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 struct wait_queue_entry wait;
187 struct work_struct work;
191 * NOTE! Each of the iocb union members has the file pointer
192 * as the first entry in their struct definition. So you can
193 * access the file pointer through any of the sub-structs,
194 * or directly as just 'ki_filp' in this struct.
198 struct file *ki_filp;
200 struct fsync_iocb fsync;
201 struct poll_iocb poll;
204 struct kioctx *ki_ctx;
205 kiocb_cancel_fn *ki_cancel;
207 struct io_event ki_res;
209 struct list_head ki_list; /* the aio core uses this
210 * for cancellation */
211 refcount_t ki_refcnt;
214 * If the aio_resfd field of the userspace iocb is not zero,
215 * this is the underlying eventfd context to deliver events to.
217 struct eventfd_ctx *ki_eventfd;
220 /*------ sysctl variables----*/
221 static DEFINE_SPINLOCK(aio_nr_lock);
222 unsigned long aio_nr; /* current system wide number of aio requests */
223 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
224 /*----end sysctl variables---*/
226 static struct kmem_cache *kiocb_cachep;
227 static struct kmem_cache *kioctx_cachep;
229 static struct vfsmount *aio_mnt;
231 static const struct file_operations aio_ring_fops;
232 static const struct address_space_operations aio_ctx_aops;
234 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
237 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
239 return ERR_CAST(inode);
241 inode->i_mapping->a_ops = &aio_ctx_aops;
242 inode->i_mapping->private_data = ctx;
243 inode->i_size = PAGE_SIZE * nr_pages;
245 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
246 O_RDWR, &aio_ring_fops);
252 static int aio_init_fs_context(struct fs_context *fc)
254 if (!init_pseudo(fc, AIO_RING_MAGIC))
256 fc->s_iflags |= SB_I_NOEXEC;
261 * Creates the slab caches used by the aio routines, panic on
262 * failure as this is done early during the boot sequence.
264 static int __init aio_setup(void)
266 static struct file_system_type aio_fs = {
268 .init_fs_context = aio_init_fs_context,
269 .kill_sb = kill_anon_super,
271 aio_mnt = kern_mount(&aio_fs);
273 panic("Failed to create aio fs mount.");
275 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
276 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
279 __initcall(aio_setup);
281 static void put_aio_ring_file(struct kioctx *ctx)
283 struct file *aio_ring_file = ctx->aio_ring_file;
284 struct address_space *i_mapping;
287 truncate_setsize(file_inode(aio_ring_file), 0);
289 /* Prevent further access to the kioctx from migratepages */
290 i_mapping = aio_ring_file->f_mapping;
291 spin_lock(&i_mapping->private_lock);
292 i_mapping->private_data = NULL;
293 ctx->aio_ring_file = NULL;
294 spin_unlock(&i_mapping->private_lock);
300 static void aio_free_ring(struct kioctx *ctx)
304 /* Disconnect the kiotx from the ring file. This prevents future
305 * accesses to the kioctx from page migration.
307 put_aio_ring_file(ctx);
309 for (i = 0; i < ctx->nr_pages; i++) {
311 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
312 page_count(ctx->ring_pages[i]));
313 page = ctx->ring_pages[i];
316 ctx->ring_pages[i] = NULL;
320 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
321 kfree(ctx->ring_pages);
322 ctx->ring_pages = NULL;
326 static int aio_ring_mremap(struct vm_area_struct *vma)
328 struct file *file = vma->vm_file;
329 struct mm_struct *mm = vma->vm_mm;
330 struct kioctx_table *table;
331 int i, res = -EINVAL;
333 spin_lock(&mm->ioctx_lock);
335 table = rcu_dereference(mm->ioctx_table);
336 for (i = 0; i < table->nr; i++) {
339 ctx = rcu_dereference(table->table[i]);
340 if (ctx && ctx->aio_ring_file == file) {
341 if (!atomic_read(&ctx->dead)) {
342 ctx->user_id = ctx->mmap_base = vma->vm_start;
350 spin_unlock(&mm->ioctx_lock);
354 static const struct vm_operations_struct aio_ring_vm_ops = {
355 .mremap = aio_ring_mremap,
356 #if IS_ENABLED(CONFIG_MMU)
357 .fault = filemap_fault,
358 .map_pages = filemap_map_pages,
359 .page_mkwrite = filemap_page_mkwrite,
363 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
365 vma->vm_flags |= VM_DONTEXPAND;
366 vma->vm_ops = &aio_ring_vm_ops;
370 static const struct file_operations aio_ring_fops = {
371 .mmap = aio_ring_mmap,
374 #if IS_ENABLED(CONFIG_MIGRATION)
375 static int aio_migratepage(struct address_space *mapping, struct page *new,
376 struct page *old, enum migrate_mode mode)
384 * We cannot support the _NO_COPY case here, because copy needs to
385 * happen under the ctx->completion_lock. That does not work with the
386 * migration workflow of MIGRATE_SYNC_NO_COPY.
388 if (mode == MIGRATE_SYNC_NO_COPY)
393 /* mapping->private_lock here protects against the kioctx teardown. */
394 spin_lock(&mapping->private_lock);
395 ctx = mapping->private_data;
401 /* The ring_lock mutex. The prevents aio_read_events() from writing
402 * to the ring's head, and prevents page migration from mucking in
403 * a partially initialized kiotx.
405 if (!mutex_trylock(&ctx->ring_lock)) {
411 if (idx < (pgoff_t)ctx->nr_pages) {
412 /* Make sure the old page hasn't already been changed */
413 if (ctx->ring_pages[idx] != old)
421 /* Writeback must be complete */
422 BUG_ON(PageWriteback(old));
425 rc = migrate_page_move_mapping(mapping, new, old, 1);
426 if (rc != MIGRATEPAGE_SUCCESS) {
431 /* Take completion_lock to prevent other writes to the ring buffer
432 * while the old page is copied to the new. This prevents new
433 * events from being lost.
