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[] __counted_by(nr);
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 static unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct ctl_table aio_sysctls[] = {
229 .procname = "aio-nr",
231 .maxlen = sizeof(aio_nr),
233 .proc_handler = proc_doulongvec_minmax,
236 .procname = "aio-max-nr",
238 .maxlen = sizeof(aio_max_nr),
240 .proc_handler = proc_doulongvec_minmax,
245 static void __init aio_sysctl_init(void)
247 register_sysctl_init("fs", aio_sysctls);
250 #define aio_sysctl_init() do { } while (0)
253 static struct kmem_cache *kiocb_cachep;
254 static struct kmem_cache *kioctx_cachep;
256 static struct vfsmount *aio_mnt;
258 static const struct file_operations aio_ring_fops;
259 static const struct address_space_operations aio_ctx_aops;
261 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
264 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
266 return ERR_CAST(inode);
268 inode->i_mapping->a_ops = &aio_ctx_aops;
269 inode->i_mapping->private_data = ctx;
270 inode->i_size = PAGE_SIZE * nr_pages;
272 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
273 O_RDWR, &aio_ring_fops);
279 static int aio_init_fs_context(struct fs_context *fc)
281 if (!init_pseudo(fc, AIO_RING_MAGIC))
283 fc->s_iflags |= SB_I_NOEXEC;
288 * Creates the slab caches used by the aio routines, panic on
289 * failure as this is done early during the boot sequence.
291 static int __init aio_setup(void)
293 static struct file_system_type aio_fs = {
295 .init_fs_context = aio_init_fs_context,
296 .kill_sb = kill_anon_super,
298 aio_mnt = kern_mount(&aio_fs);
300 panic("Failed to create aio fs mount.");
302 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
303 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
307 __initcall(aio_setup);
309 static void put_aio_ring_file(struct kioctx *ctx)
311 struct file *aio_ring_file = ctx->aio_ring_file;
312 struct address_space *i_mapping;
315 truncate_setsize(file_inode(aio_ring_file), 0);
317 /* Prevent further access to the kioctx from migratepages */
318 i_mapping = aio_ring_file->f_mapping;
319 spin_lock(&i_mapping->private_lock);
320 i_mapping->private_data = NULL;
321 ctx->aio_ring_file = NULL;
322 spin_unlock(&i_mapping->private_lock);
328 static void aio_free_ring(struct kioctx *ctx)
332 /* Disconnect the kiotx from the ring file. This prevents future
333 * accesses to the kioctx from page migration.
335 put_aio_ring_file(ctx);
337 for (i = 0; i < ctx->nr_pages; i++) {
339 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
340 page_count(ctx->ring_pages[i]));
341 page = ctx->ring_pages[i];
344 ctx->ring_pages[i] = NULL;
348 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
349 kfree(ctx->ring_pages);
350 ctx->ring_pages = NULL;
354 static int aio_ring_mremap(struct vm_area_struct *vma)
356 struct file *file = vma->vm_file;
357 struct mm_struct *mm = vma->vm_mm;
358 struct kioctx_table *table;
359 int i, res = -EINVAL;
361 spin_lock(&mm->ioctx_lock);
363 table = rcu_dereference(mm->ioctx_table);
367 for (i = 0; i < table->nr; i++) {
370 ctx = rcu_dereference(table->table[i]);
371 if (ctx && ctx->aio_ring_file == file) {
372 if (!atomic_read(&ctx->dead)) {
373 ctx->user_id = ctx->mmap_base = vma->vm_start;
382 spin_unlock(&mm->ioctx_lock);
386 static const struct vm_operations_struct aio_ring_vm_ops = {
387 .mremap = aio_ring_mremap,
388 #if IS_ENABLED(CONFIG_MMU)
389 .fault = filemap_fault,
390 .map_pages = filemap_map_pages,
391 .page_mkwrite = filemap_page_mkwrite,
395 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
397 vm_flags_set(vma, VM_DONTEXPAND);
398 vma->vm_ops = &aio_ring_vm_ops;
402 static const struct file_operations aio_ring_fops = {
403 .mmap = aio_ring_mmap,
406 #if IS_ENABLED(CONFIG_MIGRATION)
407 static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
408 struct folio *src, enum migrate_mode mode)
416 * We cannot support the _NO_COPY case here, because copy needs to
417 * happen under the ctx->completion_lock. That does not work with the
418 * migration workflow of MIGRATE_SYNC_NO_COPY.
420 if (mode == MIGRATE_SYNC_NO_COPY)
425 /* mapping->private_lock here protects against the kioctx teardown. */
426 spin_lock(&mapping->private_lock);
427 ctx = mapping->private_data;
433 /* The ring_lock mutex. The prevents aio_read_events() from writing
434 * to the ring's head, and prevents page migration from mucking in
435 * a partially initialized kiotx.
437 if (!mutex_trylock(&ctx->ring_lock)) {
443 if (idx < (pgoff_t)ctx->nr_pages) {
444 /* Make sure the old folio hasn't already been changed */
445 if (ctx->ring_pages[idx] != &src->page)
453 /* Writeback must be complete */
454 BUG_ON(folio_test_writeback(src));
457 rc = folio_migrate_mapping(mapping, dst, src, 1);
458 if (rc != MIGRATEPAGE_SUCCESS) {
463 /* Take completion_lock to prevent other writes to the ring buffer
464 * while the old folio is copied to the new. This prevents new
465 * events from being lost.
