2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig.
10 * See ../COPYING for licensing terms.
12 #define pr_fmt(fmt) "%s: " fmt, __func__
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
25 #include <linux/sched/signal.h>
27 #include <linux/file.h>
29 #include <linux/mman.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.h>
46 #include <linux/uaccess.h>
47 #include <linux/nospec.h>
53 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0
57 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by aio_read_events_ring(). */
64 unsigned compat_features;
65 unsigned incompat_features;
66 unsigned header_length; /* size of aio_ring */
69 struct io_event io_events[];
70 }; /* 128 bytes + ring size */
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
76 #define AIO_PLUG_THRESHOLD 2
78 #define AIO_RING_PAGES 8
83 struct kioctx __rcu *table[];
87 unsigned reqs_available;
91 struct completion comp;
96 struct percpu_ref users;
99 struct percpu_ref reqs;
101 unsigned long user_id;
103 struct __percpu kioctx_cpu *cpu;
106 * For percpu reqs_available, number of slots we move to/from global
111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the global max_reqs quota.
114 * The real limit is nr_events - 1, which will be larger (see
119 /* Size of ringbuffer, in units of struct io_event */
122 unsigned long mmap_base;
123 unsigned long mmap_size;
125 struct page **ring_pages;
128 struct rcu_work free_rwork; /* see free_ioctx() */
131 * signals when all in-flight requests are done
133 struct ctx_rq_wait *rq_wait;
137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the ringbuffer.
142 * We batch accesses to it with a percpu version.
144 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp;
149 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp;
153 struct mutex ring_lock;
154 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp;
159 unsigned completed_events;
160 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp;
163 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file;
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
175 struct work_struct work;
182 struct wait_queue_head *head;
186 bool work_need_resched;
187 struct wait_queue_entry wait;
188 struct work_struct work;
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
199 struct file *ki_filp;
201 struct fsync_iocb fsync;
202 struct poll_iocb poll;
205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel;
208 struct io_event ki_res;
210 struct list_head ki_list; /* the aio core uses this
211 * for cancellation */
212 refcount_t ki_refcnt;
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
218 struct eventfd_ctx *ki_eventfd;
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock);
223 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);
364 for (i = 0; i < table->nr; i++) {
367 ctx = rcu_dereference(table->table[i]);
368 if (ctx && ctx->aio_ring_file == file) {
369 if (!atomic_read(&ctx->dead)) {
370 ctx->user_id = ctx->mmap_base = vma->vm_start;
378 spin_unlock(&mm->ioctx_lock);
382 static const struct vm_operations_struct aio_ring_vm_ops = {
383 .mremap = aio_ring_mremap,
384 #if IS_ENABLED(CONFIG_MMU)
385 .fault = filemap_fault,
386 .map_pages = filemap_map_pages,
387 .page_mkwrite = filemap_page_mkwrite,
391 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
393 vma->vm_flags |= VM_DONTEXPAND;
394 vma->vm_ops = &aio_ring_vm_ops;
398 static const struct file_operations aio_ring_fops = {
399 .mmap = aio_ring_mmap,
402 #if IS_ENABLED(CONFIG_MIGRATION)
403 static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
404 struct folio *src, enum migrate_mode mode)
412 * We cannot support the _NO_COPY case here, because copy needs to
413 * happen under the ctx->completion_lock. That does not work with the
414 * migration workflow of MIGRATE_SYNC_NO_COPY.
416 if (mode == MIGRATE_SYNC_NO_COPY)
421 /* mapping->private_lock here protects against the kioctx teardown. */
422 spin_lock(&mapping->private_lock);
423 ctx = mapping->private_data;
429 /* The ring_lock mutex. The prevents aio_read_events() from writing
430 * to the ring's head, and prevents page migration from mucking in
431 * a partially initialized kiotx.
433 if (!mutex_trylock(&ctx->ring_lock)) {
439 if (idx < (pgoff_t)ctx->nr_pages) {
440 /* Make sure the old folio hasn't already been changed */
441 if (ctx->ring_pages[idx] != &src->page)
449 /* Writeback must be complete */
450 BUG_ON(folio_test_writeback(src));
453 rc = folio_migrate_mapping(mapping, dst, src, 1);
454 if (rc != MIGRATEPAGE_SUCCESS) {
459 /* Take completion_lock to prevent other writes to the ring buffer
460 * while the old folio is copied to the new. This prevents new
461 * events from being lost.
463 spin_lock_irqsave(&ctx->completion_lock, flags);
464 folio_migrate_copy(dst, src);
465 BUG_ON(ctx->ring_pages[idx] != &src->page);
466 ctx->ring_pages[idx] = &dst->page;
467 spin_unlock_irqrestore(&ctx->completion_lock, flags);
469 /* The old folio is no longer accessible. */
473 mutex_unlock(&ctx->ring_lock);
475 spin_unlock(&mapping->private_lock);
479 #define aio_migrate_folio NULL
482 static const struct address_space_operations aio_ctx_aops = {
483 .dirty_folio = noop_dirty_folio,
484 .migrate_folio = aio_migrate_folio,
487 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
489 struct aio_ring *ring;
490 struct mm_struct *mm = current->mm;
491 unsigned long size, unused;
496 /* Compensate for the ring buffer's head/tail overlap entry */
497 nr_events += 2; /* 1 is required, 2 for good luck */
499 size = sizeof(struct aio_ring);
500 size += sizeof(struct io_event) * nr_events;
502 nr_pages = PFN_UP(size);
506 file = aio_private_file(ctx, nr_pages);
508 ctx->aio_ring_file = NULL;
512 ctx->aio_ring_file = file;
513 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
514 / sizeof(struct io_event);
516 ctx->ring_pages = ctx->internal_pages;
517 if (nr_pages > AIO_RING_PAGES) {
518 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
520 if (!ctx->ring_pages) {
521 put_aio_ring_file(ctx);
526 for (i = 0; i < nr_pages; i++) {
528 page = find_or_create_page(file->f_mapping,
529 i, GFP_HIGHUSER | __GFP_ZERO);
532 pr_debug("pid(%d) page[%d]->count=%d\n",
533 current->pid, i, page_count(page));
534 SetPageUptodate(page);
537 ctx->ring_pages[i] = page;
541 if (unlikely(i != nr_pages)) {
546 ctx->mmap_size = nr_pages * PAGE_SIZE;
547 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
549 if (mmap_write_lock_killable(mm)) {
555 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
556 PROT_READ | PROT_WRITE,
557 MAP_SHARED, 0, &unused, NULL);
558 mmap_write_unlock(mm);
559 if (IS_ERR((void *)ctx->mmap_base)) {
565 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
567 ctx->user_id = ctx->mmap_base;
568 ctx->nr_events = nr_events; /* trusted copy */
570 ring = kmap_atomic(ctx->ring_pages[0]);
571 ring->nr = nr_events; /* user copy */
573 ring->head = ring->tail = 0;
574 ring->magic = AIO_RING_MAGIC;
575 ring->compat_features = AIO_RING_COMPAT_FEATURES;
576 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
577 ring->header_length = sizeof(struct aio_ring);
579 flush_dcache_page(ctx->ring_pages[0]);
584 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
585 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
586 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
588 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
590 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
591 struct kioctx *ctx = req->ki_ctx;
594 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
597 spin_lock_irqsave(&ctx->ctx_lock, flags);
598 list_add_tail(&req->ki_list, &ctx->active_reqs);
599 req->ki_cancel = cancel;
600 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
602 EXPORT_SYMBOL(kiocb_set_cancel_fn);
605 * free_ioctx() should be RCU delayed to synchronize against the RCU
606 * protected lookup_ioctx() and also needs process context to call
607 * aio_free_ring(). Use rcu_work.
