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/mmu_context.h>
31 #include <linux/percpu.h>
32 #include <linux/slab.h>
33 #include <linux/timer.h>
34 #include <linux/aio.h>
35 #include <linux/highmem.h>
36 #include <linux/workqueue.h>
37 #include <linux/security.h>
38 #include <linux/eventfd.h>
39 #include <linux/blkdev.h>
40 #include <linux/compat.h>
41 #include <linux/migrate.h>
42 #include <linux/ramfs.h>
43 #include <linux/percpu-refcount.h>
44 #include <linux/mount.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[0];
71 }; /* 128 bytes + ring size */
73 #define AIO_RING_PAGES 8
78 struct kioctx __rcu *table[];
82 unsigned reqs_available;
86 struct completion comp;
91 struct percpu_ref users;
94 struct percpu_ref reqs;
96 unsigned long user_id;
98 struct __percpu kioctx_cpu *cpu;
101 * For percpu reqs_available, number of slots we move to/from global
106 * This is what userspace passed to io_setup(), it's not used for
107 * anything but counting against the global max_reqs quota.
109 * The real limit is nr_events - 1, which will be larger (see
114 /* Size of ringbuffer, in units of struct io_event */
117 unsigned long mmap_base;
118 unsigned long mmap_size;
120 struct page **ring_pages;
123 struct rcu_work free_rwork; /* see free_ioctx() */
126 * signals when all in-flight requests are done
128 struct ctx_rq_wait *rq_wait;
132 * This counts the number of available slots in the ringbuffer,
133 * so we avoid overflowing it: it's decremented (if positive)
134 * when allocating a kiocb and incremented when the resulting
135 * io_event is pulled off the ringbuffer.
137 * We batch accesses to it with a percpu version.
139 atomic_t reqs_available;
140 } ____cacheline_aligned_in_smp;
144 struct list_head active_reqs; /* used for cancellation */
145 } ____cacheline_aligned_in_smp;
148 struct mutex ring_lock;
149 wait_queue_head_t wait;
150 } ____cacheline_aligned_in_smp;
154 unsigned completed_events;
155 spinlock_t completion_lock;
156 } ____cacheline_aligned_in_smp;
158 struct page *internal_pages[AIO_RING_PAGES];
159 struct file *aio_ring_file;
165 * First field must be the file pointer in all the
166 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
170 struct work_struct work;
177 struct wait_queue_head *head;
181 bool work_need_resched;
182 struct wait_queue_entry wait;
183 struct work_struct work;
187 * NOTE! Each of the iocb union members has the file pointer
188 * as the first entry in their struct definition. So you can
189 * access the file pointer through any of the sub-structs,
190 * or directly as just 'ki_filp' in this struct.
194 struct file *ki_filp;
196 struct fsync_iocb fsync;
197 struct poll_iocb poll;
200 struct kioctx *ki_ctx;
201 kiocb_cancel_fn *ki_cancel;
203 struct io_event ki_res;
205 struct list_head ki_list; /* the aio core uses this
206 * for cancellation */
207 refcount_t ki_refcnt;
210 * If the aio_resfd field of the userspace iocb is not zero,
211 * this is the underlying eventfd context to deliver events to.
213 struct eventfd_ctx *ki_eventfd;
216 /*------ sysctl variables----*/
217 static DEFINE_SPINLOCK(aio_nr_lock);
218 unsigned long aio_nr; /* current system wide number of aio requests */
219 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
220 /*----end sysctl variables---*/
222 static struct kmem_cache *kiocb_cachep;
223 static struct kmem_cache *kioctx_cachep;
225 static struct vfsmount *aio_mnt;
227 static const struct file_operations aio_ring_fops;
228 static const struct address_space_operations aio_ctx_aops;
230 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
233 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
235 return ERR_CAST(inode);
237 inode->i_mapping->a_ops = &aio_ctx_aops;
238 inode->i_mapping->private_data = ctx;
239 inode->i_size = PAGE_SIZE * nr_pages;
241 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
242 O_RDWR, &aio_ring_fops);
248 static struct dentry *aio_mount(struct file_system_type *fs_type,
249 int flags, const char *dev_name, void *data)
251 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
255 root->d_sb->s_iflags |= SB_I_NOEXEC;
260 * Creates the slab caches used by the aio routines, panic on
261 * failure as this is done early during the boot sequence.
263 static int __init aio_setup(void)
265 static struct file_system_type aio_fs = {
268 .kill_sb = kill_anon_super,
270 aio_mnt = kern_mount(&aio_fs);
272 panic("Failed to create aio fs mount.");
274 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
275 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278 __initcall(aio_setup);
280 static void put_aio_ring_file(struct kioctx *ctx)
282 struct file *aio_ring_file = ctx->aio_ring_file;
283 struct address_space *i_mapping;
286 truncate_setsize(file_inode(aio_ring_file), 0);
288 /* Prevent further access to the kioctx from migratepages */
289 i_mapping = aio_ring_file->f_mapping;
290 spin_lock(&i_mapping->private_lock);
291 i_mapping->private_data = NULL;
292 ctx->aio_ring_file = NULL;
293 spin_unlock(&i_mapping->private_lock);
299 static void aio_free_ring(struct kioctx *ctx)
303 /* Disconnect the kiotx from the ring file. This prevents future
304 * accesses to the kioctx from page migration.
306 put_aio_ring_file(ctx);
308 for (i = 0; i < ctx->nr_pages; i++) {
310 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
311 page_count(ctx->ring_pages[i]));
312 page = ctx->ring_pages[i];
315 ctx->ring_pages[i] = NULL;
319 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
320 kfree(ctx->ring_pages);
321 ctx->ring_pages = NULL;
325 static int aio_ring_mremap(struct vm_area_struct *vma)
327 struct file *file = vma->vm_file;
328 struct mm_struct *mm = vma->vm_mm;
329 struct kioctx_table *table;
330 int i, res = -EINVAL;
332 spin_lock(&mm->ioctx_lock);
334 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;
353 spin_unlock(&mm->ioctx_lock);
357 static const struct vm_operations_struct aio_ring_vm_ops = {
358 .mremap = aio_ring_mremap,
359 #if IS_ENABLED(CONFIG_MMU)
360 .fault = filemap_fault,
361 .map_pages = filemap_map_pages,
362 .page_mkwrite = filemap_page_mkwrite,
366 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
368 vma->vm_flags |= VM_DONTEXPAND;
369 vma->vm_ops = &aio_ring_vm_ops;
373 static const struct file_operations aio_ring_fops = {
374 .mmap = aio_ring_mmap,
377 #if IS_ENABLED(CONFIG_MIGRATION)
378 static int aio_migratepage(struct address_space *mapping, struct page *new,
379 struct page *old, enum migrate_mode mode)
387 * We cannot support the _NO_COPY case here, because copy needs to
388 * happen under the ctx->completion_lock. That does not work with the
389 * migration workflow of MIGRATE_SYNC_NO_COPY.