467 spin_lock_irqsave(&ctx->completion_lock, flags);
468 folio_migrate_copy(dst, src);
469 BUG_ON(ctx->ring_pages[idx] != &src->page);
470 ctx->ring_pages[idx] = &dst->page;
471 spin_unlock_irqrestore(&ctx->completion_lock, flags);
473 /* The old folio is no longer accessible. */
477 mutex_unlock(&ctx->ring_lock);
479 spin_unlock(&mapping->private_lock);
483 #define aio_migrate_folio NULL
486 static const struct address_space_operations aio_ctx_aops = {
487 .dirty_folio = noop_dirty_folio,
488 .migrate_folio = aio_migrate_folio,
491 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
493 struct aio_ring *ring;
494 struct mm_struct *mm = current->mm;
495 unsigned long size, unused;
500 /* Compensate for the ring buffer's head/tail overlap entry */
501 nr_events += 2; /* 1 is required, 2 for good luck */
503 size = sizeof(struct aio_ring);
504 size += sizeof(struct io_event) * nr_events;
506 nr_pages = PFN_UP(size);
510 file = aio_private_file(ctx, nr_pages);
512 ctx->aio_ring_file = NULL;
516 ctx->aio_ring_file = file;
517 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
518 / sizeof(struct io_event);
520 ctx->ring_pages = ctx->internal_pages;
521 if (nr_pages > AIO_RING_PAGES) {
522 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
524 if (!ctx->ring_pages) {
525 put_aio_ring_file(ctx);
530 for (i = 0; i < nr_pages; i++) {
532 page = find_or_create_page(file->f_mapping,
533 i, GFP_USER | __GFP_ZERO);
536 pr_debug("pid(%d) page[%d]->count=%d\n",
537 current->pid, i, page_count(page));
538 SetPageUptodate(page);
541 ctx->ring_pages[i] = page;
545 if (unlikely(i != nr_pages)) {
550 ctx->mmap_size = nr_pages * PAGE_SIZE;
551 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
553 if (mmap_write_lock_killable(mm)) {
559 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
560 PROT_READ | PROT_WRITE,
561 MAP_SHARED, 0, 0, &unused, NULL);
562 mmap_write_unlock(mm);
563 if (IS_ERR((void *)ctx->mmap_base)) {
569 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
571 ctx->user_id = ctx->mmap_base;
572 ctx->nr_events = nr_events; /* trusted copy */
574 ring = page_address(ctx->ring_pages[0]);
575 ring->nr = nr_events; /* user copy */
577 ring->head = ring->tail = 0;
578 ring->magic = AIO_RING_MAGIC;
579 ring->compat_features = AIO_RING_COMPAT_FEATURES;
580 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
581 ring->header_length = sizeof(struct aio_ring);
582 flush_dcache_page(ctx->ring_pages[0]);
587 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
588 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
589 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
591 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
593 struct aio_kiocb *req;
598 * kiocb didn't come from aio or is neither a read nor a write, hence
601 if (!(iocb->ki_flags & IOCB_AIO_RW))
604 req = container_of(iocb, struct aio_kiocb, rw);
606 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
611 spin_lock_irqsave(&ctx->ctx_lock, flags);
612 list_add_tail(&req->ki_list, &ctx->active_reqs);
613 req->ki_cancel = cancel;
614 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
616 EXPORT_SYMBOL(kiocb_set_cancel_fn);
619 * free_ioctx() should be RCU delayed to synchronize against the RCU
620 * protected lookup_ioctx() and also needs process context to call
621 * aio_free_ring(). Use rcu_work.
623 static void free_ioctx(struct work_struct *work)
625 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
627 pr_debug("freeing %p\n", ctx);
630 free_percpu(ctx->cpu);
631 percpu_ref_exit(&ctx->reqs);
632 percpu_ref_exit(&ctx->users);
633 kmem_cache_free(kioctx_cachep, ctx);
636 static void free_ioctx_reqs(struct percpu_ref *ref)
638 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
640 /* At this point we know that there are no any in-flight requests */
641 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
642 complete(&ctx->rq_wait->comp);
644 /* Synchronize against RCU protected table->table[] dereferences */
645 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
646 queue_rcu_work(system_wq, &ctx->free_rwork);
650 * When this function runs, the kioctx has been removed from the "hash table"
651 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
652 * now it's safe to cancel any that need to be.
654 static void free_ioctx_users(struct percpu_ref *ref)
656 struct kioctx *ctx = container_of(ref, struct kioctx, users);
657 struct aio_kiocb *req;
659 spin_lock_irq(&ctx->ctx_lock);
661 while (!list_empty(&ctx->active_reqs)) {
662 req = list_first_entry(&ctx->active_reqs,
663 struct aio_kiocb, ki_list);
664 req->ki_cancel(&req->rw);
665 list_del_init(&req->ki_list);
668 spin_unlock_irq(&ctx->ctx_lock);
670 percpu_ref_kill(&ctx->reqs);
671 percpu_ref_put(&ctx->reqs);
674 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
677 struct kioctx_table *table, *old;
678 struct aio_ring *ring;
680 spin_lock(&mm->ioctx_lock);
681 table = rcu_dereference_raw(mm->ioctx_table);
685 for (i = 0; i < table->nr; i++)
686 if (!rcu_access_pointer(table->table[i])) {
688 rcu_assign_pointer(table->table[i], ctx);
689 spin_unlock(&mm->ioctx_lock);
691 /* While kioctx setup is in progress,
692 * we are protected from page migration
693 * changes ring_pages by ->ring_lock.
695 ring = page_address(ctx->ring_pages[0]);
700 new_nr = (table ? table->nr : 1) * 4;
701 spin_unlock(&mm->ioctx_lock);
703 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
709 spin_lock(&mm->ioctx_lock);
710 old = rcu_dereference_raw(mm->ioctx_table);
713 rcu_assign_pointer(mm->ioctx_table, table);
714 } else if (table->nr > old->nr) {
715 memcpy(table->table, old->table,
716 old->nr * sizeof(struct kioctx *));
718 rcu_assign_pointer(mm->ioctx_table, table);
727 static void aio_nr_sub(unsigned nr)
729 spin_lock(&aio_nr_lock);
730 if (WARN_ON(aio_nr - nr > aio_nr))
734 spin_unlock(&aio_nr_lock);
738 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
740 static struct kioctx *ioctx_alloc(unsigned nr_events)
742 struct mm_struct *mm = current->mm;
747 * Store the original nr_events -- what userspace passed to io_setup(),
748 * for counting against the global limit -- before it changes.
750 unsigned int max_reqs = nr_events;
753 * We keep track of the number of available ringbuffer slots, to prevent
754 * overflow (reqs_available), and we also use percpu counters for this.