609 static void free_ioctx(struct work_struct *work)
611 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
613 pr_debug("freeing %p\n", ctx);
616 free_percpu(ctx->cpu);
617 percpu_ref_exit(&ctx->reqs);
618 percpu_ref_exit(&ctx->users);
619 kmem_cache_free(kioctx_cachep, ctx);
622 static void free_ioctx_reqs(struct percpu_ref *ref)
624 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
626 /* At this point we know that there are no any in-flight requests */
627 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
628 complete(&ctx->rq_wait->comp);
630 /* Synchronize against RCU protected table->table[] dereferences */
631 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
632 queue_rcu_work(system_wq, &ctx->free_rwork);
636 * When this function runs, the kioctx has been removed from the "hash table"
637 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
638 * now it's safe to cancel any that need to be.
640 static void free_ioctx_users(struct percpu_ref *ref)
642 struct kioctx *ctx = container_of(ref, struct kioctx, users);
643 struct aio_kiocb *req;
645 spin_lock_irq(&ctx->ctx_lock);
647 while (!list_empty(&ctx->active_reqs)) {
648 req = list_first_entry(&ctx->active_reqs,
649 struct aio_kiocb, ki_list);
650 req->ki_cancel(&req->rw);
651 list_del_init(&req->ki_list);
654 spin_unlock_irq(&ctx->ctx_lock);
656 percpu_ref_kill(&ctx->reqs);
657 percpu_ref_put(&ctx->reqs);
660 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
663 struct kioctx_table *table, *old;
664 struct aio_ring *ring;
666 spin_lock(&mm->ioctx_lock);
667 table = rcu_dereference_raw(mm->ioctx_table);
671 for (i = 0; i < table->nr; i++)
672 if (!rcu_access_pointer(table->table[i])) {
674 rcu_assign_pointer(table->table[i], ctx);
675 spin_unlock(&mm->ioctx_lock);
677 /* While kioctx setup is in progress,
678 * we are protected from page migration
679 * changes ring_pages by ->ring_lock.
681 ring = kmap_atomic(ctx->ring_pages[0]);
687 new_nr = (table ? table->nr : 1) * 4;
688 spin_unlock(&mm->ioctx_lock);
690 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
696 spin_lock(&mm->ioctx_lock);
697 old = rcu_dereference_raw(mm->ioctx_table);
700 rcu_assign_pointer(mm->ioctx_table, table);
701 } else if (table->nr > old->nr) {
702 memcpy(table->table, old->table,
703 old->nr * sizeof(struct kioctx *));
705 rcu_assign_pointer(mm->ioctx_table, table);
714 static void aio_nr_sub(unsigned nr)
716 spin_lock(&aio_nr_lock);
717 if (WARN_ON(aio_nr - nr > aio_nr))
721 spin_unlock(&aio_nr_lock);
725 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
727 static struct kioctx *ioctx_alloc(unsigned nr_events)
729 struct mm_struct *mm = current->mm;
734 * Store the original nr_events -- what userspace passed to io_setup(),
735 * for counting against the global limit -- before it changes.
737 unsigned int max_reqs = nr_events;
740 * We keep track of the number of available ringbuffer slots, to prevent
741 * overflow (reqs_available), and we also use percpu counters for this.
743 * So since up to half the slots might be on other cpu's percpu counters
744 * and unavailable, double nr_events so userspace sees what they
745 * expected: additionally, we move req_batch slots to/from percpu
746 * counters at a time, so make sure that isn't 0:
748 nr_events = max(nr_events, num_possible_cpus() * 4);
751 /* Prevent overflows */
752 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
753 pr_debug("ENOMEM: nr_events too high\n");
754 return ERR_PTR(-EINVAL);
757 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
758 return ERR_PTR(-EAGAIN);
760 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
762 return ERR_PTR(-ENOMEM);
764 ctx->max_reqs = max_reqs;
766 spin_lock_init(&ctx->ctx_lock);
767 spin_lock_init(&ctx->completion_lock);
768 mutex_init(&ctx->ring_lock);
769 /* Protect against page migration throughout kiotx setup by keeping
770 * the ring_lock mutex held until setup is complete. */
771 mutex_lock(&ctx->ring_lock);
772 init_waitqueue_head(&ctx->wait);
774 INIT_LIST_HEAD(&ctx->active_reqs);
776 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
779 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
782 ctx->cpu = alloc_percpu(struct kioctx_cpu);
786 err = aio_setup_ring(ctx, nr_events);
790 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
791 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
792 if (ctx->req_batch < 1)
795 /* limit the number of system wide aios */
796 spin_lock(&aio_nr_lock);
797 if (aio_nr + ctx->max_reqs > aio_max_nr ||
798 aio_nr + ctx->max_reqs < aio_nr) {
799 spin_unlock(&aio_nr_lock);
803 aio_nr += ctx->max_reqs;
804 spin_unlock(&aio_nr_lock);
806 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
807 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
809 err = ioctx_add_table(ctx, mm);
813 /* Release the ring_lock mutex now that all setup is complete. */
814 mutex_unlock(&ctx->ring_lock);
816 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
817 ctx, ctx->user_id, mm, ctx->nr_events);
821 aio_nr_sub(ctx->max_reqs);
823 atomic_set(&ctx->dead, 1);
825 vm_munmap(ctx->mmap_base, ctx->mmap_size);
828 mutex_unlock(&ctx->ring_lock);
829 free_percpu(ctx->cpu);
830 percpu_ref_exit(&ctx->reqs);
831 percpu_ref_exit(&ctx->users);
832 kmem_cache_free(kioctx_cachep, ctx);
833 pr_debug("error allocating ioctx %d\n", err);
838 * Cancels all outstanding aio requests on an aio context. Used
839 * when the processes owning a context have all exited to encourage
840 * the rapid destruction of the kioctx.
