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 percpu_ref users;
84 struct percpu_ref reqs;
86 unsigned long user_id;
88 struct __percpu kioctx_cpu *cpu;
91 * For percpu reqs_available, number of slots we move to/from global
96 * This is what userspace passed to io_setup(), it's not used for
97 * anything but counting against the global max_reqs quota.
99 * The real limit is nr_events - 1, which will be larger (see
104 /* Size of ringbuffer, in units of struct io_event */
107 unsigned long mmap_base;
108 unsigned long mmap_size;
110 struct page **ring_pages;
113 struct work_struct free_work;
116 * signals when all in-flight requests are done
118 struct completion *requests_done;
122 * This counts the number of available slots in the ringbuffer,
123 * so we avoid overflowing it: it's decremented (if positive)
124 * when allocating a kiocb and incremented when the resulting
125 * io_event is pulled off the ringbuffer.
127 * We batch accesses to it with a percpu version.
129 atomic_t reqs_available;
130 } ____cacheline_aligned_in_smp;
134 struct list_head active_reqs; /* used for cancellation */
135 } ____cacheline_aligned_in_smp;
138 struct mutex ring_lock;
139 wait_queue_head_t wait;
140 } ____cacheline_aligned_in_smp;
144 unsigned completed_events;
145 spinlock_t completion_lock;
146 } ____cacheline_aligned_in_smp;
148 struct page *internal_pages[AIO_RING_PAGES];
149 struct file *aio_ring_file;
155 * We use ki_cancel == KIOCB_CANCELLED to indicate that a kiocb has been either
156 * cancelled or completed (this makes a certain amount of sense because
157 * successful cancellation - io_cancel() - does deliver the completion to
160 * And since most things don't implement kiocb cancellation and we'd really like
161 * kiocb completion to be lockless when possible, we use ki_cancel to
162 * synchronize cancellation and completion - we only set it to KIOCB_CANCELLED
163 * with xchg() or cmpxchg(), see batch_complete_aio() and kiocb_cancel().
165 #define KIOCB_CANCELLED ((void *) (~0ULL))
170 struct kioctx *ki_ctx;
171 kiocb_cancel_fn *ki_cancel;
173 struct iocb __user *ki_user_iocb; /* user's aiocb */
174 __u64 ki_user_data; /* user's data for completion */
176 struct list_head ki_list; /* the aio core uses this
177 * for cancellation */
180 * If the aio_resfd field of the userspace iocb is not zero,
181 * this is the underlying eventfd context to deliver events to.
183 struct eventfd_ctx *ki_eventfd;
186 /*------ sysctl variables----*/
187 static DEFINE_SPINLOCK(aio_nr_lock);
188 unsigned long aio_nr; /* current system wide number of aio requests */
189 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
190 /*----end sysctl variables---*/
192 static struct kmem_cache *kiocb_cachep;
193 static struct kmem_cache *kioctx_cachep;
195 static struct vfsmount *aio_mnt;
197 static const struct file_operations aio_ring_fops;
198 static const struct address_space_operations aio_ctx_aops;
200 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
202 struct qstr this = QSTR_INIT("[aio]", 5);
205 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
207 return ERR_CAST(inode);
209 inode->i_mapping->a_ops = &aio_ctx_aops;
210 inode->i_mapping->private_data = ctx;
211 inode->i_size = PAGE_SIZE * nr_pages;
213 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
216 return ERR_PTR(-ENOMEM);
218 path.mnt = mntget(aio_mnt);
220 d_instantiate(path.dentry, inode);
221 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
227 file->f_flags = O_RDWR;
231 static struct dentry *aio_mount(struct file_system_type *fs_type,
232 int flags, const char *dev_name, void *data)
234 static const struct dentry_operations ops = {
235 .d_dname = simple_dname,
237 return mount_pseudo(fs_type, "aio:", NULL, &ops, AIO_RING_MAGIC);
241 * Creates the slab caches used by the aio routines, panic on
242 * failure as this is done early during the boot sequence.
244 static int __init aio_setup(void)
246 static struct file_system_type aio_fs = {
249 .kill_sb = kill_anon_super,
251 aio_mnt = kern_mount(&aio_fs);
253 panic("Failed to create aio fs mount.");
255 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
256 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
258 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
262 __initcall(aio_setup);
264 static void put_aio_ring_file(struct kioctx *ctx)
266 struct file *aio_ring_file = ctx->aio_ring_file;
268 truncate_setsize(aio_ring_file->f_inode, 0);
270 /* Prevent further access to the kioctx from migratepages */
271 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
272 aio_ring_file->f_inode->i_mapping->private_data = NULL;
273 ctx->aio_ring_file = NULL;
274 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
280 static void aio_free_ring(struct kioctx *ctx)
284 /* Disconnect the kiotx from the ring file. This prevents future
285 * accesses to the kioctx from page migration.
