read_events() from writing
408 * to the ring's head, and prevents page migration from mucking in
409 * a partially initialized kiotx.
411 if (!mutex_trylock(&ctx->ring_lock)) {
417 if (idx < (pgoff_t)ctx->nr_pages) {
418 /* Make sure the old page hasn't already been changed */
419 if (ctx->ring_pages[idx] != old)
427 /* Writeback must be complete */
428 BUG_ON(PageWriteback(old));
431 rc = migrate_page_move_mapping(mapping, new, old, 1);
432 if (rc != MIGRATEPAGE_SUCCESS) {
437 /* Take completion_lock to prevent other writes to the ring buffer
438 * while the old page is copied to the new. This prevents new
439 * events from being lost.
441 spin_lock_irqsave(&ctx->completion_lock, flags);
442 migrate_page_copy(new, old);
443 BUG_ON(ctx->ring_pages[idx] != old);
444 ctx->ring_pages[idx] = new;
445 spin_unlock_irqrestore(&ctx->completion_lock, flags);
447 /* The old page is no longer accessible. */
451 mutex_unlock(&ctx->ring_lock);
453 spin_unlock(&mapping->private_lock);
458 static const struct address_space_operations aio_ctx_aops = {
459 .set_page_dirty = __set_page_dirty_no_writeback,
460 #if IS_ENABLED(CONFIG_MIGRATION)
461 .migratepage = aio_migratepage,
465 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
467 struct aio_ring *ring;
468 struct mm_struct *mm = current->mm;
469 unsigned long size, unused;
474 /* Compensate for the ring buffer's head/tail overlap entry */
475 nr_events += 2; /* 1 is required, 2 for good luck */
477 size = sizeof(struct aio_ring);
478 size += sizeof(struct io_event) * nr_events;
480 nr_pages = PFN_UP(size);
484 file = aio_private_file(ctx, nr_pages);
486 ctx->aio_ring_file = NULL;
490 ctx->aio_ring_file = file;
491 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
492 / sizeof(struct io_event);
494 ctx->ring_pages = ctx->internal_pages;
495 if (nr_pages > AIO_RING_PAGES) {
496 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
498 if (!ctx->ring_pages) {
499 put_aio_ring_file(ctx);
504 for (i = 0; i < nr_pages; i++) {
506 page = find_or_create_page(file->f_mapping,
507 i, GFP_HIGHUSER | __GFP_ZERO);
510 pr_debug("pid(%d) page[%d]->count=%d\n",
511 current->pid, i, page_count(page));
512 SetPageUptodate(page);
515 ctx->ring_pages[i] = page;
519 if (unlikely(i != nr_pages)) {
524 ctx->mmap_size = nr_pages * PAGE_SIZE;
525 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
527 if (mmap_write_lock_killable(mm)) {
533 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
534 PROT_READ | PROT_WRITE,
535 MAP_SHARED, 0, &unused, NULL);
536 mmap_write_unlock(mm);
537 if (IS_ERR((void *)ctx->mmap_base)) {
543 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
545 ctx->user_id = ctx->mmap_base;
546 ctx->nr_events = nr_events; /* trusted copy */
548 ring = kmap_atomic(ctx->ring_pages[0]);
549 ring->nr = nr_events; /* user copy */
551 ring->head = ring->tail = 0;
552 ring->magic = AIO_RING_MAGIC;
553 ring->compat_features = AIO_RING_COMPAT_FEATURES;
554 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
555 ring->header_length = sizeof(struct aio_ring);
557 flush_dcache_page(ctx->ring_pages[0]);
562 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
563 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
564 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
566 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
568 struct aio_kiocb *req;
573 * kiocb didn't come from aio or is neither a read nor a write, hence
576 if (!(iocb->ki_flags & IOCB_AIO_RW))
579 req = container_of(iocb, struct aio_kiocb, rw);
581 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
586 spin_lock_irqsave(&ctx->ctx_lock, flags);
587 list_add_tail(&req->ki_list, &ctx->active_reqs);
588 req->ki_cancel = cancel;
589 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
591 EXPORT_SYMBOL(kiocb_set_cancel_fn);
594 * free_ioctx() should be RCU delayed to synchronize against the RCU
595 * protected lookup_ioctx() and also needs process context to call
596 * aio_free_ring(). Use rcu_work.
598 static void free_ioctx(struct work_struct *work)
600 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
602 pr_debug("freeing %p\n", ctx);
605 free_percpu(ctx->cpu);
606 percpu_ref_exit(&ctx->reqs);
607 percpu_ref_exit(&ctx->users);
608 kmem_cache_free(kioctx_cachep, ctx);
611 static void free_ioctx_reqs(struct percpu_ref *ref)
613 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
615 /* At this point we know that there are no any in-flight requests */
616 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
617 complete(&ctx->rq_wait->comp);
619 /* Synchronize against RCU protected table->table[] dereferences */
620 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
621 queue_rcu_work(system_wq, &ctx->free_rwork);
625 * When this function runs, the kioctx has been removed from the "hash table"
626 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
627 * now it's safe to cancel any that need to be.
629 static void free_ioctx_users(struct percpu_ref *ref)
631 struct kioctx *ctx = container_of(ref, struct kioctx, users);
632 struct aio_kiocb *req;
634 spin_lock_irq(&ctx->ctx_lock);
636 while (!list_empty(&ctx->active_reqs)) {
637 req = list_first_entry(&ctx->active_reqs,
638 struct aio_kiocb, ki_list);
639 req->ki_cancel(&req->rw);
640 list_del_init(&req->ki_list);
643 spin_unlock_irq(&ctx->ctx_lock);
645 percpu_ref_kill(&ctx->reqs);
646 percpu_ref_put(&ctx->reqs);
649 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
652 struct kioctx_table *table, *old;
653 struct aio_ring *ring;
655 spin_lock(&mm->ioctx_lock);
656 table = rcu_dereference_raw(mm->ioctx_table);
660 for (i = 0; i < table->nr; i++)
661 if (!rcu_access_pointer(table->table[i])) {
663 rcu_assign_pointer(table->table[i], ctx);
664 spin_unlock(&mm->ioctx_lock);
666 /* While kioctx setup is in progress,
667 * we are protected from page migration
668 * changes ring_pages by ->ring_lock.