842 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
843 struct ctx_rq_wait *wait)
845 struct kioctx_table *table;
847 spin_lock(&mm->ioctx_lock);
848 if (atomic_xchg(&ctx->dead, 1)) {
849 spin_unlock(&mm->ioctx_lock);
853 table = rcu_dereference_raw(mm->ioctx_table);
854 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
855 RCU_INIT_POINTER(table->table[ctx->id], NULL);
856 spin_unlock(&mm->ioctx_lock);
858 /* free_ioctx_reqs() will do the necessary RCU synchronization */
859 wake_up_all(&ctx->wait);
862 * It'd be more correct to do this in free_ioctx(), after all
863 * the outstanding kiocbs have finished - but by then io_destroy
864 * has already returned, so io_setup() could potentially return
865 * -EAGAIN with no ioctxs actually in use (as far as userspace
868 aio_nr_sub(ctx->max_reqs);
871 vm_munmap(ctx->mmap_base, ctx->mmap_size);
874 percpu_ref_kill(&ctx->users);
879 * exit_aio: called when the last user of mm goes away. At this point, there is
880 * no way for any new requests to be submited or any of the io_* syscalls to be
881 * called on the context.
883 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
886 void exit_aio(struct mm_struct *mm)
888 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
889 struct ctx_rq_wait wait;
895 atomic_set(&wait.count, table->nr);
896 init_completion(&wait.comp);
899 for (i = 0; i < table->nr; ++i) {
901 rcu_dereference_protected(table->table[i], true);
909 * We don't need to bother with munmap() here - exit_mmap(mm)
910 * is coming and it'll unmap everything. And we simply can't,
911 * this is not necessarily our ->mm.
912 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
913 * that it needs to unmap the area, just set it to 0.
916 kill_ioctx(mm, ctx, &wait);
919 if (!atomic_sub_and_test(skipped, &wait.count)) {
920 /* Wait until all IO for the context are done. */
921 wait_for_completion(&wait.comp);
924 RCU_INIT_POINTER(mm->ioctx_table, NULL);
928 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
930 struct kioctx_cpu *kcpu;
933 local_irq_save(flags);
934 kcpu = this_cpu_ptr(ctx->cpu);
935 kcpu->reqs_available += nr;
937 while (kcpu->reqs_available >= ctx->req_batch * 2) {
938 kcpu->reqs_available -= ctx->req_batch;
939 atomic_add(ctx->req_batch, &ctx->reqs_available);
942 local_irq_restore(flags);
945 static bool __get_reqs_available(struct kioctx *ctx)
947 struct kioctx_cpu *kcpu;
951 local_irq_save(flags);
952 kcpu = this_cpu_ptr(ctx->cpu);
953 if (!kcpu->reqs_available) {
954 int avail = atomic_read(&ctx->reqs_available);
957 if (avail < ctx->req_batch)
959 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
960 &avail, avail - ctx->req_batch));
962 kcpu->reqs_available += ctx->req_batch;
966 kcpu->reqs_available--;
968 local_irq_restore(flags);
972 /* refill_reqs_available
973 * Updates the reqs_available reference counts used for tracking the
974 * number of free slots in the completion ring. This can be called
975 * from aio_complete() (to optimistically update reqs_available) or
976 * from aio_get_req() (the we're out of events case). It must be
977 * called holding ctx->completion_lock.
979 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
982 unsigned events_in_ring, completed;
984 /* Clamp head since userland can write to it. */
985 head %= ctx->nr_events;
987 events_in_ring = tail - head;
989 events_in_ring = ctx->nr_events - (head - tail);
991 completed = ctx->completed_events;
992 if (events_in_ring < completed)
993 completed -= events_in_ring;
1000 ctx->completed_events -= completed;
1001 put_reqs_available(ctx, completed);
1004 /* user_refill_reqs_available
1005 * Called to refill reqs_available when aio_get_req() encounters an
1006 * out of space in the completion ring.
1008 static void user_refill_reqs_available(struct kioctx *ctx)
1010 spin_lock_irq(&ctx->completion_lock);
1011 if (ctx->completed_events) {
1012 struct aio_ring *ring;
1015 /* Access of ring->head may race with aio_read_events_ring()
1016 * here, but that's okay since whether we read the old version
1017 * or the new version, and either will be valid. The important
1018 * part is that head cannot pass tail since we prevent
1019 * aio_complete() from updating tail by holding
1020 * ctx->completion_lock. Even if head is invalid, the check
1021 * against ctx->completed_events below will make sure we do the
1024 ring = kmap_atomic(ctx->ring_pages[0]);
1026 kunmap_atomic(ring);
1028 refill_reqs_available(ctx, head, ctx->tail);
1031 spin_unlock_irq(&ctx->completion_lock);
1034 static bool get_reqs_available(struct kioctx *ctx)
1036 if (__get_reqs_available(ctx))
1038 user_refill_reqs_available(ctx);
1039 return __get_reqs_available(ctx);
1043 * Allocate a slot for an aio request.
1044 * Returns NULL if no requests are free.
