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>
44 #include <asm/kmap_types.h>
45 #include <asm/uaccess.h>
49 #define AIO_RING_MAGIC 0xa10a10a1
50 #define AIO_RING_COMPAT_FEATURES 1
51 #define AIO_RING_INCOMPAT_FEATURES 0
53 unsigned id; /* kernel internal index number */
54 unsigned nr; /* number of io_events */
55 unsigned head; /* Written to by userland or under ring_lock
56 * mutex by aio_read_events_ring(). */
60 unsigned compat_features;
61 unsigned incompat_features;
62 unsigned header_length; /* size of aio_ring */
65 struct io_event io_events[0];
66 }; /* 128 bytes + ring size */
68 #define AIO_RING_PAGES 8
73 struct kioctx *table[];
77 unsigned reqs_available;
81 struct completion comp;
86 struct percpu_ref users;
89 struct percpu_ref reqs;
91 unsigned long user_id;
93 struct __percpu kioctx_cpu *cpu;
96 * For percpu reqs_available, number of slots we move to/from global
101 * This is what userspace passed to io_setup(), it's not used for
102 * anything but counting against the global max_reqs quota.
104 * The real limit is nr_events - 1, which will be larger (see
109 /* Size of ringbuffer, in units of struct io_event */
112 unsigned long mmap_base;
113 unsigned long mmap_size;
115 struct page **ring_pages;
118 struct work_struct free_work;
121 * signals when all in-flight requests are done
123 struct ctx_rq_wait *rq_wait;
127 * This counts the number of available slots in the ringbuffer,
128 * so we avoid overflowing it: it's decremented (if positive)
129 * when allocating a kiocb and incremented when the resulting
130 * io_event is pulled off the ringbuffer.
132 * We batch accesses to it with a percpu version.
134 atomic_t reqs_available;
135 } ____cacheline_aligned_in_smp;
139 struct list_head active_reqs; /* used for cancellation */
140 } ____cacheline_aligned_in_smp;
143 struct mutex ring_lock;
144 wait_queue_head_t wait;
145 } ____cacheline_aligned_in_smp;
149 unsigned completed_events;
150 spinlock_t completion_lock;
151 } ____cacheline_aligned_in_smp;
153 struct page *internal_pages[AIO_RING_PAGES];
154 struct file *aio_ring_file;
160 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
161 * cancelled or completed (this makes a certain amount of sense because
162 * successful cancellation - io_cancel() - does deliver the completion to
165 * And since most things don't implement kiocb cancellation and we'd really like
166 * kiocb completion to be lockless when possible, we use ki_cancel to
167 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
168 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
170 #define KIOCB_CANCELLED ((void *) (~0ULL))
175 struct kioctx *ki_ctx;
176 kiocb_cancel_fn *ki_cancel;
178 struct iocb __user *ki_user_iocb; /* user's aiocb */
179 __u64 ki_user_data; /* user's data for completion */
181 struct list_head ki_list; /* the aio core uses this
182 * for cancellation */
185 * If the aio_resfd field of the userspace iocb is not zero,
186 * this is the underlying eventfd context to deliver events to.
188 struct eventfd_ctx *ki_eventfd;
191 /*------ sysctl variables----*/
192 static DEFINE_SPINLOCK(aio_nr_lock);
193 unsigned long aio_nr; /* current system wide number of aio requests */
194 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
195 /*----end sysctl variables---*/
197 static struct kmem_cache *kiocb_cachep;
198 static struct kmem_cache *kioctx_cachep;
200 static struct vfsmount *aio_mnt;
202 static const struct file_operations aio_ring_fops;
203 static const struct address_space_operations aio_ctx_aops;
205 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
207 struct qstr this = QSTR_INIT("[aio]", 5);
210 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
212 return ERR_CAST(inode);
214 inode->i_mapping->a_ops = &aio_ctx_aops;
215 inode->i_mapping->private_data = ctx;
216 inode->i_size = PAGE_SIZE * nr_pages;
218 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
221 return ERR_PTR(-ENOMEM);
223 path.mnt = mntget(aio_mnt);
225 d_instantiate(path.dentry, inode);
226 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
232 file->f_flags = O_RDWR;
236 static struct dentry *aio_mount(struct file_system_type *fs_type,
237 int flags, const char *dev_name, void *data)
239 static const struct dentry_operations ops = {
240 .d_dname = simple_dname,
242 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
246 * Creates the slab caches used by the aio routines, panic on
247 * failure as this is done early during the boot sequence.