670 ring = kmap_atomic(ctx->ring_pages[0]);
676 new_nr = (table ? table->nr : 1) * 4;
677 spin_unlock(&mm->ioctx_lock);
679 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
686 spin_lock(&mm->ioctx_lock);
687 old = rcu_dereference_raw(mm->ioctx_table);
690 rcu_assign_pointer(mm->ioctx_table, table);
691 } else if (table->nr > old->nr) {
692 memcpy(table->table, old->table,
693 old->nr * sizeof(struct kioctx *));
695 rcu_assign_pointer(mm->ioctx_table, table);
704 static void aio_nr_sub(unsigned nr)
706 spin_lock(&aio_nr_lock);
707 if (WARN_ON(aio_nr - nr > aio_nr))
711 spin_unlock(&aio_nr_lock);
715 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
717 static struct kioctx *ioctx_alloc(unsigned nr_events)
719 struct mm_struct *mm = current->mm;
724 * Store the original nr_events -- what userspace passed to io_setup(),
725 * for counting against the global limit -- before it changes.
727 unsigned int max_reqs = nr_events;
730 * We keep track of the number of available ringbuffer slots, to prevent
731 * overflow (reqs_available), and we also use percpu counters for this.
733 * So since up to half the slots might be on other cpu's percpu counters
734 * and unavailable, double nr_events so userspace sees what they
735 * expected: additionally, we move req_batch slots to/from percpu
736 * counters at a time, so make sure that isn't 0:
738 nr_events = max(nr_events, num_possible_cpus() * 4);
741 /* Prevent overflows */
742 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
743 pr_debug("ENOMEM: nr_events too high\n");
744 return ERR_PTR(-EINVAL);
747 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
748 return ERR_PTR(-EAGAIN);
750 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
752 return ERR_PTR(-ENOMEM);
754 ctx->max_reqs = max_reqs;
756 spin_lock_init(&ctx->ctx_lock);
757 spin_lock_init(&ctx->completion_lock);
758 mutex_init(&ctx->ring_lock);
759 /* Protect against page migration throughout kiotx setup by keeping
760 * the ring_lock mutex held until setup is complete. */
761 mutex_lock(&ctx->ring_lock);
762 init_waitqueue_head(&ctx->wait);
764 INIT_LIST_HEAD(&ctx->active_reqs);
766 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
769 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
772 ctx->cpu = alloc_percpu(struct kioctx_cpu);
776 err = aio_setup_ring(ctx, nr_events);
780 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
781 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
782 if (ctx->req_batch < 1)
785 /* limit the number of system wide aios */
786 spin_lock(&aio_nr_lock);
787 if (aio_nr + ctx->max_reqs > aio_max_nr ||
788 aio_nr + ctx->max_reqs < aio_nr) {
789 spin_unlock(&aio_nr_lock);
793 aio_nr += ctx->max_reqs;
794 spin_unlock(&aio_nr_lock);
796 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
797 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
799 err = ioctx_add_table(ctx, mm);
803 /* Release the ring_lock mutex now that all setup is complete. */
804 mutex_unlock(&ctx->ring_lock);
806 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
807 ctx, ctx->user_id, mm, ctx->nr_events);
811 aio_nr_sub(ctx->max_reqs);
813 atomic_set(&ctx->dead, 1);
815 vm_munmap(ctx->mmap_base, ctx->mmap_size);
818 mutex_unlock(&ctx->ring_lock);
819 free_percpu(ctx->cpu);
820 percpu_ref_exit(&ctx->reqs);
821 percpu_ref_exit(&ctx->users);
822 kmem_cache_free(kioctx_cachep, ctx);
823 pr_debug("error allocating ioctx %d\n", err);
828 * Cancels all outstanding aio requests on an aio context. Used
829 * when the processes owning a context have all exited to encourage
830 * the rapid destruction of the kioctx.
832 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
833 struct ctx_rq_wait *wait)
835 struct kioctx_table *table;
837 spin_lock(&mm->ioctx_lock);
838 if (atomic_xchg(&ctx->dead, 1)) {
839 spin_unlock(&mm->ioctx_lock);
843 table = rcu_dereference_raw(mm->ioctx_table);
844 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
845 RCU_INIT_POINTER(table->table[ctx->id], NULL);
846 spin_unlock(&mm->ioctx_lock);
848 /* free_ioctx_reqs() will do the necessary RCU synchronization */
849 wake_up_all(&ctx->wait);
852 * It'd be more correct to do this in free_ioctx(), after all
853 * the outstanding kiocbs have finished - but by then io_destroy
854 * has already returned, so io_setup() could potentially return
855 * -EAGAIN with no ioctxs actually in use (as far as userspace
858 aio_nr_sub(ctx->max_reqs);
861 vm_munmap(ctx->mmap_base, ctx->mmap_size);
864 percpu_ref_kill(&ctx->users);
869 * exit_aio: called when the last user of mm goes away. At this point, there is
870 * no way for any new requests to be submited or any of the io_* syscalls to be
871 * called on the context.
873 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
876 void exit_aio(struct mm_struct *mm)
878 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
879 struct ctx_rq_wait wait;
885 atomic_set(&wait.count, table->nr);
886 init_completion(&wait.comp);
889 for (i = 0; i < table->nr; ++i) {
891 rcu_dereference_protected(table->table[i], true);
899 * We don't need to bother with munmap() here - exit_mmap(mm)
900 * is coming and it'll unmap everything. And we simply can't,
901 * this is not necessarily our ->mm.