1046 * The refcount is initialized to 2 - one for the async op completion,
1047 * one for the synchronous code that does this.
1049 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1051 struct aio_kiocb *req;
1053 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1057 if (unlikely(!get_reqs_available(ctx))) {
1058 kmem_cache_free(kiocb_cachep, req);
1062 percpu_ref_get(&ctx->reqs);
1064 INIT_LIST_HEAD(&req->ki_list);
1065 refcount_set(&req->ki_refcnt, 2);
1066 req->ki_eventfd = NULL;
1070 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1072 struct aio_ring __user *ring = (void __user *)ctx_id;
1073 struct mm_struct *mm = current->mm;
1074 struct kioctx *ctx, *ret = NULL;
1075 struct kioctx_table *table;
1078 if (get_user(id, &ring->id))
1082 table = rcu_dereference(mm->ioctx_table);
1084 if (!table || id >= table->nr)
1087 id = array_index_nospec(id, table->nr);
1088 ctx = rcu_dereference(table->table[id]);
1089 if (ctx && ctx->user_id == ctx_id) {
1090 if (percpu_ref_tryget_live(&ctx->users))
1098 static inline void iocb_destroy(struct aio_kiocb *iocb)
1100 if (iocb->ki_eventfd)
1101 eventfd_ctx_put(iocb->ki_eventfd);
1103 fput(iocb->ki_filp);
1104 percpu_ref_put(&iocb->ki_ctx->reqs);
1105 kmem_cache_free(kiocb_cachep, iocb);
1109 * Called when the io request on the given iocb is complete.
1111 static void aio_complete(struct aio_kiocb *iocb)
1113 struct kioctx *ctx = iocb->ki_ctx;
1114 struct aio_ring *ring;
1115 struct io_event *ev_page, *event;
1116 unsigned tail, pos, head;
1117 unsigned long flags;
1120 * Add a completion event to the ring buffer. Must be done holding
1121 * ctx->completion_lock to prevent other code from messing with the tail
1122 * pointer since we might be called from irq context.
1124 spin_lock_irqsave(&ctx->completion_lock, flags);
1127 pos = tail + AIO_EVENTS_OFFSET;
1129 if (++tail >= ctx->nr_events)
1132 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1133 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1135 *event = iocb->ki_res;
1137 kunmap_atomic(ev_page);
1138 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1140 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1141 (void __user *)(unsigned long)iocb->ki_res.obj,
1142 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1144 /* after flagging the request as done, we
1145 * must never even look at it again
1147 smp_wmb(); /* make event visible before updating tail */
1151 ring = kmap_atomic(ctx->ring_pages[0]);
1154 kunmap_atomic(ring);
1155 flush_dcache_page(ctx->ring_pages[0]);
1157 ctx->completed_events++;
1158 if (ctx->completed_events > 1)
1159 refill_reqs_available(ctx, head, tail);
1160 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1162 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1165 * Check if the user asked us to deliver the result through an
1166 * eventfd. The eventfd_signal() function is safe to be called
1169 if (iocb->ki_eventfd)
1170 eventfd_signal(iocb->ki_eventfd, 1);
1173 * We have to order our ring_info tail store above and test
1174 * of the wait list below outside the wait lock. This is
1175 * like in wake_up_bit() where clearing a bit has to be
1176 * ordered with the unlocked test.
1180 if (waitqueue_active(&ctx->wait))
1181 wake_up(&ctx->wait);
1184 static inline void iocb_put(struct aio_kiocb *iocb)
1186 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1192 /* aio_read_events_ring
1193 * Pull an event off of the ioctx's event ring. Returns the number of
1196 static long aio_read_events_ring(struct kioctx *ctx,
1197 struct io_event __user *event, long nr)
1199 struct aio_ring *ring;
1200 unsigned head, tail, pos;
1205 * The mutex can block and wake us up and that will cause
1206 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1207 * and repeat. This should be rare enough that it doesn't cause
1208 * peformance issues. See the comment in read_events() for more detail.
1210 sched_annotate_sleep();
1211 mutex_lock(&ctx->ring_lock);
1213 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1214 ring = kmap_atomic(ctx->ring_pages[0]);
1217 kunmap_atomic(ring);
1220 * Ensure that once we've read the current tail pointer, that
1221 * we also see the events that were stored up to the tail.
1225 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1230 head %= ctx->nr_events;
1231 tail %= ctx->nr_events;
1235 struct io_event *ev;
1238 avail = (head <= tail ? tail : ctx->nr_events) - head;
1242 pos = head + AIO_EVENTS_OFFSET;
1243 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1244 pos %= AIO_EVENTS_PER_PAGE;
1246 avail = min(avail, nr - ret);
1247 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1250 copy_ret = copy_to_user(event + ret, ev + pos,
1251 sizeof(*ev) * avail);
1254 if (unlikely(copy_ret)) {
1261 head %= ctx->nr_events;
1264 ring = kmap_atomic(ctx->ring_pages[0]);
1266 kunmap_atomic(ring);
1267 flush_dcache_page(ctx->ring_pages[0]);
1269 pr_debug("%li h%u t%u\n", ret, head, tail);
1271 mutex_unlock(&ctx->ring_lock);
1276 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1277 struct io_event __user *event, long *i)
1279 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1284 if (unlikely(atomic_read(&ctx->dead)))
1290 return ret < 0 || *i >= min_nr;
1293 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1294 struct io_event __user *event,
1300 * Note that aio_read_events() is being called as the conditional - i.e.
1301 * we're calling it after prepare_to_wait() has set task state to
1302 * TASK_INTERRUPTIBLE.
1304 * But aio_read_events() can block, and if it blocks it's going to flip
1305 * the task state back to TASK_RUNNING.
1307 * This should be ok, provided it doesn't flip the state back to
1308 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1309 * will only happen if the mutex_lock() call blocks, and we then find
1310 * the ringbuffer empty. So in practice we should be ok, but it's
1311 * something to be aware of when touching this code.
1314 aio_read_events(ctx, min_nr, nr, event, &ret);
1316 wait_event_interruptible_hrtimeout(ctx->wait,
1317 aio_read_events(ctx, min_nr, nr, event, &ret),
1323 * Create an aio_context capable of receiving at least nr_events.