249 static int __init aio_setup(void)
251 static struct file_system_type aio_fs = {
254 .kill_sb = kill_anon_super,
256 aio_mnt = kern_mount(&aio_fs);
258 panic("Failed to create aio fs mount.");
260 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
261 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
263 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
267 __initcall(aio_setup);
269 static void put_aio_ring_file(struct kioctx *ctx)
271 struct file *aio_ring_file = ctx->aio_ring_file;
273 truncate_setsize(aio_ring_file->f_inode, 0);
275 /* Prevent further access to the kioctx from migratepages */
276 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
277 aio_ring_file->f_inode->i_mapping->private_data = NULL;
278 ctx->aio_ring_file = NULL;
279 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
285 static void aio_free_ring(struct kioctx *ctx)
289 /* Disconnect the kiotx from the ring file. This prevents future
290 * accesses to the kioctx from page migration.
292 put_aio_ring_file(ctx);
294 for (i = 0; i < ctx->nr_pages; i++) {
296 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
297 page_count(ctx->ring_pages[i]));
298 page = ctx->ring_pages[i];
301 ctx->ring_pages[i] = NULL;
305 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
306 kfree(ctx->ring_pages);
307 ctx->ring_pages = NULL;
311 static int aio_ring_mremap(struct vm_area_struct *vma)
313 struct file *file = vma->vm_file;
314 struct mm_struct *mm = vma->vm_mm;
315 struct kioctx_table *table;
316 int i, res = -EINVAL;
318 spin_lock(&mm->ioctx_lock);
320 table = rcu_dereference(mm->ioctx_table);
321 for (i = 0; i < table->nr; i++) {
324 ctx = table->table[i];
325 if (ctx && ctx->aio_ring_file == file) {
326 if (!atomic_read(&ctx->dead)) {
327 ctx->user_id = ctx->mmap_base = vma->vm_start;
335 spin_unlock(&mm->ioctx_lock);
339 static const struct vm_operations_struct aio_ring_vm_ops = {
340 .mremap = aio_ring_mremap,
341 #if IS_ENABLED(CONFIG_MMU)
342 .fault = filemap_fault,
343 .map_pages = filemap_map_pages,
344 .page_mkwrite = filemap_page_mkwrite,
348 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
350 vma->vm_flags |= VM_DONTEXPAND;
351 vma->vm_ops = &aio_ring_vm_ops;
355 static const struct file_operations aio_ring_fops = {
356 .mmap = aio_ring_mmap,
359 #if IS_ENABLED(CONFIG_MIGRATION)
360 static int aio_migratepage(struct address_space *mapping, struct page *new,
361 struct page *old, enum migrate_mode mode)
370 /* mapping->private_lock here protects against the kioctx teardown. */
371 spin_lock(&mapping->private_lock);
372 ctx = mapping->private_data;
378 /* The ring_lock mutex. The prevents aio_read_events() from writing
379 * to the ring's head, and prevents page migration from mucking in
380 * a partially initialized kiotx.
382 if (!mutex_trylock(&ctx->ring_lock)) {
388 if (idx < (pgoff_t)ctx->nr_pages) {
389 /* Make sure the old page hasn't already been changed */
390 if (ctx->ring_pages[idx] != old)
398 /* Writeback must be complete */
399 BUG_ON(PageWriteback(old));
402 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
403 if (rc != MIGRATEPAGE_SUCCESS) {
408 /* Take completion_lock to prevent other writes to the ring buffer
409 * while the old page is copied to the new. This prevents new
410 * events from being lost.
412 spin_lock_irqsave(&ctx->completion_lock, flags);
413 migrate_page_copy(new, old);
414 BUG_ON(ctx->ring_pages[idx] != old);
415 ctx->ring_pages[idx] = new;
416 spin_unlock_irqrestore(&ctx->completion_lock, flags);
418 /* The old page is no longer accessible. */
422 mutex_unlock(&ctx->ring_lock);
424 spin_unlock(&mapping->private_lock);
429 static const struct address_space_operations aio_ctx_aops = {
430 .set_page_dirty = __set_page_dirty_no_writeback,
431 #if IS_ENABLED(CONFIG_MIGRATION)
432 .migratepage = aio_migratepage,
436 static int aio_setup_ring(struct kioctx *ctx)
438 struct aio_ring *ring;
439 unsigned nr_events = ctx->max_reqs;
440 struct mm_struct *mm = current->mm;
441 unsigned long size, unused;
446 /* Compensate for the ring buffer's head/tail overlap entry */
447 nr_events += 2; /* 1 is required, 2 for good luck */
449 size = sizeof(struct aio_ring);
450 size += sizeof(struct io_event) * nr_events;
452 nr_pages = PFN_UP(size);