902 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
903 * that it needs to unmap the area, just set it to 0.
906 kill_ioctx(mm, ctx, &wait);
909 if (!atomic_sub_and_test(skipped, &wait.count)) {
910 /* Wait until all IO for the context are done. */
911 wait_for_completion(&wait.comp);
914 RCU_INIT_POINTER(mm->ioctx_table, NULL);
918 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
920 struct kioctx_cpu *kcpu;
923 local_irq_save(flags);
924 kcpu = this_cpu_ptr(ctx->cpu);
925 kcpu->reqs_available += nr;
927 while (kcpu->reqs_available >= ctx->req_batch * 2) {
928 kcpu->reqs_available -= ctx->req_batch;
929 atomic_add(ctx->req_batch, &ctx->reqs_available);
932 local_irq_restore(flags);
935 static bool __get_reqs_available(struct kioctx *ctx)
937 struct kioctx_cpu *kcpu;
941 local_irq_save(flags);
942 kcpu = this_cpu_ptr(ctx->cpu);
943 if (!kcpu->reqs_available) {
944 int old, avail = atomic_read(&ctx->reqs_available);
947 if (avail < ctx->req_batch)
951 avail = atomic_cmpxchg(&ctx->reqs_available,
952 avail, avail - ctx->req_batch);
953 } while (avail != old);
955 kcpu->reqs_available += ctx->req_batch;
959 kcpu->reqs_available--;
961 local_irq_restore(flags);
965 /* refill_reqs_available
966 * Updates the reqs_available reference counts used for tracking the
967 * number of free slots in the completion ring. This can be called
968 * from aio_complete() (to optimistically update reqs_available) or
969 * from aio_get_req() (the we're out of events case). It must be
970 * called holding ctx->completion_lock.
972 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
975 unsigned events_in_ring, completed;
977 /* Clamp head since userland can write to it. */
978 head %= ctx->nr_events;
980 events_in_ring = tail - head;
982 events_in_ring = ctx->nr_events - (head - tail);
984 completed = ctx->completed_events;
985 if (events_in_ring < completed)
986 completed -= events_in_ring;
993 ctx->completed_events -= completed;
994 put_reqs_available(ctx, completed);
997 /* user_refill_reqs_available
998 * Called to refill reqs_available when aio_get_req() encounters an
999 * out of space in the completion ring.
1001 static void user_refill_reqs_available(struct kioctx *ctx)
1003 spin_lock_irq(&ctx->completion_lock);
1004 if (ctx->completed_events) {
1005 struct aio_ring *ring;
1008 /* Access of ring->head may race with aio_read_events_ring()
1009 * here, but that's okay since whether we read the old version
1010 * or the new version, and either will be valid. The important
1011 * part is that head cannot pass tail since we prevent
1012 * aio_complete() from updating tail by holding
1013 * ctx->completion_lock. Even if head is invalid, the check
1014 * against ctx->completed_events below will make sure we do the
1017 ring = kmap_atomic(ctx->ring_pages[0]);
1019 kunmap_atomic(ring);
1021 refill_reqs_available(ctx, head, ctx->tail);
1024 spin_unlock_irq(&ctx->completion_lock);
1027 static bool get_reqs_available(struct kioctx *ctx)
1029 if (__get_reqs_available(ctx))
1031 user_refill_reqs_available(ctx);
1032 return __get_reqs_available(ctx);
1036 * Allocate a slot for an aio request.
1037 * Returns NULL if no requests are free.
1039 * The refcount is initialized to 2 - one for the async op completion,
1040 * one for the synchronous code that does this.
1042 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1044 struct aio_kiocb *req;
1046 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1050 if (unlikely(!get_reqs_available(ctx))) {
1051 kmem_cache_free(kiocb_cachep, req);
1055 percpu_ref_get(&ctx->reqs);
1057 INIT_LIST_HEAD(&req->ki_list);
1058 refcount_set(&req->ki_refcnt, 2);
1059 req->ki_eventfd = NULL;
1063 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1065 struct aio_ring __user *ring = (void __user *)ctx_id;
1066 struct mm_struct *mm = current->mm;
1067 struct kioctx *ctx, *ret = NULL;
1068 struct kioctx_table *table;
1071 if (get_user(id, &ring->id))
1075 table = rcu_dereference(mm->ioctx_table);
1077 if (!table || id >= table->nr)
1080 id = array_index_nospec(id, table->nr);
1081 ctx = rcu_dereference(table->table[id]);
1082 if (ctx && ctx->user_id == ctx_id) {
1083 if (percpu_ref_tryget_live(&ctx->users))
1091 static inline void iocb_destroy(struct aio_kiocb *iocb)
1093 if (iocb->ki_eventfd)
1094 eventfd_ctx_put(iocb->ki_eventfd);
1096 fput(iocb->ki_filp);
1097 percpu_ref_put(&iocb->ki_ctx->reqs);
1098 kmem_cache_free(kiocb_cachep, iocb);
1102 * Called when the io request on the given iocb is complete.
1104 static void aio_complete(struct aio_kiocb *iocb)
1106 struct kioctx *ctx = iocb->ki_ctx;
1107 struct aio_ring *ring;
1108 struct io_event *ev_page, *event;
1109 unsigned tail, pos, head;
1110 unsigned long flags;
1113 * Add a completion event to the ring buffer. Must be done holding
1114 * ctx->completion_lock to prevent other code from messing with the tail
1115 * pointer since we might be called from irq context.