435 spin_lock_irqsave(&ctx->completion_lock, flags);
436 migrate_page_copy(new, old);
437 BUG_ON(ctx->ring_pages[idx] != old);
438 ctx->ring_pages[idx] = new;
439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
441 /* The old page is no longer accessible. */
445 mutex_unlock(&ctx->ring_lock);
447 spin_unlock(&mapping->private_lock);
452 static const struct address_space_operations aio_ctx_aops = {
453 .set_page_dirty = __set_page_dirty_no_writeback,
454 #if IS_ENABLED(CONFIG_MIGRATION)
455 .migratepage = aio_migratepage,
459 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
461 struct aio_ring *ring;
462 struct mm_struct *mm = current->mm;
463 unsigned long size, unused;
468 /* Compensate for the ring buffer's head/tail overlap entry */
469 nr_events += 2; /* 1 is required, 2 for good luck */
471 size = sizeof(struct aio_ring);
472 size += sizeof(struct io_event) * nr_events;
474 nr_pages = PFN_UP(size);
478 file = aio_private_file(ctx, nr_pages);
480 ctx->aio_ring_file = NULL;
484 ctx->aio_ring_file = file;
485 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
486 / sizeof(struct io_event);
488 ctx->ring_pages = ctx->internal_pages;
489 if (nr_pages > AIO_RING_PAGES) {
490 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
492 if (!ctx->ring_pages) {
493 put_aio_ring_file(ctx);
498 for (i = 0; i < nr_pages; i++) {
500 page = find_or_create_page(file->f_mapping,
501 i, GFP_HIGHUSER | __GFP_ZERO);
504 pr_debug("pid(%d) page[%d]->count=%d\n",
505 current->pid, i, page_count(page));
506 SetPageUptodate(page);
509 ctx->ring_pages[i] = page;
513 if (unlikely(i != nr_pages)) {
518 ctx->mmap_size = nr_pages * PAGE_SIZE;
519 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
521 if (mmap_write_lock_killable(mm)) {
527 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
528 PROT_READ | PROT_WRITE,
529 MAP_SHARED, 0, &unused, NULL);
530 mmap_write_unlock(mm);
531 if (IS_ERR((void *)ctx->mmap_base)) {
537 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
539 ctx->user_id = ctx->mmap_base;
540 ctx->nr_events = nr_events; /* trusted copy */
542 ring = kmap_atomic(ctx->ring_pages[0]);
543 ring->nr = nr_events; /* user copy */
545 ring->head = ring->tail = 0;
546 ring->magic = AIO_RING_MAGIC;
547 ring->compat_features = AIO_RING_COMPAT_FEATURES;
548 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
549 ring->header_length = sizeof(struct aio_ring);
551 flush_dcache_page(ctx->ring_pages[0]);
556 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
557 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
558 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
560 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
562 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
563 struct kioctx *ctx = req->ki_ctx;
566 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
569 spin_lock_irqsave(&ctx->ctx_lock, flags);
570 list_add_tail(&req->ki_list, &ctx->active_reqs);
571 req->ki_cancel = cancel;
572 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
574 EXPORT_SYMBOL(kiocb_set_cancel_fn);
577 * free_ioctx() should be RCU delayed to synchronize against the RCU
578 * protected lookup_ioctx() and also needs process context to call
579 * aio_free_ring(). Use rcu_work.
581 static void free_ioctx(struct work_struct *work)
583 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
585 pr_debug("freeing %p\n", ctx);
588 free_percpu(ctx->cpu);
589 percpu_ref_exit(&ctx->reqs);
590 percpu_ref_exit(&ctx->users);
591 kmem_cache_free(kioctx_cachep, ctx);
594 static void free_ioctx_reqs(struct percpu_ref *ref)
596 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
598 /* At this point we know that there are no any in-flight requests */
599 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
600 complete(&ctx->rq_wait->comp);
602 /* Synchronize against RCU protected table->table[] dereferences */
603 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
604 queue_rcu_work(system_wq, &ctx->free_rwork);
608 * When this function runs, the kioctx has been removed from the "hash table"
609 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
610 * now it's safe to cancel any that need to be.
612 static void free_ioctx_users(struct percpu_ref *ref)
614 struct kioctx *ctx = container_of(ref, struct kioctx, users);
615 struct aio_kiocb *req;
617 spin_lock_irq(&ctx->ctx_lock);
619 while (!list_empty(&ctx->active_reqs)) {
620 req = list_first_entry(&ctx->active_reqs,
621 struct aio_kiocb, ki_list);
622 req->ki_cancel(&req->rw);
623 list_del_init(&req->ki_list);
626 spin_unlock_irq(&ctx->ctx_lock);
628 percpu_ref_kill(&ctx->reqs);
629 percpu_ref_put(&ctx->reqs);
632 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
635 struct kioctx_table *table, *old;
636 struct aio_ring *ring;
638 spin_lock(&mm->ioctx_lock);
639 table = rcu_dereference_raw(mm->ioctx_table);
643 for (i = 0; i < table->nr; i++)
644 if (!rcu_access_pointer(table->table[i])) {
646 rcu_assign_pointer(table->table[i], ctx);
647 spin_unlock(&mm->ioctx_lock);
649 /* While kioctx setup is in progress,
650 * we are protected from page migration
651 * changes ring_pages by ->ring_lock.
653 ring = kmap_atomic(ctx->ring_pages[0]);
659 new_nr = (table ? table->nr : 1) * 4;
660 spin_unlock(&mm->ioctx_lock);
662 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
668 spin_lock(&mm->ioctx_lock);
669 old = rcu_dereference_raw(mm->ioctx_table);
672 rcu_assign_pointer(mm->ioctx_table, table);
673 } else if (table->nr > old->nr) {
674 memcpy(table->table, old->table,
675 old->nr * sizeof(struct kioctx *));
677 rcu_assign_pointer(mm->ioctx_table, table);
686 static void aio_nr_sub(unsigned nr)
688 spin_lock(&aio_nr_lock);
689 if (WARN_ON(aio_nr - nr > aio_nr))
693 spin_unlock(&aio_nr_lock);
697 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
699 static struct kioctx *ioctx_alloc(unsigned nr_events)
701 struct mm_struct *mm = current->mm;
706 * Store the original nr_events -- what userspace passed to io_setup(),
707 * for counting against the global limit -- before it changes.