756 * So since up to half the slots might be on other cpu's percpu counters
757 * and unavailable, double nr_events so userspace sees what they
758 * expected: additionally, we move req_batch slots to/from percpu
759 * counters at a time, so make sure that isn't 0:
761 nr_events = max(nr_events, num_possible_cpus() * 4);
764 /* Prevent overflows */
765 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
766 pr_debug("ENOMEM: nr_events too high\n");
767 return ERR_PTR(-EINVAL);
770 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
771 return ERR_PTR(-EAGAIN);
773 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
775 return ERR_PTR(-ENOMEM);
777 ctx->max_reqs = max_reqs;
779 spin_lock_init(&ctx->ctx_lock);
780 spin_lock_init(&ctx->completion_lock);
781 mutex_init(&ctx->ring_lock);
782 /* Protect against page migration throughout kiotx setup by keeping
783 * the ring_lock mutex held until setup is complete. */
784 mutex_lock(&ctx->ring_lock);
785 init_waitqueue_head(&ctx->wait);
787 INIT_LIST_HEAD(&ctx->active_reqs);
789 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
792 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
795 ctx->cpu = alloc_percpu(struct kioctx_cpu);
799 err = aio_setup_ring(ctx, nr_events);
803 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
804 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
805 if (ctx->req_batch < 1)
808 /* limit the number of system wide aios */
809 spin_lock(&aio_nr_lock);
810 if (aio_nr + ctx->max_reqs > aio_max_nr ||
811 aio_nr + ctx->max_reqs < aio_nr) {
812 spin_unlock(&aio_nr_lock);
816 aio_nr += ctx->max_reqs;
817 spin_unlock(&aio_nr_lock);
819 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
820 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
822 err = ioctx_add_table(ctx, mm);
826 /* Release the ring_lock mutex now that all setup is complete. */
827 mutex_unlock(&ctx->ring_lock);
829 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
830 ctx, ctx->user_id, mm, ctx->nr_events);
834 aio_nr_sub(ctx->max_reqs);
836 atomic_set(&ctx->dead, 1);
838 vm_munmap(ctx->mmap_base, ctx->mmap_size);
841 mutex_unlock(&ctx->ring_lock);
842 free_percpu(ctx->cpu);
843 percpu_ref_exit(&ctx->reqs);
844 percpu_ref_exit(&ctx->users);
845 kmem_cache_free(kioctx_cachep, ctx);
846 pr_debug("error allocating ioctx %d\n", err);
851 * Cancels all outstanding aio requests on an aio context. Used
852 * when the processes owning a context have all exited to encourage
853 * the rapid destruction of the kioctx.
855 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
856 struct ctx_rq_wait *wait)
858 struct kioctx_table *table;
860 spin_lock(&mm->ioctx_lock);
861 if (atomic_xchg(&ctx->dead, 1)) {
862 spin_unlock(&mm->ioctx_lock);
866 table = rcu_dereference_raw(mm->ioctx_table);
867 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
868 RCU_INIT_POINTER(table->table[ctx->id], NULL);
869 spin_unlock(&mm->ioctx_lock);
871 /* free_ioctx_reqs() will do the necessary RCU synchronization */
872 wake_up_all(&ctx->wait);
875 * It'd be more correct to do this in free_ioctx(), after all
876 * the outstanding kiocbs have finished - but by then io_destroy
877 * has already returned, so io_setup() could potentially return
878 * -EAGAIN with no ioctxs actually in use (as far as userspace
881 aio_nr_sub(ctx->max_reqs);
884 vm_munmap(ctx->mmap_base, ctx->mmap_size);
887 percpu_ref_kill(&ctx->users);
892 * exit_aio: called when the last user of mm goes away. At this point, there is
893 * no way for any new requests to be submited or any of the io_* syscalls to be
894 * called on the context.
896 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
899 void exit_aio(struct mm_struct *mm)
901 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
902 struct ctx_rq_wait wait;
908 atomic_set(&wait.count, table->nr);
909 init_completion(&wait.comp);
912 for (i = 0; i < table->nr; ++i) {
914 rcu_dereference_protected(table->table[i], true);
922 * We don't need to bother with munmap() here - exit_mmap(mm)
923 * is coming and it'll unmap everything. And we simply can't,
924 * this is not necessarily our ->mm.
925 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
926 * that it needs to unmap the area, just set it to 0.
929 kill_ioctx(mm, ctx, &wait);
932 if (!atomic_sub_and_test(skipped, &wait.count)) {
933 /* Wait until all IO for the context are done. */
934 wait_for_completion(&wait.comp);
937 RCU_INIT_POINTER(mm->ioctx_table, NULL);
941 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
943 struct kioctx_cpu *kcpu;
946 local_irq_save(flags);
947 kcpu = this_cpu_ptr(ctx->cpu);
948 kcpu->reqs_available += nr;
950 while (kcpu->reqs_available >= ctx->req_batch * 2) {
951 kcpu->reqs_available -= ctx->req_batch;
952 atomic_add(ctx->req_batch, &ctx->reqs_available);
955 local_irq_restore(flags);
958 static bool __get_reqs_available(struct kioctx *ctx)
960 struct kioctx_cpu *kcpu;
964 local_irq_save(flags);
965 kcpu = this_cpu_ptr(ctx->cpu);
966 if (!kcpu->reqs_available) {
967 int avail = atomic_read(&ctx->reqs_available);
970 if (avail < ctx->req_batch)
972 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
973 &avail, avail - ctx->req_batch));
975 kcpu->reqs_available += ctx->req_batch;
979 kcpu->reqs_available--;
981 local_irq_restore(flags);
985 /* refill_reqs_available
986 * Updates the reqs_available reference counts used for tracking the
987 * number of free slots in the completion ring. This can be called
988 * from aio_complete() (to optimistically update reqs_available) or
989 * from aio_get_req() (the we're out of events case). It must be
990 * called holding ctx->completion_lock.
992 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
995 unsigned events_in_ring, completed;
997 /* Clamp head since userland can write to it. */
998 head %= ctx->nr_events;
1000 events_in_ring = tail - head;
1002 events_in_ring = ctx->nr_events - (head - tail);
1004 completed = ctx->completed_events;
1005 if (events_in_ring < completed)
1006 completed -= events_in_ring;
1013 ctx->completed_events -= completed;
1014 put_reqs_available(ctx, completed);
1017 /* user_refill_reqs_available
1018 * Called to refill reqs_available when aio_get_req() encounters an
1019 * out of space in the completion ring.
