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.
9 * See ../COPYING for licensing terms.
11 #define pr_fmt(fmt) "%s: " fmt, __func__
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/errno.h>
16 #include <linux/time.h>
17 #include <linux/aio_abi.h>
18 #include <linux/export.h>
19 #include <linux/syscalls.h>
20 #include <linux/backing-dev.h>
21 #include <linux/uio.h>
23 #include <linux/sched.h>
25 #include <linux/file.h>
27 #include <linux/mman.h>
28 #include <linux/mmu_context.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/timer.h>
32 #include <linux/aio.h>
33 #include <linux/highmem.h>
34 #include <linux/workqueue.h>
35 #include <linux/security.h>
36 #include <linux/eventfd.h>
37 #include <linux/blkdev.h>
38 #include <linux/compat.h>
39 #include <linux/migrate.h>
40 #include <linux/ramfs.h>
41 #include <linux/percpu-refcount.h>
42 #include <linux/mount.h>
43 #include <linux/work-simple.h>
45 #include <asm/kmap_types.h>
46 #include <asm/uaccess.h>
50 #define AIO_RING_MAGIC 0xa10a10a1
51 #define AIO_RING_COMPAT_FEATURES 1
52 #define AIO_RING_INCOMPAT_FEATURES 0
54 unsigned id; /* kernel internal index number */
55 unsigned nr; /* number of io_events */
56 unsigned head; /* Written to by userland or under ring_lock
57 * mutex by aio_read_events_ring(). */
61 unsigned compat_features;
62 unsigned incompat_features;
63 unsigned header_length; /* size of aio_ring */
66 struct io_event io_events[0];
67 }; /* 128 bytes + ring size */
69 #define AIO_RING_PAGES 8
74 struct kioctx *table[];
78 unsigned reqs_available;
82 struct completion comp;
87 struct percpu_ref users;
90 struct percpu_ref reqs;
92 unsigned long user_id;
94 struct __percpu kioctx_cpu *cpu;
97 * For percpu reqs_available, number of slots we move to/from global
102 * This is what userspace passed to io_setup(), it's not used for
103 * anything but counting against the global max_reqs quota.
105 * The real limit is nr_events - 1, which will be larger (see
110 /* Size of ringbuffer, in units of struct io_event */
113 unsigned long mmap_base;
114 unsigned long mmap_size;
116 struct page **ring_pages;
119 struct swork_event free_work;
122 * signals when all in-flight requests are done
124 struct ctx_rq_wait *rq_wait;
128 * This counts the number of available slots in the ringbuffer,
129 * so we avoid overflowing it: it's decremented (if positive)
130 * when allocating a kiocb and incremented when the resulting
131 * io_event is pulled off the ringbuffer.
133 * We batch accesses to it with a percpu version.
135 atomic_t reqs_available;
136 } ____cacheline_aligned_in_smp;
140 struct list_head active_reqs; /* used for cancellation */
141 } ____cacheline_aligned_in_smp;
144 struct mutex ring_lock;
145 wait_queue_head_t wait;
146 } ____cacheline_aligned_in_smp;
150 unsigned completed_events;
151 spinlock_t completion_lock;
152 } ____cacheline_aligned_in_smp;
154 struct page *internal_pages[AIO_RING_PAGES];
155 struct file *aio_ring_file;
161 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
162 * cancelled or completed (this makes a certain amount of sense because
163 * successful cancellation - io_cancel() - does deliver the completion to
166 * And since most things don't implement kiocb cancellation and we'd really like
167 * kiocb completion to be lockless when possible, we use ki_cancel to
168 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
169 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
171 #define KIOCB_CANCELLED ((void *) (~0ULL))
176 struct kioctx *ki_ctx;
177 kiocb_cancel_fn *ki_cancel;
179 struct iocb __user *ki_user_iocb; /* user's aiocb */
180 __u64 ki_user_data; /* user's data for completion */
182 struct list_head ki_list; /* the aio core uses this
183 * for cancellation */
186 * If the aio_resfd field of the userspace iocb is not zero,
187 * this is the underlying eventfd context to deliver events to.
189 struct eventfd_ctx *ki_eventfd;
192 /*------ sysctl variables----*/
193 static DEFINE_SPINLOCK(aio_nr_lock);
194 unsigned long aio_nr; /* current system wide number of aio requests */
195 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
196 /*----end sysctl variables---*/
198 static struct kmem_cache *kiocb_cachep;
199 static struct kmem_cache *kioctx_cachep;
201 static struct vfsmount *aio_mnt;
203 static const struct file_operations aio_ring_fops;
204 static const struct address_space_operations aio_ctx_aops;
206 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
208 struct qstr this = QSTR_INIT("[aio]", 5);
211 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
213 return ERR_CAST(inode);
215 inode->i_mapping->a_ops = &aio_ctx_aops;
216 inode->i_mapping->private_data = ctx;
217 inode->i_size = PAGE_SIZE * nr_pages;
219 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
222 return ERR_PTR(-ENOMEM);
224 path.mnt = mntget(aio_mnt);
226 d_instantiate(path.dentry, inode);
227 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
233 file->f_flags = O_RDWR;
237 static struct dentry *aio_mount(struct file_system_type *fs_type,
238 int flags, const char *dev_name, void *data)
240 static const struct dentry_operations ops = {
241 .d_dname = simple_dname,
243 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
247 * Creates the slab caches used by the aio routines, panic on
248 * failure as this is done early during the boot sequence.
