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 <linux/uaccess.h>
47 #include <linux/nospec.h>
53 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0
57 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by aio_read_events_ring(). */
64 unsigned compat_features;
65 unsigned incompat_features;
66 unsigned header_length; /* size of aio_ring */
69 struct io_event io_events[];
70 }; /* 128 bytes + ring size */
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
76 #define AIO_PLUG_THRESHOLD 2
78 #define AIO_RING_PAGES 8
83 struct kioctx __rcu *table[];
87 unsigned reqs_available;
91 struct completion comp;
96 struct percpu_ref users;
99 struct percpu_ref reqs;
101 unsigned long user_id;
103 struct __percpu kioctx_cpu *cpu;
106 * For percpu reqs_available, number of slots we move to/from global
111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the global max_reqs quota.
114 * The real limit is nr_events - 1, which will be larger (see
119 /* Size of ringbuffer, in units of struct io_event */
122 unsigned long mmap_base;
123 unsigned long mmap_size;
125 struct page **ring_pages;
128 struct rcu_work free_rwork; /* see free_ioctx() */
131 * signals when all in-flight requests are done
133 struct ctx_rq_wait *rq_wait;
137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the ringbuffer.
142 * We batch accesses to it with a percpu version.
144 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp;
149 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp;
153 struct mutex ring_lock;
154 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp;
159 unsigned completed_events;
160 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp;
163 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file;
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
175 struct work_struct work;
182 struct wait_queue_head *head;
186 bool work_need_resched;
187 struct wait_queue_entry wait;
188 struct work_struct work;
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
199 struct file *ki_filp;
201 struct fsync_iocb fsync;
202 struct poll_iocb poll;
205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel;
208 struct io_event ki_res;
210 struct list_head ki_list; /* the aio core uses this
211 * for cancellation */
212 refcount_t ki_refcnt;
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
218 struct eventfd_ctx *ki_eventfd;
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock);
223 unsigned long aio_nr; /* current system wide number of aio requests */
224 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct kmem_cache *kiocb_cachep;
228 static struct kmem_cache *kioctx_cachep;
230 static struct vfsmount *aio_mnt;
232 static const struct file_operations aio_ring_fops;
233 static const struct address_space_operations aio_ctx_aops;
235 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
238 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
240 return ERR_CAST(inode);
242 inode->i_mapping->a_ops = &aio_ctx_aops;
243 inode->i_mapping->private_data = ctx;
244 inode->i_size = PAGE_SIZE * nr_pages;
246 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
247 O_RDWR, &aio_ring_fops);
253 static int aio_init_fs_context(struct fs_context *fc)
255 if (!init_pseudo(fc, AIO_RING_MAGIC))
257 fc->s_iflags |= SB_I_NOEXEC;
262 * Creates the slab caches used by the aio routines, panic on
263 * failure as this is done early during the boot sequence.
265 static int __init aio_setup(void)
267 static struct file_system_type aio_fs = {
269 .init_fs_context = aio_init_fs_context,
270 .kill_sb = kill_anon_super,
272 aio_mnt = kern_mount(&aio_fs);
274 panic("Failed to create aio fs mount.");
276 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
280 __initcall(aio_setup);
282 static void put_aio_ring_file(struct kioctx *ctx)
284 struct file *aio_ring_file = ctx->aio_ring_file;
285 struct address_space *i_mapping;
288 truncate_setsize(file_inode(aio_ring_file), 0);
290 /* Prevent further access to the kioctx from migratepages */
291 i_mapping = aio_ring_file->f_mapping;
292 spin_lock(&i_mapping->private_lock);
293 i_mapping->private_data = NULL;
294 ctx->aio_ring_file = NULL;
295 spin_unlock(&i_mapping->private_lock);
301 static void aio_free_ring(struct kioctx *ctx)
305 /* Disconnect the kiotx from the ring file. This prevents future
306 * accesses to the kioctx from page migration.
308 put_aio_ring_file(ctx);
310 for (i = 0; i < ctx->nr_pages; i++) {
312 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
313 page_count(ctx->ring_pages[i]));
314 page = ctx->ring_pages[i];
317 ctx->ring_pages[i] = NULL;
321 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
322 kfree(ctx->ring_pages);
323 ctx->ring_pages = NULL;
327 static int aio_ring_mremap(struct vm_area_struct *vma)
329 struct file *file = vma->vm_file;
330 struct mm_struct *mm = vma->vm_mm;
331 struct kioctx_table *table;
332 int i, res = -EINVAL;
334 spin_lock(&mm->ioctx_lock);
336 table = rcu_dereference(mm->ioctx_table);
340 for (i = 0; i < table->nr; i++) {
343 ctx = rcu_dereference(table->table[i]);
344 if (ctx && ctx->aio_ring_file == file) {
345 if (!atomic_read(&ctx->dead)) {
346 ctx->user_id = ctx->mmap_base = vma->vm_start;
355 spin_unlock(&mm->ioctx_lock);
359 static const struct vm_operations_struct aio_ring_vm_ops = {
360 .mremap = aio_ring_mremap,
361 #if IS_ENABLED(CONFIG_MMU)
362 .fault = filemap_fault,
363 .map_pages = filemap_map_pages,
364 .page_mkwrite = filemap_page_mkwrite,
368 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
370 vma->vm_flags |= VM_DONTEXPAND;
371 vma->vm_ops = &aio_ring_vm_ops;
375 static const struct file_operations aio_ring_fops = {
376 .mmap = aio_ring_mmap,
379 #if IS_ENABLED(CONFIG_MIGRATION)
380 static int aio_migratepage(struct address_space *mapping, struct page *new,
381 struct page *old, enum migrate_mode mode)
389 * We cannot support the _NO_COPY case here, because copy needs to
390 * happen under the ctx->completion_lock. That does not work with the
391 * migration workflow of MIGRATE_SYNC_NO_COPY.
393 if (mode == MIGRATE_SYNC_NO_COPY)
398 /* mapping->private_lock here protects against the kioctx teardown. */
399 spin_lock(&mapping->private_lock);
400 ctx = mapping->private_data;
406 /* The ring_lock mutex. The prevents aio_read_events() from writing
407 * to the ring's head, and prevents page migration from mucking in
408 * a partially initialized kiotx.