391 if (mode == MIGRATE_SYNC_NO_COPY)
396 /* mapping->private_lock here protects against the kioctx teardown. */
397 spin_lock(&mapping->private_lock);
398 ctx = mapping->private_data;
404 /* The ring_lock mutex. The prevents aio_read_events() from writing
405 * to the ring's head, and prevents page migration from mucking in
406 * a partially initialized kiotx.
408 if (!mutex_trylock(&ctx->ring_lock)) {
414 if (idx < (pgoff_t)ctx->nr_pages) {
415 /* Make sure the old page hasn't already been changed */
416 if (ctx->ring_pages[idx] != old)
424 /* Writeback must be complete */
425 BUG_ON(PageWriteback(old));
428 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
429 if (rc != MIGRATEPAGE_SUCCESS) {
434 /* Take completion_lock to prevent other writes to the ring buffer
435 * while the old page is copied to the new. This prevents new
436 * events from being lost.
438 spin_lock_irqsave(&ctx->completion_lock, flags);
439 migrate_page_copy(new, old);
440 BUG_ON(ctx->ring_pages[idx] != old);
441 ctx->ring_pages[idx] = new;
442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
444 /* The old page is no longer accessible. */
448 mutex_unlock(&ctx->ring_lock);
450 spin_unlock(&mapping->private_lock);
455 static const struct address_space_operations aio_ctx_aops = {
456 .set_page_dirty = __set_page_dirty_no_writeback,
457 #if IS_ENABLED(CONFIG_MIGRATION)
458 .migratepage = aio_migratepage,
462 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
464 struct aio_ring *ring;
465 struct mm_struct *mm = current->mm;
466 unsigned long size, unused;
471 /* Compensate for the ring buffer's head/tail overlap entry */
472 nr_events += 2; /* 1 is required, 2 for good luck */
474 size = sizeof(struct aio_ring);
475 size += sizeof(struct io_event) * nr_events;
477 nr_pages = PFN_UP(size);
481 file = aio_private_file(ctx, nr_pages);
483 ctx->aio_ring_file = NULL;
487 ctx->aio_ring_file = file;
488 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
489 / sizeof(struct io_event);
491 ctx->ring_pages = ctx->internal_pages;
492 if (nr_pages > AIO_RING_PAGES) {
493 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
495 if (!ctx->ring_pages) {
496 put_aio_ring_file(ctx);
501 for (i = 0; i < nr_pages; i++) {
503 page = find_or_create_page(file->f_mapping,
504 i, GFP_HIGHUSER | __GFP_ZERO);
507 pr_debug("pid(%d) page[%d]->count=%d\n",
508 current->pid, i, page_count(page));
509 SetPageUptodate(page);
512 ctx->ring_pages[i] = page;
516 if (unlikely(i != nr_pages)) {
521 ctx->mmap_size = nr_pages * PAGE_SIZE;
522 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
524 if (down_write_killable(&mm->mmap_sem)) {
530 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
531 PROT_READ | PROT_WRITE,
532 MAP_SHARED, 0, &unused, NULL);
533 up_write(&mm->mmap_sem);
534 if (IS_ERR((void *)ctx->mmap_base)) {
540 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
542 ctx->user_id = ctx->mmap_base;
543 ctx->nr_events = nr_events; /* trusted copy */
545 ring = kmap_atomic(ctx->ring_pages[0]);
546 ring->nr = nr_events; /* user copy */
548 ring->head = ring->tail = 0;
549 ring->magic = AIO_RING_MAGIC;
550 ring->compat_features = AIO_RING_COMPAT_FEATURES;
551 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
552 ring->header_length = sizeof(struct aio_ring);
554 flush_dcache_page(ctx->ring_pages[0]);
559 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
560 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
561 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
563 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
565 struct aio_kiocb *req;
570 * kiocb didn't come from aio or is neither a read nor a write, hence
573 if (!(iocb->ki_flags & IOCB_AIO_RW))
576 req = container_of(iocb, struct aio_kiocb, rw);
578 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
583 spin_lock_irqsave(&ctx->ctx_lock, flags);
584 list_add_tail(&req->ki_list, &ctx->active_reqs);
585 req->ki_cancel = cancel;
586 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
588 EXPORT_SYMBOL(kiocb_set_cancel_fn);
591 * free_ioctx() should be RCU delayed to synchronize against the RCU
592 * protected lookup_ioctx() and also needs process context to call
593 * aio_free_ring(). Use rcu_work.
595 static void free_ioctx(struct work_struct *work)
597 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
599 pr_debug("freeing %p\n", ctx);
602 free_percpu(ctx->cpu);
603 percpu_ref_exit(&ctx->reqs);
604 percpu_ref_exit(&ctx->users);
605 kmem_cache_free(kioctx_cachep, ctx);
608 static void free_ioctx_reqs(struct percpu_ref *ref)
610 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
612 /* At this point we know that there are no any in-flight requests */
613 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
614 complete(&ctx->rq_wait->comp);
616 /* Synchronize against RCU protected table->table[] dereferences */
617 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
618 queue_rcu_work(system_wq, &ctx->free_rwork);
622 * When this function runs, the kioctx has been removed from the "hash table"
623 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
624 * now it's safe to cancel any that need to be.
626 static void free_ioctx_users(struct percpu_ref *ref)
628 struct kioctx *ctx = container_of(ref, struct kioctx, users);
629 struct aio_kiocb *req;
631 spin_lock_irq(&ctx->ctx_lock);
633 while (!list_empty(&ctx->active_reqs)) {
634 req = list_first_entry(&ctx->active_reqs,
635 struct aio_kiocb, ki_list);
636 req->ki_cancel(&req->rw);
637 list_del_init(&req->ki_list);
640 spin_unlock_irq(&ctx->ctx_lock);
642 percpu_ref_kill(&ctx->reqs);
643 percpu_ref_put(&ctx->reqs);
646 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
649 struct kioctx_table *table, *old;
650 struct aio_ring *ring;
652 spin_lock(&mm->ioctx_lock);
653 table = rcu_dereference_raw(mm->ioctx_table);
657 for (i = 0; i < table->nr; i++)
658 if (!rcu_access_pointer(table->table[i])) {
660 rcu_assign_pointer(table->table[i], ctx);
661 spin_unlock(&mm->ioctx_lock);
663 /* While kioctx setup is in progress,
664 * we are protected from page migration
665 * changes ring_pages by ->ring_lock.