456 file = aio_private_file(ctx, nr_pages);
458 ctx->aio_ring_file = NULL;
462 ctx->aio_ring_file = file;
463 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
464 / sizeof(struct io_event);
466 ctx->ring_pages = ctx->internal_pages;
467 if (nr_pages > AIO_RING_PAGES) {
468 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
470 if (!ctx->ring_pages) {
471 put_aio_ring_file(ctx);
476 for (i = 0; i < nr_pages; i++) {
478 page = find_or_create_page(file->f_inode->i_mapping,
479 i, GFP_HIGHUSER | __GFP_ZERO);
482 pr_debug("pid(%d) page[%d]->count=%d\n",
483 current->pid, i, page_count(page));
484 SetPageUptodate(page);
487 ctx->ring_pages[i] = page;
491 if (unlikely(i != nr_pages)) {
496 ctx->mmap_size = nr_pages * PAGE_SIZE;
497 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
499 if (down_write_killable(&mm->mmap_sem)) {
505 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
506 PROT_READ | PROT_WRITE,
507 MAP_SHARED, 0, &unused);
508 up_write(&mm->mmap_sem);
509 if (IS_ERR((void *)ctx->mmap_base)) {
515 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
517 ctx->user_id = ctx->mmap_base;
518 ctx->nr_events = nr_events; /* trusted copy */
520 ring = kmap_atomic(ctx->ring_pages[0]);
521 ring->nr = nr_events; /* user copy */
523 ring->head = ring->tail = 0;
524 ring->magic = AIO_RING_MAGIC;
525 ring->compat_features = AIO_RING_COMPAT_FEATURES;
526 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
527 ring->header_length = sizeof(struct aio_ring);
529 flush_dcache_page(ctx->ring_pages[0]);
534 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
535 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
536 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
538 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
540 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
541 struct kioctx *ctx = req->ki_ctx;
544 spin_lock_irqsave(&ctx->ctx_lock, flags);
546 if (!req->ki_list.next)
547 list_add(&req->ki_list, &ctx->active_reqs);
549 req->ki_cancel = cancel;
551 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
553 EXPORT_SYMBOL(kiocb_set_cancel_fn);
555 static int kiocb_cancel(struct aio_kiocb *kiocb)
557 kiocb_cancel_fn *old, *cancel;
560 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
561 * actually has a cancel function, hence the cmpxchg()
564 cancel = ACCESS_ONCE(kiocb->ki_cancel);
566 if (!cancel || cancel == KIOCB_CANCELLED)
570 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
571 } while (cancel != old);
573 return cancel(&kiocb->common);
576 static void free_ioctx(struct work_struct *work)
578 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
580 pr_debug("freeing %p\n", ctx);
583 free_percpu(ctx->cpu);
584 percpu_ref_exit(&ctx->reqs);
585 percpu_ref_exit(&ctx->users);
586 kmem_cache_free(kioctx_cachep, ctx);
589 static void free_ioctx_reqs(struct percpu_ref *ref)
591 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
593 /* At this point we know that there are no any in-flight requests */
594 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
595 complete(&ctx->rq_wait->comp);
597 INIT_WORK(&ctx->free_work, free_ioctx);
598 schedule_work(&ctx->free_work);
602 * When this function runs, the kioctx has been removed from the "hash table"
603 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
604 * now it's safe to cancel any that need to be.
606 static void free_ioctx_users(struct percpu_ref *ref)
608 struct kioctx *ctx = container_of(ref, struct kioctx, users);
609 struct aio_kiocb *req;
611 spin_lock_irq(&ctx->ctx_lock);
613 while (!list_empty(&ctx->active_reqs)) {
614 req = list_first_entry(&ctx->active_reqs,
615 struct aio_kiocb, ki_list);
617 list_del_init(&req->ki_list);
621 spin_unlock_irq(&ctx->ctx_lock);
623 percpu_ref_kill(&ctx->reqs);
624 percpu_ref_put(&ctx->reqs);
627 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
630 struct kioctx_table *table, *old;
631 struct aio_ring *ring;
633 spin_lock(&mm->ioctx_lock);
634 table = rcu_dereference_raw(mm->ioctx_table);
638 for (i = 0; i < table->nr; i++)
639 if (!table->table[i]) {
641 table->table[i] = ctx;
642 spin_unlock(&mm->ioctx_lock);
644 /* While kioctx setup is in progress,
645 * we are protected from page migration
646 * changes ring_pages by ->ring_lock.