1021 static void user_refill_reqs_available(struct kioctx *ctx)
1023 spin_lock_irq(&ctx->completion_lock);
1024 if (ctx->completed_events) {
1025 struct aio_ring *ring;
1028 /* Access of ring->head may race with aio_read_events_ring()
1029 * here, but that's okay since whether we read the old version
1030 * or the new version, and either will be valid. The important
1031 * part is that head cannot pass tail since we prevent
1032 * aio_complete() from updating tail by holding
1033 * ctx->completion_lock. Even if head is invalid, the check
1034 * against ctx->completed_events below will make sure we do the
1037 ring = page_address(ctx->ring_pages[0]);
1040 refill_reqs_available(ctx, head, ctx->tail);
1043 spin_unlock_irq(&ctx->completion_lock);
1046 static bool get_reqs_available(struct kioctx *ctx)
1048 if (__get_reqs_available(ctx))
1050 user_refill_reqs_available(ctx);
1051 return __get_reqs_available(ctx);
1055 * Allocate a slot for an aio request.
1056 * Returns NULL if no requests are free.
1058 * The refcount is initialized to 2 - one for the async op completion,
1059 * one for the synchronous code that does this.
1061 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1063 struct aio_kiocb *req;
1065 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1069 if (unlikely(!get_reqs_available(ctx))) {
1070 kmem_cache_free(kiocb_cachep, req);
1074 percpu_ref_get(&ctx->reqs);
1076 INIT_LIST_HEAD(&req->ki_list);
1077 refcount_set(&req->ki_refcnt, 2);
1078 req->ki_eventfd = NULL;
1082 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1084 struct aio_ring __user *ring = (void __user *)ctx_id;
1085 struct mm_struct *mm = current->mm;
1086 struct kioctx *ctx, *ret = NULL;
1087 struct kioctx_table *table;
1090 if (get_user(id, &ring->id))
1094 table = rcu_dereference(mm->ioctx_table);
1096 if (!table || id >= table->nr)
1099 id = array_index_nospec(id, table->nr);
1100 ctx = rcu_dereference(table->table[id]);
1101 if (ctx && ctx->user_id == ctx_id) {
1102 if (percpu_ref_tryget_live(&ctx->users))
1110 static inline void iocb_destroy(struct aio_kiocb *iocb)
1112 if (iocb->ki_eventfd)
1113 eventfd_ctx_put(iocb->ki_eventfd);
1115 fput(iocb->ki_filp);
1116 percpu_ref_put(&iocb->ki_ctx->reqs);
1117 kmem_cache_free(kiocb_cachep, iocb);
1121 * Called when the io request on the given iocb is complete.
1123 static void aio_complete(struct aio_kiocb *iocb)
1125 struct kioctx *ctx = iocb->ki_ctx;
1126 struct aio_ring *ring;
1127 struct io_event *ev_page, *event;
1128 unsigned tail, pos, head;
1129 unsigned long flags;
1132 * Add a completion event to the ring buffer. Must be done holding
1133 * ctx->completion_lock to prevent other code from messing with the tail
1134 * pointer since we might be called from irq context.
1136 spin_lock_irqsave(&ctx->completion_lock, flags);
1139 pos = tail + AIO_EVENTS_OFFSET;
1141 if (++tail >= ctx->nr_events)
1144 ev_page = page_address(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1145 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1147 *event = iocb->ki_res;
1149 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1151 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1152 (void __user *)(unsigned long)iocb->ki_res.obj,
1153 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1155 /* after flagging the request as done, we
1156 * must never even look at it again
1158 smp_wmb(); /* make event visible before updating tail */
1162 ring = page_address(ctx->ring_pages[0]);
1165 flush_dcache_page(ctx->ring_pages[0]);
1167 ctx->completed_events++;
1168 if (ctx->completed_events > 1)
1169 refill_reqs_available(ctx, head, tail);
1170 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1172 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1175 * Check if the user asked us to deliver the result through an
1176 * eventfd. The eventfd_signal() function is safe to be called
1179 if (iocb->ki_eventfd)
1180 eventfd_signal(iocb->ki_eventfd, 1);
1183 * We have to order our ring_info tail store above and test
1184 * of the wait list below outside the wait lock. This is
1185 * like in wake_up_bit() where clearing a bit has to be
1186 * ordered with the unlocked test.
1190 if (waitqueue_active(&ctx->wait))
1191 wake_up(&ctx->wait);
1194 static inline void iocb_put(struct aio_kiocb *iocb)
1196 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1202 /* aio_read_events_ring
1203 * Pull an event off of the ioctx's event ring. Returns the number of
1206 static long aio_read_events_ring(struct kioctx *ctx,
1207 struct io_event __user *event, long nr)
1209 struct aio_ring *ring;
1210 unsigned head, tail, pos;
1215 * The mutex can block and wake us up and that will cause
1216 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1217 * and repeat. This should be rare enough that it doesn't cause
1218 * peformance issues. See the comment in read_events() for more detail.
1220 sched_annotate_sleep();
1221 mutex_lock(&ctx->ring_lock);
1223 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1224 ring = page_address(ctx->ring_pages[0]);
1229 * Ensure that once we've read the current tail pointer, that
1230 * we also see the events that were stored up to the tail.
1234 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1239 head %= ctx->nr_events;
1240 tail %= ctx->nr_events;
1244 struct io_event *ev;
1247 avail = (head <= tail ? tail : ctx->nr_events) - head;
1251 pos = head + AIO_EVENTS_OFFSET;
1252 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1253 pos %= AIO_EVENTS_PER_PAGE;
1255 avail = min(avail, nr - ret);
1256 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1258 ev = page_address(page);
1259 copy_ret = copy_to_user(event + ret, ev + pos,
1260 sizeof(*ev) * avail);
1262 if (unlikely(copy_ret)) {
1269 head %= ctx->nr_events;
1272 ring = page_address(ctx->ring_pages[0]);
1274 flush_dcache_page(ctx->ring_pages[0]);
1276 pr_debug("%li h%u t%u\n", ret, head, tail);
1278 mutex_unlock(&ctx->ring_lock);
1283 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1284 struct io_event __user *event, long *i)
1286 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1291 if (unlikely(atomic_read(&ctx->dead)))
1297 return ret < 0 || *i >= min_nr;
1300 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1301 struct io_event __user *event,
1307 * Note that aio_read_events() is being called as the conditional - i.e.