667 ring = kmap_atomic(ctx->ring_pages[0]);
673 new_nr = (table ? table->nr : 1) * 4;
674 spin_unlock(&mm->ioctx_lock);
676 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
683 spin_lock(&mm->ioctx_lock);
684 old = rcu_dereference_raw(mm->ioctx_table);
687 rcu_assign_pointer(mm->ioctx_table, table);
688 } else if (table->nr > old->nr) {
689 memcpy(table->table, old->table,
690 old->nr * sizeof(struct kioctx *));
692 rcu_assign_pointer(mm->ioctx_table, table);
701 static void aio_nr_sub(unsigned nr)
703 spin_lock(&aio_nr_lock);
704 if (WARN_ON(aio_nr - nr > aio_nr))
708 spin_unlock(&aio_nr_lock);
712 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
714 static struct kioctx *ioctx_alloc(unsigned nr_events)
716 struct mm_struct *mm = current->mm;
721 * Store the original nr_events -- what userspace passed to io_setup(),
722 * for counting against the global limit -- before it changes.
724 unsigned int max_reqs = nr_events;
727 * We keep track of the number of available ringbuffer slots, to prevent
728 * overflow (reqs_available), and we also use percpu counters for this.
730 * So since up to half the slots might be on other cpu's percpu counters
731 * and unavailable, double nr_events so userspace sees what they
732 * expected: additionally, we move req_batch slots to/from percpu
733 * counters at a time, so make sure that isn't 0:
735 nr_events = max(nr_events, num_possible_cpus() * 4);
738 /* Prevent overflows */
739 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
740 pr_debug("ENOMEM: nr_events too high\n");
741 return ERR_PTR(-EINVAL);
744 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
745 return ERR_PTR(-EAGAIN);
747 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
749 return ERR_PTR(-ENOMEM);
751 ctx->max_reqs = max_reqs;
753 spin_lock_init(&ctx->ctx_lock);
754 spin_lock_init(&ctx->completion_lock);
755 mutex_init(&ctx->ring_lock);
756 /* Protect against page migration throughout kiotx setup by keeping
757 * the ring_lock mutex held until setup is complete. */
758 mutex_lock(&ctx->ring_lock);
759 init_waitqueue_head(&ctx->wait);
761 INIT_LIST_HEAD(&ctx->active_reqs);
763 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
766 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
769 ctx->cpu = alloc_percpu(struct kioctx_cpu);
773 err = aio_setup_ring(ctx, nr_events);
777 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
778 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
779 if (ctx->req_batch < 1)
782 /* limit the number of system wide aios */
783 spin_lock(&aio_nr_lock);
784 if (aio_nr + ctx->max_reqs > aio_max_nr ||
785 aio_nr + ctx->max_reqs < aio_nr) {
786 spin_unlock(&aio_nr_lock);
790 aio_nr += ctx->max_reqs;
791 spin_unlock(&aio_nr_lock);
793 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
794 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
796 err = ioctx_add_table(ctx, mm);
800 /* Release the ring_lock mutex now that all setup is complete. */
801 mutex_unlock(&ctx->ring_lock);
803 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
804 ctx, ctx->user_id, mm, ctx->nr_events);
808 aio_nr_sub(ctx->max_reqs);
810 atomic_set(&ctx->dead, 1);
812 vm_munmap(ctx->mmap_base, ctx->mmap_size);
815 mutex_unlock(&ctx->ring_lock);
816 free_percpu(ctx->cpu);
817 percpu_ref_exit(&ctx->reqs);
818 percpu_ref_exit(&ctx->users);
819 kmem_cache_free(kioctx_cachep, ctx);
820 pr_debug("error allocating ioctx %d\n", err);
825 * Cancels all outstanding aio requests on an aio context. Used
826 * when the processes owning a context have all exited to encourage
827 * the rapid destruction of the kioctx.
829 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
830 struct ctx_rq_wait *wait)
832 struct kioctx_table *table;
834 spin_lock(&mm->ioctx_lock);
835 if (atomic_xchg(&ctx->dead, 1)) {
836 spin_unlock(&mm->ioctx_lock);
840 table = rcu_dereference_raw(mm->ioctx_table);
841 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
842 RCU_INIT_POINTER(table->table[ctx->id], NULL);
843 spin_unlock(&mm->ioctx_lock);
845 /* free_ioctx_reqs() will do the necessary RCU synchronization */
846 wake_up_all(&ctx->wait);
849 * It'd be more correct to do this in free_ioctx(), after all
850 * the outstanding kiocbs have finished - but by then io_destroy
851 * has already returned, so io_setup() could potentially return
852 * -EAGAIN with no ioctxs actually in use (as far as userspace
855 aio_nr_sub(ctx->max_reqs);
858 vm_munmap(ctx->mmap_base, ctx->mmap_size);
861 percpu_ref_kill(&ctx->users);
866 * exit_aio: called when the last user of mm goes away. At this point, there is
867 * no way for any new requests to be submited or any of the io_* syscalls to be
868 * called on the context.
870 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
873 void exit_aio(struct mm_struct *mm)
875 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
876 struct ctx_rq_wait wait;
882 atomic_set(&wait.count, table->nr);
883 init_completion(&wait.comp);
886 for (i = 0; i < table->nr; ++i) {
888 rcu_dereference_protected(table->table[i], true);
896 * We don't need to bother with munmap() here - exit_mmap(mm)
897 * is coming and it'll unmap everything. And we simply can't,
898 * this is not necessarily our ->mm.
899 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
900 * that it needs to unmap the area, just set it to 0.