410 if (!mutex_trylock(&ctx->ring_lock)) {
416 if (idx < (pgoff_t)ctx->nr_pages) {
417 /* Make sure the old page hasn't already been changed */
418 if (ctx->ring_pages[idx] != old)
426 /* Writeback must be complete */
427 BUG_ON(PageWriteback(old));
430 rc = migrate_page_move_mapping(mapping, new, old, 1);
431 if (rc != MIGRATEPAGE_SUCCESS) {
436 /* Take completion_lock to prevent other writes to the ring buffer
437 * while the old page is copied to the new. This prevents new
438 * events from being lost.
440 spin_lock_irqsave(&ctx->completion_lock, flags);
441 migrate_page_copy(new, old);
442 BUG_ON(ctx->ring_pages[idx] != old);
443 ctx->ring_pages[idx] = new;
444 spin_unlock_irqrestore(&ctx->completion_lock, flags);
446 /* The old page is no longer accessible. */
450 mutex_unlock(&ctx->ring_lock);
452 spin_unlock(&mapping->private_lock);
457 static const struct address_space_operations aio_ctx_aops = {
458 .set_page_dirty = __set_page_dirty_no_writeback,
459 #if IS_ENABLED(CONFIG_MIGRATION)
460 .migratepage = aio_migratepage,
464 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
466 struct aio_ring *ring;
467 struct mm_struct *mm = current->mm;
468 unsigned long size, unused;
473 /* Compensate for the ring buffer's head/tail overlap entry */
474 nr_events += 2; /* 1 is required, 2 for good luck */
476 size = sizeof(struct aio_ring);
477 size += sizeof(struct io_event) * nr_events;
479 nr_pages = PFN_UP(size);
483 file = aio_private_file(ctx, nr_pages);
485 ctx->aio_ring_file = NULL;
489 ctx->aio_ring_file = file;
490 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
491 / sizeof(struct io_event);
493 ctx->ring_pages = ctx->internal_pages;
494 if (nr_pages > AIO_RING_PAGES) {
495 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
497 if (!ctx->ring_pages) {
498 put_aio_ring_file(ctx);
503 for (i = 0; i < nr_pages; i++) {
505 page = find_or_create_page(file->f_mapping,
506 i, GFP_HIGHUSER | __GFP_ZERO);
509 pr_debug("pid(%d) page[%d]->count=%d\n",
510 current->pid, i, page_count(page));
511 SetPageUptodate(page);
514 ctx->ring_pages[i] = page;
518 if (unlikely(i != nr_pages)) {
523 ctx->mmap_size = nr_pages * PAGE_SIZE;
524 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
526 if (mmap_write_lock_killable(mm)) {
532 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
533 PROT_READ | PROT_WRITE,
534 MAP_SHARED, 0, &unused, NULL);
535 mmap_write_unlock(mm);
536 if (IS_ERR((void *)ctx->mmap_base)) {
542 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
544 ctx->user_id = ctx->mmap_base;
545 ctx->nr_events = nr_events; /* trusted copy */
547 ring = kmap_atomic(ctx->ring_pages[0]);
548 ring->nr = nr_events; /* user copy */
550 ring->head = ring->tail = 0;
551 ring->magic = AIO_RING_MAGIC;
552 ring->compat_features = AIO_RING_COMPAT_FEATURES;
553 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
554 ring->header_length = sizeof(struct aio_ring);
556 flush_dcache_page(ctx->ring_pages[0]);
561 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
562 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
563 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
565 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
567 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
568 struct kioctx *ctx = req->ki_ctx;
571 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
574 spin_lock_irqsave(&ctx->ctx_lock, flags);
575 list_add_tail(&req->ki_list, &ctx->active_reqs);
576 req->ki_cancel = cancel;
577 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
579 EXPORT_SYMBOL(kiocb_set_cancel_fn);
582 * free_ioctx() should be RCU delayed to synchronize against the RCU
583 * protected lookup_ioctx() and also needs process context to call
584 * aio_free_ring(). Use rcu_work.
586 static void free_ioctx(struct work_struct *work)
588 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
590 pr_debug("freeing %p\n", ctx);
593 free_percpu(ctx->cpu);
594 percpu_ref_exit(&ctx->reqs);
595 percpu_ref_exit(&ctx->users);
596 kmem_cache_free(kioctx_cachep, ctx);
599 static void free_ioctx_reqs(struct percpu_ref *ref)
601 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
603 /* At this point we know that there are no any in-flight requests */
604 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
605 complete(&ctx->rq_wait->comp);
607 /* Synchronize against RCU protected table->table[] dereferences */
608 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
609 queue_rcu_work(system_wq, &ctx->free_rwork);
613 * When this function runs, the kioctx has been removed from the "hash table"
614 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
615 * now it's safe to cancel any that need to be.
617 static void free_ioctx_users(struct percpu_ref *ref)
619 struct kioctx *ctx = container_of(ref, struct kioctx, users);
620 struct aio_kiocb *req;
622 spin_lock_irq(&ctx->ctx_lock);
624 while (!list_empty(&ctx->active_reqs)) {
625 req = list_first_entry(&ctx->active_reqs,
626 struct aio_kiocb, ki_list);
627 req->ki_cancel(&req->rw);
628 list_del_init(&req->ki_list);
631 spin_unlock_irq(&ctx->ctx_lock);
633 percpu_ref_kill(&ctx->reqs);
634 percpu_ref_put(&ctx->reqs);
637 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
640 struct kioctx_table *table, *old;
641 struct aio_ring *ring;
643 spin_lock(&mm->ioctx_lock);
644 table = rcu_dereference_raw(mm->ioctx_table);
648 for (i = 0; i < table->nr; i++)
649 if (!rcu_access_pointer(table->table[i])) {
651 rcu_assign_pointer(table->table[i], ctx);
652 spin_unlock(&mm->ioctx_lock);
654 /* While kioctx setup is in progress,
655 * we are protected from page migration
656 * changes ring_pages by ->ring_lock.
658 ring = kmap_atomic(ctx->ring_pages[0]);
664 new_nr = (table ? table->nr : 1) * 4;
665 spin_unlock(&mm->ioctx_lock);
667 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
674 spin_lock(&mm->ioctx_lock);
675 old = rcu_dereference_raw(mm->ioctx_table);
678 rcu_assign_pointer(mm->ioctx_table, table);
679 } else if (table->nr > old->nr) {
680 memcpy(table->table, old->table,
681 old->nr * sizeof(struct kioctx *));
683 rcu_assign_pointer(mm->ioctx_table, table);
692 static void aio_nr_sub(unsigned nr)
694 spin_lock(&aio_nr_lock);
695 if (WARN_ON(aio_nr - nr > aio_nr))
699 spin_unlock(&aio_nr_lock);
703 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
705 static struct kioctx *ioctx_alloc(unsigned nr_events)
707 struct mm_struct *mm = current->mm;
712 * Store the original nr_events -- what userspace passed to io_setup(),
713 * for counting against the global limit -- before it changes.