1324 * ctxp must not point to an aio_context that already exists, and
1325 * must be initialized to 0 prior to the call. On successful
1326 * creation of the aio_context, *ctxp is filled in with the resulting
1327 * handle. May fail with -EINVAL if *ctxp is not initialized,
1328 * if the specified nr_events exceeds internal limits. May fail
1329 * with -EAGAIN if the specified nr_events exceeds the user's limit
1330 * of available events. May fail with -ENOMEM if insufficient kernel
1331 * resources are available. May fail with -EFAULT if an invalid
1332 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1335 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1337 struct kioctx *ioctx = NULL;
1341 ret = get_user(ctx, ctxp);
1346 if (unlikely(ctx || nr_events == 0)) {
1347 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1352 ioctx = ioctx_alloc(nr_events);
1353 ret = PTR_ERR(ioctx);
1354 if (!IS_ERR(ioctx)) {
1355 ret = put_user(ioctx->user_id, ctxp);
1357 kill_ioctx(current->mm, ioctx, NULL);
1358 percpu_ref_put(&ioctx->users);
1365 #ifdef CONFIG_COMPAT
1366 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1368 struct kioctx *ioctx = NULL;
1372 ret = get_user(ctx, ctx32p);
1377 if (unlikely(ctx || nr_events == 0)) {
1378 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1383 ioctx = ioctx_alloc(nr_events);
1384 ret = PTR_ERR(ioctx);
1385 if (!IS_ERR(ioctx)) {
1386 /* truncating is ok because it's a user address */
1387 ret = put_user((u32)ioctx->user_id, ctx32p);
1389 kill_ioctx(current->mm, ioctx, NULL);
1390 percpu_ref_put(&ioctx->users);
1399 * Destroy the aio_context specified. May cancel any outstanding
1400 * AIOs and block on completion. Will fail with -ENOSYS if not
1401 * implemented. May fail with -EINVAL if the context pointed to
1404 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1406 struct kioctx *ioctx = lookup_ioctx(ctx);
1407 if (likely(NULL != ioctx)) {
1408 struct ctx_rq_wait wait;
1411 init_completion(&wait.comp);
1412 atomic_set(&wait.count, 1);
1414 /* Pass requests_done to kill_ioctx() where it can be set
1415 * in a thread-safe way. If we try to set it here then we have
1416 * a race condition if two io_destroy() called simultaneously.
1418 ret = kill_ioctx(current->mm, ioctx, &wait);
1419 percpu_ref_put(&ioctx->users);
1421 /* Wait until all IO for the context are done. Otherwise kernel
1422 * keep using user-space buffers even if user thinks the context
1426 wait_for_completion(&wait.comp);
1430 pr_debug("EINVAL: invalid context id\n");
1434 static void aio_remove_iocb(struct aio_kiocb *iocb)
1436 struct kioctx *ctx = iocb->ki_ctx;
1437 unsigned long flags;
1439 spin_lock_irqsave(&ctx->ctx_lock, flags);
1440 list_del(&iocb->ki_list);
1441 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1444 static void aio_complete_rw(struct kiocb *kiocb, long res)
1446 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1448 if (!list_empty_careful(&iocb->ki_list))
1449 aio_remove_iocb(iocb);
1451 if (kiocb->ki_flags & IOCB_WRITE) {
1452 struct inode *inode = file_inode(kiocb->ki_filp);
1455 * Tell lockdep we inherited freeze protection from submission
1458 if (S_ISREG(inode->i_mode))
1459 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1460 file_end_write(kiocb->ki_filp);
1463 iocb->ki_res.res = res;
1464 iocb->ki_res.res2 = 0;
1468 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1472 req->ki_complete = aio_complete_rw;
1473 req->private = NULL;
1474 req->ki_pos = iocb->aio_offset;
1475 req->ki_flags = req->ki_filp->f_iocb_flags;
1476 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1477 req->ki_flags |= IOCB_EVENTFD;
1478 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1480 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1481 * aio_reqprio is interpreted as an I/O scheduling
1482 * class and priority.
1484 ret = ioprio_check_cap(iocb->aio_reqprio);
1486 pr_debug("aio ioprio check cap error: %d\n", ret);
1490 req->ki_ioprio = iocb->aio_reqprio;
1492 req->ki_ioprio = get_current_ioprio();
1494 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1498 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1502 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1503 struct iovec **iovec, bool vectored, bool compat,
1504 struct iov_iter *iter)
1506 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1507 size_t len = iocb->aio_nbytes;
1510 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1515 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1518 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1524 case -ERESTARTNOINTR:
1525 case -ERESTARTNOHAND:
1526 case -ERESTART_RESTARTBLOCK:
1528 * There's no easy way to restart the syscall since other AIO's
1529 * may be already running. Just fail this IO with EINTR.
1534 req->ki_complete(req, ret);
1538 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1539 bool vectored, bool compat)
1541 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1542 struct iov_iter iter;
1546 ret = aio_prep_rw(req, iocb);
1549 file = req->ki_filp;
1550 if (unlikely(!(file->f_mode & FMODE_READ)))
1552 if (unlikely(!file->f_op->read_iter))
1555 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1558 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1560 aio_rw_done(req, call_read_iter(file, req, &iter));
1565 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1566 bool vectored, bool compat)
1568 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1569 struct iov_iter iter;
1573 ret = aio_prep_rw(req, iocb);
1576 file = req->ki_filp;
1578 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1580 if (unlikely(!file->f_op->write_iter))
1583 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1586 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1589 * Open-code file_start_write here to grab freeze protection,
1590 * which will be released by another thread in
1591 * aio_complete_rw(). Fool lockdep by telling it the lock got
1592 * released so that it doesn't complain about the held lock when
1593 * we return to userspace.