648 ring = kmap_atomic(ctx->ring_pages[0]);
654 new_nr = (table ? table->nr : 1) * 4;
655 spin_unlock(&mm->ioctx_lock);
657 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
664 spin_lock(&mm->ioctx_lock);
665 old = rcu_dereference_raw(mm->ioctx_table);
668 rcu_assign_pointer(mm->ioctx_table, table);
669 } else if (table->nr > old->nr) {
670 memcpy(table->table, old->table,
671 old->nr * sizeof(struct kioctx *));
673 rcu_assign_pointer(mm->ioctx_table, table);
682 static void aio_nr_sub(unsigned nr)
684 spin_lock(&aio_nr_lock);
685 if (WARN_ON(aio_nr - nr > aio_nr))
689 spin_unlock(&aio_nr_lock);
693 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
695 static struct kioctx *ioctx_alloc(unsigned nr_events)
697 struct mm_struct *mm = current->mm;
702 * We keep track of the number of available ringbuffer slots, to prevent
703 * overflow (reqs_available), and we also use percpu counters for this.
705 * So since up to half the slots might be on other cpu's percpu counters
706 * and unavailable, double nr_events so userspace sees what they
707 * expected: additionally, we move req_batch slots to/from percpu
708 * counters at a time, so make sure that isn't 0:
710 nr_events = max(nr_events, num_possible_cpus() * 4);
713 /* Prevent overflows */
714 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
715 pr_debug("ENOMEM: nr_events too high\n");
716 return ERR_PTR(-EINVAL);
719 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
720 return ERR_PTR(-EAGAIN);
722 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
724 return ERR_PTR(-ENOMEM);
726 ctx->max_reqs = nr_events;
728 spin_lock_init(&ctx->ctx_lock);
729 spin_lock_init(&ctx->completion_lock);
730 mutex_init(&ctx->ring_lock);
731 /* Protect against page migration throughout kiotx setup by keeping
732 * the ring_lock mutex held until setup is complete. */
733 mutex_lock(&ctx->ring_lock);
734 init_waitqueue_head(&ctx->wait);
736 INIT_LIST_HEAD(&ctx->active_reqs);
738 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
741 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
744 ctx->cpu = alloc_percpu(struct kioctx_cpu);
748 err = aio_setup_ring(ctx);
752 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
753 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
754 if (ctx->req_batch < 1)
757 /* limit the number of system wide aios */
758 spin_lock(&aio_nr_lock);
759 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
760 aio_nr + nr_events < aio_nr) {
761 spin_unlock(&aio_nr_lock);
765 aio_nr += ctx->max_reqs;
766 spin_unlock(&aio_nr_lock);
768 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
769 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
771 err = ioctx_add_table(ctx, mm);
775 /* Release the ring_lock mutex now that all setup is complete. */
776 mutex_unlock(&ctx->ring_lock);
778 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
779 ctx, ctx->user_id, mm, ctx->nr_events);
783 aio_nr_sub(ctx->max_reqs);
785 atomic_set(&ctx->dead, 1);
787 vm_munmap(ctx->mmap_base, ctx->mmap_size);
790 mutex_unlock(&ctx->ring_lock);
791 free_percpu(ctx->cpu);
792 percpu_ref_exit(&ctx->reqs);
793 percpu_ref_exit(&ctx->users);
794 kmem_cache_free(kioctx_cachep, ctx);
795 pr_debug("error allocating ioctx %d\n", err);
800 * Cancels all outstanding aio requests on an aio context. Used
801 * when the processes owning a context have all exited to encourage
802 * the rapid destruction of the kioctx.
804 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
805 struct ctx_rq_wait *wait)
807 struct kioctx_table *table;
809 spin_lock(&mm->ioctx_lock);
810 if (atomic_xchg(&ctx->dead, 1)) {
811 spin_unlock(&mm->ioctx_lock);
815 table = rcu_dereference_raw(mm->ioctx_table);
816 WARN_ON(ctx != table->table[ctx->id]);
817 table->table[ctx->id] = NULL;
818 spin_unlock(&mm->ioctx_lock);
820 /* percpu_ref_kill() will do the necessary call_rcu() */
821 wake_up_all(&ctx->wait);
824 * It'd be more correct to do this in free_ioctx(), after all
825 * the outstanding kiocbs have finished - but by then io_destroy
826 * has already returned, so io_setup() could potentially return
827 * -EAGAIN with no ioctxs actually in use (as far as userspace
830 aio_nr_sub(ctx->max_reqs);
833 vm_munmap(ctx->mmap_base, ctx->mmap_size);
836 percpu_ref_kill(&ctx->users);
841 * exit_aio: called when the last user of mm goes away. At this point, there is
842 * no way for any new requests to be submited or any of the io_* syscalls to be
843 * called on the context.
845 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
848 void exit_aio(struct mm_struct *mm)
850 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
851 struct ctx_rq_wait wait;
857 atomic_set(&wait.count, table->nr);
858 init_completion(&wait.comp);
861 for (i = 0; i < table->nr; ++i) {
862 struct kioctx *ctx = table->table[i];
870 * We don't need to bother with munmap() here - exit_mmap(mm)
871 * is coming and it'll unmap everything. And we simply can't,
872 * this is not necessarily our ->mm.