715 unsigned int max_reqs = nr_events;
718 * We keep track of the number of available ringbuffer slots, to prevent
719 * overflow (reqs_available), and we also use percpu counters for this.
721 * So since up to half the slots might be on other cpu's percpu counters
722 * and unavailable, double nr_events so userspace sees what they
723 * expected: additionally, we move req_batch slots to/from percpu
724 * counters at a time, so make sure that isn't 0:
726 nr_events = max(nr_events, num_possible_cpus() * 4);
729 /* Prevent overflows */
730 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
731 pr_debug("ENOMEM: nr_events too high\n");
732 return ERR_PTR(-EINVAL);
735 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
736 return ERR_PTR(-EAGAIN);
738 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
740 return ERR_PTR(-ENOMEM);
742 ctx->max_reqs = max_reqs;
744 spin_lock_init(&ctx->ctx_lock);
745 spin_lock_init(&ctx->completion_lock);
746 mutex_init(&ctx->ring_lock);
747 /* Protect against page migration throughout kiotx setup by keeping
748 * the ring_lock mutex held until setup is complete. */
749 mutex_lock(&ctx->ring_lock);
750 init_waitqueue_head(&ctx->wait);
752 INIT_LIST_HEAD(&ctx->active_reqs);
754 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
757 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
760 ctx->cpu = alloc_percpu(struct kioctx_cpu);
764 err = aio_setup_ring(ctx, nr_events);
768 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
769 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
770 if (ctx->req_batch < 1)
773 /* limit the number of system wide aios */
774 spin_lock(&aio_nr_lock);
775 if (aio_nr + ctx->max_reqs > aio_max_nr ||
776 aio_nr + ctx->max_reqs < aio_nr) {
777 spin_unlock(&aio_nr_lock);
781 aio_nr += ctx->max_reqs;
782 spin_unlock(&aio_nr_lock);
784 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
785 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
787 err = ioctx_add_table(ctx, mm);
791 /* Release the ring_lock mutex now that all setup is complete. */
792 mutex_unlock(&ctx->ring_lock);
794 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
795 ctx, ctx->user_id, mm, ctx->nr_events);
799 aio_nr_sub(ctx->max_reqs);
801 atomic_set(&ctx->dead, 1);
803 vm_munmap(ctx->mmap_base, ctx->mmap_size);
806 mutex_unlock(&ctx->ring_lock);
807 free_percpu(ctx->cpu);
808 percpu_ref_exit(&ctx->reqs);
809 percpu_ref_exit(&ctx->users);
810 kmem_cache_free(kioctx_cachep, ctx);
811 pr_debug("error allocating ioctx %d\n", err);
816 * Cancels all outstanding aio requests on an aio context. Used
817 * when the processes owning a context have all exited to encourage
818 * the rapid destruction of the kioctx.
820 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
821 struct ctx_rq_wait *wait)
823 struct kioctx_table *table;
825 spin_lock(&mm->ioctx_lock);
826 if (atomic_xchg(&ctx->dead, 1)) {
827 spin_unlock(&mm->ioctx_lock);
831 table = rcu_dereference_raw(mm->ioctx_table);
832 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
833 RCU_INIT_POINTER(table->table[ctx->id], NULL);
834 spin_unlock(&mm->ioctx_lock);
836 /* free_ioctx_reqs() will do the necessary RCU synchronization */
837 wake_up_all(&ctx->wait);
840 * It'd be more correct to do this in free_ioctx(), after all
841 * the outstanding kiocbs have finished - but by then io_destroy
842 * has already returned, so io_setup() could potentially return
843 * -EAGAIN with no ioctxs actually in use (as far as userspace
846 aio_nr_sub(ctx->max_reqs);
849 vm_munmap(ctx->mmap_base, ctx->mmap_size);
852 percpu_ref_kill(&ctx->users);
857 * exit_aio: called when the last user of mm goes away. At this point, there is
858 * no way for any new requests to be submited or any of the io_* syscalls to be
859 * called on the context.
861 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
864 void exit_aio(struct mm_struct *mm)
866 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
867 struct ctx_rq_wait wait;
873 atomic_set(&wait.count, table->nr);
874 init_completion(&wait.comp);
877 for (i = 0; i < table->nr; ++i) {
879 rcu_dereference_protected(table->table[i], true);
887 * We don't need to bother with munmap() here - exit_mmap(mm)
888 * is coming and it'll unmap everything. And we simply can't,
889 * this is not necessarily our ->mm.
890 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
891 * that it needs to unmap the area, just set it to 0.
894 kill_ioctx(mm, ctx, &wait);
897 if (!atomic_sub_and_test(skipped, &wait.count)) {
898 /* Wait until all IO for the context are done. */
899 wait_for_completion(&wait.comp);
902 RCU_INIT_POINTER(mm->ioctx_table, NULL);
906 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
908 struct kioctx_cpu *kcpu;
911 local_irq_save(flags);
912 kcpu = this_cpu_ptr(ctx->cpu);
913 kcpu->reqs_available += nr;
915 while (kcpu->reqs_available >= ctx->req_batch * 2) {
916 kcpu->reqs_available -= ctx->req_batch;
917 atomic_add(ctx->req_batch, &ctx->reqs_available);
920 local_irq_restore(flags);
923 static bool __get_reqs_available(struct kioctx *ctx)
925 struct kioctx_cpu *kcpu;
929 local_irq_save(flags);
930 kcpu = this_cpu_ptr(ctx->cpu);
931 if (!kcpu->reqs_available) {
932 int old, avail = atomic_read(&ctx->reqs_available);
935 if (avail < ctx->req_batch)
939 avail = atomic_cmpxchg(&ctx->reqs_available,
940 avail, avail - ctx->req_batch);
941 } while (avail != old);
943 kcpu->reqs_available += ctx->req_batch;
947 kcpu->reqs_available--;
949 local_irq_restore(flags);
953 /* refill_reqs_available
954 * Updates the reqs_available reference counts used for tracking the
955 * number of free slots in the completion ring. This can be called
956 * from aio_complete() (to optimistically update reqs_available) or
957 * from aio_get_req() (the we're out of events case). It must be
958 * called holding ctx->completion_lock.