250 static int __init aio_setup(void)
252 static struct file_system_type aio_fs = {
255 .kill_sb = kill_anon_super,
258 aio_mnt = kern_mount(&aio_fs);
260 panic("Failed to create aio fs mount.");
262 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
263 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
265 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
269 __initcall(aio_setup);
271 static void put_aio_ring_file(struct kioctx *ctx)
273 struct file *aio_ring_file = ctx->aio_ring_file;
275 truncate_setsize(aio_ring_file->f_inode, 0);
277 /* Prevent further access to the kioctx from migratepages */
278 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
279 aio_ring_file->f_inode->i_mapping->private_data = NULL;
280 ctx->aio_ring_file = NULL;
281 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
287 static void aio_free_ring(struct kioctx *ctx)
291 /* Disconnect the kiotx from the ring file. This prevents future
292 * accesses to the kioctx from page migration.
294 put_aio_ring_file(ctx);
296 for (i = 0; i < ctx->nr_pages; i++) {
298 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
299 page_count(ctx->ring_pages[i]));
300 page = ctx->ring_pages[i];
303 ctx->ring_pages[i] = NULL;
307 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
308 kfree(ctx->ring_pages);
309 ctx->ring_pages = NULL;
313 static int aio_ring_mremap(struct vm_area_struct *vma)
315 struct file *file = vma->vm_file;
316 struct mm_struct *mm = vma->vm_mm;
317 struct kioctx_table *table;
318 int i, res = -EINVAL;
320 spin_lock(&mm->ioctx_lock);
322 table = rcu_dereference(mm->ioctx_table);
323 for (i = 0; i < table->nr; i++) {
326 ctx = table->table[i];
327 if (ctx && ctx->aio_ring_file == file) {
328 if (!atomic_read(&ctx->dead)) {
329 ctx->user_id = ctx->mmap_base = vma->vm_start;
337 spin_unlock(&mm->ioctx_lock);
341 static const struct vm_operations_struct aio_ring_vm_ops = {
342 .mremap = aio_ring_mremap,
343 #if IS_ENABLED(CONFIG_MMU)
344 .fault = filemap_fault,
345 .map_pages = filemap_map_pages,
346 .page_mkwrite = filemap_page_mkwrite,
350 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
352 vma->vm_flags |= VM_DONTEXPAND;
353 vma->vm_ops = &aio_ring_vm_ops;
357 static const struct file_operations aio_ring_fops = {
358 .mmap = aio_ring_mmap,
361 #if IS_ENABLED(CONFIG_MIGRATION)
362 static int aio_migratepage(struct address_space *mapping, struct page *new,
363 struct page *old, enum migrate_mode mode)
372 /* mapping->private_lock here protects against the kioctx teardown. */
373 spin_lock(&mapping->private_lock);
374 ctx = mapping->private_data;
380 /* The ring_lock mutex. The prevents aio_read_events() from writing
381 * to the ring's head, and prevents page migration from mucking in
382 * a partially initialized kiotx.
384 if (!mutex_trylock(&ctx->ring_lock)) {
390 if (idx < (pgoff_t)ctx->nr_pages) {
391 /* Make sure the old page hasn't already been changed */
392 if (ctx->ring_pages[idx] != old)
400 /* Writeback must be complete */
401 BUG_ON(PageWriteback(old));
404 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
405 if (rc != MIGRATEPAGE_SUCCESS) {
410 /* Take completion_lock to prevent other writes to the ring buffer
411 * while the old page is copied to the new. This prevents new
412 * events from being lost.