903 kill_ioctx(mm, ctx, &wait);
906 if (!atomic_sub_and_test(skipped, &wait.count)) {
907 /* Wait until all IO for the context are done. */
908 wait_for_completion(&wait.comp);
911 RCU_INIT_POINTER(mm->ioctx_table, NULL);
915 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
917 struct kioctx_cpu *kcpu;
920 local_irq_save(flags);
921 kcpu = this_cpu_ptr(ctx->cpu);
922 kcpu->reqs_available += nr;
924 while (kcpu->reqs_available >= ctx->req_batch * 2) {
925 kcpu->reqs_available -= ctx->req_batch;
926 atomic_add(ctx->req_batch, &ctx->reqs_available);
929 local_irq_restore(flags);
932 static bool __get_reqs_available(struct kioctx *ctx)
934 struct kioctx_cpu *kcpu;
938 local_irq_save(flags);
939 kcpu = this_cpu_ptr(ctx->cpu);
940 if (!kcpu->reqs_available) {
941 int old, avail = atomic_read(&ctx->reqs_available);
944 if (avail < ctx->req_batch)
948 avail = atomic_cmpxchg(&ctx->reqs_available,
949 avail, avail - ctx->req_batch);
950 } while (avail != old);
952 kcpu->reqs_available += ctx->req_batch;
956 kcpu->reqs_available--;
958 local_irq_restore(flags);
962 /* refill_reqs_available
963 * Updates the reqs_available reference counts used for tracking the
964 * number of free slots in the completion ring. This can be called
965 * from aio_complete() (to optimistically update reqs_available) or
966 * from aio_get_req() (the we're out of events case). It must be
967 * called holding ctx->completion_lock.
969 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
972 unsigned events_in_ring, completed;
974 /* Clamp head since userland can write to it. */
975 head %= ctx->nr_events;
977 events_in_ring = tail - head;
979 events_in_ring = ctx->nr_events - (head - tail);
981 completed = ctx->completed_events;
982 if (events_in_ring < completed)
983 completed -= events_in_ring;
990 ctx->completed_events -= completed;
991 put_reqs_available(ctx, completed);
994 /* user_refill_reqs_available
995 * Called to refill reqs_available when aio_get_req() encounters an
996 * out of space in the completion ring.
998 static void user_refill_reqs_available(struct kioctx *ctx)
1000 spin_lock_irq(&ctx->completion_lock);
1001 if (ctx->completed_events) {
1002 struct aio_ring *ring;
1005 /* Access of ring->head may race with aio_read_events_ring()
1006 * here, but that's okay since whether we read the old version
1007 * or the new version, and either will be valid. The important
1008 * part is that head cannot pass tail since we prevent
1009 * aio_complete() from updating tail by holding
1010 * ctx->completion_lock. Even if head is invalid, the check
1011 * against ctx->completed_events below will make sure we do the
1014 ring = kmap_atomic(ctx->ring_pages[0]);
1016 kunmap_atomic(ring);
1018 refill_reqs_available(ctx, head, ctx->tail);
1021 spin_unlock_irq(&ctx->completion_lock);
1024 static bool get_reqs_available(struct kioctx *ctx)
1026 if (__get_reqs_available(ctx))
1028 user_refill_reqs_available(ctx);
1029 return __get_reqs_available(ctx);
1033 * Allocate a slot for an aio request.
1034 * Returns NULL if no requests are free.
1036 * The refcount is initialized to 2 - one for the async op completion,
1037 * one for the synchronous code that does this.
1039 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1041 struct aio_kiocb *req;
1043 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1047 percpu_ref_get(&ctx->reqs);
1049 INIT_LIST_HEAD(&req->ki_list);
1050 refcount_set(&req->ki_refcnt, 2);
1051 req->ki_eventfd = NULL;
1055 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1057 struct aio_ring __user *ring = (void __user *)ctx_id;
1058 struct mm_struct *mm = current->mm;
1059 struct kioctx *ctx, *ret = NULL;
1060 struct kioctx_table *table;
1063 if (get_user(id, &ring->id))
1067 table = rcu_dereference(mm->ioctx_table);
1069 if (!table || id >= table->nr)
1072 id = array_index_nospec(id, table->nr);
1073 ctx = rcu_dereference(table->table[id]);
1074 if (ctx && ctx->user_id == ctx_id) {
1075 if (percpu_ref_tryget_live(&ctx->users))
1083 static inline void iocb_destroy(struct aio_kiocb *iocb)
1086 fput(iocb->ki_filp);
1087 percpu_ref_put(&iocb->ki_ctx->reqs);
1088 kmem_cache_free(kiocb_cachep, iocb);
1092 * Called when the io request on the given iocb is complete.
1094 static void aio_complete(struct aio_kiocb *iocb)
1096 struct kioctx *ctx = iocb->ki_ctx;
1097 struct aio_ring *ring;
1098 struct io_event *ev_page, *event;
1099 unsigned tail, pos, head;
1100 unsigned long flags;
1103 * Add a completion event to the ring buffer. Must be done holding
1104 * ctx->completion_lock to prevent other code from messing with the tail
1105 * pointer since we might be called from irq context.
1107 spin_lock_irqsave(&ctx->completion_lock, flags);
1110 pos = tail + AIO_EVENTS_OFFSET;
1112 if (++tail >= ctx->nr_events)
1115 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1116 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1118 *event = iocb->ki_res;
1120 kunmap_atomic(ev_page);
1121 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1123 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1124 (void __user *)(unsigned long)iocb->ki_res.obj,
1125 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1127 /* after flagging the request as done, we
1128 * must never even look at it again
1130 smp_wmb(); /* make event visible before updating tail */
1134 ring = kmap_atomic(ctx->ring_pages[0]);
1137 kunmap_atomic(ring);
1138 flush_dcache_page(ctx->ring_pages[0]);
1140 ctx->completed_events++;
1141 if (ctx->completed_events > 1)
1142 refill_reqs_available(ctx, head, tail);
1143 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1145 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1148 * Check if the user asked us to deliver the result through an
1149 * eventfd. The eventfd_signal() function is safe to be called
1152 if (iocb->ki_eventfd) {
1153 eventfd_signal(iocb->ki_eventfd, 1);
1154 eventfd_ctx_put(iocb->ki_eventfd);
1158 * We have to order our ring_info tail store above and test
1159 * of the wait list below outside the wait lock. This is
1160 * like in wake_up_bit() where clearing a bit has to be
1161 * ordered with the unlocked test.
1165 if (waitqueue_active(&ctx->wait))
1166 wake_up(&ctx->wait);
1169 static inline void iocb_put(struct aio_kiocb *iocb)
1171 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1177 /* aio_read_events_ring
1178 * Pull an event off of the ioctx's event ring. Returns the number of
1181 static long aio_read_events_ring(struct kioctx *ctx,
1182 struct io_event __user *event, long nr)
1184 struct aio_ring *ring;
1185 unsigned head, tail, pos;
1190 * The mutex can block and wake us up and that will cause
1191 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1192 * and repeat. This should be rare enough that it doesn't cause
1193 * peformance issues. See the comment in read_events() for more detail.