960 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
963 unsigned events_in_ring, completed;
965 /* Clamp head since userland can write to it. */
966 head %= ctx->nr_events;
968 events_in_ring = tail - head;
970 events_in_ring = ctx->nr_events - (head - tail);
972 completed = ctx->completed_events;
973 if (events_in_ring < completed)
974 completed -= events_in_ring;
981 ctx->completed_events -= completed;
982 put_reqs_available(ctx, completed);
985 /* user_refill_reqs_available
986 * Called to refill reqs_available when aio_get_req() encounters an
987 * out of space in the completion ring.
989 static void user_refill_reqs_available(struct kioctx *ctx)
991 spin_lock_irq(&ctx->completion_lock);
992 if (ctx->completed_events) {
993 struct aio_ring *ring;
996 /* Access of ring->head may race with aio_read_events_ring()
997 * here, but that's okay since whether we read the old version
998 * or the new version, and either will be valid. The important
999 * part is that head cannot pass tail since we prevent
1000 * aio_complete() from updating tail by holding
1001 * ctx->completion_lock. Even if head is invalid, the check
1002 * against ctx->completed_events below will make sure we do the
1005 ring = kmap_atomic(ctx->ring_pages[0]);
1007 kunmap_atomic(ring);
1009 refill_reqs_available(ctx, head, ctx->tail);
1012 spin_unlock_irq(&ctx->completion_lock);
1015 static bool get_reqs_available(struct kioctx *ctx)
1017 if (__get_reqs_available(ctx))
1019 user_refill_reqs_available(ctx);
1020 return __get_reqs_available(ctx);
1024 * Allocate a slot for an aio request.
1025 * Returns NULL if no requests are free.
1027 * The refcount is initialized to 2 - one for the async op completion,
1028 * one for the synchronous code that does this.
1030 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1032 struct aio_kiocb *req;
1034 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1038 if (unlikely(!get_reqs_available(ctx))) {
1039 kmem_cache_free(kiocb_cachep, req);
1043 percpu_ref_get(&ctx->reqs);
1045 INIT_LIST_HEAD(&req->ki_list);
1046 refcount_set(&req->ki_refcnt, 2);
1047 req->ki_eventfd = NULL;
1051 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1053 struct aio_ring __user *ring = (void __user *)ctx_id;
1054 struct mm_struct *mm = current->mm;
1055 struct kioctx *ctx, *ret = NULL;
1056 struct kioctx_table *table;
1059 if (get_user(id, &ring->id))
1063 table = rcu_dereference(mm->ioctx_table);
1065 if (!table || id >= table->nr)
1068 id = array_index_nospec(id, table->nr);
1069 ctx = rcu_dereference(table->table[id]);
1070 if (ctx && ctx->user_id == ctx_id) {
1071 if (percpu_ref_tryget_live(&ctx->users))
1079 static inline void iocb_destroy(struct aio_kiocb *iocb)
1081 if (iocb->ki_eventfd)
1082 eventfd_ctx_put(iocb->ki_eventfd);
1084 fput(iocb->ki_filp);
1085 percpu_ref_put(&iocb->ki_ctx->reqs);
1086 kmem_cache_free(kiocb_cachep, iocb);
1090 * Called when the io request on the given iocb is complete.
1092 static void aio_complete(struct aio_kiocb *iocb)
1094 struct kioctx *ctx = iocb->ki_ctx;
1095 struct aio_ring *ring;
1096 struct io_event *ev_page, *event;
1097 unsigned tail, pos, head;
1098 unsigned long flags;
1101 * Add a completion event to the ring buffer. Must be done holding
1102 * ctx->completion_lock to prevent other code from messing with the tail
1103 * pointer since we might be called from irq context.
1105 spin_lock_irqsave(&ctx->completion_lock, flags);
1108 pos = tail + AIO_EVENTS_OFFSET;
1110 if (++tail >= ctx->nr_events)
1113 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1114 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1116 *event = iocb->ki_res;
1118 kunmap_atomic(ev_page);
1119 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1121 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1122 (void __user *)(unsigned long)iocb->ki_res.obj,
1123 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1125 /* after flagging the request as done, we
1126 * must never even look at it again
1128 smp_wmb(); /* make event visible before updating tail */
1132 ring = kmap_atomic(ctx->ring_pages[0]);
1135 kunmap_atomic(ring);
1136 flush_dcache_page(ctx->ring_pages[0]);
1138 ctx->completed_events++;
1139 if (ctx->completed_events > 1)
1140 refill_reqs_available(ctx, head, tail);
1141 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1143 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1146 * Check if the user asked us to deliver the result through an
1147 * eventfd. The eventfd_signal() function is safe to be called
1150 if (iocb->ki_eventfd)
1151 eventfd_signal(iocb->ki_eventfd, 1);
1154 * We have to order our ring_info tail store above and test
1155 * of the wait list below outside the wait lock. This is
1156 * like in wake_up_bit() where clearing a bit has to be
1157 * ordered with the unlocked test.
1161 if (waitqueue_active(&ctx->wait))
1162 wake_up(&ctx->wait);
1165 static inline void iocb_put(struct aio_kiocb *iocb)
1167 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1173 /* aio_read_events_ring
1174 * Pull an event off of the ioctx's event ring. Returns the number of
1177 static long aio_read_events_ring(struct kioctx *ctx,
1178 struct io_event __user *event, long nr)
1180 struct aio_ring *ring;
1181 unsigned head, tail, pos;
1186 * The mutex can block and wake us up and that will cause
1187 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1188 * and repeat. This should be rare enough that it doesn't cause
1189 * peformance issues. See the comment in read_events() for more detail.
1191 sched_annotate_sleep();
1192 mutex_lock(&ctx->ring_lock);
1194 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1195 ring = kmap_atomic(ctx->ring_pages[0]);
1198 kunmap_atomic(ring);
1201 * Ensure that once we've read the current tail pointer, that
1202 * we also see the events that were stored up to the tail.
1206 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1211 head %= ctx->nr_events;
1212 tail %= ctx->nr_events;
1216 struct io_event *ev;
1219 avail = (head <= tail ? tail : ctx->nr_events) - head;
1223 pos = head + AIO_EVENTS_OFFSET;
1224 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1225 pos %= AIO_EVENTS_PER_PAGE;
1227 avail = min(avail, nr - ret);
1228 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1231 copy_ret = copy_to_user(event + ret, ev + pos,
1232 sizeof(*ev) * avail);
1235 if (unlikely(copy_ret)) {
1242 head %= ctx->nr_events;
1245 ring = kmap_atomic(ctx->ring_pages[0]);
1247 kunmap_atomic(ring);
1248 flush_dcache_page(ctx->ring_pages[0]);
1250 pr_debug("%li h%u t%u\n", ret, head, tail);
1252 mutex_unlock(&ctx->ring_lock);
1257 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1258 struct io_event __user *event, long *i)
1260 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1265 if (unlikely(atomic_read(&ctx->dead)))
1271 return ret < 0 || *i >= min_nr;
1274 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1275 struct io_event __user *event,
1281 * Note that aio_read_events() is being called as the conditional - i.e.