414 spin_lock_irqsave(&ctx->completion_lock, flags);
415 migrate_page_copy(new, old);
416 BUG_ON(ctx->ring_pages[idx] != old);
417 ctx->ring_pages[idx] = new;
418 spin_unlock_irqrestore(&ctx->completion_lock, flags);
420 /* The old page is no longer accessible. */
424 mutex_unlock(&ctx->ring_lock);
426 spin_unlock(&mapping->private_lock);
431 static const struct address_space_operations aio_ctx_aops = {
432 .set_page_dirty = __set_page_dirty_no_writeback,
433 #if IS_ENABLED(CONFIG_MIGRATION)
434 .migratepage = aio_migratepage,
438 static int aio_setup_ring(struct kioctx *ctx)
440 struct aio_ring *ring;
441 unsigned nr_events = ctx->max_reqs;
442 struct mm_struct *mm = current->mm;
443 unsigned long size, unused;
448 /* Compensate for the ring buffer's head/tail overlap entry */
449 nr_events += 2; /* 1 is required, 2 for good luck */
451 size = sizeof(struct aio_ring);
452 size += sizeof(struct io_event) * nr_events;
454 nr_pages = PFN_UP(size);
458 file = aio_private_file(ctx, nr_pages);
460 ctx->aio_ring_file = NULL;
464 ctx->aio_ring_file = file;
465 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
466 / sizeof(struct io_event);
468 ctx->ring_pages = ctx->internal_pages;
469 if (nr_pages > AIO_RING_PAGES) {
470 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
472 if (!ctx->ring_pages) {
473 put_aio_ring_file(ctx);
478 for (i = 0; i < nr_pages; i++) {
480 page = find_or_create_page(file->f_inode->i_mapping,
481 i, GFP_HIGHUSER | __GFP_ZERO);
484 pr_debug("pid(%d) page[%d]->count=%d\n",
485 current->pid, i, page_count(page));
486 SetPageUptodate(page);
489 ctx->ring_pages[i] = page;
493 if (unlikely(i != nr_pages)) {
498 ctx->mmap_size = nr_pages * PAGE_SIZE;
499 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
501 down_write(&mm->mmap_sem);
502 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
503 PROT_READ | PROT_WRITE,
504 MAP_SHARED, 0, &unused);
505 up_write(&mm->mmap_sem);
506 if (IS_ERR((void *)ctx->mmap_base)) {
512 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
514 ctx->user_id = ctx->mmap_base;
515 ctx->nr_events = nr_events; /* trusted copy */
517 ring = kmap_atomic(ctx->ring_pages[0]);
518 ring->nr = nr_events; /* user copy */
520 ring->head = ring->tail = 0;
521 ring->magic = AIO_RING_MAGIC;
522 ring->compat_features = AIO_RING_COMPAT_FEATURES;
523 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
524 ring->header_length = sizeof(struct aio_ring);
526 flush_dcache_page(ctx->ring_pages[0]);
531 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
532 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
533 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
535 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
537 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
538 struct kioctx *ctx = req->ki_ctx;
541 spin_lock_irqsave(&ctx->ctx_lock, flags);
543 if (!req->ki_list.next)
544 list_add(&req->ki_list, &ctx->active_reqs);
546 req->ki_cancel = cancel;
548 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
550 EXPORT_SYMBOL(kiocb_set_cancel_fn);
552 static int kiocb_cancel(struct aio_kiocb *kiocb)
554 kiocb_cancel_fn *old, *cancel;
557 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
558 * actually has a cancel function, hence the cmpxchg()
561 cancel = ACCESS_ONCE(kiocb->ki_cancel);
563 if (!cancel || cancel == KIOCB_CANCELLED)
567 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
568 } while (cancel != old);
570 return cancel(&kiocb->common);
573 static void free_ioctx(struct swork_event *sev)
575 struct kioctx *ctx = container_of(sev, struct kioctx, free_work);
577 pr_debug("freeing %p\n", ctx);
580 free_percpu(ctx->cpu);
581 percpu_ref_exit(&ctx->reqs);
582 percpu_ref_exit(&ctx->users);
583 kmem_cache_free(kioctx_cachep, ctx);
586 static void free_ioctx_reqs(struct percpu_ref *ref)
588 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
590 /* At this point we know that there are no any in-flight requests */
591 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
592 complete(&ctx->rq_wait->comp);
594 INIT_SWORK(&ctx->free_work, free_ioctx);
595 swork_queue(&ctx->free_work);
599 * When this function runs, the kioctx has been removed from the "hash table"
600 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
601 * now it's safe to cancel any that need to be.
603 static void free_ioctx_users_work(struct swork_event *sev)
605 struct kioctx *ctx = container_of(sev, struct kioctx, free_work);
606 struct aio_kiocb *req;
608 spin_lock_irq(&ctx->ctx_lock);
610 while (!list_empty(&ctx->active_reqs)) {
611 req = list_first_entry(&ctx->active_reqs,
612 struct aio_kiocb, ki_list);
614 list_del_init(&req->ki_list);
618 spin_unlock_irq(&ctx->ctx_lock);
620 percpu_ref_kill(&ctx->reqs);
621 percpu_ref_put(&ctx->reqs);
624 static void free_ioctx_users(struct percpu_ref *ref)
626 struct kioctx *ctx = container_of(ref, struct kioctx, users);
628 INIT_SWORK(&ctx->free_work, free_ioctx_users_work);
629 swork_queue(&ctx->free_work);
632 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
635 struct kioctx_table *table, *old;
636 struct aio_ring *ring;
638 spin_lock(&mm->ioctx_lock);
639 table = rcu_dereference_raw(mm->ioctx_table);
643 for (i = 0; i < table->nr; i++)
644 if (!table->table[i]) {
646 table->table[i] = ctx;
647 spin_unlock(&mm->ioctx_lock);
649 /* While kioctx setup is in progress,
650 * we are protected from page migration
651 * changes ring_pages by ->ring_lock.