1117 spin_lock_irqsave(&ctx->completion_lock, flags);
1120 pos = tail + AIO_EVENTS_OFFSET;
1122 if (++tail >= ctx->nr_events)
1125 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1126 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1128 *event = iocb->ki_res;
1130 kunmap_atomic(ev_page);
1131 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1133 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1134 (void __user *)(unsigned long)iocb->ki_res.obj,
1135 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1137 /* after flagging the request as done, we
1138 * must never even look at it again
1140 smp_wmb(); /* make event visible before updating tail */
1144 ring = kmap_atomic(ctx->ring_pages[0]);
1147 kunmap_atomic(ring);
1148 flush_dcache_page(ctx->ring_pages[0]);
1150 ctx->completed_events++;
1151 if (ctx->completed_events > 1)
1152 refill_reqs_available(ctx, head, tail);
1153 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1155 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1158 * Check if the user asked us to deliver the result through an
1159 * eventfd. The eventfd_signal() function is safe to be called
1162 if (iocb->ki_eventfd)
1163 eventfd_signal(iocb->ki_eventfd, 1);
1166 * We have to order our ring_info tail store above and test
1167 * of the wait list below outside the wait lock. This is
1168 * like in wake_up_bit() where clearing a bit has to be
1169 * ordered with the unlocked test.
1173 if (waitqueue_active(&ctx->wait))
1174 wake_up(&ctx->wait);
1177 static inline void iocb_put(struct aio_kiocb *iocb)
1179 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1185 /* aio_read_events_ring
1186 * Pull an event off of the ioctx's event ring. Returns the number of
1189 static long aio_read_events_ring(struct kioctx *ctx,
1190 struct io_event __user *event, long nr)
1192 struct aio_ring *ring;
1193 unsigned head, tail, pos;
1198 * The mutex can block and wake us up and that will cause
1199 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1200 * and repeat. This should be rare enough that it doesn't cause
1201 * peformance issues. See the comment in read_events() for more detail.
1203 sched_annotate_sleep();
1204 mutex_lock(&ctx->ring_lock);
1206 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1207 ring = kmap_atomic(ctx->ring_pages[0]);
1210 kunmap_atomic(ring);
1213 * Ensure that once we've read the current tail pointer, that
1214 * we also see the events that were stored up to the tail.
1218 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1223 head %= ctx->nr_events;
1224 tail %= ctx->nr_events;
1228 struct io_event *ev;
1231 avail = (head <= tail ? tail : ctx->nr_events) - head;
1235 pos = head + AIO_EVENTS_OFFSET;
1236 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1237 pos %= AIO_EVENTS_PER_PAGE;
1239 avail = min(avail, nr - ret);
1240 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1243 copy_ret = copy_to_user(event + ret, ev + pos,
1244 sizeof(*ev) * avail);
1247 if (unlikely(copy_ret)) {
1254 head %= ctx->nr_events;
1257 ring = kmap_atomic(ctx->ring_pages[0]);
1259 kunmap_atomic(ring);
1260 flush_dcache_page(ctx->ring_pages[0]);
1262 pr_debug("%li h%u t%u\n", ret, head, tail);
1264 mutex_unlock(&ctx->ring_lock);
1269 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1270 struct io_event __user *event, long *i)
1272 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1277 if (unlikely(atomic_read(&ctx->dead)))
1283 return ret < 0 || *i >= min_nr;
1286 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1287 struct io_event __user *event,
1293 * Note that aio_read_events() is being called as the conditional - i.e.
1294 * we're calling it after prepare_to_wait() has set task state to
1295 * TASK_INTERRUPTIBLE.
1297 * But aio_read_events() can block, and if it blocks it's going to flip
1298 * the task state back to TASK_RUNNING.
1300 * This should be ok, provided it doesn't flip the state back to
1301 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1302 * will only happen if the mutex_lock() call blocks, and we then find
1303 * the ringbuffer empty. So in practice we should be ok, but it's
1304 * something to be aware of when touching this code.
1307 aio_read_events(ctx, min_nr, nr, event, &ret);
1309 wait_event_interruptible_hrtimeout(ctx->wait,
1310 aio_read_events(ctx, min_nr, nr, event, &ret),
1316 * Create an aio_context capable of receiving at least nr_events.
1317 * ctxp must not point to an aio_context that already exists, and
1318 * must be initialized to 0 prior to the call. On successful
1319 * creation of the aio_context, *ctxp is filled in with the resulting
1320 * handle. May fail with -EINVAL if *ctxp is not initialized,
1321 * if the specified nr_events exceeds internal limits. May fail
1322 * with -EAGAIN if the specified nr_events exceeds the user's limit
1323 * of available events. May fail with -ENOMEM if insufficient kernel
1324 * resources are available. May fail with -EFAULT if an invalid
1325 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1328 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1330 struct kioctx *ioctx = NULL;
1334 ret = get_user(ctx, ctxp);
1339 if (unlikely(ctx || nr_events == 0)) {
1340 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1345 ioctx = ioctx_alloc(nr_events);
1346 ret = PTR_ERR(ioctx);
1347 if (!IS_ERR(ioctx)) {
1348 ret = put_user(ioctx->user_id, ctxp);
1350 kill_ioctx(current->mm, ioctx, NULL);
1351 percpu_ref_put(&ioctx->users);
1358 #ifdef CONFIG_COMPAT
1359 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1361 struct kioctx *ioctx = NULL;
1365 ret = get_user(ctx, ctx32p);
1370 if (unlikely(ctx || nr_events == 0)) {
1371 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1376 ioctx = ioctx_alloc(nr_events);
1377 ret = PTR_ERR(ioctx);
1378 if (!IS_ERR(ioctx)) {
1379 /* truncating is ok because it's a user address */
1380 ret = put_user((u32)ioctx->user_id, ctx32p);
1382 kill_ioctx(current->mm, ioctx, NULL);
1383 percpu_ref_put(&ioctx->users);
1392 * Destroy the aio_context specified. May cancel any outstanding
1393 * AIOs and block on completion. Will fail with -ENOSYS if not
1394 * implemented. May fail with -EINVAL if the context pointed to
1397 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1399 struct kioctx *ioctx = lookup_ioctx(ctx);
1400 if (likely(NULL != ioctx)) {
1401 struct ctx_rq_wait wait;
1404 init_completion(&wait.comp);
1405 atomic_set(&wait.count, 1);
1407 /* Pass requests_done to kill_ioctx() where it can be set
1408 * in a thread-safe way. If we try to set it here then we have
1409 * a race condition if two io_destroy() called simultaneously.