1282 * we're calling it after prepare_to_wait() has set task state to
1283 * TASK_INTERRUPTIBLE.
1285 * But aio_read_events() can block, and if it blocks it's going to flip
1286 * the task state back to TASK_RUNNING.
1288 * This should be ok, provided it doesn't flip the state back to
1289 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1290 * will only happen if the mutex_lock() call blocks, and we then find
1291 * the ringbuffer empty. So in practice we should be ok, but it's
1292 * something to be aware of when touching this code.
1295 aio_read_events(ctx, min_nr, nr, event, &ret);
1297 wait_event_interruptible_hrtimeout(ctx->wait,
1298 aio_read_events(ctx, min_nr, nr, event, &ret),
1304 * Create an aio_context capable of receiving at least nr_events.
1305 * ctxp must not point to an aio_context that already exists, and
1306 * must be initialized to 0 prior to the call. On successful
1307 * creation of the aio_context, *ctxp is filled in with the resulting
1308 * handle. May fail with -EINVAL if *ctxp is not initialized,
1309 * if the specified nr_events exceeds internal limits. May fail
1310 * with -EAGAIN if the specified nr_events exceeds the user's limit
1311 * of available events. May fail with -ENOMEM if insufficient kernel
1312 * resources are available. May fail with -EFAULT if an invalid
1313 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1316 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1318 struct kioctx *ioctx = NULL;
1322 ret = get_user(ctx, ctxp);
1327 if (unlikely(ctx || nr_events == 0)) {
1328 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1333 ioctx = ioctx_alloc(nr_events);
1334 ret = PTR_ERR(ioctx);
1335 if (!IS_ERR(ioctx)) {
1336 ret = put_user(ioctx->user_id, ctxp);
1338 kill_ioctx(current->mm, ioctx, NULL);
1339 percpu_ref_put(&ioctx->users);
1346 #ifdef CONFIG_COMPAT
1347 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1349 struct kioctx *ioctx = NULL;
1353 ret = get_user(ctx, ctx32p);
1358 if (unlikely(ctx || nr_events == 0)) {
1359 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1364 ioctx = ioctx_alloc(nr_events);
1365 ret = PTR_ERR(ioctx);
1366 if (!IS_ERR(ioctx)) {
1367 /* truncating is ok because it's a user address */
1368 ret = put_user((u32)ioctx->user_id, ctx32p);
1370 kill_ioctx(current->mm, ioctx, NULL);
1371 percpu_ref_put(&ioctx->users);
1380 * Destroy the aio_context specified. May cancel any outstanding
1381 * AIOs and block on completion. Will fail with -ENOSYS if not
1382 * implemented. May fail with -EINVAL if the context pointed to
1385 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1387 struct kioctx *ioctx = lookup_ioctx(ctx);
1388 if (likely(NULL != ioctx)) {
1389 struct ctx_rq_wait wait;
1392 init_completion(&wait.comp);
1393 atomic_set(&wait.count, 1);
1395 /* Pass requests_done to kill_ioctx() where it can be set
1396 * in a thread-safe way. If we try to set it here then we have
1397 * a race condition if two io_destroy() called simultaneously.
1399 ret = kill_ioctx(current->mm, ioctx, &wait);
1400 percpu_ref_put(&ioctx->users);
1402 /* Wait until all IO for the context are done. Otherwise kernel
1403 * keep using user-space buffers even if user thinks the context
1407 wait_for_completion(&wait.comp);
1411 pr_debug("EINVAL: invalid context id\n");
1415 static void aio_remove_iocb(struct aio_kiocb *iocb)
1417 struct kioctx *ctx = iocb->ki_ctx;
1418 unsigned long flags;
1420 spin_lock_irqsave(&ctx->ctx_lock, flags);
1421 list_del(&iocb->ki_list);
1422 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1425 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1427 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1429 if (!list_empty_careful(&iocb->ki_list))
1430 aio_remove_iocb(iocb);
1432 if (kiocb->ki_flags & IOCB_WRITE) {
1433 struct inode *inode = file_inode(kiocb->ki_filp);
1436 * Tell lockdep we inherited freeze protection from submission
1439 if (S_ISREG(inode->i_mode))
1440 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1441 file_end_write(kiocb->ki_filp);
1444 iocb->ki_res.res = res;
1445 iocb->ki_res.res2 = res2;
1449 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1453 req->ki_complete = aio_complete_rw;
1454 req->private = NULL;
1455 req->ki_pos = iocb->aio_offset;
1456 req->ki_flags = iocb_flags(req->ki_filp);
1457 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1458 req->ki_flags |= IOCB_EVENTFD;
1459 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1460 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1462 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1463 * aio_reqprio is interpreted as an I/O scheduling
1464 * class and priority.
1466 ret = ioprio_check_cap(iocb->aio_reqprio);
1468 pr_debug("aio ioprio check cap error: %d\n", ret);
1472 req->ki_ioprio = iocb->aio_reqprio;
1474 req->ki_ioprio = get_current_ioprio();
1476 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1480 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1484 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1485 struct iovec **iovec, bool vectored, bool compat,
1486 struct iov_iter *iter)
1488 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1489 size_t len = iocb->aio_nbytes;
1492 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1497 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1500 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1506 case -ERESTARTNOINTR:
1507 case -ERESTARTNOHAND:
1508 case -ERESTART_RESTARTBLOCK:
1510 * There's no easy way to restart the syscall since other AIO's
1511 * may be already running. Just fail this IO with EINTR.
1516 req->ki_complete(req, ret, 0);
1520 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1521 bool vectored, bool compat)
1523 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1524 struct iov_iter iter;
1528 ret = aio_prep_rw(req, iocb);
1531 file = req->ki_filp;
1532 if (unlikely(!(file->f_mode & FMODE_READ)))
1535 if (unlikely(!file->f_op->read_iter))
1538 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1541 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1543 aio_rw_done(req, call_read_iter(file, req, &iter));
1548 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1549 bool vectored, bool compat)
1551 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1552 struct iov_iter iter;
1556 ret = aio_prep_rw(req, iocb);
1559 file = req->ki_filp;
1561 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1563 if (unlikely(!file->f_op->write_iter))
1566 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1569 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1572 * Open-code file_start_write here to grab freeze protection,
1573 * which will be released by another thread in
1574 * aio_complete_rw(). Fool lockdep by telling it the lock got
1575 * released so that it doesn't complain about the held lock when
1576 * we return to userspace.