653 ring = kmap_atomic(ctx->ring_pages[0]);
659 new_nr = (table ? table->nr : 1) * 4;
660 spin_unlock(&mm->ioctx_lock);
662 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
669 spin_lock(&mm->ioctx_lock);
670 old = rcu_dereference_raw(mm->ioctx_table);
673 rcu_assign_pointer(mm->ioctx_table, table);
674 } else if (table->nr > old->nr) {
675 memcpy(table->table, old->table,
676 old->nr * sizeof(struct kioctx *));
678 rcu_assign_pointer(mm->ioctx_table, table);
687 static void aio_nr_sub(unsigned nr)
689 spin_lock(&aio_nr_lock);
690 if (WARN_ON(aio_nr - nr > aio_nr))
694 spin_unlock(&aio_nr_lock);
698 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
700 static struct kioctx *ioctx_alloc(unsigned nr_events)
702 struct mm_struct *mm = current->mm;
707 * We keep track of the number of available ringbuffer slots, to prevent
708 * overflow (reqs_available), and we also use percpu counters for this.
710 * So since up to half the slots might be on other cpu's percpu counters
711 * and unavailable, double nr_events so userspace sees what they
712 * expected: additionally, we move req_batch slots to/from percpu
713 * counters at a time, so make sure that isn't 0:
715 nr_events = max(nr_events, num_possible_cpus() * 4);
718 /* Prevent overflows */
719 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
720 pr_debug("ENOMEM: nr_events too high\n");
721 return ERR_PTR(-EINVAL);
724 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
725 return ERR_PTR(-EAGAIN);
727 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
729 return ERR_PTR(-ENOMEM);
731 ctx->max_reqs = nr_events;
733 spin_lock_init(&ctx->ctx_lock);
734 spin_lock_init(&ctx->completion_lock);
735 mutex_init(&ctx->ring_lock);
736 /* Protect against page migration throughout kiotx setup by keeping
737 * the ring_lock mutex held until setup is complete. */
738 mutex_lock(&ctx->ring_lock);
739 init_waitqueue_head(&ctx->wait);
741 INIT_LIST_HEAD(&ctx->active_reqs);
743 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
746 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
749 ctx->cpu = alloc_percpu(struct kioctx_cpu);
753 err = aio_setup_ring(ctx);
757 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
758 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
759 if (ctx->req_batch < 1)
762 /* limit the number of system wide aios */
763 spin_lock(&aio_nr_lock);
764 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
765 aio_nr + nr_events < aio_nr) {
766 spin_unlock(&aio_nr_lock);
770 aio_nr += ctx->max_reqs;
771 spin_unlock(&aio_nr_lock);
773 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
774 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
776 err = ioctx_add_table(ctx, mm);
780 /* Release the ring_lock mutex now that all setup is complete. */
781 mutex_unlock(&ctx->ring_lock);
783 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
784 ctx, ctx->user_id, mm, ctx->nr_events);
788 aio_nr_sub(ctx->max_reqs);
790 atomic_set(&ctx->dead, 1);
792 vm_munmap(ctx->mmap_base, ctx->mmap_size);
795 mutex_unlock(&ctx->ring_lock);
796 free_percpu(ctx->cpu);
797 percpu_ref_exit(&ctx->reqs);
798 percpu_ref_exit(&ctx->users);
799 kmem_cache_free(kioctx_cachep, ctx);
800 pr_debug("error allocating ioctx %d\n", err);
805 * Cancels all outstanding aio requests on an aio context. Used
806 * when the processes owning a context have all exited to encourage
807 * the rapid destruction of the kioctx.
809 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
810 struct ctx_rq_wait *wait)
812 struct kioctx_table *table;
814 spin_lock(&mm->ioctx_lock);
815 if (atomic_xchg(&ctx->dead, 1)) {
816 spin_unlock(&mm->ioctx_lock);
820 table = rcu_dereference_raw(mm->ioctx_table);
821 WARN_ON(ctx != table->table[ctx->id]);
822 table->table[ctx->id] = NULL;
823 spin_unlock(&mm->ioctx_lock);
825 /* percpu_ref_kill() will do the necessary call_rcu() */
826 wake_up_all(&ctx->wait);
829 * It'd be more correct to do this in free_ioctx(), after all
830 * the outstanding kiocbs have finished - but by then io_destroy
831 * has already returned, so io_setup() could potentially return
832 * -EAGAIN with no ioctxs actually in use (as far as userspace
835 aio_nr_sub(ctx->max_reqs);
838 vm_munmap(ctx->mmap_base, ctx->mmap_size);
841 percpu_ref_kill(&ctx->users);
846 * exit_aio: called when the last user of mm goes away. At this point, there is
847 * no way for any new requests to be submited or any of the io_* syscalls to be
848 * called on the context.