1411 ret = kill_ioctx(current->mm, ioctx, &wait);
1412 percpu_ref_put(&ioctx->users);
1414 /* Wait until all IO for the context are done. Otherwise kernel
1415 * keep using user-space buffers even if user thinks the context
1419 wait_for_completion(&wait.comp);
1423 pr_debug("EINVAL: invalid context id\n");
1427 static void aio_remove_iocb(struct aio_kiocb *iocb)
1429 struct kioctx *ctx = iocb->ki_ctx;
1430 unsigned long flags;
1432 spin_lock_irqsave(&ctx->ctx_lock, flags);
1433 list_del(&iocb->ki_list);
1434 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1437 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1439 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1441 if (!list_empty_careful(&iocb->ki_list))
1442 aio_remove_iocb(iocb);
1444 if (kiocb->ki_flags & IOCB_WRITE) {
1445 struct inode *inode = file_inode(kiocb->ki_filp);
1448 * Tell lockdep we inherited freeze protection from submission
1451 if (S_ISREG(inode->i_mode))
1452 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1453 file_end_write(kiocb->ki_filp);
1456 iocb->ki_res.res = res;
1457 iocb->ki_res.res2 = res2;
1461 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1465 req->ki_complete = aio_complete_rw;
1466 req->private = NULL;
1467 req->ki_pos = iocb->aio_offset;
1468 req->ki_flags = iocb_flags(req->ki_filp) | IOCB_AIO_RW;
1469 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1470 req->ki_flags |= IOCB_EVENTFD;
1471 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1472 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1474 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1475 * aio_reqprio is interpreted as an I/O scheduling
1476 * class and priority.
1478 ret = ioprio_check_cap(iocb->aio_reqprio);
1480 pr_debug("aio ioprio check cap error: %d\n", ret);
1484 req->ki_ioprio = iocb->aio_reqprio;
1486 req->ki_ioprio = get_current_ioprio();
1488 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1492 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1496 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1497 struct iovec **iovec, bool vectored, bool compat,
1498 struct iov_iter *iter)
1500 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1501 size_t len = iocb->aio_nbytes;
1504 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1509 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1512 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1518 case -ERESTARTNOINTR:
1519 case -ERESTARTNOHAND:
1520 case -ERESTART_RESTARTBLOCK:
1522 * There's no easy way to restart the syscall since other AIO's
1523 * may be already running. Just fail this IO with EINTR.
1528 req->ki_complete(req, ret, 0);
1532 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1533 bool vectored, bool compat)
1535 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1536 struct iov_iter iter;
1540 ret = aio_prep_rw(req, iocb);
1543 file = req->ki_filp;
1544 if (unlikely(!(file->f_mode & FMODE_READ)))
1547 if (unlikely(!file->f_op->read_iter))
1550 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1553 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1555 aio_rw_done(req, call_read_iter(file, req, &iter));
1560 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1561 bool vectored, bool compat)
1563 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1564 struct iov_iter iter;
1568 ret = aio_prep_rw(req, iocb);
1571 file = req->ki_filp;
1573 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1575 if (unlikely(!file->f_op->write_iter))
1578 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1581 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1584 * Open-code file_start_write here to grab freeze protection,
1585 * which will be released by another thread in
1586 * aio_complete_rw(). Fool lockdep by telling it the lock got
1587 * released so that it doesn't complain about the held lock when
1588 * we return to userspace.
1590 if (S_ISREG(file_inode(file)->i_mode)) {
1591 sb_start_write(file_inode(file)->i_sb);
1592 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1594 req->ki_flags |= IOCB_WRITE;
1595 aio_rw_done(req, call_write_iter(file, req, &iter));
1601 static void aio_fsync_work(struct work_struct *work)
1603 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1604 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1606 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1607 revert_creds(old_cred);
1608 put_cred(iocb->fsync.creds);
1612 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1615 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1616 iocb->aio_rw_flags))
1619 if (unlikely(!req->file->f_op->fsync))
1622 req->creds = prepare_creds();
1626 req->datasync = datasync;
1627 INIT_WORK(&req->work, aio_fsync_work);
1628 schedule_work(&req->work);
1632 static void aio_poll_put_work(struct work_struct *work)
1634 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1635 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1641 * Safely lock the waitqueue which the request is on, synchronizing with the
1642 * case where the ->poll() provider decides to free its waitqueue early.
1644 * Returns true on success, meaning that req->head->lock was locked, req->wait
1645 * is on req->head, and an RCU read lock was taken. Returns false if the
1646 * request was already removed from its waitqueue (which might no longer exist).
1648 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1650 wait_queue_head_t *head;
1653 * While we hold the waitqueue lock and the waitqueue is nonempty,
1654 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1655 * lock in the first place can race with the waitqueue being freed.
1657 * We solve this as eventpoll does: by taking advantage of the fact that
1658 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1659 * we enter rcu_read_lock() and see that the pointer to the queue is
1660 * non-NULL, we can then lock it without the memory being freed out from
1661 * under us, then check whether the request is still on the queue.
1663 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1664 * case the caller deletes the entry from the queue, leaving it empty.
1665 * In that case, only RCU prevents the queue memory from being freed.