1595 if (S_ISREG(file_inode(file)->i_mode)) {
1596 sb_start_write(file_inode(file)->i_sb);
1597 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1599 req->ki_flags |= IOCB_WRITE;
1600 aio_rw_done(req, call_write_iter(file, req, &iter));
1606 static void aio_fsync_work(struct work_struct *work)
1608 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1609 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1611 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1612 revert_creds(old_cred);
1613 put_cred(iocb->fsync.creds);
1617 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1620 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1621 iocb->aio_rw_flags))
1624 if (unlikely(!req->file->f_op->fsync))
1627 req->creds = prepare_creds();
1631 req->datasync = datasync;
1632 INIT_WORK(&req->work, aio_fsync_work);
1633 schedule_work(&req->work);
1637 static void aio_poll_put_work(struct work_struct *work)
1639 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1640 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1646 * Safely lock the waitqueue which the request is on, synchronizing with the
1647 * case where the ->poll() provider decides to free its waitqueue early.
1649 * Returns true on success, meaning that req->head->lock was locked, req->wait
1650 * is on req->head, and an RCU read lock was taken. Returns false if the
1651 * request was already removed from its waitqueue (which might no longer exist).
1653 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1655 wait_queue_head_t *head;
1658 * While we hold the waitqueue lock and the waitqueue is nonempty,
1659 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1660 * lock in the first place can race with the waitqueue being freed.
1662 * We solve this as eventpoll does: by taking advantage of the fact that
1663 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1664 * we enter rcu_read_lock() and see that the pointer to the queue is
1665 * non-NULL, we can then lock it without the memory being freed out from
1666 * under us, then check whether the request is still on the queue.
1668 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1669 * case the caller deletes the entry from the queue, leaving it empty.
1670 * In that case, only RCU prevents the queue memory from being freed.
1673 head = smp_load_acquire(&req->head);
1675 spin_lock(&head->lock);
1676 if (!list_empty(&req->wait.entry))
1678 spin_unlock(&head->lock);
1684 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1686 spin_unlock(&req->head->lock);
1690 static void aio_poll_complete_work(struct work_struct *work)
1692 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1693 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1694 struct poll_table_struct pt = { ._key = req->events };
1695 struct kioctx *ctx = iocb->ki_ctx;
1698 if (!READ_ONCE(req->cancelled))
1699 mask = vfs_poll(req->file, &pt) & req->events;
1702 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1703 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1704 * synchronize with them. In the cancellation case the list_del_init
1705 * itself is not actually needed, but harmless so we keep it in to
1706 * avoid further branches in the fast path.
1708 spin_lock_irq(&ctx->ctx_lock);
1709 if (poll_iocb_lock_wq(req)) {
1710 if (!mask && !READ_ONCE(req->cancelled)) {
1712 * The request isn't actually ready to be completed yet.
1713 * Reschedule completion if another wakeup came in.
1715 if (req->work_need_resched) {
1716 schedule_work(&req->work);
1717 req->work_need_resched = false;
1719 req->work_scheduled = false;
1721 poll_iocb_unlock_wq(req);
1722 spin_unlock_irq(&ctx->ctx_lock);
1725 list_del_init(&req->wait.entry);
1726 poll_iocb_unlock_wq(req);
1727 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1728 list_del_init(&iocb->ki_list);
1729 iocb->ki_res.res = mangle_poll(mask);
1730 spin_unlock_irq(&ctx->ctx_lock);
1735 /* assumes we are called with irqs disabled */
1736 static int aio_poll_cancel(struct kiocb *iocb)
1738 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1739 struct poll_iocb *req = &aiocb->poll;
1741 if (poll_iocb_lock_wq(req)) {
1742 WRITE_ONCE(req->cancelled, true);
1743 if (!req->work_scheduled) {
1744 schedule_work(&aiocb->poll.work);
1745 req->work_scheduled = true;
1747 poll_iocb_unlock_wq(req);
1748 } /* else, the request was force-cancelled by POLLFREE already */
1753 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1756 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1757 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1758 __poll_t mask = key_to_poll(key);
1759 unsigned long flags;
1761 /* for instances that support it check for an event match first: */
1762 if (mask && !(mask & req->events))
1766 * Complete the request inline if possible. This requires that three
1767 * conditions be met:
1768 * 1. An event mask must have been passed. If a plain wakeup was done
1769 * instead, then mask == 0 and we have to call vfs_poll() to get
1770 * the events, so inline completion isn't possible.
1771 * 2. The completion work must not have already been scheduled.
1772 * 3. ctx_lock must not be busy. We have to use trylock because we
1773 * already hold the waitqueue lock, so this inverts the normal
1774 * locking order. Use irqsave/irqrestore because not all
1775 * filesystems (e.g. fuse) call this function with IRQs disabled,
1776 * yet IRQs have to be disabled before ctx_lock is obtained.
1778 if (mask && !req->work_scheduled &&
1779 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1780 struct kioctx *ctx = iocb->ki_ctx;
1782 list_del_init(&req->wait.entry);
1783 list_del(&iocb->ki_list);
1784 iocb->ki_res.res = mangle_poll(mask);
1785 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1787 INIT_WORK(&req->work, aio_poll_put_work);
1788 schedule_work(&req->work);
1790 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1795 * Schedule the completion work if needed. If it was already
1796 * scheduled, record that another wakeup came in.
1798 * Don't remove the request from the waitqueue here, as it might
1799 * not actually be complete yet (we won't know until vfs_poll()
1800 * is called), and we must not miss any wakeups. POLLFREE is an
1801 * exception to this; see below.
1803 if (req->work_scheduled) {
1804 req->work_need_resched = true;
1806 schedule_work(&req->work);
1807 req->work_scheduled = true;
1811 * If the waitqueue is being freed early but we can't complete
1812 * the request inline, we have to tear down the request as best
1813 * we can. That means immediately removing the request from its
1814 * waitqueue and preventing all further accesses to the
1815 * waitqueue via the request. We also need to schedule the
1816 * completion work (done above). Also mark the request as
1817 * cancelled, to potentially skip an unneeded call to ->poll().
1819 if (mask & POLLFREE) {
1820 WRITE_ONCE(req->cancelled, true);
1821 list_del_init(&req->wait.entry);
1824 * Careful: this *must* be the last step, since as soon
1825 * as req->head is NULL'ed out, the request can be
1826 * completed and freed, since aio_poll_complete_work()
1827 * will no longer need to take the waitqueue lock.