709 unsigned int max_reqs = nr_events;
712 * We keep track of the number of available ringbuffer slots, to prevent
713 * overflow (reqs_available), and we also use percpu counters for this.
715 * So since up to half the slots might be on other cpu's percpu counters
716 * and unavailable, double nr_events so userspace sees what they
717 * expected: additionally, we move req_batch slots to/from percpu
718 * counters at a time, so make sure that isn't 0:
720 nr_events = max(nr_events, num_possible_cpus() * 4);
723 /* Prevent overflows */
724 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
725 pr_debug("ENOMEM: nr_events too high\n");
726 return ERR_PTR(-EINVAL);
729 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
730 return ERR_PTR(-EAGAIN);
732 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
734 return ERR_PTR(-ENOMEM);
736 ctx->max_reqs = max_reqs;
738 spin_lock_init(&ctx->ctx_lock);
739 spin_lock_init(&ctx->completion_lock);
740 mutex_init(&ctx->ring_lock);
741 /* Protect against page migration throughout kiotx setup by keeping
742 * the ring_lock mutex held until setup is complete. */
743 mutex_lock(&ctx->ring_lock);
744 init_waitqueue_head(&ctx->wait);
746 INIT_LIST_HEAD(&ctx->active_reqs);
748 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754 ctx->cpu = alloc_percpu(struct kioctx_cpu);
758 err = aio_setup_ring(ctx, nr_events);
762 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
763 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
764 if (ctx->req_batch < 1)
767 /* limit the number of system wide aios */
768 spin_lock(&aio_nr_lock);
769 if (aio_nr + ctx->max_reqs > aio_max_nr ||
770 aio_nr + ctx->max_reqs < aio_nr) {
771 spin_unlock(&aio_nr_lock);
775 aio_nr += ctx->max_reqs;
776 spin_unlock(&aio_nr_lock);
778 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
779 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
781 err = ioctx_add_table(ctx, mm);
785 /* Release the ring_lock mutex now that all setup is complete. */
786 mutex_unlock(&ctx->ring_lock);
788 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
789 ctx, ctx->user_id, mm, ctx->nr_events);
793 aio_nr_sub(ctx->max_reqs);
795 atomic_set(&ctx->dead, 1);
797 vm_munmap(ctx->mmap_base, ctx->mmap_size);
800 mutex_unlock(&ctx->ring_lock);
801 free_percpu(ctx->cpu);
802 percpu_ref_exit(&ctx->reqs);
803 percpu_ref_exit(&ctx->users);
804 kmem_cache_free(kioctx_cachep, ctx);
805 pr_debug("error allocating ioctx %d\n", err);
810 * Cancels all outstanding aio requests on an aio context. Used
811 * when the processes owning a context have all exited to encourage
812 * the rapid destruction of the kioctx.
814 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
815 struct ctx_rq_wait *wait)
817 struct kioctx_table *table;
819 spin_lock(&mm->ioctx_lock);
820 if (atomic_xchg(&ctx->dead, 1)) {
821 spin_unlock(&mm->ioctx_lock);
825 table = rcu_dereference_raw(mm->ioctx_table);
826 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
827 RCU_INIT_POINTER(table->table[ctx->id], NULL);
828 spin_unlock(&mm->ioctx_lock);
830 /* free_ioctx_reqs() will do the necessary RCU synchronization */
831 wake_up_all(&ctx->wait);
834 * It'd be more correct to do this in free_ioctx(), after all
835 * the outstanding kiocbs have finished - but by then io_destroy
836 * has already returned, so io_setup() could potentially return
837 * -EAGAIN with no ioctxs actually in use (as far as userspace
840 aio_nr_sub(ctx->max_reqs);
843 vm_munmap(ctx->mmap_base, ctx->mmap_size);
846 percpu_ref_kill(&ctx->users);
851 * exit_aio: called when the last user of mm goes away. At this point, there is
852 * no way for any new requests to be submited or any of the io_* syscalls to be
853 * called on the context.
855 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
858 void exit_aio(struct mm_struct *mm)
860 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
861 struct ctx_rq_wait wait;
867 atomic_set(&wait.count, table->nr);
868 init_completion(&wait.comp);
871 for (i = 0; i < table->nr; ++i) {
873 rcu_dereference_protected(table->table[i], true);
881 * We don't need to bother with munmap() here - exit_mmap(mm)
882 * is coming and it'll unmap everything. And we simply can't,
883 * this is not necessarily our ->mm.
884 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
885 * that it needs to unmap the area, just set it to 0.
888 kill_ioctx(mm, ctx, &wait);
891 if (!atomic_sub_and_test(skipped, &wait.count)) {
892 /* Wait until all IO for the context are done. */
893 wait_for_completion(&wait.comp);
896 RCU_INIT_POINTER(mm->ioctx_table, NULL);
900 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
902 struct kioctx_cpu *kcpu;
905 local_irq_save(flags);
906 kcpu = this_cpu_ptr(ctx->cpu);
907 kcpu->reqs_available += nr;
909 while (kcpu->reqs_available >= ctx->req_batch * 2) {
910 kcpu->reqs_available -= ctx->req_batch;
911 atomic_add(ctx->req_batch, &ctx->reqs_available);
914 local_irq_restore(flags);
917 static bool __get_reqs_available(struct kioctx *ctx)
919 struct kioctx_cpu *kcpu;
923 local_irq_save(flags);
924 kcpu = this_cpu_ptr(ctx->cpu);
925 if (!kcpu->reqs_available) {
926 int old, avail = atomic_read(&ctx->reqs_available);
929 if (avail < ctx->req_batch)
933 avail = atomic_cmpxchg(&ctx->reqs_available,
934 avail, avail - ctx->req_batch);
935 } while (avail != old);
937 kcpu->reqs_available += ctx->req_batch;
941 kcpu->reqs_available--;
943 local_irq_restore(flags);
947 /* refill_reqs_available
948 * Updates the reqs_available reference counts used for tracking the
949 * number of free slots in the completion ring. This can be called
950 * from aio_complete() (to optimistically update reqs_available) or
951 * from aio_get_req() (the we're out of events case). It must be
952 * called holding ctx->completion_lock.