1668 head = smp_load_acquire(&req->head);
1670 spin_lock(&head->lock);
1671 if (!list_empty(&req->wait.entry))
1673 spin_unlock(&head->lock);
1679 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1681 spin_unlock(&req->head->lock);
1685 static void aio_poll_complete_work(struct work_struct *work)
1687 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1688 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1689 struct poll_table_struct pt = { ._key = req->events };
1690 struct kioctx *ctx = iocb->ki_ctx;
1693 if (!READ_ONCE(req->cancelled))
1694 mask = vfs_poll(req->file, &pt) & req->events;
1697 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1698 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1699 * synchronize with them. In the cancellation case the list_del_init
1700 * itself is not actually needed, but harmless so we keep it in to
1701 * avoid further branches in the fast path.
1703 spin_lock_irq(&ctx->ctx_lock);
1704 if (poll_iocb_lock_wq(req)) {
1705 if (!mask && !READ_ONCE(req->cancelled)) {
1707 * The request isn't actually ready to be completed yet.
1708 * Reschedule completion if another wakeup came in.
1710 if (req->work_need_resched) {
1711 schedule_work(&req->work);
1712 req->work_need_resched = false;
1714 req->work_scheduled = false;
1716 poll_iocb_unlock_wq(req);
1717 spin_unlock_irq(&ctx->ctx_lock);
1720 list_del_init(&req->wait.entry);
1721 poll_iocb_unlock_wq(req);
1722 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1723 list_del_init(&iocb->ki_list);
1724 iocb->ki_res.res = mangle_poll(mask);
1725 spin_unlock_irq(&ctx->ctx_lock);
1730 /* assumes we are called with irqs disabled */
1731 static int aio_poll_cancel(struct kiocb *iocb)
1733 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1734 struct poll_iocb *req = &aiocb->poll;
1736 if (poll_iocb_lock_wq(req)) {
1737 WRITE_ONCE(req->cancelled, true);
1738 if (!req->work_scheduled) {
1739 schedule_work(&aiocb->poll.work);
1740 req->work_scheduled = true;
1742 poll_iocb_unlock_wq(req);
1743 } /* else, the request was force-cancelled by POLLFREE already */
1748 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1751 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1752 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1753 __poll_t mask = key_to_poll(key);
1754 unsigned long flags;
1756 /* for instances that support it check for an event match first: */
1757 if (mask && !(mask & req->events))
1761 * Complete the request inline if possible. This requires that three
1762 * conditions be met:
1763 * 1. An event mask must have been passed. If a plain wakeup was done
1764 * instead, then mask == 0 and we have to call vfs_poll() to get
1765 * the events, so inline completion isn't possible.
1766 * 2. The completion work must not have already been scheduled.
1767 * 3. ctx_lock must not be busy. We have to use trylock because we
1768 * already hold the waitqueue lock, so this inverts the normal
1769 * locking order. Use irqsave/irqrestore because not all
1770 * filesystems (e.g. fuse) call this function with IRQs disabled,
1771 * yet IRQs have to be disabled before ctx_lock is obtained.
1773 if (mask && !req->work_scheduled &&
1774 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1775 struct kioctx *ctx = iocb->ki_ctx;
1777 list_del_init(&req->wait.entry);
1778 list_del(&iocb->ki_list);
1779 iocb->ki_res.res = mangle_poll(mask);
1780 if (iocb->ki_eventfd && eventfd_signal_count()) {
1782 INIT_WORK(&req->work, aio_poll_put_work);
1783 schedule_work(&req->work);
1785 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1790 * Schedule the completion work if needed. If it was already
1791 * scheduled, record that another wakeup came in.
1793 * Don't remove the request from the waitqueue here, as it might
1794 * not actually be complete yet (we won't know until vfs_poll()
1795 * is called), and we must not miss any wakeups. POLLFREE is an
1796 * exception to this; see below.
1798 if (req->work_scheduled) {
1799 req->work_need_resched = true;
1801 schedule_work(&req->work);
1802 req->work_scheduled = true;
1806 * If the waitqueue is being freed early but we can't complete
1807 * the request inline, we have to tear down the request as best
1808 * we can. That means immediately removing the request from its
1809 * waitqueue and preventing all further accesses to the
1810 * waitqueue via the request. We also need to schedule the
1811 * completion work (done above). Also mark the request as
1812 * cancelled, to potentially skip an unneeded call to ->poll().
1814 if (mask & POLLFREE) {
1815 WRITE_ONCE(req->cancelled, true);
1816 list_del_init(&req->wait.entry);
1819 * Careful: this *must* be the last step, since as soon
1820 * as req->head is NULL'ed out, the request can be
1821 * completed and freed, since aio_poll_complete_work()
1822 * will no longer need to take the waitqueue lock.
1824 smp_store_release(&req->head, NULL);
1830 struct aio_poll_table {
1831 struct poll_table_struct pt;
1832 struct aio_kiocb *iocb;
1838 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1839 struct poll_table_struct *p)
1841 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1843 /* multiple wait queues per file are not supported */
1844 if (unlikely(pt->queued)) {
1845 pt->error = -EINVAL;
1851 pt->iocb->poll.head = head;
1852 add_wait_queue(head, &pt->iocb->poll.wait);
1855 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1857 struct kioctx *ctx = aiocb->ki_ctx;
1858 struct poll_iocb *req = &aiocb->poll;
1859 struct aio_poll_table apt;
1860 bool cancel = false;
1863 /* reject any unknown events outside the normal event mask. */
1864 if ((u16)iocb->aio_buf != iocb->aio_buf)
1866 /* reject fields that are not defined for poll */
1867 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1870 INIT_WORK(&req->work, aio_poll_complete_work);
1871 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1874 req->cancelled = false;
1875 req->work_scheduled = false;
1876 req->work_need_resched = false;
1878 apt.pt._qproc = aio_poll_queue_proc;
1879 apt.pt._key = req->events;
1882 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1884 /* initialized the list so that we can do list_empty checks */
1885 INIT_LIST_HEAD(&req->wait.entry);
1886 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1888 mask = vfs_poll(req->file, &apt.pt) & req->events;
1889 spin_lock_irq(&ctx->ctx_lock);
1890 if (likely(apt.queued)) {
1891 bool on_queue = poll_iocb_lock_wq(req);
1893 if (!on_queue || req->work_scheduled) {
1895 * aio_poll_wake() already either scheduled the async
1896 * completion work, or completed the request inline.