873 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
874 * that it needs to unmap the area, just set it to 0.
877 kill_ioctx(mm, ctx, &wait);
880 if (!atomic_sub_and_test(skipped, &wait.count)) {
881 /* Wait until all IO for the context are done. */
882 wait_for_completion(&wait.comp);
885 RCU_INIT_POINTER(mm->ioctx_table, NULL);
889 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
891 struct kioctx_cpu *kcpu;
894 local_irq_save(flags);
895 kcpu = this_cpu_ptr(ctx->cpu);
896 kcpu->reqs_available += nr;
898 while (kcpu->reqs_available >= ctx->req_batch * 2) {
899 kcpu->reqs_available -= ctx->req_batch;
900 atomic_add(ctx->req_batch, &ctx->reqs_available);
903 local_irq_restore(flags);
906 static bool get_reqs_available(struct kioctx *ctx)
908 struct kioctx_cpu *kcpu;
912 local_irq_save(flags);
913 kcpu = this_cpu_ptr(ctx->cpu);
914 if (!kcpu->reqs_available) {
915 int old, avail = atomic_read(&ctx->reqs_available);
918 if (avail < ctx->req_batch)
922 avail = atomic_cmpxchg(&ctx->reqs_available,
923 avail, avail - ctx->req_batch);
924 } while (avail != old);
926 kcpu->reqs_available += ctx->req_batch;
930 kcpu->reqs_available--;
932 local_irq_restore(flags);
936 /* refill_reqs_available
937 * Updates the reqs_available reference counts used for tracking the
938 * number of free slots in the completion ring. This can be called
939 * from aio_complete() (to optimistically update reqs_available) or
940 * from aio_get_req() (the we're out of events case). It must be
941 * called holding ctx->completion_lock.
943 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
946 unsigned events_in_ring, completed;
948 /* Clamp head since userland can write to it. */
949 head %= ctx->nr_events;
951 events_in_ring = tail - head;
953 events_in_ring = ctx->nr_events - (head - tail);
955 completed = ctx->completed_events;
956 if (events_in_ring < completed)
957 completed -= events_in_ring;
964 ctx->completed_events -= completed;
965 put_reqs_available(ctx, completed);
968 /* user_refill_reqs_available
969 * Called to refill reqs_available when aio_get_req() encounters an
970 * out of space in the completion ring.
972 static void user_refill_reqs_available(struct kioctx *ctx)
974 spin_lock_irq(&ctx->completion_lock);
975 if (ctx->completed_events) {
976 struct aio_ring *ring;
979 /* Access of ring->head may race with aio_read_events_ring()
980 * here, but that's okay since whether we read the old version
981 * or the new version, and either will be valid. The important
982 * part is that head cannot pass tail since we prevent
983 * aio_complete() from updating tail by holding
984 * ctx->completion_lock. Even if head is invalid, the check
985 * against ctx->completed_events below will make sure we do the
988 ring = kmap_atomic(ctx->ring_pages[0]);
992 refill_reqs_available(ctx, head, ctx->tail);
995 spin_unlock_irq(&ctx->completion_lock);
999 * Allocate a slot for an aio request.
1000 * Returns NULL if no requests are free.
1002 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1004 struct aio_kiocb *req;
1006 if (!get_reqs_available(ctx)) {
1007 user_refill_reqs_available(ctx);
1008 if (!get_reqs_available(ctx))
1012 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1016 percpu_ref_get(&ctx->reqs);
1021 put_reqs_available(ctx, 1);
1025 static void kiocb_free(struct aio_kiocb *req)
1027 if (req->common.ki_filp)
1028 fput(req->common.ki_filp);
1029 if (req->ki_eventfd != NULL)
1030 eventfd_ctx_put(req->ki_eventfd);
1031 kmem_cache_free(kiocb_cachep, req);
1034 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1036 struct aio_ring __user *ring = (void __user *)ctx_id;
1037 struct mm_struct *mm = current->mm;
1038 struct kioctx *ctx, *ret = NULL;
1039 struct kioctx_table *table;
1042 if (get_user(id, &ring->id))
1046 table = rcu_dereference(mm->ioctx_table);
1048 if (!table || id >= table->nr)
1051 ctx = table->table[id];
1052 if (ctx && ctx->user_id == ctx_id) {
1053 percpu_ref_get(&ctx->users);
1062 * Called when the io request on the given iocb is complete.