1308 * we're calling it after prepare_to_wait() has set task state to
1309 * TASK_INTERRUPTIBLE.
1311 * But aio_read_events() can block, and if it blocks it's going to flip
1312 * the task state back to TASK_RUNNING.
1314 * This should be ok, provided it doesn't flip the state back to
1315 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1316 * will only happen if the mutex_lock() call blocks, and we then find
1317 * the ringbuffer empty. So in practice we should be ok, but it's
1318 * something to be aware of when touching this code.
1321 aio_read_events(ctx, min_nr, nr, event, &ret);
1323 wait_event_interruptible_hrtimeout(ctx->wait,
1324 aio_read_events(ctx, min_nr, nr, event, &ret),
1330 * Create an aio_context capable of receiving at least nr_events.
1331 * ctxp must not point to an aio_context that already exists, and
1332 * must be initialized to 0 prior to the call. On successful
1333 * creation of the aio_context, *ctxp is filled in with the resulting
1334 * handle. May fail with -EINVAL if *ctxp is not initialized,
1335 * if the specified nr_events exceeds internal limits. May fail
1336 * with -EAGAIN if the specified nr_events exceeds the user's limit
1337 * of available events. May fail with -ENOMEM if insufficient kernel
1338 * resources are available. May fail with -EFAULT if an invalid
1339 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1342 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1344 struct kioctx *ioctx = NULL;
1348 ret = get_user(ctx, ctxp);
1353 if (unlikely(ctx || nr_events == 0)) {
1354 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1359 ioctx = ioctx_alloc(nr_events);
1360 ret = PTR_ERR(ioctx);
1361 if (!IS_ERR(ioctx)) {
1362 ret = put_user(ioctx->user_id, ctxp);
1364 kill_ioctx(current->mm, ioctx, NULL);
1365 percpu_ref_put(&ioctx->users);
1372 #ifdef CONFIG_COMPAT
1373 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1375 struct kioctx *ioctx = NULL;
1379 ret = get_user(ctx, ctx32p);
1384 if (unlikely(ctx || nr_events == 0)) {
1385 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1390 ioctx = ioctx_alloc(nr_events);
1391 ret = PTR_ERR(ioctx);
1392 if (!IS_ERR(ioctx)) {
1393 /* truncating is ok because it's a user address */
1394 ret = put_user((u32)ioctx->user_id, ctx32p);
1396 kill_ioctx(current->mm, ioctx, NULL);
1397 percpu_ref_put(&ioctx->users);
1406 * Destroy the aio_context specified. May cancel any outstanding
1407 * AIOs and block on completion. Will fail with -ENOSYS if not
1408 * implemented. May fail with -EINVAL if the context pointed to
1411 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1413 struct kioctx *ioctx = lookup_ioctx(ctx);
1414 if (likely(NULL != ioctx)) {
1415 struct ctx_rq_wait wait;
1418 init_completion(&wait.comp);
1419 atomic_set(&wait.count, 1);
1421 /* Pass requests_done to kill_ioctx() where it can be set
1422 * in a thread-safe way. If we try to set it here then we have
1423 * a race condition if two io_destroy() called simultaneously.
1425 ret = kill_ioctx(current->mm, ioctx, &wait);
1426 percpu_ref_put(&ioctx->users);
1428 /* Wait until all IO for the context are done. Otherwise kernel
1429 * keep using user-space buffers even if user thinks the context
1433 wait_for_completion(&wait.comp);
1437 pr_debug("EINVAL: invalid context id\n");
1441 static void aio_remove_iocb(struct aio_kiocb *iocb)
1443 struct kioctx *ctx = iocb->ki_ctx;
1444 unsigned long flags;
1446 spin_lock_irqsave(&ctx->ctx_lock, flags);
1447 list_del(&iocb->ki_list);
1448 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1451 static void aio_complete_rw(struct kiocb *kiocb, long res)
1453 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1455 if (!list_empty_careful(&iocb->ki_list))
1456 aio_remove_iocb(iocb);
1458 if (kiocb->ki_flags & IOCB_WRITE) {
1459 struct inode *inode = file_inode(kiocb->ki_filp);
1461 if (S_ISREG(inode->i_mode))
1462 kiocb_end_write(kiocb);
1465 iocb->ki_res.res = res;
1466 iocb->ki_res.res2 = 0;
1470 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1474 req->ki_complete = aio_complete_rw;
1475 req->private = NULL;
1476 req->ki_pos = iocb->aio_offset;
1477 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1478 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1479 req->ki_flags |= IOCB_EVENTFD;
1480 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1482 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1483 * aio_reqprio is interpreted as an I/O scheduling
1484 * class and priority.
1486 ret = ioprio_check_cap(iocb->aio_reqprio);
1488 pr_debug("aio ioprio check cap error: %d\n", ret);
1492 req->ki_ioprio = iocb->aio_reqprio;
1494 req->ki_ioprio = get_current_ioprio();
1496 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1500 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1504 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1505 struct iovec **iovec, bool vectored, bool compat,
1506 struct iov_iter *iter)
1508 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1509 size_t len = iocb->aio_nbytes;
1512 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1517 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1520 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1526 case -ERESTARTNOINTR:
1527 case -ERESTARTNOHAND:
1528 case -ERESTART_RESTARTBLOCK:
1530 * There's no easy way to restart the syscall since other AIO's
1531 * may be already running. Just fail this IO with EINTR.
1536 req->ki_complete(req, ret);
1540 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1541 bool vectored, bool compat)
1543 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1544 struct iov_iter iter;
1548 ret = aio_prep_rw(req, iocb);
1551 file = req->ki_filp;
1552 if (unlikely(!(file->f_mode & FMODE_READ)))
1554 if (unlikely(!file->f_op->read_iter))
1557 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1560 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1562 aio_rw_done(req, call_read_iter(file, req, &iter));
1567 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1568 bool vectored, bool compat)
1570 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1571 struct iov_iter iter;
1575 ret = aio_prep_rw(req, iocb);
1578 file = req->ki_filp;
1580 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1582 if (unlikely(!file->f_op->write_iter))
1585 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1588 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1590 if (S_ISREG(file_inode(file)->i_mode))
1591 kiocb_start_write(req);
1592 req->ki_flags |= IOCB_WRITE;
1593 aio_rw_done(req, call_write_iter(file, req, &iter));
1599 static void aio_fsync_work(struct work_struct *work)
1601 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1602 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1604 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1605 revert_creds(old_cred);
1606 put_cred(iocb->fsync.creds);
1610 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1613 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1614 iocb->aio_rw_flags))
1617 if (unlikely(!req->file->f_op->fsync))
1620 req->creds = prepare_creds();
1624 req->datasync = datasync;
1625 INIT_WORK(&req->work, aio_fsync_work);
1626 schedule_work(&req->work);
1630 static void aio_poll_put_work(struct work_struct *work)
1632 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1633 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1639 * Safely lock the waitqueue which the request is on, synchronizing with the
1640 * case where the ->poll() provider decides to free its waitqueue early.