1898 if (apt.error) /* unsupported case: multiple queues */
1903 if (mask || apt.error) {
1904 /* Steal to complete synchronously. */
1905 list_del_init(&req->wait.entry);
1906 } else if (cancel) {
1907 /* Cancel if possible (may be too late though). */
1908 WRITE_ONCE(req->cancelled, true);
1909 } else if (on_queue) {
1911 * Actually waiting for an event, so add the request to
1912 * active_reqs so that it can be cancelled if needed.
1914 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1915 aiocb->ki_cancel = aio_poll_cancel;
1918 poll_iocb_unlock_wq(req);
1920 if (mask) { /* no async, we'd stolen it */
1921 aiocb->ki_res.res = mangle_poll(mask);
1924 spin_unlock_irq(&ctx->ctx_lock);
1930 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1931 struct iocb __user *user_iocb, struct aio_kiocb *req,
1934 req->ki_filp = fget(iocb->aio_fildes);
1935 if (unlikely(!req->ki_filp))
1938 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1939 struct eventfd_ctx *eventfd;
1941 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1942 * instance of the file* now. The file descriptor must be
1943 * an eventfd() fd, and will be signaled for each completed
1944 * event using the eventfd_signal() function.
1946 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1947 if (IS_ERR(eventfd))
1948 return PTR_ERR(eventfd);
1950 req->ki_eventfd = eventfd;
1953 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1954 pr_debug("EFAULT: aio_key\n");
1958 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1959 req->ki_res.data = iocb->aio_data;
1960 req->ki_res.res = 0;
1961 req->ki_res.res2 = 0;
1963 switch (iocb->aio_lio_opcode) {
1964 case IOCB_CMD_PREAD:
1965 return aio_read(&req->rw, iocb, false, compat);
1966 case IOCB_CMD_PWRITE:
1967 return aio_write(&req->rw, iocb, false, compat);
1968 case IOCB_CMD_PREADV:
1969 return aio_read(&req->rw, iocb, true, compat);
1970 case IOCB_CMD_PWRITEV:
1971 return aio_write(&req->rw, iocb, true, compat);
1972 case IOCB_CMD_FSYNC:
1973 return aio_fsync(&req->fsync, iocb, false);
1974 case IOCB_CMD_FDSYNC:
1975 return aio_fsync(&req->fsync, iocb, true);
1977 return aio_poll(req, iocb);
1979 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1984 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1987 struct aio_kiocb *req;
1991 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1994 /* enforce forwards compatibility on users */
1995 if (unlikely(iocb.aio_reserved2)) {
1996 pr_debug("EINVAL: reserve field set\n");
2000 /* prevent overflows */
2002 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2003 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2004 ((ssize_t)iocb.aio_nbytes < 0)
2006 pr_debug("EINVAL: overflow check\n");
2010 req = aio_get_req(ctx);
2014 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2016 /* Done with the synchronous reference */
2020 * If err is 0, we'd either done aio_complete() ourselves or have
2021 * arranged for that to be done asynchronously. Anything non-zero
2022 * means that we need to destroy req ourselves.
2024 if (unlikely(err)) {
2026 put_reqs_available(ctx, 1);
2032 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2033 * the number of iocbs queued. May return -EINVAL if the aio_context
2034 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2035 * *iocbpp[0] is not properly initialized, if the operation specified
2036 * is invalid for the file descriptor in the iocb. May fail with
2037 * -EFAULT if any of the data structures point to invalid data. May
2038 * fail with -EBADF if the file descriptor specified in the first
2039 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2040 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2041 * fail with -ENOSYS if not implemented.
2043 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2044 struct iocb __user * __user *, iocbpp)
2049 struct blk_plug plug;
2051 if (unlikely(nr < 0))
2054 ctx = lookup_ioctx(ctx_id);
2055 if (unlikely(!ctx)) {
2056 pr_debug("EINVAL: invalid context id\n");
2060 if (nr > ctx->nr_events)
2061 nr = ctx->nr_events;
2063 if (nr > AIO_PLUG_THRESHOLD)
2064 blk_start_plug(&plug);
2065 for (i = 0; i < nr; i++) {
2066 struct iocb __user *user_iocb;
2068 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2073 ret = io_submit_one(ctx, user_iocb, false);
2077 if (nr > AIO_PLUG_THRESHOLD)
2078 blk_finish_plug(&plug);
2080 percpu_ref_put(&ctx->users);
2084 #ifdef CONFIG_COMPAT
2085 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2086 int, nr, compat_uptr_t __user *, iocbpp)
2091 struct blk_plug plug;
2093 if (unlikely(nr < 0))
2096 ctx = lookup_ioctx(ctx_id);
2097 if (unlikely(!ctx)) {
2098 pr_debug("EINVAL: invalid context id\n");
2102 if (nr > ctx->nr_events)
2103 nr = ctx->nr_events;
2105 if (nr > AIO_PLUG_THRESHOLD)
2106 blk_start_plug(&plug);
2107 for (i = 0; i < nr; i++) {
2108 compat_uptr_t user_iocb;
2110 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2115 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2119 if (nr > AIO_PLUG_THRESHOLD)
2120 blk_finish_plug(&plug);
2122 percpu_ref_put(&ctx->users);
2128 * Attempts to cancel an iocb previously passed to io_submit. If
2129 * the operation is successfully cancelled, the resulting event is
2130 * copied into the memory pointed to by result without being placed
2131 * into the completion queue and 0 is returned. May fail with
2132 * -EFAULT if any of the data structures pointed to are invalid.