954 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
957 unsigned events_in_ring, completed;
959 /* Clamp head since userland can write to it. */
960 head %= ctx->nr_events;
962 events_in_ring = tail - head;
964 events_in_ring = ctx->nr_events - (head - tail);
966 completed = ctx->completed_events;
967 if (events_in_ring < completed)
968 completed -= events_in_ring;
975 ctx->completed_events -= completed;
976 put_reqs_available(ctx, completed);
979 /* user_refill_reqs_available
980 * Called to refill reqs_available when aio_get_req() encounters an
981 * out of space in the completion ring.
983 static void user_refill_reqs_available(struct kioctx *ctx)
985 spin_lock_irq(&ctx->completion_lock);
986 if (ctx->completed_events) {
987 struct aio_ring *ring;
990 /* Access of ring->head may race with aio_read_events_ring()
991 * here, but that's okay since whether we read the old version
992 * or the new version, and either will be valid. The important
993 * part is that head cannot pass tail since we prevent
994 * aio_complete() from updating tail by holding
995 * ctx->completion_lock. Even if head is invalid, the check
996 * against ctx->completed_events below will make sure we do the
999 ring = kmap_atomic(ctx->ring_pages[0]);
1001 kunmap_atomic(ring);
1003 refill_reqs_available(ctx, head, ctx->tail);
1006 spin_unlock_irq(&ctx->completion_lock);
1009 static bool get_reqs_available(struct kioctx *ctx)
1011 if (__get_reqs_available(ctx))
1013 user_refill_reqs_available(ctx);
1014 return __get_reqs_available(ctx);
1018 * Allocate a slot for an aio request.
1019 * Returns NULL if no requests are free.
1021 * The refcount is initialized to 2 - one for the async op completion,
1022 * one for the synchronous code that does this.
1024 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1026 struct aio_kiocb *req;
1028 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1032 if (unlikely(!get_reqs_available(ctx))) {
1033 kmem_cache_free(kiocb_cachep, req);
1037 percpu_ref_get(&ctx->reqs);
1039 INIT_LIST_HEAD(&req->ki_list);
1040 refcount_set(&req->ki_refcnt, 2);
1041 req->ki_eventfd = NULL;
1045 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1047 struct aio_ring __user *ring = (void __user *)ctx_id;
1048 struct mm_struct *mm = current->mm;
1049 struct kioctx *ctx, *ret = NULL;
1050 struct kioctx_table *table;
1053 if (get_user(id, &ring->id))
1057 table = rcu_dereference(mm->ioctx_table);
1059 if (!table || id >= table->nr)
1062 id = array_index_nospec(id, table->nr);
1063 ctx = rcu_dereference(table->table[id]);
1064 if (ctx && ctx->user_id == ctx_id) {
1065 if (percpu_ref_tryget_live(&ctx->users))
1073 static inline void iocb_destroy(struct aio_kiocb *iocb)
1075 if (iocb->ki_eventfd)
1076 eventfd_ctx_put(iocb->ki_eventfd);
1078 fput(iocb->ki_filp);
1079 percpu_ref_put(&iocb->ki_ctx->reqs);
1080 kmem_cache_free(kiocb_cachep, iocb);
1084 * Called when the io request on the given iocb is complete.
1086 static void aio_complete(struct aio_kiocb *iocb)
1088 struct kioctx *ctx = iocb->ki_ctx;
1089 struct aio_ring *ring;
1090 struct io_event *ev_page, *event;
1091 unsigned tail, pos, head;
1092 unsigned long flags;
1095 * Add a completion event to the ring buffer. Must be done holding
1096 * ctx->completion_lock to prevent other code from messing with the tail
1097 * pointer since we might be called from irq context.
1099 spin_lock_irqsave(&ctx->completion_lock, flags);
1102 pos = tail + AIO_EVENTS_OFFSET;
1104 if (++tail >= ctx->nr_events)
1107 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1108 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1110 *event = iocb->ki_res;
1112 kunmap_atomic(ev_page);
1113 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1115 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1116 (void __user *)(unsigned long)iocb->ki_res.obj,
1117 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1119 /* after flagging the request as done, we
1120 * must never even look at it again
1122 smp_wmb(); /* make event visible before updating tail */
1126 ring = kmap_atomic(ctx->ring_pages[0]);
1129 kunmap_atomic(ring);
1130 flush_dcache_page(ctx->ring_pages[0]);
1132 ctx->completed_events++;
1133 if (ctx->completed_events > 1)
1134 refill_reqs_available(ctx, head, tail);
1135 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1137 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1140 * Check if the user asked us to deliver the result through an
1141 * eventfd. The eventfd_signal() function is safe to be called
1144 if (iocb->ki_eventfd)
1145 eventfd_signal(iocb->ki_eventfd, 1);
1148 * We have to order our ring_info tail store above and test
1149 * of the wait list below outside the wait lock. This is
1150 * like in wake_up_bit() where clearing a bit has to be
1151 * ordered with the unlocked test.