1064 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1066 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1067 struct kioctx *ctx = iocb->ki_ctx;
1068 struct aio_ring *ring;
1069 struct io_event *ev_page, *event;
1070 unsigned tail, pos, head;
1071 unsigned long flags;
1074 * Special case handling for sync iocbs:
1075 * - events go directly into the iocb for fast handling
1076 * - the sync task with the iocb in its stack holds the single iocb
1077 * ref, no other paths have a way to get another ref
1078 * - the sync task helpfully left a reference to itself in the iocb
1080 BUG_ON(is_sync_kiocb(kiocb));
1082 if (iocb->ki_list.next) {
1083 unsigned long flags;
1085 spin_lock_irqsave(&ctx->ctx_lock, flags);
1086 list_del(&iocb->ki_list);
1087 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1091 * Add a completion event to the ring buffer. Must be done holding
1092 * ctx->completion_lock to prevent other code from messing with the tail
1093 * pointer since we might be called from irq context.
1095 spin_lock_irqsave(&ctx->completion_lock, flags);
1098 pos = tail + AIO_EVENTS_OFFSET;
1100 if (++tail >= ctx->nr_events)
1103 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1104 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1106 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1107 event->data = iocb->ki_user_data;
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",
1115 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
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 != NULL)
1144 eventfd_signal(iocb->ki_eventfd, 1);
1146 /* everything turned out well, dispose of the aiocb. */
1150 * We have to order our ring_info tail store above and test
1151 * of the wait list below outside the wait lock. This is
1152 * like in wake_up_bit() where clearing a bit has to be
1153 * ordered with the unlocked test.
1157 if (waitqueue_active(&ctx->wait))
1158 wake_up(&ctx->wait);
1160 percpu_ref_put(&ctx->reqs);
1163 /* aio_read_events_ring
1164 * Pull an event off of the ioctx's event ring. Returns the number of
1167 static long aio_read_events_ring(struct kioctx *ctx,
1168 struct io_event __user *event, long nr)
1170 struct aio_ring *ring;
1171 unsigned head, tail, pos;
1176 * The mutex can block and wake us up and that will cause
1177 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1178 * and repeat. This should be rare enough that it doesn't cause
1179 * peformance issues. See the comment in read_events() for more detail.
1181 sched_annotate_sleep();
1182 mutex_lock(&ctx->ring_lock);
1184 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1185 ring = kmap_atomic(ctx->ring_pages[0]);
1188 kunmap_atomic(ring);
1191 * Ensure that once we've read the current tail pointer, that
1192 * we also see the events that were stored up to the tail.
1196 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1201 head %= ctx->nr_events;
1202 tail %= ctx->nr_events;
1206 struct io_event *ev;
1209 avail = (head <= tail ? tail : ctx->nr_events) - head;
1213 avail = min(avail, nr - ret);
1214 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1215 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1217 pos = head + AIO_EVENTS_OFFSET;
1218 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1219 pos %= AIO_EVENTS_PER_PAGE;
1222 copy_ret = copy_to_user(event + ret, ev + pos,
1223 sizeof(*ev) * avail);
1226 if (unlikely(copy_ret)) {
1233 head %= ctx->nr_events;
1236 ring = kmap_atomic(ctx->ring_pages[0]);
1238 kunmap_atomic(ring);
1239 flush_dcache_page(ctx->ring_pages[0]);
1241 pr_debug("%li h%u t%u\n", ret, head, tail);
1243 mutex_unlock(&ctx->ring_lock);
1248 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1249 struct io_event __user *event, long *i)
1251 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1256 if (unlikely(atomic_read(&ctx->dead)))
1262 return ret < 0 || *i >= min_nr;
1265 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1266 struct io_event __user *event,
1267 struct timespec __user *timeout)
1269 ktime_t until = { .tv64 = KTIME_MAX };
1275 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1278 until = timespec_to_ktime(ts);
1282 * Note that aio_read_events() is being called as the conditional - i.e.
1283 * we're calling it after prepare_to_wait() has set task state to
1284 * TASK_INTERRUPTIBLE.
1286 * But aio_read_events() can block, and if it blocks it's going to flip
1287 * the task state back to TASK_RUNNING.
1289 * This should be ok, provided it doesn't flip the state back to
1290 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1291 * will only happen if the mutex_lock() call blocks, and we then find
1292 * the ringbuffer empty. So in practice we should be ok, but it's
1293 * something to be aware of when touching this code.
1295 if (until.tv64 == 0)
1296 aio_read_events(ctx, min_nr, nr, event, &ret);
1298 wait_event_interruptible_hrtimeout(ctx->wait,
1299 aio_read_events(ctx, min_nr, nr, event, &ret),
1302 if (!ret && signal_pending(current))
1309 * Create an aio_context capable of receiving at least nr_events.
