2 * An async IO implementation for Linux
3 * Written by Benjamin LaHaise <bcrl@kvack.org>
5 * Implements an efficient asynchronous io interface.
7 * Copyright 2000, 2001, 2002 Red Hat, Inc. All Rights Reserved.
8 * Copyright 2018 Christoph Hellwig.
10 * See ../COPYING for licensing terms.
12 #define pr_fmt(fmt) "%s: " fmt, __func__
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
25 #include <linux/sched/signal.h>
27 #include <linux/file.h>
29 #include <linux/mman.h>
30 #include <linux/mmu_context.h>
31 #include <linux/percpu.h>
32 #include <linux/slab.h>
33 #include <linux/timer.h>
34 #include <linux/aio.h>
35 #include <linux/highmem.h>
36 #include <linux/workqueue.h>
37 #include <linux/security.h>
38 #include <linux/eventfd.h>
39 #include <linux/blkdev.h>
40 #include <linux/compat.h>
41 #include <linux/migrate.h>
42 #include <linux/ramfs.h>
43 #include <linux/percpu-refcount.h>
44 #include <linux/mount.h>
46 #include <asm/kmap_types.h>
47 #include <linux/uaccess.h>
48 #include <linux/nospec.h>
54 #define AIO_RING_MAGIC 0xa10a10a1
55 #define AIO_RING_COMPAT_FEATURES 1
56 #define AIO_RING_INCOMPAT_FEATURES 0
58 unsigned id; /* kernel internal index number */
59 unsigned nr; /* number of io_events */
60 unsigned head; /* Written to by userland or under ring_lock
61 * mutex by aio_read_events_ring(). */
65 unsigned compat_features;
66 unsigned incompat_features;
67 unsigned header_length; /* size of aio_ring */
70 struct io_event io_events[0];
71 }; /* 128 bytes + ring size */
74 * Plugging is meant to work with larger batches of IOs. If we don't
75 * have more than the below, then don't bother setting up a plug.
77 #define AIO_PLUG_THRESHOLD 2
79 #define AIO_RING_PAGES 8
84 struct kioctx __rcu *table[];
88 unsigned reqs_available;
92 struct completion comp;
97 struct percpu_ref users;
100 struct percpu_ref reqs;
102 unsigned long user_id;
104 struct __percpu kioctx_cpu *cpu;
107 * For percpu reqs_available, number of slots we move to/from global
112 * This is what userspace passed to io_setup(), it's not used for
113 * anything but counting against the global max_reqs quota.
115 * The real limit is nr_events - 1, which will be larger (see
120 /* Size of ringbuffer, in units of struct io_event */
123 unsigned long mmap_base;
124 unsigned long mmap_size;
126 struct page **ring_pages;
129 struct rcu_work free_rwork; /* see free_ioctx() */
132 * signals when all in-flight requests are done
134 struct ctx_rq_wait *rq_wait;
138 * This counts the number of available slots in the ringbuffer,
139 * so we avoid overflowing it: it's decremented (if positive)
140 * when allocating a kiocb and incremented when the resulting
141 * io_event is pulled off the ringbuffer.
143 * We batch accesses to it with a percpu version.
145 atomic_t reqs_available;
146 } ____cacheline_aligned_in_smp;
150 struct list_head active_reqs; /* used for cancellation */
151 } ____cacheline_aligned_in_smp;
154 struct mutex ring_lock;
155 wait_queue_head_t wait;
156 } ____cacheline_aligned_in_smp;
160 unsigned completed_events;
161 spinlock_t completion_lock;
162 } ____cacheline_aligned_in_smp;
164 struct page *internal_pages[AIO_RING_PAGES];
165 struct file *aio_ring_file;
171 * First field must be the file pointer in all the
172 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
176 struct work_struct work;
182 struct wait_queue_head *head;
186 struct wait_queue_entry wait;
187 struct work_struct work;
191 * NOTE! Each of the iocb union members has the file pointer
192 * as the first entry in their struct definition. So you can
193 * access the file pointer through any of the sub-structs,
194 * or directly as just 'ki_filp' in this struct.
198 struct file *ki_filp;
200 struct fsync_iocb fsync;
201 struct poll_iocb poll;
204 struct kioctx *ki_ctx;
205 kiocb_cancel_fn *ki_cancel;
207 struct io_event ki_res;
209 struct list_head ki_list; /* the aio core uses this
210 * for cancellation */
211 refcount_t ki_refcnt;
214 * If the aio_resfd field of the userspace iocb is not zero,
215 * this is the underlying eventfd context to deliver events to.
217 struct eventfd_ctx *ki_eventfd;
220 /*------ sysctl variables----*/
221 static DEFINE_SPINLOCK(aio_nr_lock);
222 unsigned long aio_nr; /* current system wide number of aio requests */
223 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
224 /*----end sysctl variables---*/
226 static struct kmem_cache *kiocb_cachep;
227 static struct kmem_cache *kioctx_cachep;
229 static struct vfsmount *aio_mnt;
231 static const struct file_operations aio_ring_fops;
232 static const struct address_space_operations aio_ctx_aops;
234 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
237 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
239 return ERR_CAST(inode);
241 inode->i_mapping->a_ops = &aio_ctx_aops;
242 inode->i_mapping->private_data = ctx;
243 inode->i_size = PAGE_SIZE * nr_pages;
245 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
246 O_RDWR, &aio_ring_fops);
252 static struct dentry *aio_mount(struct file_system_type *fs_type,
253 int flags, const char *dev_name, void *data)
255 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
259 root->d_sb->s_iflags |= SB_I_NOEXEC;
264 * Creates the slab caches used by the aio routines, panic on
265 * failure as this is done early during the boot sequence.
267 static int __init aio_setup(void)
269 static struct file_system_type aio_fs = {
272 .kill_sb = kill_anon_super,
274 aio_mnt = kern_mount(&aio_fs);
276 panic("Failed to create aio fs mount.");
278 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
279 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
282 __initcall(aio_setup);
284 static void put_aio_ring_file(struct kioctx *ctx)
286 struct file *aio_ring_file = ctx->aio_ring_file;
287 struct address_space *i_mapping;
290 truncate_setsize(file_inode(aio_ring_file), 0);
292 /* Prevent further access to the kioctx from migratepages */
293 i_mapping = aio_ring_file->f_mapping;
294 spin_lock(&i_mapping->private_lock);
295 i_mapping->private_data = NULL;
296 ctx->aio_ring_file = NULL;
297 spin_unlock(&i_mapping->private_lock);
303 static void aio_free_ring(struct kioctx *ctx)
307 /* Disconnect the kiotx from the ring file. This prevents future
308 * accesses to the kioctx from page migration.