850 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
853 void exit_aio(struct mm_struct *mm)
855 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
856 struct ctx_rq_wait wait;
862 atomic_set(&wait.count, table->nr);
863 init_completion(&wait.comp);
866 for (i = 0; i < table->nr; ++i) {
867 struct kioctx *ctx = table->table[i];
875 * We don't need to bother with munmap() here - exit_mmap(mm)
876 * is coming and it'll unmap everything. And we simply can't,
877 * this is not necessarily our ->mm.
878 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
879 * that it needs to unmap the area, just set it to 0.
882 kill_ioctx(mm, ctx, &wait);
885 if (!atomic_sub_and_test(skipped, &wait.count)) {
886 /* Wait until all IO for the context are done. */
887 wait_for_completion(&wait.comp);
890 RCU_INIT_POINTER(mm->ioctx_table, NULL);
894 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
896 struct kioctx_cpu *kcpu;
899 local_irq_save(flags);
900 kcpu = this_cpu_ptr(ctx->cpu);
901 kcpu->reqs_available += nr;
903 while (kcpu->reqs_available >= ctx->req_batch * 2) {
904 kcpu->reqs_available -= ctx->req_batch;
905 atomic_add(ctx->req_batch, &ctx->reqs_available);
908 local_irq_restore(flags);
911 static bool get_reqs_available(struct kioctx *ctx)
913 struct kioctx_cpu *kcpu;
917 local_irq_save(flags);
918 kcpu = this_cpu_ptr(ctx->cpu);
919 if (!kcpu->reqs_available) {
920 int old, avail = atomic_read(&ctx->reqs_available);
923 if (avail < ctx->req_batch)
927 avail = atomic_cmpxchg(&ctx->reqs_available,
928 avail, avail - ctx->req_batch);
929 } while (avail != old);
931 kcpu->reqs_available += ctx->req_batch;
935 kcpu->reqs_available--;
937 local_irq_restore(flags);
941 /* refill_reqs_available
942 * Updates the reqs_available reference counts used for tracking the
943 * number of free slots in the completion ring. This can be called
944 * from aio_complete() (to optimistically update reqs_available) or
945 * from aio_get_req() (the we're out of events case). It must be
946 * called holding ctx->completion_lock.
948 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
951 unsigned events_in_ring, completed;
953 /* Clamp head since userland can write to it. */
954 head %= ctx->nr_events;
956 events_in_ring = tail - head;
958 events_in_ring = ctx->nr_events - (head - tail);
960 completed = ctx->completed_events;
961 if (events_in_ring < completed)
962 completed -= events_in_ring;
969 ctx->completed_events -= completed;
970 put_reqs_available(ctx, completed);
973 /* user_refill_reqs_available
974 * Called to refill reqs_available when aio_get_req() encounters an
975 * out of space in the completion ring.
977 static void user_refill_reqs_available(struct kioctx *ctx)
979 spin_lock_irq(&ctx->completion_lock);
980 if (ctx->completed_events) {
981 struct aio_ring *ring;
984 /* Access of ring->head may race with aio_read_events_ring()
985 * here, but that's okay since whether we read the old version
986 * or the new version, and either will be valid. The important
987 * part is that head cannot pass tail since we prevent
988 * aio_complete() from updating tail by holding
989 * ctx->completion_lock. Even if head is invalid, the check
990 * against ctx->completed_events below will make sure we do the
993 ring = kmap_atomic(ctx->ring_pages[0]);
997 refill_reqs_available(ctx, head, ctx->tail);
1000 spin_unlock_irq(&ctx->completion_lock);
1004 * Allocate a slot for an aio request.
1005 * Returns NULL if no requests are free.
1007 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1009 struct aio_kiocb *req;
1011 if (!get_reqs_available(ctx)) {
1012 user_refill_reqs_available(ctx);
1013 if (!get_reqs_available(ctx))
1017 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1021 percpu_ref_get(&ctx->reqs);
1026 put_reqs_available(ctx, 1);
1030 static void kiocb_free(struct aio_kiocb *req)
1032 if (req->common.ki_filp)
1033 fput(req->common.ki_filp);
1034 if (req->ki_eventfd != NULL)
1035 eventfd_ctx_put(req->ki_eventfd);
1036 kmem_cache_free(kiocb_cachep, req);
1039 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1041 struct aio_ring __user *ring = (void __user *)ctx_id;
1042 struct mm_struct *mm = current->mm;
1043 struct kioctx *ctx, *ret = NULL;
1044 struct kioctx_table *table;
1047 if (get_user(id, &ring->id))
1051 table = rcu_dereference(mm->ioctx_table);
1053 if (!table || id >= table->nr)
1056 ctx = table->table[id];
1057 if (ctx && ctx->user_id == ctx_id) {
1058 percpu_ref_get(&ctx->users);
1067 * Called when the io request on the given iocb is complete.