1829 smp_store_release(&req->head, NULL);
1835 struct aio_poll_table {
1836 struct poll_table_struct pt;
1837 struct aio_kiocb *iocb;
1843 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1844 struct poll_table_struct *p)
1846 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1848 /* multiple wait queues per file are not supported */
1849 if (unlikely(pt->queued)) {
1850 pt->error = -EINVAL;
1856 pt->iocb->poll.head = head;
1857 add_wait_queue(head, &pt->iocb->poll.wait);
1860 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1862 struct kioctx *ctx = aiocb->ki_ctx;
1863 struct poll_iocb *req = &aiocb->poll;
1864 struct aio_poll_table apt;
1865 bool cancel = false;
1868 /* reject any unknown events outside the normal event mask. */
1869 if ((u16)iocb->aio_buf != iocb->aio_buf)
1871 /* reject fields that are not defined for poll */
1872 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1875 INIT_WORK(&req->work, aio_poll_complete_work);
1876 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1879 req->cancelled = false;
1880 req->work_scheduled = false;
1881 req->work_need_resched = false;
1883 apt.pt._qproc = aio_poll_queue_proc;
1884 apt.pt._key = req->events;
1887 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1889 /* initialized the list so that we can do list_empty checks */
1890 INIT_LIST_HEAD(&req->wait.entry);
1891 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1893 mask = vfs_poll(req->file, &apt.pt) & req->events;
1894 spin_lock_irq(&ctx->ctx_lock);
1895 if (likely(apt.queued)) {
1896 bool on_queue = poll_iocb_lock_wq(req);
1898 if (!on_queue || req->work_scheduled) {
1900 * aio_poll_wake() already either scheduled the async
1901 * completion work, or completed the request inline.
1903 if (apt.error) /* unsupported case: multiple queues */
1908 if (mask || apt.error) {
1909 /* Steal to complete synchronously. */
1910 list_del_init(&req->wait.entry);
1911 } else if (cancel) {
1912 /* Cancel if possible (may be too late though). */
1913 WRITE_ONCE(req->cancelled, true);
1914 } else if (on_queue) {
1916 * Actually waiting for an event, so add the request to
1917 * active_reqs so that it can be cancelled if needed.
1919 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1920 aiocb->ki_cancel = aio_poll_cancel;
1923 poll_iocb_unlock_wq(req);
1925 if (mask) { /* no async, we'd stolen it */
1926 aiocb->ki_res.res = mangle_poll(mask);
1929 spin_unlock_irq(&ctx->ctx_lock);
1935 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1936 struct iocb __user *user_iocb, struct aio_kiocb *req,
1939 req->ki_filp = fget(iocb->aio_fildes);
1940 if (unlikely(!req->ki_filp))
1943 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1944 struct eventfd_ctx *eventfd;
1946 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1947 * instance of the file* now. The file descriptor must be
1948 * an eventfd() fd, and will be signaled for each completed
1949 * event using the eventfd_signal() function.
1951 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1952 if (IS_ERR(eventfd))
1953 return PTR_ERR(eventfd);
1955 req->ki_eventfd = eventfd;
1958 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1959 pr_debug("EFAULT: aio_key\n");
1963 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1964 req->ki_res.data = iocb->aio_data;
1965 req->ki_res.res = 0;
1966 req->ki_res.res2 = 0;
1968 switch (iocb->aio_lio_opcode) {
1969 case IOCB_CMD_PREAD:
1970 return aio_read(&req->rw, iocb, false, compat);
1971 case IOCB_CMD_PWRITE:
1972 return aio_write(&req->rw, iocb, false, compat);
1973 case IOCB_CMD_PREADV:
1974 return aio_read(&req->rw, iocb, true, compat);
1975 case IOCB_CMD_PWRITEV:
1976 return aio_write(&req->rw, iocb, true, compat);
1977 case IOCB_CMD_FSYNC:
1978 return aio_fsync(&req->fsync, iocb, false);
1979 case IOCB_CMD_FDSYNC:
1980 return aio_fsync(&req->fsync, iocb, true);
1982 return aio_poll(req, iocb);
1984 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1989 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1992 struct aio_kiocb *req;
1996 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1999 /* enforce forwards compatibility on users */
2000 if (unlikely(iocb.aio_reserved2)) {
2001 pr_debug("EINVAL: reserve field set\n");
2005 /* prevent overflows */
2007 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2008 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2009 ((ssize_t)iocb.aio_nbytes < 0)
2011 pr_debug("EINVAL: overflow check\n");
2015 req = aio_get_req(ctx);
2019 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2021 /* Done with the synchronous reference */
2025 * If err is 0, we'd either done aio_complete() ourselves or have
2026 * arranged for that to be done asynchronously. Anything non-zero
2027 * means that we need to destroy req ourselves.
2029 if (unlikely(err)) {
2031 put_reqs_available(ctx, 1);
2037 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2038 * the number of iocbs queued. May return -EINVAL if the aio_context
2039 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2040 * *iocbpp[0] is not properly initialized, if the operation specified
2041 * is invalid for the file descriptor in the iocb. May fail with
2042 * -EFAULT if any of the data structures point to invalid data. May
2043 * fail with -EBADF if the file descriptor specified in the first
2044 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2045 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2046 * fail with -ENOSYS if not implemented.
2048 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2049 struct iocb __user * __user *, iocbpp)
2054 struct blk_plug plug;
2056 if (unlikely(nr < 0))
2059 ctx = lookup_ioctx(ctx_id);
2060 if (unlikely(!ctx)) {
2061 pr_debug("EINVAL: invalid context id\n");
2065 if (nr > ctx->nr_events)
2066 nr = ctx->nr_events;
2068 if (nr > AIO_PLUG_THRESHOLD)
2069 blk_start_plug(&plug);
2070 for (i = 0; i < nr; i++) {
2071 struct iocb __user *user_iocb;
2073 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2078 ret = io_submit_one(ctx, user_iocb, false);
2082 if (nr > AIO_PLUG_THRESHOLD)
2083 blk_finish_plug(&plug);
2085 percpu_ref_put(&ctx->users);
2089 #ifdef CONFIG_COMPAT
2090 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2091 int, nr, compat_uptr_t __user *, iocbpp)
2096 struct blk_plug plug;
2098 if (unlikely(nr < 0))
2101 ctx = lookup_ioctx(ctx_id);
2102 if (unlikely(!ctx)) {
2103 pr_debug("EINVAL: invalid context id\n");
2107 if (nr > ctx->nr_events)
2108 nr = ctx->nr_events;
2110 if (nr > AIO_PLUG_THRESHOLD)
2111 blk_start_plug(&plug);
2112 for (i = 0; i < nr; i++) {
2113 compat_uptr_t user_iocb;
2115 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2120 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2124 if (nr > AIO_PLUG_THRESHOLD)
2125 blk_finish_plug(&plug);
2127 percpu_ref_put(&ctx->users);
2133 * Attempts to cancel an iocb previously passed to io_submit. If
2134 * the operation is successfully cancelled, the resulting event is
2135 * copied into the memory pointed to by result without being placed
2136 * into the completion queue and 0 is returned. May fail with
2137 * -EFAULT if any of the data structures pointed to are invalid.