2133 * May fail with -EINVAL if aio_context specified by ctx_id is
2134 * invalid. May fail with -EAGAIN if the iocb specified was not
2135 * cancelled. Will fail with -ENOSYS if not implemented.
2137 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2138 struct io_event __user *, result)
2141 struct aio_kiocb *kiocb;
2144 u64 obj = (u64)(unsigned long)iocb;
2146 if (unlikely(get_user(key, &iocb->aio_key)))
2148 if (unlikely(key != KIOCB_KEY))
2151 ctx = lookup_ioctx(ctx_id);
2155 spin_lock_irq(&ctx->ctx_lock);
2156 /* TODO: use a hash or array, this sucks. */
2157 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2158 if (kiocb->ki_res.obj == obj) {
2159 ret = kiocb->ki_cancel(&kiocb->rw);
2160 list_del_init(&kiocb->ki_list);
2164 spin_unlock_irq(&ctx->ctx_lock);
2168 * The result argument is no longer used - the io_event is
2169 * always delivered via the ring buffer. -EINPROGRESS indicates
2170 * cancellation is progress:
2175 percpu_ref_put(&ctx->users);
2180 static long do_io_getevents(aio_context_t ctx_id,
2183 struct io_event __user *events,
2184 struct timespec64 *ts)
2186 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2187 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2190 if (likely(ioctx)) {
2191 if (likely(min_nr <= nr && min_nr >= 0))
2192 ret = read_events(ioctx, min_nr, nr, events, until);
2193 percpu_ref_put(&ioctx->users);
2200 * Attempts to read at least min_nr events and up to nr events from
2201 * the completion queue for the aio_context specified by ctx_id. If
2202 * it succeeds, the number of read events is returned. May fail with
2203 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2204 * out of range, if timeout is out of range. May fail with -EFAULT
2205 * if any of the memory specified is invalid. May return 0 or
2206 * < min_nr if the timeout specified by timeout has elapsed
2207 * before sufficient events are available, where timeout == NULL
2208 * specifies an infinite timeout. Note that the timeout pointed to by
2209 * timeout is relative. Will fail with -ENOSYS if not implemented.
2213 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2216 struct io_event __user *, events,
2217 struct __kernel_timespec __user *, timeout)
2219 struct timespec64 ts;
2222 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2225 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2226 if (!ret && signal_pending(current))
2233 struct __aio_sigset {
2234 const sigset_t __user *sigmask;
2238 SYSCALL_DEFINE6(io_pgetevents,
2239 aio_context_t, ctx_id,
2242 struct io_event __user *, events,
2243 struct __kernel_timespec __user *, timeout,
2244 const struct __aio_sigset __user *, usig)
2246 struct __aio_sigset ksig = { NULL, };
2247 struct timespec64 ts;
2251 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2254 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2257 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2261 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2263 interrupted = signal_pending(current);
2264 restore_saved_sigmask_unless(interrupted);
2265 if (interrupted && !ret)
2266 ret = -ERESTARTNOHAND;
2271 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2273 SYSCALL_DEFINE6(io_pgetevents_time32,
2274 aio_context_t, ctx_id,
2277 struct io_event __user *, events,
2278 struct old_timespec32 __user *, timeout,
2279 const struct __aio_sigset __user *, usig)
2281 struct __aio_sigset ksig = { NULL, };
2282 struct timespec64 ts;
2286 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2289 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2293 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2297 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2299 interrupted = signal_pending(current);
2300 restore_saved_sigmask_unless(interrupted);
2301 if (interrupted && !ret)
2302 ret = -ERESTARTNOHAND;
2309 #if defined(CONFIG_COMPAT_32BIT_TIME)
2311 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2314 struct io_event __user *, events,
2315 struct old_timespec32 __user *, timeout)
2317 struct timespec64 t;
2320 if (timeout && get_old_timespec32(&t, timeout))
2323 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2324 if (!ret && signal_pending(current))
2331 #ifdef CONFIG_COMPAT
2333 struct __compat_aio_sigset {
2334 compat_uptr_t sigmask;
2335 compat_size_t sigsetsize;
2338 #if defined(CONFIG_COMPAT_32BIT_TIME)
2340 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2341 compat_aio_context_t, ctx_id,
2342 compat_long_t, min_nr,
2344 struct io_event __user *, events,
2345 struct old_timespec32 __user *, timeout,
2346 const struct __compat_aio_sigset __user *, usig)
2348 struct __compat_aio_sigset ksig = { 0, };
2349 struct timespec64 t;
2353 if (timeout && get_old_timespec32(&t, timeout))
2356 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2359 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2363 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2365 interrupted = signal_pending(current);
2366 restore_saved_sigmask_unless(interrupted);
2367 if (interrupted && !ret)
2368 ret = -ERESTARTNOHAND;
2375 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2376 compat_aio_context_t, ctx_id,
2377 compat_long_t, min_nr,
2379 struct io_event __user *, events,
2380 struct __kernel_timespec __user *, timeout,
2381 const struct __compat_aio_sigset __user *, usig)
2383 struct __compat_aio_sigset ksig = { 0, };
2384 struct timespec64 t;
2388 if (timeout && get_timespec64(&t, timeout))
2391 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2394 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2398 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2400 interrupted = signal_pending(current);
2401 restore_saved_sigmask_unless(interrupted);
2402 if (interrupted && !ret)
2403 ret = -ERESTARTNOHAND;