1069 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1071 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1072 struct kioctx *ctx = iocb->ki_ctx;
1073 struct aio_ring *ring;
1074 struct io_event *ev_page, *event;
1075 unsigned tail, pos, head;
1076 unsigned long flags;
1079 * Special case handling for sync iocbs:
1080 * - events go directly into the iocb for fast handling
1081 * - the sync task with the iocb in its stack holds the single iocb
1082 * ref, no other paths have a way to get another ref
1083 * - the sync task helpfully left a reference to itself in the iocb
1085 BUG_ON(is_sync_kiocb(kiocb));
1087 if (iocb->ki_list.next) {
1088 unsigned long flags;
1090 spin_lock_irqsave(&ctx->ctx_lock, flags);
1091 list_del(&iocb->ki_list);
1092 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1096 * Add a completion event to the ring buffer. Must be done holding
1097 * ctx->completion_lock to prevent other code from messing with the tail
1098 * pointer since we might be called from irq context.
1100 spin_lock_irqsave(&ctx->completion_lock, flags);
1103 pos = tail + AIO_EVENTS_OFFSET;
1105 if (++tail >= ctx->nr_events)
1108 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1109 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1111 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1112 event->data = iocb->ki_user_data;
1116 kunmap_atomic(ev_page);
1117 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1119 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1120 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1123 /* after flagging the request as done, we
1124 * must never even look at it again
1126 smp_wmb(); /* make event visible before updating tail */
1130 ring = kmap_atomic(ctx->ring_pages[0]);
1133 kunmap_atomic(ring);
1134 flush_dcache_page(ctx->ring_pages[0]);
1136 ctx->completed_events++;
1137 if (ctx->completed_events > 1)
1138 refill_reqs_available(ctx, head, tail);
1139 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1141 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1144 * Check if the user asked us to deliver the result through an
1145 * eventfd. The eventfd_signal() function is safe to be called
1148 if (iocb->ki_eventfd != NULL)
1149 eventfd_signal(iocb->ki_eventfd, 1);
1151 /* everything turned out well, dispose of the aiocb. */
1155 * We have to order our ring_info tail store above and test
1156 * of the wait list below outside the wait lock. This is
1157 * like in wake_up_bit() where clearing a bit has to be
1158 * ordered with the unlocked test.
1162 if (waitqueue_active(&ctx->wait))
1163 wake_up(&ctx->wait);
1165 percpu_ref_put(&ctx->reqs);
1168 /* aio_read_events_ring
1169 * Pull an event off of the ioctx's event ring. Returns the number of
1172 static long aio_read_events_ring(struct kioctx *ctx,
1173 struct io_event __user *event, long nr)
1175 struct aio_ring *ring;
1176 unsigned head, tail, pos;
1181 * The mutex can block and wake us up and that will cause
1182 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1183 * and repeat. This should be rare enough that it doesn't cause
1184 * peformance issues. See the comment in read_events() for more detail.
1186 sched_annotate_sleep();
1187 mutex_lock(&ctx->ring_lock);
1189 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1190 ring = kmap_atomic(ctx->ring_pages[0]);
1193 kunmap_atomic(ring);
1196 * Ensure that once we've read the current tail pointer, that
1197 * we also see the events that were stored up to the tail.
1201 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1206 head %= ctx->nr_events;
1207 tail %= ctx->nr_events;
1211 struct io_event *ev;
1214 avail = (head <= tail ? tail : ctx->nr_events) - head;
1218 avail = min(avail, nr - ret);
1219 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1220 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1222 pos = head + AIO_EVENTS_OFFSET;
1223 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1224 pos %= AIO_EVENTS_PER_PAGE;
1227 copy_ret = copy_to_user(event + ret, ev + pos,
1228 sizeof(*ev) * avail);
1231 if (unlikely(copy_ret)) {
1238 head %= ctx->nr_events;
1241 ring = kmap_atomic(ctx->ring_pages[0]);
1243 kunmap_atomic(ring);
1244 flush_dcache_page(ctx->ring_pages[0]);
1246 pr_debug("%li h%u t%u\n", ret, head, tail);
1248 mutex_unlock(&ctx->ring_lock);
1253 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1254 struct io_event __user *event, long *i)
1256 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1261 if (unlikely(atomic_read(&ctx->dead)))
1267 return ret < 0 || *i >= min_nr;
1270 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1271 struct io_event __user *event,
1272 struct timespec __user *timeout)
1274 ktime_t until = { .tv64 = KTIME_MAX };
1280 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1283 until = timespec_to_ktime(ts);
1287 * Note that aio_read_events() is being called as the conditional - i.e.
1288 * we're calling it after prepare_to_wait() has set task state to
1289 * TASK_INTERRUPTIBLE.
1291 * But aio_read_events() can block, and if it blocks it's going to flip
1292 * the task state back to TASK_RUNNING.
1294 * This should be ok, provided it doesn't flip the state back to
1295 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1296 * will only happen if the mutex_lock() call blocks, and we then find
1297 * the ringbuffer empty. So in practice we should be ok, but it's
1298 * something to be aware of when touching this code.