1155 if (waitqueue_active(&ctx->wait))
1156 wake_up(&ctx->wait);
1159 static inline void iocb_put(struct aio_kiocb *iocb)
1161 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1167 /* aio_read_events_ring
1168 * Pull an event off of the ioctx's event ring. Returns the number of
1171 static long aio_read_events_ring(struct kioctx *ctx,
1172 struct io_event __user *event, long nr)
1174 struct aio_ring *ring;
1175 unsigned head, tail, pos;
1180 * The mutex can block and wake us up and that will cause
1181 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1182 * and repeat. This should be rare enough that it doesn't cause
1183 * peformance issues. See the comment in read_events() for more detail.
1185 sched_annotate_sleep();
1186 mutex_lock(&ctx->ring_lock);
1188 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1189 ring = kmap_atomic(ctx->ring_pages[0]);
1192 kunmap_atomic(ring);
1195 * Ensure that once we've read the current tail pointer, that
1196 * we also see the events that were stored up to the tail.
1200 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1205 head %= ctx->nr_events;
1206 tail %= ctx->nr_events;
1210 struct io_event *ev;
1213 avail = (head <= tail ? tail : ctx->nr_events) - head;
1217 pos = head + AIO_EVENTS_OFFSET;
1218 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1219 pos %= AIO_EVENTS_PER_PAGE;
1221 avail = min(avail, nr - ret);
1222 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1225 copy_ret = copy_to_user(event + ret, ev + pos,
1226 sizeof(*ev) * avail);
1229 if (unlikely(copy_ret)) {
1236 head %= ctx->nr_events;
1239 ring = kmap_atomic(ctx->ring_pages[0]);
1241 kunmap_atomic(ring);
1242 flush_dcache_page(ctx->ring_pages[0]);
1244 pr_debug("%li h%u t%u\n", ret, head, tail);
1246 mutex_unlock(&ctx->ring_lock);
1251 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1252 struct io_event __user *event, long *i)
1254 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1259 if (unlikely(atomic_read(&ctx->dead)))
1265 return ret < 0 || *i >= min_nr;
1268 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1269 struct io_event __user *event,
1275 * Note that aio_read_events() is being called as the conditional - i.e.
1276 * we're calling it after prepare_to_wait() has set task state to
1277 * TASK_INTERRUPTIBLE.
1279 * But aio_read_events() can block, and if it blocks it's going to flip
1280 * the task state back to TASK_RUNNING.
1282 * This should be ok, provided it doesn't flip the state back to
1283 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1284 * will only happen if the mutex_lock() call blocks, and we then find
1285 * the ringbuffer empty. So in practice we should be ok, but it's
1286 * something to be aware of when touching this code.
1289 aio_read_events(ctx, min_nr, nr, event, &ret);
1291 wait_event_interruptible_hrtimeout(ctx->wait,
1292 aio_read_events(ctx, min_nr, nr, event, &ret),
1298 * Create an aio_context capable of receiving at least nr_events.
1299 * ctxp must not point to an aio_context that already exists, and
1300 * must be initialized to 0 prior to the call. On successful
1301 * creation of the aio_context, *ctxp is filled in with the resulting
1302 * handle. May fail with -EINVAL if *ctxp is not initialized,
1303 * if the specified nr_events exceeds internal limits. May fail
1304 * with -EAGAIN if the specified nr_events exceeds the user's limit
1305 * of available events. May fail with -ENOMEM if insufficient kernel
1306 * resources are available. May fail with -EFAULT if an invalid
1307 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1310 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1312 struct kioctx *ioctx = NULL;
1316 ret = get_user(ctx, ctxp);
1321 if (unlikely(ctx || nr_events == 0)) {
1322 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1327 ioctx = ioctx_alloc(nr_events);
1328 ret = PTR_ERR(ioctx);
1329 if (!IS_ERR(ioctx)) {
1330 ret = put_user(ioctx->user_id, ctxp);
1332 kill_ioctx(current->mm, ioctx, NULL);
1333 percpu_ref_put(&ioctx->users);
1340 #ifdef CONFIG_COMPAT
1341 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1343 struct kioctx *ioctx = NULL;
1347 ret = get_user(ctx, ctx32p);
1352 if (unlikely(ctx || nr_events == 0)) {
1353 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1358 ioctx = ioctx_alloc(nr_events);
1359 ret = PTR_ERR(ioctx);
1360 if (!IS_ERR(ioctx)) {
1361 /* truncating is ok because it's a user address */
1362 ret = put_user((u32)ioctx->user_id, ctx32p);
1364 kill_ioctx(current->mm, ioctx, NULL);
1365 percpu_ref_put(&ioctx->users);
1374 * Destroy the aio_context specified. May cancel any outstanding
1375 * AIOs and block on completion. Will fail with -ENOSYS if not
1376 * implemented. May fail with -EINVAL if the context pointed to
1379 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1381 struct kioctx *ioctx = lookup_ioctx(ctx);
1382 if (likely(NULL != ioctx)) {
1383 struct ctx_rq_wait wait;
1386 init_completion(&wait.comp);
1387 atomic_set(&wait.count, 1);
1389 /* Pass requests_done to kill_ioctx() where it can be set
1390 * in a thread-safe way. If we try to set it here then we have
1391 * a race condition if two io_destroy() called simultaneously.
1393 ret = kill_ioctx(current->mm, ioctx, &wait);
1394 percpu_ref_put(&ioctx->users);
1396 /* Wait until all IO for the context are done. Otherwise kernel
1397 * keep using user-space buffers even if user thinks the context
1401 wait_for_completion(&wait.comp);
1405 pr_debug("EINVAL: invalid context id\n");
1409 static void aio_remove_iocb(struct aio_kiocb *iocb)
1411 struct kioctx *ctx = iocb->ki_ctx;
1412 unsigned long flags;
1414 spin_lock_irqsave(&ctx->ctx_lock, flags);
1415 list_del(&iocb->ki_list);
1416 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1419 static void aio_complete_rw(struct kiocb *kiocb, long res)
1421 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1423 if (!list_empty_careful(&iocb->ki_list))
1424 aio_remove_iocb(iocb);
1426 if (kiocb->ki_flags & IOCB_WRITE) {
1427 struct inode *inode = file_inode(kiocb->ki_filp);
1430 * Tell lockdep we inherited freeze protection from submission
1433 if (S_ISREG(inode->i_mode))
1434 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1435 file_end_write(kiocb->ki_filp);
1438 iocb->ki_res.res = res;
1439 iocb->ki_res.res2 = 0;
1443 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1447 req->ki_complete = aio_complete_rw;
1448 req->private = NULL;
1449 req->ki_pos = iocb->aio_offset;
1450 req->ki_flags = iocb_flags(req->ki_filp);
1451 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1452 req->ki_flags |= IOCB_EVENTFD;
1453 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1454 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1456 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1457 * aio_reqprio is interpreted as an I/O scheduling
1458 * class and priority.