310 put_aio_ring_file(ctx);
312 for (i = 0; i < ctx->nr_pages; i++) {
314 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
315 page_count(ctx->ring_pages[i]));
316 page = ctx->ring_pages[i];
319 ctx->ring_pages[i] = NULL;
323 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
324 kfree(ctx->ring_pages);
325 ctx->ring_pages = NULL;
329 static int aio_ring_mremap(struct vm_area_struct *vma)
331 struct file *file = vma->vm_file;
332 struct mm_struct *mm = vma->vm_mm;
333 struct kioctx_table *table;
334 int i, res = -EINVAL;
336 spin_lock(&mm->ioctx_lock);
338 table = rcu_dereference(mm->ioctx_table);
339 for (i = 0; i < table->nr; i++) {
342 ctx = rcu_dereference(table->table[i]);
343 if (ctx && ctx->aio_ring_file == file) {
344 if (!atomic_read(&ctx->dead)) {
345 ctx->user_id = ctx->mmap_base = vma->vm_start;
353 spin_unlock(&mm->ioctx_lock);
357 static const struct vm_operations_struct aio_ring_vm_ops = {
358 .mremap = aio_ring_mremap,
359 #if IS_ENABLED(CONFIG_MMU)
360 .fault = filemap_fault,
361 .map_pages = filemap_map_pages,
362 .page_mkwrite = filemap_page_mkwrite,
366 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
368 vma->vm_flags |= VM_DONTEXPAND;
369 vma->vm_ops = &aio_ring_vm_ops;
373 static const struct file_operations aio_ring_fops = {
374 .mmap = aio_ring_mmap,
377 #if IS_ENABLED(CONFIG_MIGRATION)
378 static int aio_migratepage(struct address_space *mapping, struct page *new,
379 struct page *old, enum migrate_mode mode)
387 * We cannot support the _NO_COPY case here, because copy needs to
388 * happen under the ctx->completion_lock. That does not work with the
389 * migration workflow of MIGRATE_SYNC_NO_COPY.
391 if (mode == MIGRATE_SYNC_NO_COPY)
396 /* mapping->private_lock here protects against the kioctx teardown. */
397 spin_lock(&mapping->private_lock);
398 ctx = mapping->private_data;
404 /* The ring_lock mutex. The prevents aio_read_events() from writing
405 * to the ring's head, and prevents page migration from mucking in
406 * a partially initialized kiotx.
408 if (!mutex_trylock(&ctx->ring_lock)) {
414 if (idx < (pgoff_t)ctx->nr_pages) {
415 /* Make sure the old page hasn't already been changed */
416 if (ctx->ring_pages[idx] != old)
424 /* Writeback must be complete */
425 BUG_ON(PageWriteback(old));
428 rc = migrate_page_move_mapping(mapping, new, old, mode, 1);
429 if (rc != MIGRATEPAGE_SUCCESS) {
434 /* Take completion_lock to prevent other writes to the ring buffer
435 * while the old page is copied to the new. This prevents new
436 * events from being lost.
438 spin_lock_irqsave(&ctx->completion_lock, flags);
439 migrate_page_copy(new, old);
440 BUG_ON(ctx->ring_pages[idx] != old);
441 ctx->ring_pages[idx] = new;
442 spin_unlock_irqrestore(&ctx->completion_lock, flags);
444 /* The old page is no longer accessible. */
448 mutex_unlock(&ctx->ring_lock);
450 spin_unlock(&mapping->private_lock);
455 static const struct address_space_operations aio_ctx_aops = {
456 .set_page_dirty = __set_page_dirty_no_writeback,
457 #if IS_ENABLED(CONFIG_MIGRATION)
458 .migratepage = aio_migratepage,
462 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
464 struct aio_ring *ring;
465 struct mm_struct *mm = current->mm;
466 unsigned long size, unused;
471 /* Compensate for the ring buffer's head/tail overlap entry */
472 nr_events += 2; /* 1 is required, 2 for good luck */
474 size = sizeof(struct aio_ring);
475 size += sizeof(struct io_event) * nr_events;
477 nr_pages = PFN_UP(size);
481 file = aio_private_file(ctx, nr_pages);
483 ctx->aio_ring_file = NULL;
487 ctx->aio_ring_file = file;
488 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
489 / sizeof(struct io_event);
491 ctx->ring_pages = ctx->internal_pages;
492 if (nr_pages > AIO_RING_PAGES) {
493 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
495 if (!ctx->ring_pages) {
496 put_aio_ring_file(ctx);
501 for (i = 0; i < nr_pages; i++) {
503 page = find_or_create_page(file->f_mapping,
504 i, GFP_HIGHUSER | __GFP_ZERO);
507 pr_debug("pid(%d) page[%d]->count=%d\n",
508 current->pid, i, page_count(page));
509 SetPageUptodate(page);
512 ctx->ring_pages[i] = page;
516 if (unlikely(i != nr_pages)) {
521 ctx->mmap_size = nr_pages * PAGE_SIZE;
522 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
524 if (down_write_killable(&mm->mmap_sem)) {
530 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
531 PROT_READ | PROT_WRITE,
532 MAP_SHARED, 0, &unused, NULL);
533 up_write(&mm->mmap_sem);
534 if (IS_ERR((void *)ctx->mmap_base)) {
540 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
542 ctx->user_id = ctx->mmap_base;
543 ctx->nr_events = nr_events; /* trusted copy */
545 ring = kmap_atomic(ctx->ring_pages[0]);
546 ring->nr = nr_events; /* user copy */
548 ring->head = ring->tail = 0;
549 ring->magic = AIO_RING_MAGIC;
550 ring->compat_features = AIO_RING_COMPAT_FEATURES;
551 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
552 ring->header_length = sizeof(struct aio_ring);
554 flush_dcache_page(ctx->ring_pages[0]);
559 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
560 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
561 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
563 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
565 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
566 struct kioctx *ctx = req->ki_ctx;
569 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
572 spin_lock_irqsave(&ctx->ctx_lock, flags);
573 list_add_tail(&req->ki_list, &ctx->active_reqs);
574 req->ki_cancel = cancel;
575 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
577 EXPORT_SYMBOL(kiocb_set_cancel_fn);
580 * free_ioctx() should be RCU delayed to synchronize against the RCU
581 * protected lookup_ioctx() and also needs process context to call
582 * aio_free_ring(). Use rcu_work.