1300 if (until.tv64 == 0)
1301 aio_read_events(ctx, min_nr, nr, event, &ret);
1303 wait_event_interruptible_hrtimeout(ctx->wait,
1304 aio_read_events(ctx, min_nr, nr, event, &ret),
1307 if (!ret && signal_pending(current))
1314 * Create an aio_context capable of receiving at least nr_events.
1315 * ctxp must not point to an aio_context that already exists, and
1316 * must be initialized to 0 prior to the call. On successful
1317 * creation of the aio_context, *ctxp is filled in with the resulting
1318 * handle. May fail with -EINVAL if *ctxp is not initialized,
1319 * if the specified nr_events exceeds internal limits. May fail
1320 * with -EAGAIN if the specified nr_events exceeds the user's limit
1321 * of available events. May fail with -ENOMEM if insufficient kernel
1322 * resources are available. May fail with -EFAULT if an invalid
1323 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1326 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1328 struct kioctx *ioctx = NULL;
1332 ret = get_user(ctx, ctxp);
1337 if (unlikely(ctx || nr_events == 0)) {
1338 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1343 ioctx = ioctx_alloc(nr_events);
1344 ret = PTR_ERR(ioctx);
1345 if (!IS_ERR(ioctx)) {
1346 ret = put_user(ioctx->user_id, ctxp);
1348 kill_ioctx(current->mm, ioctx, NULL);
1349 percpu_ref_put(&ioctx->users);
1357 * Destroy the aio_context specified. May cancel any outstanding
1358 * AIOs and block on completion. Will fail with -ENOSYS if not
1359 * implemented. May fail with -EINVAL if the context pointed to
1362 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1364 struct kioctx *ioctx = lookup_ioctx(ctx);
1365 if (likely(NULL != ioctx)) {
1366 struct ctx_rq_wait wait;
1369 init_completion(&wait.comp);
1370 atomic_set(&wait.count, 1);
1372 /* Pass requests_done to kill_ioctx() where it can be set
1373 * in a thread-safe way. If we try to set it here then we have
1374 * a race condition if two io_destroy() called simultaneously.
1376 ret = kill_ioctx(current->mm, ioctx, &wait);
1377 percpu_ref_put(&ioctx->users);
1379 /* Wait until all IO for the context are done. Otherwise kernel
1380 * keep using user-space buffers even if user thinks the context
1384 wait_for_completion(&wait.comp);
1388 pr_debug("EINVAL: invalid context id\n");
1392 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1394 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1395 struct iovec **iovec,
1397 struct iov_iter *iter)
1399 #ifdef CONFIG_COMPAT
1401 return compat_import_iovec(rw,
1402 (struct compat_iovec __user *)buf,
1403 len, UIO_FASTIOV, iovec, iter);
1405 return import_iovec(rw, (struct iovec __user *)buf,
1406 len, UIO_FASTIOV, iovec, iter);
1411 * Performs the initial checks and io submission.
1413 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1414 char __user *buf, size_t len, bool compat)
1416 struct file *file = req->ki_filp;
1420 rw_iter_op *iter_op;
1421 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1422 struct iov_iter iter;
1425 case IOCB_CMD_PREAD:
1426 case IOCB_CMD_PREADV:
1429 iter_op = file->f_op->read_iter;
1432 case IOCB_CMD_PWRITE:
1433 case IOCB_CMD_PWRITEV:
1436 iter_op = file->f_op->write_iter;
1439 if (unlikely(!(file->f_mode & mode)))
1445 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1446 ret = aio_setup_vectored_rw(rw, buf, len,
1447 &iovec, compat, &iter);
1449 ret = import_single_range(rw, buf, len, iovec, &iter);
1453 ret = rw_verify_area(rw, file, &req->ki_pos,
1454 iov_iter_count(&iter));
1463 file_start_write(file);
1465 ret = iter_op(req, &iter);
1468 file_end_write(file);
1472 case IOCB_CMD_FDSYNC:
1473 if (!file->f_op->aio_fsync)
1476 ret = file->f_op->aio_fsync(req, 1);
1479 case IOCB_CMD_FSYNC:
1480 if (!file->f_op->aio_fsync)
1483 ret = file->f_op->aio_fsync(req, 0);
1487 pr_debug("EINVAL: no operation provided\n");
1491 if (ret != -EIOCBQUEUED) {
1493 * There's no easy way to restart the syscall since other AIO's
1494 * may be already running. Just fail this IO with EINTR.
1496 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1497 ret == -ERESTARTNOHAND ||
1498 ret == -ERESTART_RESTARTBLOCK))
1500 aio_complete(req, ret, 0);
1506 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1507 struct iocb *iocb, bool compat)
1509 struct aio_kiocb *req;
1512 /* enforce forwards compatibility on users */
1513 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1514 pr_debug("EINVAL: reserve field set\n");
1518 /* prevent overflows */
1520 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1521 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1522 ((ssize_t)iocb->aio_nbytes < 0)
1524 pr_debug("EINVAL: overflow check\n");
1528 req = aio_get_req(ctx);
1532 req->common.ki_filp = fget(iocb->aio_fildes);
1533 if (unlikely(!req->common.ki_filp)) {
1537 req->common.ki_pos = iocb->aio_offset;
1538 req->common.ki_complete = aio_complete;
1539 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1541 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1543 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1544 * instance of the file* now. The file descriptor must be
1545 * an eventfd() fd, and will be signaled for each completed
1546 * event using the eventfd_signal() function.