1578 if (S_ISREG(file_inode(file)->i_mode)) {
1579 sb_start_write(file_inode(file)->i_sb);
1580 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1582 req->ki_flags |= IOCB_WRITE;
1583 aio_rw_done(req, call_write_iter(file, req, &iter));
1589 static void aio_fsync_work(struct work_struct *work)
1591 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1592 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1594 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1595 revert_creds(old_cred);
1596 put_cred(iocb->fsync.creds);
1600 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1603 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1604 iocb->aio_rw_flags))
1607 if (unlikely(!req->file->f_op->fsync))
1610 req->creds = prepare_creds();
1614 req->datasync = datasync;
1615 INIT_WORK(&req->work, aio_fsync_work);
1616 schedule_work(&req->work);
1620 static void aio_poll_put_work(struct work_struct *work)
1622 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1623 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1629 * Safely lock the waitqueue which the request is on, synchronizing with the
1630 * case where the ->poll() provider decides to free its waitqueue early.
1632 * Returns true on success, meaning that req->head->lock was locked, req->wait
1633 * is on req->head, and an RCU read lock was taken. Returns false if the
1634 * request was already removed from its waitqueue (which might no longer exist).
1636 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1638 wait_queue_head_t *head;
1641 * While we hold the waitqueue lock and the waitqueue is nonempty,
1642 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1643 * lock in the first place can race with the waitqueue being freed.
1645 * We solve this as eventpoll does: by taking advantage of the fact that
1646 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1647 * we enter rcu_read_lock() and see that the pointer to the queue is
1648 * non-NULL, we can then lock it without the memory being freed out from
1649 * under us, then check whether the request is still on the queue.
1651 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1652 * case the caller deletes the entry from the queue, leaving it empty.
1653 * In that case, only RCU prevents the queue memory from being freed.
1656 head = smp_load_acquire(&req->head);
1658 spin_lock(&head->lock);
1659 if (!list_empty(&req->wait.entry))
1661 spin_unlock(&head->lock);
1667 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1669 spin_unlock(&req->head->lock);
1673 static void aio_poll_complete_work(struct work_struct *work)
1675 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1676 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1677 struct poll_table_struct pt = { ._key = req->events };
1678 struct kioctx *ctx = iocb->ki_ctx;
1681 if (!READ_ONCE(req->cancelled))
1682 mask = vfs_poll(req->file, &pt) & req->events;
1685 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1686 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1687 * synchronize with them. In the cancellation case the list_del_init
1688 * itself is not actually needed, but harmless so we keep it in to
1689 * avoid further branches in the fast path.
1691 spin_lock_irq(&ctx->ctx_lock);
1692 if (poll_iocb_lock_wq(req)) {
1693 if (!mask && !READ_ONCE(req->cancelled)) {
1695 * The request isn't actually ready to be completed yet.
1696 * Reschedule completion if another wakeup came in.
1698 if (req->work_need_resched) {
1699 schedule_work(&req->work);
1700 req->work_need_resched = false;
1702 req->work_scheduled = false;
1704 poll_iocb_unlock_wq(req);
1705 spin_unlock_irq(&ctx->ctx_lock);
1708 list_del_init(&req->wait.entry);
1709 poll_iocb_unlock_wq(req);
1710 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1711 list_del_init(&iocb->ki_list);
1712 iocb->ki_res.res = mangle_poll(mask);
1713 spin_unlock_irq(&ctx->ctx_lock);
1718 /* assumes we are called with irqs disabled */
1719 static int aio_poll_cancel(struct kiocb *iocb)
1721 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1722 struct poll_iocb *req = &aiocb->poll;
1724 if (poll_iocb_lock_wq(req)) {
1725 WRITE_ONCE(req->cancelled, true);
1726 if (!req->work_scheduled) {
1727 schedule_work(&aiocb->poll.work);
1728 req->work_scheduled = true;
1730 poll_iocb_unlock_wq(req);
1731 } /* else, the request was force-cancelled by POLLFREE already */
1736 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1739 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1740 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1741 __poll_t mask = key_to_poll(key);
1742 unsigned long flags;
1744 /* for instances that support it check for an event match first: */
1745 if (mask && !(mask & req->events))
1749 * Complete the request inline if possible. This requires that three
1750 * conditions be met:
1751 * 1. An event mask must have been passed. If a plain wakeup was done
1752 * instead, then mask == 0 and we have to call vfs_poll() to get
1753 * the events, so inline completion isn't possible.
1754 * 2. The completion work must not have already been scheduled.
1755 * 3. ctx_lock must not be busy. We have to use trylock because we
1756 * already hold the waitqueue lock, so this inverts the normal
1757 * locking order. Use irqsave/irqrestore because not all
1758 * filesystems (e.g. fuse) call this function with IRQs disabled,
1759 * yet IRQs have to be disabled before ctx_lock is obtained.
1761 if (mask && !req->work_scheduled &&
1762 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1763 struct kioctx *ctx = iocb->ki_ctx;
1765 list_del_init(&req->wait.entry);
1766 list_del(&iocb->ki_list);
1767 iocb->ki_res.res = mangle_poll(mask);
1768 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1770 INIT_WORK(&req->work, aio_poll_put_work);
1771 schedule_work(&req->work);
1773 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1778 * Schedule the completion work if needed. If it was already
1779 * scheduled, record that another wakeup came in.
1781 * Don't remove the request from the waitqueue here, as it might
1782 * not actually be complete yet (we won't know until vfs_poll()
1783 * is called), and we must not miss any wakeups. POLLFREE is an
1784 * exception to this; see below.
1786 if (req->work_scheduled) {
1787 req->work_need_resched = true;
1789 schedule_work(&req->work);
1790 req->work_scheduled = true;
1794 * If the waitqueue is being freed early but we can't complete
1795 * the request inline, we have to tear down the request as best
1796 * we can. That means immediately removing the request from its
1797 * waitqueue and preventing all further accesses to the
1798 * waitqueue via the request. We also need to schedule the
1799 * completion work (done above). Also mark the request as
1800 * cancelled, to potentially skip an unneeded call to ->poll().