584 static void free_ioctx(struct work_struct *work)
586 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
588 pr_debug("freeing %p\n", ctx);
591 free_percpu(ctx->cpu);
592 percpu_ref_exit(&ctx->reqs);
593 percpu_ref_exit(&ctx->users);
594 kmem_cache_free(kioctx_cachep, ctx);
597 static void free_ioctx_reqs(struct percpu_ref *ref)
599 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
601 /* At this point we know that there are no any in-flight requests */
602 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
603 complete(&ctx->rq_wait->comp);
605 /* Synchronize against RCU protected table->table[] dereferences */
606 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
607 queue_rcu_work(system_wq, &ctx->free_rwork);
611 * When this function runs, the kioctx has been removed from the "hash table"
612 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
613 * now it's safe to cancel any that need to be.
615 static void free_ioctx_users(struct percpu_ref *ref)
617 struct kioctx *ctx = container_of(ref, struct kioctx, users);
618 struct aio_kiocb *req;
620 spin_lock_irq(&ctx->ctx_lock);
622 while (!list_empty(&ctx->active_reqs)) {
623 req = list_first_entry(&ctx->active_reqs,
624 struct aio_kiocb, ki_list);
625 req->ki_cancel(&req->rw);
626 list_del_init(&req->ki_list);
629 spin_unlock_irq(&ctx->ctx_lock);
631 percpu_ref_kill(&ctx->reqs);
632 percpu_ref_put(&ctx->reqs);
635 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
638 struct kioctx_table *table, *old;
639 struct aio_ring *ring;
641 spin_lock(&mm->ioctx_lock);
642 table = rcu_dereference_raw(mm->ioctx_table);
646 for (i = 0; i < table->nr; i++)
647 if (!rcu_access_pointer(table->table[i])) {
649 rcu_assign_pointer(table->table[i], ctx);
650 spin_unlock(&mm->ioctx_lock);
652 /* While kioctx setup is in progress,
653 * we are protected from page migration
654 * changes ring_pages by ->ring_lock.
656 ring = kmap_atomic(ctx->ring_pages[0]);
662 new_nr = (table ? table->nr : 1) * 4;
663 spin_unlock(&mm->ioctx_lock);
665 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
672 spin_lock(&mm->ioctx_lock);
673 old = rcu_dereference_raw(mm->ioctx_table);
676 rcu_assign_pointer(mm->ioctx_table, table);
677 } else if (table->nr > old->nr) {
678 memcpy(table->table, old->table,
679 old->nr * sizeof(struct kioctx *));
681 rcu_assign_pointer(mm->ioctx_table, table);
690 static void aio_nr_sub(unsigned nr)
692 spin_lock(&aio_nr_lock);
693 if (WARN_ON(aio_nr - nr > aio_nr))
697 spin_unlock(&aio_nr_lock);
701 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
703 static struct kioctx *ioctx_alloc(unsigned nr_events)
705 struct mm_struct *mm = current->mm;
710 * Store the original nr_events -- what userspace passed to io_setup(),
711 * for counting against the global limit -- before it changes.
713 unsigned int max_reqs = nr_events;
716 * We keep track of the number of available ringbuffer slots, to prevent
717 * overflow (reqs_available), and we also use percpu counters for this.
719 * So since up to half the slots might be on other cpu's percpu counters
720 * and unavailable, double nr_events so userspace sees what they
721 * expected: additionally, we move req_batch slots to/from percpu
722 * counters at a time, so make sure that isn't 0:
724 nr_events = max(nr_events, num_possible_cpus() * 4);
727 /* Prevent overflows */
728 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
729 pr_debug("ENOMEM: nr_events too high\n");
730 return ERR_PTR(-EINVAL);
733 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
734 return ERR_PTR(-EAGAIN);
736 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
738 return ERR_PTR(-ENOMEM);
740 ctx->max_reqs = max_reqs;
742 spin_lock_init(&ctx->ctx_lock);
743 spin_lock_init(&ctx->completion_lock);
744 mutex_init(&ctx->ring_lock);
745 /* Protect against page migration throughout kiotx setup by keeping
746 * the ring_lock mutex held until setup is complete. */
747 mutex_lock(&ctx->ring_lock);
748 init_waitqueue_head(&ctx->wait);
750 INIT_LIST_HEAD(&ctx->active_reqs);
752 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
755 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
758 ctx->cpu = alloc_percpu(struct kioctx_cpu);
762 err = aio_setup_ring(ctx, nr_events);
766 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
767 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
768 if (ctx->req_batch < 1)
771 /* limit the number of system wide aios */
772 spin_lock(&aio_nr_lock);
773 if (aio_nr + ctx->max_reqs > aio_max_nr ||
774 aio_nr + ctx->max_reqs < aio_nr) {
775 spin_unlock(&aio_nr_lock);
779 aio_nr += ctx->max_reqs;
780 spin_unlock(&aio_nr_lock);
782 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
783 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
785 err = ioctx_add_table(ctx, mm);
789 /* Release the ring_lock mutex now that all setup is complete. */
790 mutex_unlock(&ctx->ring_lock);
792 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
793 ctx, ctx->user_id, mm, ctx->nr_events);
797 aio_nr_sub(ctx->max_reqs);
799 atomic_set(&ctx->dead, 1);
801 vm_munmap(ctx->mmap_base, ctx->mmap_size);
804 mutex_unlock(&ctx->ring_lock);
805 free_percpu(ctx->cpu);
806 percpu_ref_exit(&ctx->reqs);
807 percpu_ref_exit(&ctx->users);
808 kmem_cache_free(kioctx_cachep, ctx);
809 pr_debug("error allocating ioctx %d\n", err);
814 * Cancels all outstanding aio requests on an aio context. Used
815 * when the processes owning a context have all exited to encourage
816 * the rapid destruction of the kioctx.