1195 sched_annotate_sleep();
1196 mutex_lock(&ctx->ring_lock);
1198 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1199 ring = kmap_atomic(ctx->ring_pages[0]);
1202 kunmap_atomic(ring);
1205 * Ensure that once we've read the current tail pointer, that
1206 * we also see the events that were stored up to the tail.
1210 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1215 head %= ctx->nr_events;
1216 tail %= ctx->nr_events;
1220 struct io_event *ev;
1223 avail = (head <= tail ? tail : ctx->nr_events) - head;
1227 pos = head + AIO_EVENTS_OFFSET;
1228 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1229 pos %= AIO_EVENTS_PER_PAGE;
1231 avail = min(avail, nr - ret);
1232 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1235 copy_ret = copy_to_user(event + ret, ev + pos,
1236 sizeof(*ev) * avail);
1239 if (unlikely(copy_ret)) {
1246 head %= ctx->nr_events;
1249 ring = kmap_atomic(ctx->ring_pages[0]);
1251 kunmap_atomic(ring);
1252 flush_dcache_page(ctx->ring_pages[0]);
1254 pr_debug("%li h%u t%u\n", ret, head, tail);
1256 mutex_unlock(&ctx->ring_lock);
1261 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1262 struct io_event __user *event, long *i)
1264 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1269 if (unlikely(atomic_read(&ctx->dead)))
1275 return ret < 0 || *i >= min_nr;
1278 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1279 struct io_event __user *event,
1285 * Note that aio_read_events() is being called as the conditional - i.e.
1286 * we're calling it after prepare_to_wait() has set task state to
1287 * TASK_INTERRUPTIBLE.
1289 * But aio_read_events() can block, and if it blocks it's going to flip
1290 * the task state back to TASK_RUNNING.
1292 * This should be ok, provided it doesn't flip the state back to
1293 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1294 * will only happen if the mutex_lock() call blocks, and we then find
1295 * the ringbuffer empty. So in practice we should be ok, but it's
1296 * something to be aware of when touching this code.
1299 aio_read_events(ctx, min_nr, nr, event, &ret);
1301 wait_event_interruptible_hrtimeout(ctx->wait,
1302 aio_read_events(ctx, min_nr, nr, event, &ret),
1308 * Create an aio_context capable of receiving at least nr_events.
1309 * ctxp must not point to an aio_context that already exists, and
1310 * must be initialized to 0 prior to the call. On successful
1311 * creation of the aio_context, *ctxp is filled in with the resulting
1312 * handle. May fail with -EINVAL if *ctxp is not initialized,
1313 * if the specified nr_events exceeds internal limits. May fail
1314 * with -EAGAIN if the specified nr_events exceeds the user's limit
1315 * of available events. May fail with -ENOMEM if insufficient kernel
1316 * resources are available. May fail with -EFAULT if an invalid
1317 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1320 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1322 struct kioctx *ioctx = NULL;
1326 ret = get_user(ctx, ctxp);
1331 if (unlikely(ctx || nr_events == 0)) {
1332 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1337 ioctx = ioctx_alloc(nr_events);
1338 ret = PTR_ERR(ioctx);
1339 if (!IS_ERR(ioctx)) {
1340 ret = put_user(ioctx->user_id, ctxp);
1342 kill_ioctx(current->mm, ioctx, NULL);
1343 percpu_ref_put(&ioctx->users);
1350 #ifdef CONFIG_COMPAT
1351 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1353 struct kioctx *ioctx = NULL;
1357 ret = get_user(ctx, ctx32p);
1362 if (unlikely(ctx || nr_events == 0)) {
1363 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1368 ioctx = ioctx_alloc(nr_events);
1369 ret = PTR_ERR(ioctx);
1370 if (!IS_ERR(ioctx)) {
1371 /* truncating is ok because it's a user address */
1372 ret = put_user((u32)ioctx->user_id, ctx32p);
1374 kill_ioctx(current->mm, ioctx, NULL);
1375 percpu_ref_put(&ioctx->users);
1384 * Destroy the aio_context specified. May cancel any outstanding
1385 * AIOs and block on completion. Will fail with -ENOSYS if not
1386 * implemented. May fail with -EINVAL if the context pointed to
1389 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1391 struct kioctx *ioctx = lookup_ioctx(ctx);
1392 if (likely(NULL != ioctx)) {
1393 struct ctx_rq_wait wait;
1396 init_completion(&wait.comp);
1397 atomic_set(&wait.count, 1);
1399 /* Pass requests_done to kill_ioctx() where it can be set
1400 * in a thread-safe way. If we try to set it here then we have
1401 * a race condition if two io_destroy() called simultaneously.
1403 ret = kill_ioctx(current->mm, ioctx, &wait);
1404 percpu_ref_put(&ioctx->users);
1406 /* Wait until all IO for the context are done. Otherwise kernel
1407 * keep using user-space buffers even if user thinks the context
1411 wait_for_completion(&wait.comp);
1415 pr_debug("EINVAL: invalid context id\n");
1419 static void aio_remove_iocb(struct aio_kiocb *iocb)
1421 struct kioctx *ctx = iocb->ki_ctx;
1422 unsigned long flags;
1424 spin_lock_irqsave(&ctx->ctx_lock, flags);
1425 list_del(&iocb->ki_list);
1426 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1429 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1431 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1433 if (!list_empty_careful(&iocb->ki_list))
1434 aio_remove_iocb(iocb);
1436 if (kiocb->ki_flags & IOCB_WRITE) {
1437 struct inode *inode = file_inode(kiocb->ki_filp);
1440 * Tell lockdep we inherited freeze protection from submission
1443 if (S_ISREG(inode->i_mode))
1444 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1445 file_end_write(kiocb->ki_filp);
1448 iocb->ki_res.res = res;
1449 iocb->ki_res.res2 = res2;
1453 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1457 req->ki_complete = aio_complete_rw;
1458 req->private = NULL;
1459 req->ki_pos = iocb->aio_offset;
1460 req->ki_flags = iocb_flags(req->ki_filp) | IOCB_AIO_RW;
1461 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1462 req->ki_flags |= IOCB_EVENTFD;
1463 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1464 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1466 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1467 * aio_reqprio is interpreted as an I/O scheduling
1468 * class and priority.
1470 ret = ioprio_check_cap(iocb->aio_reqprio);
1472 pr_debug("aio ioprio check cap error: %d\n", ret);
1476 req->ki_ioprio = iocb->aio_reqprio;
1478 req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0);
1480 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1484 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1488 static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
1489 bool vectored, bool compat, struct iov_iter *iter)
1491 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1492 size_t len = iocb->aio_nbytes;
1495 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1499 #ifdef CONFIG_COMPAT
1501 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1504 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1507 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1513 case -ERESTARTNOINTR:
1514 case -ERESTARTNOHAND:
1515 case -ERESTART_RESTARTBLOCK:
1517 * There's no easy way to restart the syscall since other AIO's
1518 * may be already running. Just fail this IO with EINTR.