1460 ret = ioprio_check_cap(iocb->aio_reqprio);
1462 pr_debug("aio ioprio check cap error: %d\n", ret);
1466 req->ki_ioprio = iocb->aio_reqprio;
1468 req->ki_ioprio = get_current_ioprio();
1470 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1474 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1478 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1479 struct iovec **iovec, bool vectored, bool compat,
1480 struct iov_iter *iter)
1482 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1483 size_t len = iocb->aio_nbytes;
1486 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1491 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1494 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1500 case -ERESTARTNOINTR:
1501 case -ERESTARTNOHAND:
1502 case -ERESTART_RESTARTBLOCK:
1504 * There's no easy way to restart the syscall since other AIO's
1505 * may be already running. Just fail this IO with EINTR.
1510 req->ki_complete(req, ret);
1514 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1515 bool vectored, bool compat)
1517 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1518 struct iov_iter iter;
1522 ret = aio_prep_rw(req, iocb);
1525 file = req->ki_filp;
1526 if (unlikely(!(file->f_mode & FMODE_READ)))
1529 if (unlikely(!file->f_op->read_iter))
1532 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1535 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1537 aio_rw_done(req, call_read_iter(file, req, &iter));
1542 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1543 bool vectored, bool compat)
1545 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1546 struct iov_iter iter;
1550 ret = aio_prep_rw(req, iocb);
1553 file = req->ki_filp;
1555 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1557 if (unlikely(!file->f_op->write_iter))
1560 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1563 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1566 * Open-code file_start_write here to grab freeze protection,
1567 * which will be released by another thread in
1568 * aio_complete_rw(). Fool lockdep by telling it the lock got
1569 * released so that it doesn't complain about the held lock when
1570 * we return to userspace.
1572 if (S_ISREG(file_inode(file)->i_mode)) {
1573 sb_start_write(file_inode(file)->i_sb);
1574 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1576 req->ki_flags |= IOCB_WRITE;
1577 aio_rw_done(req, call_write_iter(file, req, &iter));
1583 static void aio_fsync_work(struct work_struct *work)
1585 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1586 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1588 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1589 revert_creds(old_cred);
1590 put_cred(iocb->fsync.creds);
1594 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1597 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1598 iocb->aio_rw_flags))
1601 if (unlikely(!req->file->f_op->fsync))
1604 req->creds = prepare_creds();
1608 req->datasync = datasync;
1609 INIT_WORK(&req->work, aio_fsync_work);
1610 schedule_work(&req->work);
1614 static void aio_poll_put_work(struct work_struct *work)
1616 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1617 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1622 static void aio_poll_complete_work(struct work_struct *work)
1624 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1625 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1626 struct poll_table_struct pt = { ._key = req->events };
1627 struct kioctx *ctx = iocb->ki_ctx;
1630 if (!READ_ONCE(req->cancelled))
1631 mask = vfs_poll(req->file, &pt) & req->events;
1634 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1635 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1636 * synchronize with them. In the cancellation case the list_del_init
1637 * itself is not actually needed, but harmless so we keep it in to
1638 * avoid further branches in the fast path.
1640 spin_lock_irq(&ctx->ctx_lock);
1641 if (!mask && !READ_ONCE(req->cancelled)) {
1642 add_wait_queue(req->head, &req->wait);
1643 spin_unlock_irq(&ctx->ctx_lock);
1646 list_del_init(&iocb->ki_list);
1647 iocb->ki_res.res = mangle_poll(mask);
1649 spin_unlock_irq(&ctx->ctx_lock);
1654 /* assumes we are called with irqs disabled */
1655 static int aio_poll_cancel(struct kiocb *iocb)
1657 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1658 struct poll_iocb *req = &aiocb->poll;
1660 spin_lock(&req->head->lock);
1661 WRITE_ONCE(req->cancelled, true);
1662 if (!list_empty(&req->wait.entry)) {
1663 list_del_init(&req->wait.entry);
1664 schedule_work(&aiocb->poll.work);
1666 spin_unlock(&req->head->lock);
1671 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1674 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1675 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1676 __poll_t mask = key_to_poll(key);
1677 unsigned long flags;
1679 /* for instances that support it check for an event match first: */
1680 if (mask && !(mask & req->events))
1683 list_del_init(&req->wait.entry);
1685 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1686 struct kioctx *ctx = iocb->ki_ctx;
1689 * Try to complete the iocb inline if we can. Use
1690 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1691 * call this function with IRQs disabled and because IRQs
1692 * have to be disabled before ctx_lock is obtained.