1642 * Returns true on success, meaning that req->head->lock was locked, req->wait
1643 * is on req->head, and an RCU read lock was taken. Returns false if the
1644 * request was already removed from its waitqueue (which might no longer exist).
1646 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1648 wait_queue_head_t *head;
1651 * While we hold the waitqueue lock and the waitqueue is nonempty,
1652 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1653 * lock in the first place can race with the waitqueue being freed.
1655 * We solve this as eventpoll does: by taking advantage of the fact that
1656 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1657 * we enter rcu_read_lock() and see that the pointer to the queue is
1658 * non-NULL, we can then lock it without the memory being freed out from
1659 * under us, then check whether the request is still on the queue.
1661 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1662 * case the caller deletes the entry from the queue, leaving it empty.
1663 * In that case, only RCU prevents the queue memory from being freed.
1666 head = smp_load_acquire(&req->head);
1668 spin_lock(&head->lock);
1669 if (!list_empty(&req->wait.entry))
1671 spin_unlock(&head->lock);
1677 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1679 spin_unlock(&req->head->lock);
1683 static void aio_poll_complete_work(struct work_struct *work)
1685 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1686 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1687 struct poll_table_struct pt = { ._key = req->events };
1688 struct kioctx *ctx = iocb->ki_ctx;
1691 if (!READ_ONCE(req->cancelled))
1692 mask = vfs_poll(req->file, &pt) & req->events;
1695 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1696 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1697 * synchronize with them. In the cancellation case the list_del_init
1698 * itself is not actually needed, but harmless so we keep it in to
1699 * avoid further branches in the fast path.
1701 spin_lock_irq(&ctx->ctx_lock);
1702 if (poll_iocb_lock_wq(req)) {
1703 if (!mask && !READ_ONCE(req->cancelled)) {
1705 * The request isn't actually ready to be completed yet.
1706 * Reschedule completion if another wakeup came in.
1708 if (req->work_need_resched) {
1709 schedule_work(&req->work);
1710 req->work_need_resched = false;
1712 req->work_scheduled = false;
1714 poll_iocb_unlock_wq(req);
1715 spin_unlock_irq(&ctx->ctx_lock);
1718 list_del_init(&req->wait.entry);
1719 poll_iocb_unlock_wq(req);
1720 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1721 list_del_init(&iocb->ki_list);
1722 iocb->ki_res.res = mangle_poll(mask);
1723 spin_unlock_irq(&ctx->ctx_lock);
1728 /* assumes we are called with irqs disabled */
1729 static int aio_poll_cancel(struct kiocb *iocb)
1731 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1732 struct poll_iocb *req = &aiocb->poll;
1734 if (poll_iocb_lock_wq(req)) {
1735 WRITE_ONCE(req->cancelled, true);
1736 if (!req->work_scheduled) {
1737 schedule_work(&aiocb->poll.work);
1738 req->work_scheduled = true;
1740 poll_iocb_unlock_wq(req);
1741 } /* else, the request was force-cancelled by POLLFREE already */
1746 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1749 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1750 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1751 __poll_t mask = key_to_poll(key);
1752 unsigned long flags;
1754 /* for instances that support it check for an event match first: */
1755 if (mask && !(mask & req->events))
1759 * Complete the request inline if possible. This requires that three
1760 * conditions be met:
1761 * 1. An event mask must have been passed. If a plain wakeup was done
1762 * instead, then mask == 0 and we have to call vfs_poll() to get
1763 * the events, so inline completion isn't possible.
1764 * 2. The completion work must not have already been scheduled.
1765 * 3. ctx_lock must not be busy. We have to use trylock because we
1766 * already hold the waitqueue lock, so this inverts the normal
1767 * locking order. Use irqsave/irqrestore because not all
1768 * filesystems (e.g. fuse) call this function with IRQs disabled,
1769 * yet IRQs have to be disabled before ctx_lock is obtained.
1771 if (mask && !req->work_scheduled &&
1772 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1773 struct kioctx *ctx = iocb->ki_ctx;
1775 list_del_init(&req->wait.entry);
1776 list_del(&iocb->ki_list);
1777 iocb->ki_res.res = mangle_poll(mask);
1778 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1780 INIT_WORK(&req->work, aio_poll_put_work);
1781 schedule_work(&req->work);
1783 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1788 * Schedule the completion work if needed. If it was already
1789 * scheduled, record that another wakeup came in.
1791 * Don't remove the request from the waitqueue here, as it might
1792 * not actually be complete yet (we won't know until vfs_poll()
1793 * is called), and we must not miss any wakeups. POLLFREE is an
1794 * exception to this; see below.
1796 if (req->work_scheduled) {
1797 req->work_need_resched = true;
1799 schedule_work(&req->work);
1800 req->work_scheduled = true;
1804 * If the waitqueue is being freed early but we can't complete
1805 * the request inline, we have to tear down the request as best
1806 * we can. That means immediately removing the request from its
1807 * waitqueue and preventing all further accesses to the
1808 * waitqueue via the request. We also need to schedule the
1809 * completion work (done above). Also mark the request as
1810 * cancelled, to potentially skip an unneeded call to ->poll().
1812 if (mask & POLLFREE) {
1813 WRITE_ONCE(req->cancelled, true);
1814 list_del_init(&req->wait.entry);
1817 * Careful: this *must* be the last step, since as soon
1818 * as req->head is NULL'ed out, the request can be
1819 * completed and freed, since aio_poll_complete_work()
1820 * will no longer need to take the waitqueue lock.