818 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
819 struct ctx_rq_wait *wait)
821 struct kioctx_table *table;
823 spin_lock(&mm->ioctx_lock);
824 if (atomic_xchg(&ctx->dead, 1)) {
825 spin_unlock(&mm->ioctx_lock);
829 table = rcu_dereference_raw(mm->ioctx_table);
830 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
831 RCU_INIT_POINTER(table->table[ctx->id], NULL);
832 spin_unlock(&mm->ioctx_lock);
834 /* free_ioctx_reqs() will do the necessary RCU synchronization */
835 wake_up_all(&ctx->wait);
838 * It'd be more correct to do this in free_ioctx(), after all
839 * the outstanding kiocbs have finished - but by then io_destroy
840 * has already returned, so io_setup() could potentially return
841 * -EAGAIN with no ioctxs actually in use (as far as userspace
844 aio_nr_sub(ctx->max_reqs);
847 vm_munmap(ctx->mmap_base, ctx->mmap_size);
850 percpu_ref_kill(&ctx->users);
855 * exit_aio: called when the last user of mm goes away. At this point, there is
856 * no way for any new requests to be submited or any of the io_* syscalls to be
857 * called on the context.
859 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
862 void exit_aio(struct mm_struct *mm)
864 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
865 struct ctx_rq_wait wait;
871 atomic_set(&wait.count, table->nr);
872 init_completion(&wait.comp);
875 for (i = 0; i < table->nr; ++i) {
877 rcu_dereference_protected(table->table[i], true);
885 * We don't need to bother with munmap() here - exit_mmap(mm)
886 * is coming and it'll unmap everything. And we simply can't,
887 * this is not necessarily our ->mm.
888 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
889 * that it needs to unmap the area, just set it to 0.
892 kill_ioctx(mm, ctx, &wait);
895 if (!atomic_sub_and_test(skipped, &wait.count)) {
896 /* Wait until all IO for the context are done. */
897 wait_for_completion(&wait.comp);
900 RCU_INIT_POINTER(mm->ioctx_table, NULL);
904 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
906 struct kioctx_cpu *kcpu;
909 local_irq_save(flags);
910 kcpu = this_cpu_ptr(ctx->cpu);
911 kcpu->reqs_available += nr;
913 while (kcpu->reqs_available >= ctx->req_batch * 2) {
914 kcpu->reqs_available -= ctx->req_batch;
915 atomic_add(ctx->req_batch, &ctx->reqs_available);
918 local_irq_restore(flags);
921 static bool __get_reqs_available(struct kioctx *ctx)
923 struct kioctx_cpu *kcpu;
927 local_irq_save(flags);
928 kcpu = this_cpu_ptr(ctx->cpu);
929 if (!kcpu->reqs_available) {
930 int old, avail = atomic_read(&ctx->reqs_available);
933 if (avail < ctx->req_batch)
937 avail = atomic_cmpxchg(&ctx->reqs_available,
938 avail, avail - ctx->req_batch);
939 } while (avail != old);
941 kcpu->reqs_available += ctx->req_batch;
945 kcpu->reqs_available--;
947 local_irq_restore(flags);
951 /* refill_reqs_available
952 * Updates the reqs_available reference counts used for tracking the
953 * number of free slots in the completion ring. This can be called
954 * from aio_complete() (to optimistically update reqs_available) or
955 * from aio_get_req() (the we're out of events case). It must be
956 * called holding ctx->completion_lock.
958 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
961 unsigned events_in_ring, completed;
963 /* Clamp head since userland can write to it. */
964 head %= ctx->nr_events;
966 events_in_ring = tail - head;
968 events_in_ring = ctx->nr_events - (head - tail);
970 completed = ctx->completed_events;
971 if (events_in_ring < completed)
972 completed -= events_in_ring;
979 ctx->completed_events -= completed;
980 put_reqs_available(ctx, completed);
983 /* user_refill_reqs_available
984 * Called to refill reqs_available when aio_get_req() encounters an
985 * out of space in the completion ring.
987 static void user_refill_reqs_available(struct kioctx *ctx)
989 spin_lock_irq(&ctx->completion_lock);
990 if (ctx->completed_events) {
991 struct aio_ring *ring;
994 /* Access of ring->head may race with aio_read_events_ring()
995 * here, but that's okay since whether we read the old version
996 * or the new version, and either will be valid. The important
997 * part is that head cannot pass tail since we prevent
998 * aio_complete() from updating tail by holding
999 * ctx->completion_lock. Even if head is invalid, the check
1000 * against ctx->completed_events below will make sure we do the
1003 ring = kmap_atomic(ctx->ring_pages[0]);
1005 kunmap_atomic(ring);
1007 refill_reqs_available(ctx, head, ctx->tail);
1010 spin_unlock_irq(&ctx->completion_lock);
1013 static bool get_reqs_available(struct kioctx *ctx)
1015 if (__get_reqs_available(ctx))
1017 user_refill_reqs_available(ctx);
1018 return __get_reqs_available(ctx);
1022 * Allocate a slot for an aio request.
1023 * Returns NULL if no requests are free.
1025 * The refcount is initialized to 2 - one for the async op completion,
1026 * one for the synchronous code that does this.
1028 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1030 struct aio_kiocb *req;
1032 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1036 percpu_ref_get(&ctx->reqs);
1038 INIT_LIST_HEAD(&req->ki_list);
1039 refcount_set(&req->ki_refcnt, 2);
1040 req->ki_eventfd = NULL;
1044 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1046 struct aio_ring __user *ring = (void __user *)ctx_id;
1047 struct mm_struct *mm = current->mm;
1048 struct kioctx *ctx, *ret = NULL;
1049 struct kioctx_table *table;
1052 if (get_user(id, &ring->id))
1056 table = rcu_dereference(mm->ioctx_table);
1058 if (!table || id >= table->nr)
1061 id = array_index_nospec(id, table->nr);
1062 ctx = rcu_dereference(table->table[id]);
1063 if (ctx && ctx->user_id == ctx_id) {
1064 if (percpu_ref_tryget_live(&ctx->users))
1072 static inline void iocb_destroy(struct aio_kiocb *iocb)
1074 if (iocb->ki_eventfd)
1075 eventfd_ctx_put(iocb->ki_eventfd);
1077 fput(iocb->ki_filp);
1078 percpu_ref_put(&iocb->ki_ctx->reqs);
1079 kmem_cache_free(kiocb_cachep, iocb);
1083 * Called when the io request on the given iocb is complete.