1802 if (mask & POLLFREE) {
1803 WRITE_ONCE(req->cancelled, true);
1804 list_del_init(&req->wait.entry);
1807 * Careful: this *must* be the last step, since as soon
1808 * as req->head is NULL'ed out, the request can be
1809 * completed and freed, since aio_poll_complete_work()
1810 * will no longer need to take the waitqueue lock.
1812 smp_store_release(&req->head, NULL);
1818 struct aio_poll_table {
1819 struct poll_table_struct pt;
1820 struct aio_kiocb *iocb;
1826 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1827 struct poll_table_struct *p)
1829 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1831 /* multiple wait queues per file are not supported */
1832 if (unlikely(pt->queued)) {
1833 pt->error = -EINVAL;
1839 pt->iocb->poll.head = head;
1840 add_wait_queue(head, &pt->iocb->poll.wait);
1843 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1845 struct kioctx *ctx = aiocb->ki_ctx;
1846 struct poll_iocb *req = &aiocb->poll;
1847 struct aio_poll_table apt;
1848 bool cancel = false;
1851 /* reject any unknown events outside the normal event mask. */
1852 if ((u16)iocb->aio_buf != iocb->aio_buf)
1854 /* reject fields that are not defined for poll */
1855 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1858 INIT_WORK(&req->work, aio_poll_complete_work);
1859 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1862 req->cancelled = false;
1863 req->work_scheduled = false;
1864 req->work_need_resched = false;
1866 apt.pt._qproc = aio_poll_queue_proc;
1867 apt.pt._key = req->events;
1870 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1872 /* initialized the list so that we can do list_empty checks */
1873 INIT_LIST_HEAD(&req->wait.entry);
1874 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1876 mask = vfs_poll(req->file, &apt.pt) & req->events;
1877 spin_lock_irq(&ctx->ctx_lock);
1878 if (likely(apt.queued)) {
1879 bool on_queue = poll_iocb_lock_wq(req);
1881 if (!on_queue || req->work_scheduled) {
1883 * aio_poll_wake() already either scheduled the async
1884 * completion work, or completed the request inline.
1886 if (apt.error) /* unsupported case: multiple queues */
1891 if (mask || apt.error) {
1892 /* Steal to complete synchronously. */
1893 list_del_init(&req->wait.entry);
1894 } else if (cancel) {
1895 /* Cancel if possible (may be too late though). */
1896 WRITE_ONCE(req->cancelled, true);
1897 } else if (on_queue) {
1899 * Actually waiting for an event, so add the request to
1900 * active_reqs so that it can be cancelled if needed.
1902 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1903 aiocb->ki_cancel = aio_poll_cancel;
1906 poll_iocb_unlock_wq(req);
1908 if (mask) { /* no async, we'd stolen it */
1909 aiocb->ki_res.res = mangle_poll(mask);
1912 spin_unlock_irq(&ctx->ctx_lock);
1918 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1919 struct iocb __user *user_iocb, struct aio_kiocb *req,
1922 req->ki_filp = fget(iocb->aio_fildes);
1923 if (unlikely(!req->ki_filp))
1926 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1927 struct eventfd_ctx *eventfd;
1929 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1930 * instance of the file* now. The file descriptor must be
1931 * an eventfd() fd, and will be signaled for each completed
1932 * event using the eventfd_signal() function.
1934 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1935 if (IS_ERR(eventfd))
1936 return PTR_ERR(eventfd);
1938 req->ki_eventfd = eventfd;
1941 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1942 pr_debug("EFAULT: aio_key\n");
1946 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1947 req->ki_res.data = iocb->aio_data;
1948 req->ki_res.res = 0;
1949 req->ki_res.res2 = 0;
1951 switch (iocb->aio_lio_opcode) {
1952 case IOCB_CMD_PREAD:
1953 return aio_read(&req->rw, iocb, false, compat);
1954 case IOCB_CMD_PWRITE:
1955 return aio_write(&req->rw, iocb, false, compat);
1956 case IOCB_CMD_PREADV:
1957 return aio_read(&req->rw, iocb, true, compat);
1958 case IOCB_CMD_PWRITEV:
1959 return aio_write(&req->rw, iocb, true, compat);
1960 case IOCB_CMD_FSYNC:
1961 return aio_fsync(&req->fsync, iocb, false);
1962 case IOCB_CMD_FDSYNC:
1963 return aio_fsync(&req->fsync, iocb, true);
1965 return aio_poll(req, iocb);
1967 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1972 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1975 struct aio_kiocb *req;
1979 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1982 /* enforce forwards compatibility on users */
1983 if (unlikely(iocb.aio_reserved2)) {
1984 pr_debug("EINVAL: reserve field set\n");
1988 /* prevent overflows */
1990 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1991 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1992 ((ssize_t)iocb.aio_nbytes < 0)
1994 pr_debug("EINVAL: overflow check\n");
1998 req = aio_get_req(ctx);
2002 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2004 /* Done with the synchronous reference */
2008 * If err is 0, we'd either done aio_complete() ourselves or have
2009 * arranged for that to be done asynchronously. Anything non-zero
2010 * means that we need to destroy req ourselves.
2012 if (unlikely(err)) {
2014 put_reqs_available(ctx, 1);
2020 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2021 * the number of iocbs queued. May return -EINVAL if the aio_context
2022 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2023 * *iocbpp[0] is not properly initialized, if the operation specified
2024 * is invalid for the file descriptor in the iocb. May fail with
2025 * -EFAULT if any of the data structures point to invalid data. May
2026 * fail with -EBADF if the file descriptor specified in the first
2027 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2028 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2029 * fail with -ENOSYS if not implemented.