1822 smp_store_release(&req->head, NULL);
1828 struct aio_poll_table {
1829 struct poll_table_struct pt;
1830 struct aio_kiocb *iocb;
1836 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1837 struct poll_table_struct *p)
1839 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1841 /* multiple wait queues per file are not supported */
1842 if (unlikely(pt->queued)) {
1843 pt->error = -EINVAL;
1849 pt->iocb->poll.head = head;
1850 add_wait_queue(head, &pt->iocb->poll.wait);
1853 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1855 struct kioctx *ctx = aiocb->ki_ctx;
1856 struct poll_iocb *req = &aiocb->poll;
1857 struct aio_poll_table apt;
1858 bool cancel = false;
1861 /* reject any unknown events outside the normal event mask. */
1862 if ((u16)iocb->aio_buf != iocb->aio_buf)
1864 /* reject fields that are not defined for poll */
1865 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1868 INIT_WORK(&req->work, aio_poll_complete_work);
1869 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1872 req->cancelled = false;
1873 req->work_scheduled = false;
1874 req->work_need_resched = false;
1876 apt.pt._qproc = aio_poll_queue_proc;
1877 apt.pt._key = req->events;
1880 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1882 /* initialized the list so that we can do list_empty checks */
1883 INIT_LIST_HEAD(&req->wait.entry);
1884 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1886 mask = vfs_poll(req->file, &apt.pt) & req->events;
1887 spin_lock_irq(&ctx->ctx_lock);
1888 if (likely(apt.queued)) {
1889 bool on_queue = poll_iocb_lock_wq(req);
1891 if (!on_queue || req->work_scheduled) {
1893 * aio_poll_wake() already either scheduled the async
1894 * completion work, or completed the request inline.
1896 if (apt.error) /* unsupported case: multiple queues */
1901 if (mask || apt.error) {
1902 /* Steal to complete synchronously. */
1903 list_del_init(&req->wait.entry);
1904 } else if (cancel) {
1905 /* Cancel if possible (may be too late though). */
1906 WRITE_ONCE(req->cancelled, true);
1907 } else if (on_queue) {
1909 * Actually waiting for an event, so add the request to
1910 * active_reqs so that it can be cancelled if needed.
1912 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1913 aiocb->ki_cancel = aio_poll_cancel;
1916 poll_iocb_unlock_wq(req);
1918 if (mask) { /* no async, we'd stolen it */
1919 aiocb->ki_res.res = mangle_poll(mask);
1922 spin_unlock_irq(&ctx->ctx_lock);
1928 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1929 struct iocb __user *user_iocb, struct aio_kiocb *req,
1932 req->ki_filp = fget(iocb->aio_fildes);
1933 if (unlikely(!req->ki_filp))
1936 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1937 struct eventfd_ctx *eventfd;
1939 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1940 * instance of the file* now. The file descriptor must be
1941 * an eventfd() fd, and will be signaled for each completed
1942 * event using the eventfd_signal() function.
1944 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1945 if (IS_ERR(eventfd))
1946 return PTR_ERR(eventfd);
1948 req->ki_eventfd = eventfd;
1951 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1952 pr_debug("EFAULT: aio_key\n");
1956 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1957 req->ki_res.data = iocb->aio_data;
1958 req->ki_res.res = 0;
1959 req->ki_res.res2 = 0;
1961 switch (iocb->aio_lio_opcode) {
1962 case IOCB_CMD_PREAD:
1963 return aio_read(&req->rw, iocb, false, compat);
1964 case IOCB_CMD_PWRITE:
1965 return aio_write(&req->rw, iocb, false, compat);
1966 case IOCB_CMD_PREADV:
1967 return aio_read(&req->rw, iocb, true, compat);
1968 case IOCB_CMD_PWRITEV:
1969 return aio_write(&req->rw, iocb, true, compat);
1970 case IOCB_CMD_FSYNC:
1971 return aio_fsync(&req->fsync, iocb, false);
1972 case IOCB_CMD_FDSYNC:
1973 return aio_fsync(&req->fsync, iocb, true);
1975 return aio_poll(req, iocb);
1977 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1982 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1985 struct aio_kiocb *req;
1989 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1992 /* enforce forwards compatibility on users */
1993 if (unlikely(iocb.aio_reserved2)) {
1994 pr_debug("EINVAL: reserve field set\n");
1998 /* prevent overflows */
2000 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2001 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2002 ((ssize_t)iocb.aio_nbytes < 0)
2004 pr_debug("EINVAL: overflow check\n");
2008 req = aio_get_req(ctx);
2012 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2014 /* Done with the synchronous reference */
2018 * If err is 0, we'd either done aio_complete() ourselves or have
2019 * arranged for that to be done asynchronously. Anything non-zero
2020 * means that we need to destroy req ourselves.
2022 if (unlikely(err)) {
2024 put_reqs_available(ctx, 1);
2030 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2031 * the number of iocbs queued. May return -EINVAL if the aio_context
2032 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2033 * *iocbpp[0] is not properly initialized, if the operation specified
2034 * is invalid for the file descriptor in the iocb. May fail with
2035 * -EFAULT if any of the data structures point to invalid data. May
2036 * fail with -EBADF if the file descriptor specified in the first
2037 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2038 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2039 * fail with -ENOSYS if not implemented.
2041 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2042 struct iocb __user * __user *, iocbpp)
2047 struct blk_plug plug;
2049 if (unlikely(nr < 0))
2052 ctx = lookup_ioctx(ctx_id);
2053 if (unlikely(!ctx)) {
2054 pr_debug("EINVAL: invalid context id\n");
2058 if (nr > ctx->nr_events)
2059 nr = ctx->nr_events;
2061 if (nr > AIO_PLUG_THRESHOLD)
2062 blk_start_plug(&plug);
2063 for (i = 0; i < nr; i++) {
2064 struct iocb __user *user_iocb;
2066 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2071 ret = io_submit_one(ctx, user_iocb, false);
2075 if (nr > AIO_PLUG_THRESHOLD)
2076 blk_finish_plug(&plug);
2078 percpu_ref_put(&ctx->users);
2082 #ifdef CONFIG_COMPAT
2083 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2084 int, nr, compat_uptr_t __user *, iocbpp)
2089 struct blk_plug plug;
2091 if (unlikely(nr < 0))
2094 ctx = lookup_ioctx(ctx_id);
2095 if (unlikely(!ctx)) {
2096 pr_debug("EINVAL: invalid context id\n");
2100 if (nr > ctx->nr_events)
2101 nr = ctx->nr_events;
2103 if (nr > AIO_PLUG_THRESHOLD)
2104 blk_start_plug(&plug);
2105 for (i = 0; i < nr; i++) {
2106 compat_uptr_t user_iocb;
2108 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2113 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2117 if (nr > AIO_PLUG_THRESHOLD)
2118 blk_finish_plug(&plug);
2120 percpu_ref_put(&ctx->users);
2126 * Attempts to cancel an iocb previously passed to io_submit. If
2127 * the operation is successfully cancelled, the resulting event is
2128 * copied into the memory pointed to by result without being placed
2129 * into the completion queue and 0 is returned. May fail with
2130 * -EFAULT if any of the data structures pointed to are invalid.