1085 static void aio_complete(struct aio_kiocb *iocb)
1087 struct kioctx *ctx = iocb->ki_ctx;
1088 struct aio_ring *ring;
1089 struct io_event *ev_page, *event;
1090 unsigned tail, pos, head;
1091 unsigned long flags;
1094 * Add a completion event to the ring buffer. Must be done holding
1095 * ctx->completion_lock to prevent other code from messing with the tail
1096 * pointer since we might be called from irq context.
1098 spin_lock_irqsave(&ctx->completion_lock, flags);
1101 pos = tail + AIO_EVENTS_OFFSET;
1103 if (++tail >= ctx->nr_events)
1106 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1107 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1109 *event = iocb->ki_res;
1111 kunmap_atomic(ev_page);
1112 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1114 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1115 (void __user *)(unsigned long)iocb->ki_res.obj,
1116 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1118 /* after flagging the request as done, we
1119 * must never even look at it again
1121 smp_wmb(); /* make event visible before updating tail */
1125 ring = kmap_atomic(ctx->ring_pages[0]);
1128 kunmap_atomic(ring);
1129 flush_dcache_page(ctx->ring_pages[0]);
1131 ctx->completed_events++;
1132 if (ctx->completed_events > 1)
1133 refill_reqs_available(ctx, head, tail);
1134 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1136 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1139 * Check if the user asked us to deliver the result through an
1140 * eventfd. The eventfd_signal() function is safe to be called
1143 if (iocb->ki_eventfd)
1144 eventfd_signal(iocb->ki_eventfd, 1);
1147 * We have to order our ring_info tail store above and test
1148 * of the wait list below outside the wait lock. This is
1149 * like in wake_up_bit() where clearing a bit has to be
1150 * ordered with the unlocked test.
1154 if (waitqueue_active(&ctx->wait))
1155 wake_up(&ctx->wait);
1158 static inline void iocb_put(struct aio_kiocb *iocb)
1160 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1166 /* aio_read_events_ring
1167 * Pull an event off of the ioctx's event ring. Returns the number of
1170 static long aio_read_events_ring(struct kioctx *ctx,
1171 struct io_event __user *event, long nr)
1173 struct aio_ring *ring;
1174 unsigned head, tail, pos;
1179 * The mutex can block and wake us up and that will cause
1180 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1181 * and repeat. This should be rare enough that it doesn't cause
1182 * peformance issues. See the comment in read_events() for more detail.
1184 sched_annotate_sleep();
1185 mutex_lock(&ctx->ring_lock);
1187 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1188 ring = kmap_atomic(ctx->ring_pages[0]);
1191 kunmap_atomic(ring);
1194 * Ensure that once we've read the current tail pointer, that
1195 * we also see the events that were stored up to the tail.
1199 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1204 head %= ctx->nr_events;
1205 tail %= ctx->nr_events;
1209 struct io_event *ev;
1212 avail = (head <= tail ? tail : ctx->nr_events) - head;
1216 pos = head + AIO_EVENTS_OFFSET;
1217 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1218 pos %= AIO_EVENTS_PER_PAGE;
1220 avail = min(avail, nr - ret);
1221 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1224 copy_ret = copy_to_user(event + ret, ev + pos,
1225 sizeof(*ev) * avail);
1228 if (unlikely(copy_ret)) {
1235 head %= ctx->nr_events;
1238 ring = kmap_atomic(ctx->ring_pages[0]);
1240 kunmap_atomic(ring);
1241 flush_dcache_page(ctx->ring_pages[0]);
1243 pr_debug("%li h%u t%u\n", ret, head, tail);
1245 mutex_unlock(&ctx->ring_lock);
1250 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1251 struct io_event __user *event, long *i)
1253 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1258 if (unlikely(atomic_read(&ctx->dead)))
1264 return ret < 0 || *i >= min_nr;
1267 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1268 struct io_event __user *event,
1274 * Note that aio_read_events() is being called as the conditional - i.e.
1275 * we're calling it after prepare_to_wait() has set task state to
1276 * TASK_INTERRUPTIBLE.
1278 * But aio_read_events() can block, and if it blocks it's going to flip
1279 * the task state back to TASK_RUNNING.
1281 * This should be ok, provided it doesn't flip the state back to
1282 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1283 * will only happen if the mutex_lock() call blocks, and we then find
1284 * the ringbuffer empty. So in practice we should be ok, but it's
1285 * something to be aware of when touching this code.
1288 aio_read_events(ctx, min_nr, nr, event, &ret);
1290 wait_event_interruptible_hrtimeout(ctx->wait,
1291 aio_read_events(ctx, min_nr, nr, event, &ret),
1297 * Create an aio_context capable of receiving at least nr_events.
1298 * ctxp must not point to an aio_context that already exists, and
1299 * must be initialized to 0 prior to the call. On successful
1300 * creation of the aio_context, *ctxp is filled in with the resulting
1301 * handle. May fail with -EINVAL if *ctxp is not initialized,
1302 * if the specified nr_events exceeds internal limits. May fail
1303 * with -EAGAIN if the specified nr_events exceeds the user's limit
1304 * of available events. May fail with -ENOMEM if insufficient kernel
1305 * resources are available. May fail with -EFAULT if an invalid
1306 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1309 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1311 struct kioctx *ioctx = NULL;
1315 ret = get_user(ctx, ctxp);
1320 if (unlikely(ctx || nr_events == 0)) {
1321 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1326 ioctx = ioctx_alloc(nr_events);
1327 ret = PTR_ERR(ioctx);
1328 if (!IS_ERR(ioctx)) {
1329 ret = put_user(ioctx->user_id, ctxp);
1331 kill_ioctx(current->mm, ioctx, NULL);
1332 percpu_ref_put(&ioctx->users);
1339 #ifdef CONFIG_COMPAT
1340 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1342 struct kioctx *ioctx = NULL;
1346 ret = get_user(ctx, ctx32p);
1351 if (unlikely(ctx || nr_events == 0)) {
1352 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1357 ioctx = ioctx_alloc(nr_events);
1358 ret = PTR_ERR(ioctx);
1359 if (!IS_ERR(ioctx)) {
1360 /* truncating is ok because it's a user address */
1361 ret = put_user((u32)ioctx->user_id, ctx32p);
1363 kill_ioctx(current->mm, ioctx, NULL);
1364 percpu_ref_put(&ioctx->users);
1373 * Destroy the aio_context specified. May cancel any outstanding
1374 * AIOs and block on completion. Will fail with -ENOSYS if not
1375 * implemented. May fail with -EINVAL if the context pointed to
1378 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1380 struct kioctx *ioctx = lookup_ioctx(ctx);
1381 if (likely(NULL != ioctx)) {
1382 struct ctx_rq_wait wait;
1385 init_completion(&wait.comp);
1386 atomic_set(&wait.count, 1);
1388 /* Pass requests_done to kill_ioctx() where it can be set
1389 * in a thread-safe way. If we try to set it here then we have
1390 * a race condition if two io_destroy() called simultaneously.