1694 list_del(&iocb->ki_list);
1695 iocb->ki_res.res = mangle_poll(mask);
1697 if (iocb->ki_eventfd && eventfd_signal_allowed()) {
1699 INIT_WORK(&req->work, aio_poll_put_work);
1700 schedule_work(&req->work);
1702 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1706 schedule_work(&req->work);
1711 struct aio_poll_table {
1712 struct poll_table_struct pt;
1713 struct aio_kiocb *iocb;
1718 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1719 struct poll_table_struct *p)
1721 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1723 /* multiple wait queues per file are not supported */
1724 if (unlikely(pt->iocb->poll.head)) {
1725 pt->error = -EINVAL;
1730 pt->iocb->poll.head = head;
1731 add_wait_queue(head, &pt->iocb->poll.wait);
1734 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1736 struct kioctx *ctx = aiocb->ki_ctx;
1737 struct poll_iocb *req = &aiocb->poll;
1738 struct aio_poll_table apt;
1739 bool cancel = false;
1742 /* reject any unknown events outside the normal event mask. */
1743 if ((u16)iocb->aio_buf != iocb->aio_buf)
1745 /* reject fields that are not defined for poll */
1746 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1749 INIT_WORK(&req->work, aio_poll_complete_work);
1750 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1754 req->cancelled = false;
1756 apt.pt._qproc = aio_poll_queue_proc;
1757 apt.pt._key = req->events;
1759 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1761 /* initialized the list so that we can do list_empty checks */
1762 INIT_LIST_HEAD(&req->wait.entry);
1763 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1765 mask = vfs_poll(req->file, &apt.pt) & req->events;
1766 spin_lock_irq(&ctx->ctx_lock);
1767 if (likely(req->head)) {
1768 spin_lock(&req->head->lock);
1769 if (unlikely(list_empty(&req->wait.entry))) {
1775 if (mask || apt.error) {
1776 list_del_init(&req->wait.entry);
1777 } else if (cancel) {
1778 WRITE_ONCE(req->cancelled, true);
1779 } else if (!req->done) { /* actually waiting for an event */
1780 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1781 aiocb->ki_cancel = aio_poll_cancel;
1783 spin_unlock(&req->head->lock);
1785 if (mask) { /* no async, we'd stolen it */
1786 aiocb->ki_res.res = mangle_poll(mask);
1789 spin_unlock_irq(&ctx->ctx_lock);
1795 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1796 struct iocb __user *user_iocb, struct aio_kiocb *req,
1799 req->ki_filp = fget(iocb->aio_fildes);
1800 if (unlikely(!req->ki_filp))
1803 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1804 struct eventfd_ctx *eventfd;
1806 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1807 * instance of the file* now. The file descriptor must be
1808 * an eventfd() fd, and will be signaled for each completed
1809 * event using the eventfd_signal() function.
1811 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1812 if (IS_ERR(eventfd))
1813 return PTR_ERR(eventfd);
1815 req->ki_eventfd = eventfd;
1818 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1819 pr_debug("EFAULT: aio_key\n");
1823 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1824 req->ki_res.data = iocb->aio_data;
1825 req->ki_res.res = 0;
1826 req->ki_res.res2 = 0;
1828 switch (iocb->aio_lio_opcode) {
1829 case IOCB_CMD_PREAD:
1830 return aio_read(&req->rw, iocb, false, compat);
1831 case IOCB_CMD_PWRITE:
1832 return aio_write(&req->rw, iocb, false, compat);
1833 case IOCB_CMD_PREADV:
1834 return aio_read(&req->rw, iocb, true, compat);
1835 case IOCB_CMD_PWRITEV:
1836 return aio_write(&req->rw, iocb, true, compat);
1837 case IOCB_CMD_FSYNC:
1838 return aio_fsync(&req->fsync, iocb, false);
1839 case IOCB_CMD_FDSYNC:
1840 return aio_fsync(&req->fsync, iocb, true);
1842 return aio_poll(req, iocb);
1844 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1849 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1852 struct aio_kiocb *req;
1856 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1859 /* enforce forwards compatibility on users */
1860 if (unlikely(iocb.aio_reserved2)) {
1861 pr_debug("EINVAL: reserve field set\n");
1865 /* prevent overflows */
1867 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1868 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1869 ((ssize_t)iocb.aio_nbytes < 0)
1871 pr_debug("EINVAL: overflow check\n");
1875 req = aio_get_req(ctx);
1879 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1881 /* Done with the synchronous reference */
1885 * If err is 0, we'd either done aio_complete() ourselves or have
1886 * arranged for that to be done asynchronously. Anything non-zero
1887 * means that we need to destroy req ourselves.
1889 if (unlikely(err)) {
1891 put_reqs_available(ctx, 1);
1897 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1898 * the number of iocbs queued. May return -EINVAL if the aio_context
1899 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1900 * *iocbpp[0] is not properly initialized, if the operation specified
1901 * is invalid for the file descriptor in the iocb. May fail with
1902 * -EFAULT if any of the data structures point to invalid data. May
1903 * fail with -EBADF if the file descriptor specified in the first
1904 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1905 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1906 * fail with -ENOSYS if not implemented.
1908 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1909 struct iocb __user * __user *, iocbpp)
1914 struct blk_plug plug;
1916 if (unlikely(nr < 0))
1919 ctx = lookup_ioctx(ctx_id);
1920 if (unlikely(!ctx)) {
1921 pr_debug("EINVAL: invalid context id\n");
1925 if (nr > ctx->nr_events)
1926 nr = ctx->nr_events;
1928 if (nr > AIO_PLUG_THRESHOLD)
1929 blk_start_plug(&plug);
1930 for (i = 0; i < nr; i++) {
1931 struct iocb __user *user_iocb;
1933 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1938 ret = io_submit_one(ctx, user_iocb, false);
1942 if (nr > AIO_PLUG_THRESHOLD)
1943 blk_finish_plug(&plug);
1945 percpu_ref_put(&ctx->users);
1949 #ifdef CONFIG_COMPAT
1950 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1951 int, nr, compat_uptr_t __user *, iocbpp)
1956 struct blk_plug plug;
1958 if (unlikely(nr < 0))
1961 ctx = lookup_ioctx(ctx_id);
1962 if (unlikely(!ctx)) {
1963 pr_debug("EINVAL: invalid context id\n");
1967 if (nr > ctx->nr_events)
1968 nr = ctx->nr_events;
1970 if (nr > AIO_PLUG_THRESHOLD)
1971 blk_start_plug(&plug);
1972 for (i = 0; i < nr; i++) {
1973 compat_uptr_t user_iocb;
1975 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1980 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1984 if (nr > AIO_PLUG_THRESHOLD)
1985 blk_finish_plug(&plug);
1987 percpu_ref_put(&ctx->users);
1993 * Attempts to cancel an iocb previously passed to io_submit. If
1994 * the operation is successfully cancelled, the resulting event is
1995 * copied into the memory pointed to by result without being placed
1996 * into the completion queue and 0 is returned. May fail with
1997 * -EFAULT if any of the data structures pointed to are invalid.