2138 * May fail with -EINVAL if aio_context specified by ctx_id is
2139 * invalid. May fail with -EAGAIN if the iocb specified was not
2140 * cancelled. Will fail with -ENOSYS if not implemented.
2142 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2143 struct io_event __user *, result)
2146 struct aio_kiocb *kiocb;
2149 u64 obj = (u64)(unsigned long)iocb;
2151 if (unlikely(get_user(key, &iocb->aio_key)))
2153 if (unlikely(key != KIOCB_KEY))
2156 ctx = lookup_ioctx(ctx_id);
2160 spin_lock_irq(&ctx->ctx_lock);
2161 /* TODO: use a hash or array, this sucks. */
2162 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2163 if (kiocb->ki_res.obj == obj) {
2164 ret = kiocb->ki_cancel(&kiocb->rw);
2165 list_del_init(&kiocb->ki_list);
2169 spin_unlock_irq(&ctx->ctx_lock);
2173 * The result argument is no longer used - the io_event is
2174 * always delivered via the ring buffer. -EINPROGRESS indicates
2175 * cancellation is progress:
2180 percpu_ref_put(&ctx->users);
2185 static long do_io_getevents(aio_context_t ctx_id,
2188 struct io_event __user *events,
2189 struct timespec64 *ts)
2191 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2192 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2195 if (likely(ioctx)) {
2196 if (likely(min_nr <= nr && min_nr >= 0))
2197 ret = read_events(ioctx, min_nr, nr, events, until);
2198 percpu_ref_put(&ioctx->users);
2205 * Attempts to read at least min_nr events and up to nr events from
2206 * the completion queue for the aio_context specified by ctx_id. If
2207 * it succeeds, the number of read events is returned. May fail with
2208 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2209 * out of range, if timeout is out of range. May fail with -EFAULT
2210 * if any of the memory specified is invalid. May return 0 or
2211 * < min_nr if the timeout specified by timeout has elapsed
2212 * before sufficient events are available, where timeout == NULL
2213 * specifies an infinite timeout. Note that the timeout pointed to by
2214 * timeout is relative. Will fail with -ENOSYS if not implemented.
2218 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2221 struct io_event __user *, events,
2222 struct __kernel_timespec __user *, timeout)
2224 struct timespec64 ts;
2227 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2230 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2231 if (!ret && signal_pending(current))
2238 struct __aio_sigset {
2239 const sigset_t __user *sigmask;
2243 SYSCALL_DEFINE6(io_pgetevents,
2244 aio_context_t, ctx_id,
2247 struct io_event __user *, events,
2248 struct __kernel_timespec __user *, timeout,
2249 const struct __aio_sigset __user *, usig)
2251 struct __aio_sigset ksig = { NULL, };
2252 struct timespec64 ts;
2256 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2259 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2262 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2266 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2268 interrupted = signal_pending(current);
2269 restore_saved_sigmask_unless(interrupted);
2270 if (interrupted && !ret)
2271 ret = -ERESTARTNOHAND;
2276 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2278 SYSCALL_DEFINE6(io_pgetevents_time32,
2279 aio_context_t, ctx_id,
2282 struct io_event __user *, events,
2283 struct old_timespec32 __user *, timeout,
2284 const struct __aio_sigset __user *, usig)
2286 struct __aio_sigset ksig = { NULL, };
2287 struct timespec64 ts;
2291 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2294 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2298 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2302 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2304 interrupted = signal_pending(current);
2305 restore_saved_sigmask_unless(interrupted);
2306 if (interrupted && !ret)
2307 ret = -ERESTARTNOHAND;
2314 #if defined(CONFIG_COMPAT_32BIT_TIME)
2316 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2319 struct io_event __user *, events,
2320 struct old_timespec32 __user *, timeout)
2322 struct timespec64 t;
2325 if (timeout && get_old_timespec32(&t, timeout))
2328 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2329 if (!ret && signal_pending(current))
2336 #ifdef CONFIG_COMPAT
2338 struct __compat_aio_sigset {
2339 compat_uptr_t sigmask;
2340 compat_size_t sigsetsize;
2343 #if defined(CONFIG_COMPAT_32BIT_TIME)
2345 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2346 compat_aio_context_t, ctx_id,
2347 compat_long_t, min_nr,
2349 struct io_event __user *, events,
2350 struct old_timespec32 __user *, timeout,
2351 const struct __compat_aio_sigset __user *, usig)
2353 struct __compat_aio_sigset ksig = { 0, };
2354 struct timespec64 t;
2358 if (timeout && get_old_timespec32(&t, timeout))
2361 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2364 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2368 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2370 interrupted = signal_pending(current);
2371 restore_saved_sigmask_unless(interrupted);
2372 if (interrupted && !ret)
2373 ret = -ERESTARTNOHAND;
2380 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2381 compat_aio_context_t, ctx_id,
2382 compat_long_t, min_nr,
2384 struct io_event __user *, events,
2385 struct __kernel_timespec __user *, timeout,
2386 const struct __compat_aio_sigset __user *, usig)
2388 struct __compat_aio_sigset ksig = { 0, };
2389 struct timespec64 t;
2393 if (timeout && get_timespec64(&t, timeout))
2396 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2399 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2403 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2405 interrupted = signal_pending(current);
2406 restore_saved_sigmask_unless(interrupted);
2407 if (interrupted && !ret)
2408 ret = -ERESTARTNOHAND;