1392 ret = kill_ioctx(current->mm, ioctx, &wait);
1393 percpu_ref_put(&ioctx->users);
1395 /* Wait until all IO for the context are done. Otherwise kernel
1396 * keep using user-space buffers even if user thinks the context
1400 wait_for_completion(&wait.comp);
1404 pr_debug("EINVAL: invalid context id\n");
1408 static void aio_remove_iocb(struct aio_kiocb *iocb)
1410 struct kioctx *ctx = iocb->ki_ctx;
1411 unsigned long flags;
1413 spin_lock_irqsave(&ctx->ctx_lock, flags);
1414 list_del(&iocb->ki_list);
1415 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1418 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1420 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1422 if (!list_empty_careful(&iocb->ki_list))
1423 aio_remove_iocb(iocb);
1425 if (kiocb->ki_flags & IOCB_WRITE) {
1426 struct inode *inode = file_inode(kiocb->ki_filp);
1429 * Tell lockdep we inherited freeze protection from submission
1432 if (S_ISREG(inode->i_mode))
1433 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1434 file_end_write(kiocb->ki_filp);
1437 iocb->ki_res.res = res;
1438 iocb->ki_res.res2 = res2;
1442 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1446 req->ki_complete = aio_complete_rw;
1447 req->private = NULL;
1448 req->ki_pos = iocb->aio_offset;
1449 req->ki_flags = iocb_flags(req->ki_filp);
1450 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1451 req->ki_flags |= IOCB_EVENTFD;
1452 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1453 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1455 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1456 * aio_reqprio is interpreted as an I/O scheduling
1457 * class and priority.
1459 ret = ioprio_check_cap(iocb->aio_reqprio);
1461 pr_debug("aio ioprio check cap error: %d\n", ret);
1465 req->ki_ioprio = iocb->aio_reqprio;
1467 req->ki_ioprio = get_current_ioprio();
1469 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1473 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1477 static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
1478 bool vectored, bool compat, struct iov_iter *iter)
1480 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1481 size_t len = iocb->aio_nbytes;
1484 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1488 #ifdef CONFIG_COMPAT
1490 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1493 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1496 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1502 case -ERESTARTNOINTR:
1503 case -ERESTARTNOHAND:
1504 case -ERESTART_RESTARTBLOCK:
1506 * There's no easy way to restart the syscall since other AIO's
1507 * may be already running. Just fail this IO with EINTR.
1512 req->ki_complete(req, ret, 0);
1516 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1517 bool vectored, bool compat)
1519 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1520 struct iov_iter iter;
1524 ret = aio_prep_rw(req, iocb);
1527 file = req->ki_filp;
1528 if (unlikely(!(file->f_mode & FMODE_READ)))
1531 if (unlikely(!file->f_op->read_iter))
1534 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1537 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1539 aio_rw_done(req, call_read_iter(file, req, &iter));
1544 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1545 bool vectored, bool compat)
1547 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1548 struct iov_iter iter;
1552 ret = aio_prep_rw(req, iocb);
1555 file = req->ki_filp;
1557 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1559 if (unlikely(!file->f_op->write_iter))
1562 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1565 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1568 * Open-code file_start_write here to grab freeze protection,
1569 * which will be released by another thread in
1570 * aio_complete_rw(). Fool lockdep by telling it the lock got
1571 * released so that it doesn't complain about the held lock when
1572 * we return to userspace.
1574 if (S_ISREG(file_inode(file)->i_mode)) {
1575 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1576 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1578 req->ki_flags |= IOCB_WRITE;
1579 aio_rw_done(req, call_write_iter(file, req, &iter));
1585 static void aio_fsync_work(struct work_struct *work)
1587 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1589 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1593 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1596 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1597 iocb->aio_rw_flags))
1600 if (unlikely(!req->file->f_op->fsync))
1603 req->datasync = datasync;
1604 INIT_WORK(&req->work, aio_fsync_work);
1605 schedule_work(&req->work);
1609 static void aio_poll_complete_work(struct work_struct *work)
1611 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1612 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1613 struct poll_table_struct pt = { ._key = req->events };
1614 struct kioctx *ctx = iocb->ki_ctx;
1617 if (!READ_ONCE(req->cancelled))
1618 mask = vfs_poll(req->file, &pt) & req->events;
1621 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1622 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1623 * synchronize with them. In the cancellation case the list_del_init
1624 * itself is not actually needed, but harmless so we keep it in to
1625 * avoid further branches in the fast path.
1627 spin_lock_irq(&ctx->ctx_lock);
1628 if (!mask && !READ_ONCE(req->cancelled)) {
1629 add_wait_queue(req->head, &req->wait);
1630 spin_unlock_irq(&ctx->ctx_lock);
1633 list_del_init(&iocb->ki_list);
1634 iocb->ki_res.res = mangle_poll(mask);
1636 spin_unlock_irq(&ctx->ctx_lock);
1641 /* assumes we are called with irqs disabled */
1642 static int aio_poll_cancel(struct kiocb *iocb)
1644 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1645 struct poll_iocb *req = &aiocb->poll;
1647 spin_lock(&req->head->lock);
1648 WRITE_ONCE(req->cancelled, true);
1649 if (!list_empty(&req->wait.entry)) {
1650 list_del_init(&req->wait.entry);
1651 schedule_work(&aiocb->poll.work);
1653 spin_unlock(&req->head->lock);
1658 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1661 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1662 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1663 __poll_t mask = key_to_poll(key);
1664 unsigned long flags;
1666 /* for instances that support it check for an event match first: */
1667 if (mask && !(mask & req->events))
1670 list_del_init(&req->wait.entry);
1672 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1674 * Try to complete the iocb inline if we can. Use
1675 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1676 * call this function with IRQs disabled and because IRQs
1677 * have to be disabled before ctx_lock is obtained.