1548 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1549 if (IS_ERR(req->ki_eventfd)) {
1550 ret = PTR_ERR(req->ki_eventfd);
1551 req->ki_eventfd = NULL;
1555 req->common.ki_flags |= IOCB_EVENTFD;
1558 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1559 if (unlikely(ret)) {
1560 pr_debug("EFAULT: aio_key\n");
1564 req->ki_user_iocb = user_iocb;
1565 req->ki_user_data = iocb->aio_data;
1567 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1568 (char __user *)(unsigned long)iocb->aio_buf,
1576 put_reqs_available(ctx, 1);
1577 percpu_ref_put(&ctx->reqs);
1582 long do_io_submit(aio_context_t ctx_id, long nr,
1583 struct iocb __user *__user *iocbpp, bool compat)
1588 struct blk_plug plug;
1590 if (unlikely(nr < 0))
1593 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1594 nr = LONG_MAX/sizeof(*iocbpp);
1596 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1599 ctx = lookup_ioctx(ctx_id);
1600 if (unlikely(!ctx)) {
1601 pr_debug("EINVAL: invalid context id\n");
1605 blk_start_plug(&plug);
1608 * AKPM: should this return a partial result if some of the IOs were
1609 * successfully submitted?
1611 for (i=0; i<nr; i++) {
1612 struct iocb __user *user_iocb;
1615 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1620 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1625 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1629 blk_finish_plug(&plug);
1631 percpu_ref_put(&ctx->users);
1636 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1637 * the number of iocbs queued. May return -EINVAL if the aio_context
1638 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1639 * *iocbpp[0] is not properly initialized, if the operation specified
1640 * is invalid for the file descriptor in the iocb. May fail with
1641 * -EFAULT if any of the data structures point to invalid data. May
1642 * fail with -EBADF if the file descriptor specified in the first
1643 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1644 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1645 * fail with -ENOSYS if not implemented.
1647 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1648 struct iocb __user * __user *, iocbpp)
1650 return do_io_submit(ctx_id, nr, iocbpp, 0);
1654 * Finds a given iocb for cancellation.
1656 static struct aio_kiocb *
1657 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1659 struct aio_kiocb *kiocb;
1661 assert_spin_locked(&ctx->ctx_lock);
1663 if (key != KIOCB_KEY)
1666 /* TODO: use a hash or array, this sucks. */
1667 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1668 if (kiocb->ki_user_iocb == iocb)
1675 * Attempts to cancel an iocb previously passed to io_submit. If
1676 * the operation is successfully cancelled, the resulting event is
1677 * copied into the memory pointed to by result without being placed
1678 * into the completion queue and 0 is returned. May fail with
1679 * -EFAULT if any of the data structures pointed to are invalid.
1680 * May fail with -EINVAL if aio_context specified by ctx_id is
1681 * invalid. May fail with -EAGAIN if the iocb specified was not
1682 * cancelled. Will fail with -ENOSYS if not implemented.
1684 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1685 struct io_event __user *, result)
1688 struct aio_kiocb *kiocb;
1692 ret = get_user(key, &iocb->aio_key);
1696 ctx = lookup_ioctx(ctx_id);
1700 spin_lock_irq(&ctx->ctx_lock);
1702 kiocb = lookup_kiocb(ctx, iocb, key);
1704 ret = kiocb_cancel(kiocb);
1708 spin_unlock_irq(&ctx->ctx_lock);
1712 * The result argument is no longer used - the io_event is
1713 * always delivered via the ring buffer. -EINPROGRESS indicates
1714 * cancellation is progress:
1719 percpu_ref_put(&ctx->users);
1725 * Attempts to read at least min_nr events and up to nr events from
1726 * the completion queue for the aio_context specified by ctx_id. If
1727 * it succeeds, the number of read events is returned. May fail with
1728 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1729 * out of range, if timeout is out of range. May fail with -EFAULT
1730 * if any of the memory specified is invalid. May return 0 or
1731 * < min_nr if the timeout specified by timeout has elapsed
1732 * before sufficient events are available, where timeout == NULL
1733 * specifies an infinite timeout. Note that the timeout pointed to by
1734 * timeout is relative. Will fail with -ENOSYS if not implemented.
1736 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1739 struct io_event __user *, events,
1740 struct timespec __user *, timeout)
1742 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1745 if (likely(ioctx)) {
1746 if (likely(min_nr <= nr && min_nr >= 0))
1747 ret = read_events(ioctx, min_nr, nr, events, timeout);
1748 percpu_ref_put(&ioctx->users);