1310 * ctxp must not point to an aio_context that already exists, and
1311 * must be initialized to 0 prior to the call. On successful
1312 * creation of the aio_context, *ctxp is filled in with the resulting
1313 * handle. May fail with -EINVAL if *ctxp is not initialized,
1314 * if the specified nr_events exceeds internal limits. May fail
1315 * with -EAGAIN if the specified nr_events exceeds the user's limit
1316 * of available events. May fail with -ENOMEM if insufficient kernel
1317 * resources are available. May fail with -EFAULT if an invalid
1318 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1321 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1323 struct kioctx *ioctx = NULL;
1327 ret = get_user(ctx, ctxp);
1332 if (unlikely(ctx || nr_events == 0)) {
1333 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1338 ioctx = ioctx_alloc(nr_events);
1339 ret = PTR_ERR(ioctx);
1340 if (!IS_ERR(ioctx)) {
1341 ret = put_user(ioctx->user_id, ctxp);
1343 kill_ioctx(current->mm, ioctx, NULL);
1344 percpu_ref_put(&ioctx->users);
1352 * Destroy the aio_context specified. May cancel any outstanding
1353 * AIOs and block on completion. Will fail with -ENOSYS if not
1354 * implemented. May fail with -EINVAL if the context pointed to
1357 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1359 struct kioctx *ioctx = lookup_ioctx(ctx);
1360 if (likely(NULL != ioctx)) {
1361 struct ctx_rq_wait wait;
1364 init_completion(&wait.comp);
1365 atomic_set(&wait.count, 1);
1367 /* Pass requests_done to kill_ioctx() where it can be set
1368 * in a thread-safe way. If we try to set it here then we have
1369 * a race condition if two io_destroy() called simultaneously.
1371 ret = kill_ioctx(current->mm, ioctx, &wait);
1372 percpu_ref_put(&ioctx->users);
1374 /* Wait until all IO for the context are done. Otherwise kernel
1375 * keep using user-space buffers even if user thinks the context
1379 wait_for_completion(&wait.comp);
1383 pr_debug("EINVAL: invalid context id\n");
1387 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1389 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1390 struct iovec **iovec,
1392 struct iov_iter *iter)
1394 #ifdef CONFIG_COMPAT
1396 return compat_import_iovec(rw,
1397 (struct compat_iovec __user *)buf,
1398 len, UIO_FASTIOV, iovec, iter);
1400 return import_iovec(rw, (struct iovec __user *)buf,
1401 len, UIO_FASTIOV, iovec, iter);
1406 * Performs the initial checks and io submission.
1408 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1409 char __user *buf, size_t len, bool compat)
1411 struct file *file = req->ki_filp;
1415 rw_iter_op *iter_op;
1416 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1417 struct iov_iter iter;
1420 case IOCB_CMD_PREAD:
1421 case IOCB_CMD_PREADV:
1424 iter_op = file->f_op->read_iter;
1427 case IOCB_CMD_PWRITE:
1428 case IOCB_CMD_PWRITEV:
1431 iter_op = file->f_op->write_iter;
1434 if (unlikely(!(file->f_mode & mode)))
1440 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1441 ret = aio_setup_vectored_rw(rw, buf, len,
1442 &iovec, compat, &iter);
1444 ret = import_single_range(rw, buf, len, iovec, &iter);
1448 ret = rw_verify_area(rw, file, &req->ki_pos,
1449 iov_iter_count(&iter));
1456 file_start_write(file);
1458 ret = iter_op(req, &iter);
1461 file_end_write(file);
1465 case IOCB_CMD_FDSYNC:
1466 if (!file->f_op->aio_fsync)
1469 ret = file->f_op->aio_fsync(req, 1);
1472 case IOCB_CMD_FSYNC:
1473 if (!file->f_op->aio_fsync)
1476 ret = file->f_op->aio_fsync(req, 0);
1480 pr_debug("EINVAL: no operation provided\n");
1484 if (ret != -EIOCBQUEUED) {
1486 * There's no easy way to restart the syscall since other AIO's
1487 * may be already running. Just fail this IO with EINTR.
1489 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1490 ret == -ERESTARTNOHAND ||
1491 ret == -ERESTART_RESTARTBLOCK))
1493 aio_complete(req, ret, 0);
1499 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1500 struct iocb *iocb, bool compat)
1502 struct aio_kiocb *req;
1505 /* enforce forwards compatibility on users */
1506 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1507 pr_debug("EINVAL: reserve field set\n");
1511 /* prevent overflows */
1513 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1514 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1515 ((ssize_t)iocb->aio_nbytes < 0)
1517 pr_debug("EINVAL: overflow check\n");
1521 req = aio_get_req(ctx);
1525 req->common.ki_filp = fget(iocb->aio_fildes);
1526 if (unlikely(!req->common.ki_filp)) {
1530 req->common.ki_pos = iocb->aio_offset;
1531 req->common.ki_complete = aio_complete;
1532 req->common.ki_flags = iocb_flags(req->common.ki_filp);
1534 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1536 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1537 * instance of the file* now. The file descriptor must be
1538 * an eventfd() fd, and will be signaled for each completed
1539 * event using the eventfd_signal() function.