1679 list_del(&iocb->ki_list);
1680 iocb->ki_res.res = mangle_poll(mask);
1682 spin_unlock_irqrestore(&iocb->ki_ctx->ctx_lock, flags);
1685 schedule_work(&req->work);
1690 struct aio_poll_table {
1691 struct poll_table_struct pt;
1692 struct aio_kiocb *iocb;
1697 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1698 struct poll_table_struct *p)
1700 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1702 /* multiple wait queues per file are not supported */
1703 if (unlikely(pt->iocb->poll.head)) {
1704 pt->error = -EINVAL;
1709 pt->iocb->poll.head = head;
1710 add_wait_queue(head, &pt->iocb->poll.wait);
1713 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1715 struct kioctx *ctx = aiocb->ki_ctx;
1716 struct poll_iocb *req = &aiocb->poll;
1717 struct aio_poll_table apt;
1718 bool cancel = false;
1721 /* reject any unknown events outside the normal event mask. */
1722 if ((u16)iocb->aio_buf != iocb->aio_buf)
1724 /* reject fields that are not defined for poll */
1725 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1728 INIT_WORK(&req->work, aio_poll_complete_work);
1729 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1733 req->cancelled = false;
1735 apt.pt._qproc = aio_poll_queue_proc;
1736 apt.pt._key = req->events;
1738 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1740 /* initialized the list so that we can do list_empty checks */
1741 INIT_LIST_HEAD(&req->wait.entry);
1742 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1744 mask = vfs_poll(req->file, &apt.pt) & req->events;
1745 spin_lock_irq(&ctx->ctx_lock);
1746 if (likely(req->head)) {
1747 spin_lock(&req->head->lock);
1748 if (unlikely(list_empty(&req->wait.entry))) {
1754 if (mask || apt.error) {
1755 list_del_init(&req->wait.entry);
1756 } else if (cancel) {
1757 WRITE_ONCE(req->cancelled, true);
1758 } else if (!req->done) { /* actually waiting for an event */
1759 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1760 aiocb->ki_cancel = aio_poll_cancel;
1762 spin_unlock(&req->head->lock);
1764 if (mask) { /* no async, we'd stolen it */
1765 aiocb->ki_res.res = mangle_poll(mask);
1768 spin_unlock_irq(&ctx->ctx_lock);
1774 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1775 struct iocb __user *user_iocb, bool compat)
1777 struct aio_kiocb *req;
1780 /* enforce forwards compatibility on users */
1781 if (unlikely(iocb->aio_reserved2)) {
1782 pr_debug("EINVAL: reserve field set\n");
1786 /* prevent overflows */
1788 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1789 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1790 ((ssize_t)iocb->aio_nbytes < 0)
1792 pr_debug("EINVAL: overflow check\n");
1796 if (!get_reqs_available(ctx))
1800 req = aio_get_req(ctx);
1802 goto out_put_reqs_available;
1804 req->ki_filp = fget(iocb->aio_fildes);
1806 if (unlikely(!req->ki_filp))
1809 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1810 struct eventfd_ctx *eventfd;
1812 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1813 * instance of the file* now. The file descriptor must be
1814 * an eventfd() fd, and will be signaled for each completed
1815 * event using the eventfd_signal() function.
1817 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1818 if (IS_ERR(eventfd)) {
1819 ret = PTR_ERR(eventfd);
1822 req->ki_eventfd = eventfd;
1825 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1826 if (unlikely(ret)) {
1827 pr_debug("EFAULT: aio_key\n");
1831 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1832 req->ki_res.data = iocb->aio_data;
1833 req->ki_res.res = 0;
1834 req->ki_res.res2 = 0;
1836 switch (iocb->aio_lio_opcode) {
1837 case IOCB_CMD_PREAD:
1838 ret = aio_read(&req->rw, iocb, false, compat);
1840 case IOCB_CMD_PWRITE:
1841 ret = aio_write(&req->rw, iocb, false, compat);
1843 case IOCB_CMD_PREADV:
1844 ret = aio_read(&req->rw, iocb, true, compat);
1846 case IOCB_CMD_PWRITEV:
1847 ret = aio_write(&req->rw, iocb, true, compat);
1849 case IOCB_CMD_FSYNC:
1850 ret = aio_fsync(&req->fsync, iocb, false);
1852 case IOCB_CMD_FDSYNC:
1853 ret = aio_fsync(&req->fsync, iocb, true);
1856 ret = aio_poll(req, iocb);
1859 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1864 /* Done with the synchronous reference */
1868 * If ret is 0, we'd either done aio_complete() ourselves or have
1869 * arranged for that to be done asynchronously. Anything non-zero
1870 * means that we need to destroy req ourselves.
1877 out_put_reqs_available:
1878 put_reqs_available(ctx, 1);
1882 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1887 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1890 return __io_submit_one(ctx, &iocb, user_iocb, compat);
1894 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1895 * the number of iocbs queued. May return -EINVAL if the aio_context
1896 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1897 * *iocbpp[0] is not properly initialized, if the operation specified
1898 * is invalid for the file descriptor in the iocb. May fail with
1899 * -EFAULT if any of the data structures point to invalid data. May
1900 * fail with -EBADF if the file descriptor specified in the first
1901 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1902 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1903 * fail with -ENOSYS if not implemented.
1905 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1906 struct iocb __user * __user *, iocbpp)
1911 struct blk_plug plug;
1913 if (unlikely(nr < 0))
1916 ctx = lookup_ioctx(ctx_id);
1917 if (unlikely(!ctx)) {
1918 pr_debug("EINVAL: invalid context id\n");
1922 if (nr > ctx->nr_events)
1923 nr = ctx->nr_events;
1925 if (nr > AIO_PLUG_THRESHOLD)
1926 blk_start_plug(&plug);
1927 for (i = 0; i < nr; i++) {
1928 struct iocb __user *user_iocb;
1930 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1935 ret = io_submit_one(ctx, user_iocb, false);
1939 if (nr > AIO_PLUG_THRESHOLD)
1940 blk_finish_plug(&plug);
1942 percpu_ref_put(&ctx->users);
1946 #ifdef CONFIG_COMPAT
1947 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1948 int, nr, compat_uptr_t __user *, iocbpp)
1953 struct blk_plug plug;
1955 if (unlikely(nr < 0))
1958 ctx = lookup_ioctx(ctx_id);
1959 if (unlikely(!ctx)) {
1960 pr_debug("EINVAL: invalid context id\n");
1964 if (nr > ctx->nr_events)
1965 nr = ctx->nr_events;
1967 if (nr > AIO_PLUG_THRESHOLD)
1968 blk_start_plug(&plug);
1969 for (i = 0; i < nr; i++) {
1970 compat_uptr_t user_iocb;
1972 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1977 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1981 if (nr > AIO_PLUG_THRESHOLD)
1982 blk_finish_plug(&plug);
1984 percpu_ref_put(&ctx->users);
1990 * Attempts to cancel an iocb previously passed to io_submit. If
1991 * the operation is successfully cancelled, the resulting event is
1992 * copied into the memory pointed to by result without being placed
1993 * into the completion queue and 0 is returned. May fail with
1994 * -EFAULT if any of the data structures pointed to are invalid.