1523 req->ki_complete(req, ret, 0);
1527 static ssize_t aio_read(struct kiocb *req, const struct iocb *iocb,
1528 bool vectored, bool compat)
1530 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1531 struct iov_iter iter;
1535 ret = aio_prep_rw(req, iocb);
1538 file = req->ki_filp;
1539 if (unlikely(!(file->f_mode & FMODE_READ)))
1542 if (unlikely(!file->f_op->read_iter))
1545 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1548 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1550 aio_rw_done(req, call_read_iter(file, req, &iter));
1555 static ssize_t aio_write(struct kiocb *req, const struct iocb *iocb,
1556 bool vectored, bool compat)
1558 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1559 struct iov_iter iter;
1563 ret = aio_prep_rw(req, iocb);
1566 file = req->ki_filp;
1568 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1570 if (unlikely(!file->f_op->write_iter))
1573 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1576 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1579 * Open-code file_start_write here to grab freeze protection,
1580 * which will be released by another thread in
1581 * aio_complete_rw(). Fool lockdep by telling it the lock got
1582 * released so that it doesn't complain about the held lock when
1583 * we return to userspace.
1585 if (S_ISREG(file_inode(file)->i_mode)) {
1586 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1587 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1589 req->ki_flags |= IOCB_WRITE;
1590 aio_rw_done(req, call_write_iter(file, req, &iter));
1596 static void aio_fsync_work(struct work_struct *work)
1598 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1599 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1601 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1602 revert_creds(old_cred);
1603 put_cred(iocb->fsync.creds);
1607 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1610 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1611 iocb->aio_rw_flags))
1614 if (unlikely(!req->file->f_op->fsync))
1617 req->creds = prepare_creds();
1621 req->datasync = datasync;
1622 INIT_WORK(&req->work, aio_fsync_work);
1623 schedule_work(&req->work);
1627 static void aio_poll_put_work(struct work_struct *work)
1629 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1630 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1636 * Safely lock the waitqueue which the request is on, synchronizing with the
1637 * case where the ->poll() provider decides to free its waitqueue early.
1639 * Returns true on success, meaning that req->head->lock was locked, req->wait
1640 * is on req->head, and an RCU read lock was taken. Returns false if the
1641 * request was already removed from its waitqueue (which might no longer exist).
1643 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1645 wait_queue_head_t *head;
1648 * While we hold the waitqueue lock and the waitqueue is nonempty,
1649 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1650 * lock in the first place can race with the waitqueue being freed.
1652 * We solve this as eventpoll does: by taking advantage of the fact that
1653 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1654 * we enter rcu_read_lock() and see that the pointer to the queue is
1655 * non-NULL, we can then lock it without the memory being freed out from
1656 * under us, then check whether the request is still on the queue.
1658 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1659 * case the caller deletes the entry from the queue, leaving it empty.
1660 * In that case, only RCU prevents the queue memory from being freed.
1663 head = smp_load_acquire(&req->head);
1665 spin_lock(&head->lock);
1666 if (!list_empty(&req->wait.entry))
1668 spin_unlock(&head->lock);
1674 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1676 spin_unlock(&req->head->lock);
1680 static void aio_poll_complete_work(struct work_struct *work)
1682 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1683 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1684 struct poll_table_struct pt = { ._key = req->events };
1685 struct kioctx *ctx = iocb->ki_ctx;
1688 if (!READ_ONCE(req->cancelled))
1689 mask = vfs_poll(req->file, &pt) & req->events;
1692 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1693 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1694 * synchronize with them. In the cancellation case the list_del_init
1695 * itself is not actually needed, but harmless so we keep it in to
1696 * avoid further branches in the fast path.
1698 spin_lock_irq(&ctx->ctx_lock);
1699 if (poll_iocb_lock_wq(req)) {
1700 if (!mask && !READ_ONCE(req->cancelled)) {
1702 * The request isn't actually ready to be completed yet.
1703 * Reschedule completion if another wakeup came in.
1705 if (req->work_need_resched) {
1706 schedule_work(&req->work);
1707 req->work_need_resched = false;
1709 req->work_scheduled = false;
1711 poll_iocb_unlock_wq(req);
1712 spin_unlock_irq(&ctx->ctx_lock);
1715 list_del_init(&req->wait.entry);
1716 poll_iocb_unlock_wq(req);
1717 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1718 list_del_init(&iocb->ki_list);
1719 iocb->ki_res.res = mangle_poll(mask);
1720 spin_unlock_irq(&ctx->ctx_lock);
1725 /* assumes we are called with irqs disabled */
1726 static int aio_poll_cancel(struct kiocb *iocb)
1728 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1729 struct poll_iocb *req = &aiocb->poll;
1731 if (poll_iocb_lock_wq(req)) {
1732 WRITE_ONCE(req->cancelled, true);
1733 if (!req->work_scheduled) {
1734 schedule_work(&aiocb->poll.work);
1735 req->work_scheduled = true;
1737 poll_iocb_unlock_wq(req);
1738 } /* else, the request was force-cancelled by POLLFREE already */
1743 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1746 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1747 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1748 __poll_t mask = key_to_poll(key);
1749 unsigned long flags;
1751 /* for instances that support it check for an event match first: */
1752 if (mask && !(mask & req->events))
1756 * Complete the request inline if possible. This requires that three
1757 * conditions be met:
1758 * 1. An event mask must have been passed. If a plain wakeup was done
1759 * instead, then mask == 0 and we have to call vfs_poll() to get
1760 * the events, so inline completion isn't possible.
1761 * 2. The completion work must not have already been scheduled.
1762 * 3. ctx_lock must not be busy. We have to use trylock because we
1763 * already hold the waitqueue lock, so this inverts the normal
1764 * locking order. Use irqsave/irqrestore because not all
1765 * filesystems (e.g. fuse) call this function with IRQs disabled,
1766 * yet IRQs have to be disabled before ctx_lock is obtained.
1768 if (mask && !req->work_scheduled &&
1769 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1770 struct kioctx *ctx = iocb->ki_ctx;
1772 list_del_init(&req->wait.entry);
1773 list_del(&iocb->ki_list);
1774 iocb->ki_res.res = mangle_poll(mask);
1775 if (iocb->ki_eventfd && eventfd_signal_count()) {
1777 INIT_WORK(&req->work, aio_poll_put_work);
1778 schedule_work(&req->work);
1780 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1785 * Schedule the completion work if needed. If it was already
1786 * scheduled, record that another wakeup came in.