2131 * May fail with -EINVAL if aio_context specified by ctx_id is
2132 * invalid. May fail with -EAGAIN if the iocb specified was not
2133 * cancelled. Will fail with -ENOSYS if not implemented.
2135 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2136 struct io_event __user *, result)
2139 struct aio_kiocb *kiocb;
2142 u64 obj = (u64)(unsigned long)iocb;
2144 if (unlikely(get_user(key, &iocb->aio_key)))
2146 if (unlikely(key != KIOCB_KEY))
2149 ctx = lookup_ioctx(ctx_id);
2153 spin_lock_irq(&ctx->ctx_lock);
2154 /* TODO: use a hash or array, this sucks. */
2155 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2156 if (kiocb->ki_res.obj == obj) {
2157 ret = kiocb->ki_cancel(&kiocb->rw);
2158 list_del_init(&kiocb->ki_list);
2162 spin_unlock_irq(&ctx->ctx_lock);
2166 * The result argument is no longer used - the io_event is
2167 * always delivered via the ring buffer. -EINPROGRESS indicates
2168 * cancellation is progress:
2173 percpu_ref_put(&ctx->users);
2178 static long do_io_getevents(aio_context_t ctx_id,
2181 struct io_event __user *events,
2182 struct timespec64 *ts)
2184 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2185 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2188 if (likely(ioctx)) {
2189 if (likely(min_nr <= nr && min_nr >= 0))
2190 ret = read_events(ioctx, min_nr, nr, events, until);
2191 percpu_ref_put(&ioctx->users);
2198 * Attempts to read at least min_nr events and up to nr events from
2199 * the completion queue for the aio_context specified by ctx_id. If
2200 * it succeeds, the number of read events is returned. May fail with
2201 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2202 * out of range, if timeout is out of range. May fail with -EFAULT
2203 * if any of the memory specified is invalid. May return 0 or
2204 * < min_nr if the timeout specified by timeout has elapsed
2205 * before sufficient events are available, where timeout == NULL
2206 * specifies an infinite timeout. Note that the timeout pointed to by
2207 * timeout is relative. Will fail with -ENOSYS if not implemented.
2211 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2214 struct io_event __user *, events,
2215 struct __kernel_timespec __user *, timeout)
2217 struct timespec64 ts;
2220 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2223 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2224 if (!ret && signal_pending(current))
2231 struct __aio_sigset {
2232 const sigset_t __user *sigmask;
2236 SYSCALL_DEFINE6(io_pgetevents,
2237 aio_context_t, ctx_id,
2240 struct io_event __user *, events,
2241 struct __kernel_timespec __user *, timeout,
2242 const struct __aio_sigset __user *, usig)
2244 struct __aio_sigset ksig = { NULL, };
2245 struct timespec64 ts;
2249 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2252 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2255 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2259 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2261 interrupted = signal_pending(current);
2262 restore_saved_sigmask_unless(interrupted);
2263 if (interrupted && !ret)
2264 ret = -ERESTARTNOHAND;
2269 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2271 SYSCALL_DEFINE6(io_pgetevents_time32,
2272 aio_context_t, ctx_id,
2275 struct io_event __user *, events,
2276 struct old_timespec32 __user *, timeout,
2277 const struct __aio_sigset __user *, usig)
2279 struct __aio_sigset ksig = { NULL, };
2280 struct timespec64 ts;
2284 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2287 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2291 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2295 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2297 interrupted = signal_pending(current);
2298 restore_saved_sigmask_unless(interrupted);
2299 if (interrupted && !ret)
2300 ret = -ERESTARTNOHAND;
2307 #if defined(CONFIG_COMPAT_32BIT_TIME)
2309 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2312 struct io_event __user *, events,
2313 struct old_timespec32 __user *, timeout)
2315 struct timespec64 t;
2318 if (timeout && get_old_timespec32(&t, timeout))
2321 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2322 if (!ret && signal_pending(current))
2329 #ifdef CONFIG_COMPAT
2331 struct __compat_aio_sigset {
2332 compat_uptr_t sigmask;
2333 compat_size_t sigsetsize;
2336 #if defined(CONFIG_COMPAT_32BIT_TIME)
2338 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2339 compat_aio_context_t, ctx_id,
2340 compat_long_t, min_nr,
2342 struct io_event __user *, events,
2343 struct old_timespec32 __user *, timeout,
2344 const struct __compat_aio_sigset __user *, usig)
2346 struct __compat_aio_sigset ksig = { 0, };
2347 struct timespec64 t;
2351 if (timeout && get_old_timespec32(&t, timeout))
2354 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2357 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2361 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2363 interrupted = signal_pending(current);
2364 restore_saved_sigmask_unless(interrupted);
2365 if (interrupted && !ret)
2366 ret = -ERESTARTNOHAND;
2373 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2374 compat_aio_context_t, ctx_id,
2375 compat_long_t, min_nr,
2377 struct io_event __user *, events,
2378 struct __kernel_timespec __user *, timeout,
2379 const struct __compat_aio_sigset __user *, usig)
2381 struct __compat_aio_sigset ksig = { 0, };
2382 struct timespec64 t;
2386 if (timeout && get_timespec64(&t, timeout))
2389 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2392 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2396 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2398 interrupted = signal_pending(current);
2399 restore_saved_sigmask_unless(interrupted);
2400 if (interrupted && !ret)
2401 ret = -ERESTARTNOHAND;