1995 * May fail with -EINVAL if aio_context specified by ctx_id is
1996 * invalid. May fail with -EAGAIN if the iocb specified was not
1997 * cancelled. Will fail with -ENOSYS if not implemented.
1999 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2000 struct io_event __user *, result)
2003 struct aio_kiocb *kiocb;
2006 u64 obj = (u64)(unsigned long)iocb;
2008 if (unlikely(get_user(key, &iocb->aio_key)))
2010 if (unlikely(key != KIOCB_KEY))
2013 ctx = lookup_ioctx(ctx_id);
2017 spin_lock_irq(&ctx->ctx_lock);
2018 /* TODO: use a hash or array, this sucks. */
2019 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2020 if (kiocb->ki_res.obj == obj) {
2021 ret = kiocb->ki_cancel(&kiocb->rw);
2022 list_del_init(&kiocb->ki_list);
2026 spin_unlock_irq(&ctx->ctx_lock);
2030 * The result argument is no longer used - the io_event is
2031 * always delivered via the ring buffer. -EINPROGRESS indicates
2032 * cancellation is progress:
2037 percpu_ref_put(&ctx->users);
2042 static long do_io_getevents(aio_context_t ctx_id,
2045 struct io_event __user *events,
2046 struct timespec64 *ts)
2048 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2049 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2052 if (likely(ioctx)) {
2053 if (likely(min_nr <= nr && min_nr >= 0))
2054 ret = read_events(ioctx, min_nr, nr, events, until);
2055 percpu_ref_put(&ioctx->users);
2062 * Attempts to read at least min_nr events and up to nr events from
2063 * the completion queue for the aio_context specified by ctx_id. If
2064 * it succeeds, the number of read events is returned. May fail with
2065 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2066 * out of range, if timeout is out of range. May fail with -EFAULT
2067 * if any of the memory specified is invalid. May return 0 or
2068 * < min_nr if the timeout specified by timeout has elapsed
2069 * before sufficient events are available, where timeout == NULL
2070 * specifies an infinite timeout. Note that the timeout pointed to by
2071 * timeout is relative. Will fail with -ENOSYS if not implemented.
2073 #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2075 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2078 struct io_event __user *, events,
2079 struct __kernel_timespec __user *, timeout)
2081 struct timespec64 ts;
2084 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2087 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2088 if (!ret && signal_pending(current))
2095 struct __aio_sigset {
2096 const sigset_t __user *sigmask;
2100 SYSCALL_DEFINE6(io_pgetevents,
2101 aio_context_t, ctx_id,
2104 struct io_event __user *, events,
2105 struct __kernel_timespec __user *, timeout,
2106 const struct __aio_sigset __user *, usig)
2108 struct __aio_sigset ksig = { NULL, };
2109 sigset_t ksigmask, sigsaved;
2110 struct timespec64 ts;
2113 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2116 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2119 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2123 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2124 restore_user_sigmask(ksig.sigmask, &sigsaved);
2125 if (signal_pending(current) && !ret)
2126 ret = -ERESTARTNOHAND;
2131 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2133 SYSCALL_DEFINE6(io_pgetevents_time32,
2134 aio_context_t, ctx_id,
2137 struct io_event __user *, events,
2138 struct old_timespec32 __user *, timeout,
2139 const struct __aio_sigset __user *, usig)
2141 struct __aio_sigset ksig = { NULL, };
2142 sigset_t ksigmask, sigsaved;
2143 struct timespec64 ts;
2146 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2149 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2153 ret = set_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2157 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2158 restore_user_sigmask(ksig.sigmask, &sigsaved);
2159 if (signal_pending(current) && !ret)
2160 ret = -ERESTARTNOHAND;
2167 #if defined(CONFIG_COMPAT_32BIT_TIME)
2169 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2172 struct io_event __user *, events,
2173 struct old_timespec32 __user *, timeout)
2175 struct timespec64 t;
2178 if (timeout && get_old_timespec32(&t, timeout))
2181 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2182 if (!ret && signal_pending(current))
2189 #ifdef CONFIG_COMPAT
2191 struct __compat_aio_sigset {
2192 compat_sigset_t __user *sigmask;
2193 compat_size_t sigsetsize;
2196 #if defined(CONFIG_COMPAT_32BIT_TIME)
2198 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2199 compat_aio_context_t, ctx_id,
2200 compat_long_t, min_nr,
2202 struct io_event __user *, events,
2203 struct old_timespec32 __user *, timeout,
2204 const struct __compat_aio_sigset __user *, usig)
2206 struct __compat_aio_sigset ksig = { NULL, };
2207 sigset_t ksigmask, sigsaved;
2208 struct timespec64 t;
2211 if (timeout && get_old_timespec32(&t, timeout))
2214 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2217 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2221 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2222 restore_user_sigmask(ksig.sigmask, &sigsaved);
2223 if (signal_pending(current) && !ret)
2224 ret = -ERESTARTNOHAND;
2231 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2232 compat_aio_context_t, ctx_id,
2233 compat_long_t, min_nr,
2235 struct io_event __user *, events,
2236 struct __kernel_timespec __user *, timeout,
2237 const struct __compat_aio_sigset __user *, usig)
2239 struct __compat_aio_sigset ksig = { NULL, };
2240 sigset_t ksigmask, sigsaved;
2241 struct timespec64 t;
2244 if (timeout && get_timespec64(&t, timeout))
2247 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2250 ret = set_compat_user_sigmask(ksig.sigmask, &ksigmask, &sigsaved, ksig.sigsetsize);
2254 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2255 restore_user_sigmask(ksig.sigmask, &sigsaved);
2256 if (signal_pending(current) && !ret)
2257 ret = -ERESTARTNOHAND;