1788 * Don't remove the request from the waitqueue here, as it might
1789 * not actually be complete yet (we won't know until vfs_poll()
1790 * is called), and we must not miss any wakeups. POLLFREE is an
1791 * exception to this; see below.
1793 if (req->work_scheduled) {
1794 req->work_need_resched = true;
1796 schedule_work(&req->work);
1797 req->work_scheduled = true;
1801 * If the waitqueue is being freed early but we can't complete
1802 * the request inline, we have to tear down the request as best
1803 * we can. That means immediately removing the request from its
1804 * waitqueue and preventing all further accesses to the
1805 * waitqueue via the request. We also need to schedule the
1806 * completion work (done above). Also mark the request as
1807 * cancelled, to potentially skip an unneeded call to ->poll().
1809 if (mask & POLLFREE) {
1810 WRITE_ONCE(req->cancelled, true);
1811 list_del_init(&req->wait.entry);
1814 * Careful: this *must* be the last step, since as soon
1815 * as req->head is NULL'ed out, the request can be
1816 * completed and freed, since aio_poll_complete_work()
1817 * will no longer need to take the waitqueue lock.
1819 smp_store_release(&req->head, NULL);
1825 struct aio_poll_table {
1826 struct poll_table_struct pt;
1827 struct aio_kiocb *iocb;
1833 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1834 struct poll_table_struct *p)
1836 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1838 /* multiple wait queues per file are not supported */
1839 if (unlikely(pt->queued)) {
1840 pt->error = -EINVAL;
1846 pt->iocb->poll.head = head;
1847 add_wait_queue(head, &pt->iocb->poll.wait);
1850 static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1852 struct kioctx *ctx = aiocb->ki_ctx;
1853 struct poll_iocb *req = &aiocb->poll;
1854 struct aio_poll_table apt;
1855 bool cancel = false;
1858 /* reject any unknown events outside the normal event mask. */
1859 if ((u16)iocb->aio_buf != iocb->aio_buf)
1861 /* reject fields that are not defined for poll */
1862 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1865 INIT_WORK(&req->work, aio_poll_complete_work);
1866 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1869 req->cancelled = false;
1870 req->work_scheduled = false;
1871 req->work_need_resched = false;
1873 apt.pt._qproc = aio_poll_queue_proc;
1874 apt.pt._key = req->events;
1877 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1879 /* initialized the list so that we can do list_empty checks */
1880 INIT_LIST_HEAD(&req->wait.entry);
1881 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1883 mask = vfs_poll(req->file, &apt.pt) & req->events;
1884 spin_lock_irq(&ctx->ctx_lock);
1885 if (likely(apt.queued)) {
1886 bool on_queue = poll_iocb_lock_wq(req);
1888 if (!on_queue || req->work_scheduled) {
1890 * aio_poll_wake() already either scheduled the async
1891 * completion work, or completed the request inline.
1893 if (apt.error) /* unsupported case: multiple queues */
1898 if (mask || apt.error) {
1899 /* Steal to complete synchronously. */
1900 list_del_init(&req->wait.entry);
1901 } else if (cancel) {
1902 /* Cancel if possible (may be too late though). */
1903 WRITE_ONCE(req->cancelled, true);
1904 } else if (on_queue) {
1906 * Actually waiting for an event, so add the request to
1907 * active_reqs so that it can be cancelled if needed.
1909 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1910 aiocb->ki_cancel = aio_poll_cancel;
1913 poll_iocb_unlock_wq(req);
1915 if (mask) { /* no async, we'd stolen it */
1916 aiocb->ki_res.res = mangle_poll(mask);
1919 spin_unlock_irq(&ctx->ctx_lock);
1925 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1926 struct iocb __user *user_iocb, bool compat)
1928 struct aio_kiocb *req;
1931 /* enforce forwards compatibility on users */
1932 if (unlikely(iocb->aio_reserved2)) {
1933 pr_debug("EINVAL: reserve field set\n");
1937 /* prevent overflows */
1939 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1940 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1941 ((ssize_t)iocb->aio_nbytes < 0)
1943 pr_debug("EINVAL: overflow check\n");
1947 if (!get_reqs_available(ctx))
1951 req = aio_get_req(ctx);
1953 goto out_put_reqs_available;
1955 req->ki_filp = fget(iocb->aio_fildes);
1957 if (unlikely(!req->ki_filp))
1960 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1962 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1963 * instance of the file* now. The file descriptor must be
1964 * an eventfd() fd, and will be signaled for each completed
1965 * event using the eventfd_signal() function.
1967 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1968 if (IS_ERR(req->ki_eventfd)) {
1969 ret = PTR_ERR(req->ki_eventfd);
1970 req->ki_eventfd = NULL;
1975 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1976 if (unlikely(ret)) {
1977 pr_debug("EFAULT: aio_key\n");
1981 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1982 req->ki_res.data = iocb->aio_data;
1983 req->ki_res.res = 0;
1984 req->ki_res.res2 = 0;
1986 switch (iocb->aio_lio_opcode) {
1987 case IOCB_CMD_PREAD:
1988 ret = aio_read(&req->rw, iocb, false, compat);
1990 case IOCB_CMD_PWRITE:
1991 ret = aio_write(&req->rw, iocb, false, compat);
1993 case IOCB_CMD_PREADV:
1994 ret = aio_read(&req->rw, iocb, true, compat);
1996 case IOCB_CMD_PWRITEV:
1997 ret = aio_write(&req->rw, iocb, true, compat);
1999 case IOCB_CMD_FSYNC:
2000 ret = aio_fsync(&req->fsync, iocb, false);
2002 case IOCB_CMD_FDSYNC:
2003 ret = aio_fsync(&req->fsync, iocb, true);
2006 ret = aio_poll(req, iocb);
2009 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2014 /* Done with the synchronous reference */
2018 * If ret is 0, we'd either done aio_complete() ourselves or have
2019 * arranged for that to be done asynchronously. Anything non-zero
2020 * means that we need to destroy req ourselves.