1998 * May fail with -EINVAL if aio_context specified by ctx_id is
1999 * invalid. May fail with -EAGAIN if the iocb specified was not
2000 * cancelled. Will fail with -ENOSYS if not implemented.
2002 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2003 struct io_event __user *, result)
2006 struct aio_kiocb *kiocb;
2009 u64 obj = (u64)(unsigned long)iocb;
2011 if (unlikely(get_user(key, &iocb->aio_key)))
2013 if (unlikely(key != KIOCB_KEY))
2016 ctx = lookup_ioctx(ctx_id);
2020 spin_lock_irq(&ctx->ctx_lock);
2021 /* TODO: use a hash or array, this sucks. */
2022 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2023 if (kiocb->ki_res.obj == obj) {
2024 ret = kiocb->ki_cancel(&kiocb->rw);
2025 list_del_init(&kiocb->ki_list);
2029 spin_unlock_irq(&ctx->ctx_lock);
2033 * The result argument is no longer used - the io_event is
2034 * always delivered via the ring buffer. -EINPROGRESS indicates
2035 * cancellation is progress:
2040 percpu_ref_put(&ctx->users);
2045 static long do_io_getevents(aio_context_t ctx_id,
2048 struct io_event __user *events,
2049 struct timespec64 *ts)
2051 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2052 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2055 if (likely(ioctx)) {
2056 if (likely(min_nr <= nr && min_nr >= 0))
2057 ret = read_events(ioctx, min_nr, nr, events, until);
2058 percpu_ref_put(&ioctx->users);
2065 * Attempts to read at least min_nr events and up to nr events from
2066 * the completion queue for the aio_context specified by ctx_id. If
2067 * it succeeds, the number of read events is returned. May fail with
2068 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2069 * out of range, if timeout is out of range. May fail with -EFAULT
2070 * if any of the memory specified is invalid. May return 0 or
2071 * < min_nr if the timeout specified by timeout has elapsed
2072 * before sufficient events are available, where timeout == NULL
2073 * specifies an infinite timeout. Note that the timeout pointed to by
2074 * timeout is relative. Will fail with -ENOSYS if not implemented.
2078 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2081 struct io_event __user *, events,
2082 struct __kernel_timespec __user *, timeout)
2084 struct timespec64 ts;
2087 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2090 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2091 if (!ret && signal_pending(current))
2098 struct __aio_sigset {
2099 const sigset_t __user *sigmask;
2103 SYSCALL_DEFINE6(io_pgetevents,
2104 aio_context_t, ctx_id,
2107 struct io_event __user *, events,
2108 struct __kernel_timespec __user *, timeout,
2109 const struct __aio_sigset __user *, usig)
2111 struct __aio_sigset ksig = { NULL, };
2112 struct timespec64 ts;
2116 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2119 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2122 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2126 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2128 interrupted = signal_pending(current);
2129 restore_saved_sigmask_unless(interrupted);
2130 if (interrupted && !ret)
2131 ret = -ERESTARTNOHAND;
2136 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2138 SYSCALL_DEFINE6(io_pgetevents_time32,
2139 aio_context_t, ctx_id,
2142 struct io_event __user *, events,
2143 struct old_timespec32 __user *, timeout,
2144 const struct __aio_sigset __user *, usig)
2146 struct __aio_sigset ksig = { NULL, };
2147 struct timespec64 ts;
2151 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2154 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2158 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2162 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2164 interrupted = signal_pending(current);
2165 restore_saved_sigmask_unless(interrupted);
2166 if (interrupted && !ret)
2167 ret = -ERESTARTNOHAND;
2174 #if defined(CONFIG_COMPAT_32BIT_TIME)
2176 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2179 struct io_event __user *, events,
2180 struct old_timespec32 __user *, timeout)
2182 struct timespec64 t;
2185 if (timeout && get_old_timespec32(&t, timeout))
2188 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2189 if (!ret && signal_pending(current))
2196 #ifdef CONFIG_COMPAT
2198 struct __compat_aio_sigset {
2199 compat_uptr_t sigmask;
2200 compat_size_t sigsetsize;
2203 #if defined(CONFIG_COMPAT_32BIT_TIME)
2205 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2206 compat_aio_context_t, ctx_id,
2207 compat_long_t, min_nr,
2209 struct io_event __user *, events,
2210 struct old_timespec32 __user *, timeout,
2211 const struct __compat_aio_sigset __user *, usig)
2213 struct __compat_aio_sigset ksig = { 0, };
2214 struct timespec64 t;
2218 if (timeout && get_old_timespec32(&t, timeout))
2221 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2224 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2228 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2230 interrupted = signal_pending(current);
2231 restore_saved_sigmask_unless(interrupted);
2232 if (interrupted && !ret)
2233 ret = -ERESTARTNOHAND;
2240 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2241 compat_aio_context_t, ctx_id,
2242 compat_long_t, min_nr,
2244 struct io_event __user *, events,
2245 struct __kernel_timespec __user *, timeout,
2246 const struct __compat_aio_sigset __user *, usig)
2248 struct __compat_aio_sigset ksig = { 0, };
2249 struct timespec64 t;
2253 if (timeout && get_timespec64(&t, timeout))
2256 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2259 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2263 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2265 interrupted = signal_pending(current);
2266 restore_saved_sigmask_unless(interrupted);
2267 if (interrupted && !ret)
2268 ret = -ERESTARTNOHAND;