2031 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2032 struct iocb __user * __user *, iocbpp)
2037 struct blk_plug plug;
2039 if (unlikely(nr < 0))
2042 ctx = lookup_ioctx(ctx_id);
2043 if (unlikely(!ctx)) {
2044 pr_debug("EINVAL: invalid context id\n");
2048 if (nr > ctx->nr_events)
2049 nr = ctx->nr_events;
2051 if (nr > AIO_PLUG_THRESHOLD)
2052 blk_start_plug(&plug);
2053 for (i = 0; i < nr; i++) {
2054 struct iocb __user *user_iocb;
2056 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2061 ret = io_submit_one(ctx, user_iocb, false);
2065 if (nr > AIO_PLUG_THRESHOLD)
2066 blk_finish_plug(&plug);
2068 percpu_ref_put(&ctx->users);
2072 #ifdef CONFIG_COMPAT
2073 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2074 int, nr, compat_uptr_t __user *, iocbpp)
2079 struct blk_plug plug;
2081 if (unlikely(nr < 0))
2084 ctx = lookup_ioctx(ctx_id);
2085 if (unlikely(!ctx)) {
2086 pr_debug("EINVAL: invalid context id\n");
2090 if (nr > ctx->nr_events)
2091 nr = ctx->nr_events;
2093 if (nr > AIO_PLUG_THRESHOLD)
2094 blk_start_plug(&plug);
2095 for (i = 0; i < nr; i++) {
2096 compat_uptr_t user_iocb;
2098 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2103 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2107 if (nr > AIO_PLUG_THRESHOLD)
2108 blk_finish_plug(&plug);
2110 percpu_ref_put(&ctx->users);
2116 * Attempts to cancel an iocb previously passed to io_submit. If
2117 * the operation is successfully cancelled, the resulting event is
2118 * copied into the memory pointed to by result without being placed
2119 * into the completion queue and 0 is returned. May fail with
2120 * -EFAULT if any of the data structures pointed to are invalid.
2121 * May fail with -EINVAL if aio_context specified by ctx_id is
2122 * invalid. May fail with -EAGAIN if the iocb specified was not
2123 * cancelled. Will fail with -ENOSYS if not implemented.
2125 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2126 struct io_event __user *, result)
2129 struct aio_kiocb *kiocb;
2132 u64 obj = (u64)(unsigned long)iocb;
2134 if (unlikely(get_user(key, &iocb->aio_key)))
2136 if (unlikely(key != KIOCB_KEY))
2139 ctx = lookup_ioctx(ctx_id);
2143 spin_lock_irq(&ctx->ctx_lock);
2144 /* TODO: use a hash or array, this sucks. */
2145 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2146 if (kiocb->ki_res.obj == obj) {
2147 ret = kiocb->ki_cancel(&kiocb->rw);
2148 list_del_init(&kiocb->ki_list);
2152 spin_unlock_irq(&ctx->ctx_lock);
2156 * The result argument is no longer used - the io_event is
2157 * always delivered via the ring buffer. -EINPROGRESS indicates
2158 * cancellation is progress:
2163 percpu_ref_put(&ctx->users);
2168 static long do_io_getevents(aio_context_t ctx_id,
2171 struct io_event __user *events,
2172 struct timespec64 *ts)
2174 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2175 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2178 if (likely(ioctx)) {
2179 if (likely(min_nr <= nr && min_nr >= 0))
2180 ret = read_events(ioctx, min_nr, nr, events, until);
2181 percpu_ref_put(&ioctx->users);
2188 * Attempts to read at least min_nr events and up to nr events from
2189 * the completion queue for the aio_context specified by ctx_id. If
2190 * it succeeds, the number of read events is returned. May fail with
2191 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2192 * out of range, if timeout is out of range. May fail with -EFAULT
2193 * if any of the memory specified is invalid. May return 0 or
2194 * < min_nr if the timeout specified by timeout has elapsed
2195 * before sufficient events are available, where timeout == NULL
2196 * specifies an infinite timeout. Note that the timeout pointed to by
2197 * timeout is relative. Will fail with -ENOSYS if not implemented.
2201 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2204 struct io_event __user *, events,
2205 struct __kernel_timespec __user *, timeout)
2207 struct timespec64 ts;
2210 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2213 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2214 if (!ret && signal_pending(current))
2221 struct __aio_sigset {
2222 const sigset_t __user *sigmask;
2226 SYSCALL_DEFINE6(io_pgetevents,
2227 aio_context_t, ctx_id,
2230 struct io_event __user *, events,
2231 struct __kernel_timespec __user *, timeout,
2232 const struct __aio_sigset __user *, usig)
2234 struct __aio_sigset ksig = { NULL, };
2235 struct timespec64 ts;
2239 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2242 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2245 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2249 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2251 interrupted = signal_pending(current);
2252 restore_saved_sigmask_unless(interrupted);
2253 if (interrupted && !ret)
2254 ret = -ERESTARTNOHAND;
2259 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2261 SYSCALL_DEFINE6(io_pgetevents_time32,
2262 aio_context_t, ctx_id,
2265 struct io_event __user *, events,
2266 struct old_timespec32 __user *, timeout,
2267 const struct __aio_sigset __user *, usig)
2269 struct __aio_sigset ksig = { NULL, };
2270 struct timespec64 ts;
2274 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2277 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2281 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2285 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2287 interrupted = signal_pending(current);
2288 restore_saved_sigmask_unless(interrupted);
2289 if (interrupted && !ret)
2290 ret = -ERESTARTNOHAND;
2297 #if defined(CONFIG_COMPAT_32BIT_TIME)
2299 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2302 struct io_event __user *, events,
2303 struct old_timespec32 __user *, timeout)
2305 struct timespec64 t;
2308 if (timeout && get_old_timespec32(&t, timeout))
2311 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2312 if (!ret && signal_pending(current))
2319 #ifdef CONFIG_COMPAT
2321 struct __compat_aio_sigset {
2322 compat_uptr_t sigmask;
2323 compat_size_t sigsetsize;
2326 #if defined(CONFIG_COMPAT_32BIT_TIME)
2328 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2329 compat_aio_context_t, ctx_id,
2330 compat_long_t, min_nr,
2332 struct io_event __user *, events,
2333 struct old_timespec32 __user *, timeout,
2334 const struct __compat_aio_sigset __user *, usig)
2336 struct __compat_aio_sigset ksig = { 0, };
2337 struct timespec64 t;
2341 if (timeout && get_old_timespec32(&t, timeout))
2344 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2347 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2351 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2353 interrupted = signal_pending(current);
2354 restore_saved_sigmask_unless(interrupted);
2355 if (interrupted && !ret)
2356 ret = -ERESTARTNOHAND;
2363 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2364 compat_aio_context_t, ctx_id,
2365 compat_long_t, min_nr,
2367 struct io_event __user *, events,
2368 struct __kernel_timespec __user *, timeout,
2369 const struct __compat_aio_sigset __user *, usig)
2371 struct __compat_aio_sigset ksig = { 0, };
2372 struct timespec64 t;
2376 if (timeout && get_timespec64(&t, timeout))
2379 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2382 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2386 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2388 interrupted = signal_pending(current);
2389 restore_saved_sigmask_unless(interrupted);
2390 if (interrupted && !ret)
2391 ret = -ERESTARTNOHAND;