2026 if (req->ki_eventfd)
2027 eventfd_ctx_put(req->ki_eventfd);
2029 out_put_reqs_available:
2030 put_reqs_available(ctx, 1);
2034 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2039 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2042 return __io_submit_one(ctx, &iocb, user_iocb, compat);
2046 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2047 * the number of iocbs queued. May return -EINVAL if the aio_context
2048 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2049 * *iocbpp[0] is not properly initialized, if the operation specified
2050 * is invalid for the file descriptor in the iocb. May fail with
2051 * -EFAULT if any of the data structures point to invalid data. May
2052 * fail with -EBADF if the file descriptor specified in the first
2053 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2054 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2055 * fail with -ENOSYS if not implemented.
2057 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2058 struct iocb __user * __user *, iocbpp)
2063 struct blk_plug plug;
2065 if (unlikely(nr < 0))
2068 ctx = lookup_ioctx(ctx_id);
2069 if (unlikely(!ctx)) {
2070 pr_debug("EINVAL: invalid context id\n");
2074 if (nr > ctx->nr_events)
2075 nr = ctx->nr_events;
2077 blk_start_plug(&plug);
2078 for (i = 0; i < nr; i++) {
2079 struct iocb __user *user_iocb;
2081 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2086 ret = io_submit_one(ctx, user_iocb, false);
2090 blk_finish_plug(&plug);
2092 percpu_ref_put(&ctx->users);
2096 #ifdef CONFIG_COMPAT
2097 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2098 int, nr, compat_uptr_t __user *, iocbpp)
2103 struct blk_plug plug;
2105 if (unlikely(nr < 0))
2108 ctx = lookup_ioctx(ctx_id);
2109 if (unlikely(!ctx)) {
2110 pr_debug("EINVAL: invalid context id\n");
2114 if (nr > ctx->nr_events)
2115 nr = ctx->nr_events;
2117 blk_start_plug(&plug);
2118 for (i = 0; i < nr; i++) {
2119 compat_uptr_t user_iocb;
2121 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2126 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2130 blk_finish_plug(&plug);
2132 percpu_ref_put(&ctx->users);
2138 * Attempts to cancel an iocb previously passed to io_submit. If
2139 * the operation is successfully cancelled, the resulting event is
2140 * copied into the memory pointed to by result without being placed
2141 * into the completion queue and 0 is returned. May fail with
2142 * -EFAULT if any of the data structures pointed to are invalid.
2143 * May fail with -EINVAL if aio_context specified by ctx_id is
2144 * invalid. May fail with -EAGAIN if the iocb specified was not
2145 * cancelled. Will fail with -ENOSYS if not implemented.
2147 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2148 struct io_event __user *, result)
2151 struct aio_kiocb *kiocb;
2154 u64 obj = (u64)(unsigned long)iocb;
2156 if (unlikely(get_user(key, &iocb->aio_key)))
2158 if (unlikely(key != KIOCB_KEY))
2161 ctx = lookup_ioctx(ctx_id);
2165 spin_lock_irq(&ctx->ctx_lock);
2166 /* TODO: use a hash or array, this sucks. */
2167 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2168 if (kiocb->ki_res.obj == obj) {
2169 ret = kiocb->ki_cancel(&kiocb->rw);
2170 list_del_init(&kiocb->ki_list);
2174 spin_unlock_irq(&ctx->ctx_lock);
2178 * The result argument is no longer used - the io_event is
2179 * always delivered via the ring buffer. -EINPROGRESS indicates
2180 * cancellation is progress:
2185 percpu_ref_put(&ctx->users);
2190 static long do_io_getevents(aio_context_t ctx_id,
2193 struct io_event __user *events,
2194 struct timespec64 *ts)
2196 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2197 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2200 if (likely(ioctx)) {
2201 if (likely(min_nr <= nr && min_nr >= 0))
2202 ret = read_events(ioctx, min_nr, nr, events, until);
2203 percpu_ref_put(&ioctx->users);
2210 * Attempts to read at least min_nr events and up to nr events from
2211 * the completion queue for the aio_context specified by ctx_id. If
2212 * it succeeds, the number of read events is returned. May fail with
2213 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2214 * out of range, if timeout is out of range. May fail with -EFAULT
2215 * if any of the memory specified is invalid. May return 0 or
2216 * < min_nr if the timeout specified by timeout has elapsed
2217 * before sufficient events are available, where timeout == NULL
2218 * specifies an infinite timeout. Note that the timeout pointed to by
2219 * timeout is relative. Will fail with -ENOSYS if not implemented.
2221 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2224 struct io_event __user *, events,
2225 struct timespec __user *, timeout)
2227 struct timespec64 ts;
2230 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2233 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2234 if (!ret && signal_pending(current))
2239 struct __aio_sigset {
2240 const sigset_t __user *sigmask;
2244 SYSCALL_DEFINE6(io_pgetevents,
2245 aio_context_t, ctx_id,
2248 struct io_event __user *, events,
2249 struct timespec __user *, timeout,
2250 const struct __aio_sigset __user *, usig)
2252 struct __aio_sigset ksig = { NULL, };
2253 sigset_t ksigmask, sigsaved;
2254 struct timespec64 ts;
2257 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2260 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2264 if (ksig.sigsetsize != sizeof(sigset_t))
2266 if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask)))
2268 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2269 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2272 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2273 if (signal_pending(current)) {
2275 current->saved_sigmask = sigsaved;
2276 set_restore_sigmask();
2280 ret = -ERESTARTNOHAND;
2283 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2289 #ifdef CONFIG_COMPAT
2290 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
2291 compat_long_t, min_nr,
2293 struct io_event __user *, events,
2294 struct compat_timespec __user *, timeout)
2296 struct timespec64 t;
2299 if (timeout && compat_get_timespec64(&t, timeout))
2302 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2303 if (!ret && signal_pending(current))
2309 struct __compat_aio_sigset {
2310 compat_sigset_t __user *sigmask;
2311 compat_size_t sigsetsize;
2314 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2315 compat_aio_context_t, ctx_id,
2316 compat_long_t, min_nr,
2318 struct io_event __user *, events,
2319 struct compat_timespec __user *, timeout,
2320 const struct __compat_aio_sigset __user *, usig)
2322 struct __compat_aio_sigset ksig = { NULL, };
2323 sigset_t ksigmask, sigsaved;
2324 struct timespec64 t;
2327 if (timeout && compat_get_timespec64(&t, timeout))
2330 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2334 if (ksig.sigsetsize != sizeof(compat_sigset_t))
2336 if (get_compat_sigset(&ksigmask, ksig.sigmask))
2338 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2339 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2342 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2343 if (signal_pending(current)) {
2345 current->saved_sigmask = sigsaved;
2346 set_restore_sigmask();
2349 ret = -ERESTARTNOHAND;
2352 sigprocmask(SIG_SETMASK, &sigsaved, NULL);