1541 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1542 if (IS_ERR(req->ki_eventfd)) {
1543 ret = PTR_ERR(req->ki_eventfd);
1544 req->ki_eventfd = NULL;
1548 req->common.ki_flags |= IOCB_EVENTFD;
1551 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1552 if (unlikely(ret)) {
1553 pr_debug("EFAULT: aio_key\n");
1557 req->ki_user_iocb = user_iocb;
1558 req->ki_user_data = iocb->aio_data;
1560 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1561 (char __user *)(unsigned long)iocb->aio_buf,
1569 put_reqs_available(ctx, 1);
1570 percpu_ref_put(&ctx->reqs);
1575 long do_io_submit(aio_context_t ctx_id, long nr,
1576 struct iocb __user *__user *iocbpp, bool compat)
1581 struct blk_plug plug;
1583 if (unlikely(nr < 0))
1586 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1587 nr = LONG_MAX/sizeof(*iocbpp);
1589 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1592 ctx = lookup_ioctx(ctx_id);
1593 if (unlikely(!ctx)) {
1594 pr_debug("EINVAL: invalid context id\n");
1598 blk_start_plug(&plug);
1601 * AKPM: should this return a partial result if some of the IOs were
1602 * successfully submitted?
1604 for (i=0; i<nr; i++) {
1605 struct iocb __user *user_iocb;
1608 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1613 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1618 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1622 blk_finish_plug(&plug);
1624 percpu_ref_put(&ctx->users);
1629 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1630 * the number of iocbs queued. May return -EINVAL if the aio_context
1631 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1632 * *iocbpp[0] is not properly initialized, if the operation specified
1633 * is invalid for the file descriptor in the iocb. May fail with
1634 * -EFAULT if any of the data structures point to invalid data. May
1635 * fail with -EBADF if the file descriptor specified in the first
1636 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1637 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1638 * fail with -ENOSYS if not implemented.
1640 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1641 struct iocb __user * __user *, iocbpp)
1643 return do_io_submit(ctx_id, nr, iocbpp, 0);
1647 * Finds a given iocb for cancellation.
1649 static struct aio_kiocb *
1650 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1652 struct aio_kiocb *kiocb;
1654 assert_spin_locked(&ctx->ctx_lock);
1656 if (key != KIOCB_KEY)
1659 /* TODO: use a hash or array, this sucks. */
1660 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1661 if (kiocb->ki_user_iocb == iocb)
1668 * Attempts to cancel an iocb previously passed to io_submit. If
1669 * the operation is successfully cancelled, the resulting event is
1670 * copied into the memory pointed to by result without being placed
1671 * into the completion queue and 0 is returned. May fail with
1672 * -EFAULT if any of the data structures pointed to are invalid.
1673 * May fail with -EINVAL if aio_context specified by ctx_id is
1674 * invalid. May fail with -EAGAIN if the iocb specified was not
1675 * cancelled. Will fail with -ENOSYS if not implemented.
1677 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1678 struct io_event __user *, result)
1681 struct aio_kiocb *kiocb;
1685 ret = get_user(key, &iocb->aio_key);
1689 ctx = lookup_ioctx(ctx_id);
1693 spin_lock_irq(&ctx->ctx_lock);
1695 kiocb = lookup_kiocb(ctx, iocb, key);
1697 ret = kiocb_cancel(kiocb);
1701 spin_unlock_irq(&ctx->ctx_lock);
1705 * The result argument is no longer used - the io_event is
1706 * always delivered via the ring buffer. -EINPROGRESS indicates
1707 * cancellation is progress:
1712 percpu_ref_put(&ctx->users);
1718 * Attempts to read at least min_nr events and up to nr events from
1719 * the completion queue for the aio_context specified by ctx_id. If
1720 * it succeeds, the number of read events is returned. May fail with
1721 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1722 * out of range, if timeout is out of range. May fail with -EFAULT
1723 * if any of the memory specified is invalid. May return 0 or
1724 * < min_nr if the timeout specified by timeout has elapsed
1725 * before sufficient events are available, where timeout == NULL
1726 * specifies an infinite timeout. Note that the timeout pointed to by
1727 * timeout is relative. Will fail with -ENOSYS if not implemented.
1729 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1732 struct io_event __user *, events,
1733 struct timespec __user *, timeout)
1735 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1738 if (likely(ioctx)) {
1739 if (likely(min_nr <= nr && min_nr >= 0))
1740 ret = read_events(ioctx, min_nr, nr, events, timeout);
1741 percpu_ref_put(&ioctx->users);