287 put_aio_ring_file(ctx);
289 for (i = 0; i < ctx->nr_pages; i++) {
291 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
292 page_count(ctx->ring_pages[i]));
293 page = ctx->ring_pages[i];
296 ctx->ring_pages[i] = NULL;
300 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
301 kfree(ctx->ring_pages);
302 ctx->ring_pages = NULL;
306 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
308 vma->vm_flags |= VM_DONTEXPAND;
309 vma->vm_ops = &generic_file_vm_ops;
313 static void aio_ring_remap(struct file *file, struct vm_area_struct *vma)
315 struct mm_struct *mm = vma->vm_mm;
316 struct kioctx_table *table;
319 spin_lock(&mm->ioctx_lock);
321 table = rcu_dereference(mm->ioctx_table);
322 for (i = 0; i < table->nr; i++) {
325 ctx = table->table[i];
326 if (ctx && ctx->aio_ring_file == file) {
327 ctx->user_id = ctx->mmap_base = vma->vm_start;
333 spin_unlock(&mm->ioctx_lock);
336 static const struct file_operations aio_ring_fops = {
337 .mmap = aio_ring_mmap,
338 .mremap = aio_ring_remap,
341 #if IS_ENABLED(CONFIG_MIGRATION)
342 static int aio_migratepage(struct address_space *mapping, struct page *new,
343 struct page *old, enum migrate_mode mode)
352 /* mapping->private_lock here protects against the kioctx teardown. */
353 spin_lock(&mapping->private_lock);
354 ctx = mapping->private_data;
360 /* The ring_lock mutex. The prevents aio_read_events() from writing
361 * to the ring's head, and prevents page migration from mucking in
362 * a partially initialized kiotx.
364 if (!mutex_trylock(&ctx->ring_lock)) {
370 if (idx < (pgoff_t)ctx->nr_pages) {
371 /* Make sure the old page hasn't already been changed */
372 if (ctx->ring_pages[idx] != old)
380 /* Writeback must be complete */
381 BUG_ON(PageWriteback(old));
384 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
385 if (rc != MIGRATEPAGE_SUCCESS) {
390 /* Take completion_lock to prevent other writes to the ring buffer
391 * while the old page is copied to the new. This prevents new
392 * events from being lost.
394 spin_lock_irqsave(&ctx->completion_lock, flags);
395 migrate_page_copy(new, old);
396 BUG_ON(ctx->ring_pages[idx] != old);
397 ctx->ring_pages[idx] = new;
398 spin_unlock_irqrestore(&ctx->completion_lock, flags);
400 /* The old page is no longer accessible. */
404 mutex_unlock(&ctx->ring_lock);
406 spin_unlock(&mapping->private_lock);
411 static const struct address_space_operations aio_ctx_aops = {
412 .set_page_dirty = __set_page_dirty_no_writeback,
413 #if IS_ENABLED(CONFIG_MIGRATION)
414 .migratepage = aio_migratepage,
418 static int aio_setup_ring(struct kioctx *ctx)
420 struct aio_ring *ring;
421 unsigned nr_events = ctx->max_reqs;
422 struct mm_struct *mm = current->mm;
423 unsigned long size, unused;
428 /* Compensate for the ring buffer's head/tail overlap entry */
429 nr_events += 2; /* 1 is required, 2 for good luck */
431 size = sizeof(struct aio_ring);
432 size += sizeof(struct io_event) * nr_events;
434 nr_pages = PFN_UP(size);
438 file = aio_private_file(ctx, nr_pages);
440 ctx->aio_ring_file = NULL;
444 ctx->aio_ring_file = file;
445 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
446 / sizeof(struct io_event);
448 ctx->ring_pages = ctx->internal_pages;
449 if (nr_pages > AIO_RING_PAGES) {
450 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
452 if (!ctx->ring_pages) {
453 put_aio_ring_file(ctx);
458 for (i = 0; i < nr_pages; i++) {
460 page = find_or_create_page(file->f_inode->i_mapping,
461 i, GFP_HIGHUSER | __GFP_ZERO);
464 pr_debug("pid(%d) page[%d]->count=%d\n",
465 current->pid, i, page_count(page));
466 SetPageUptodate(page);
469 ctx->ring_pages[i] = page;
473 if (unlikely(i != nr_pages)) {
478 ctx->mmap_size = nr_pages * PAGE_SIZE;
479 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
481 down_write(&mm->mmap_sem);
482 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
483 PROT_READ | PROT_WRITE,
484 MAP_SHARED, 0, &unused);
485 up_write(&mm->mmap_sem);
486 if (IS_ERR((void *)ctx->mmap_base)) {
492 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
494 ctx->user_id = ctx->mmap_base;
495 ctx->nr_events = nr_events; /* trusted copy */
497 ring = kmap_atomic(ctx->ring_pages[0]);
498 ring->nr = nr_events; /* user copy */
500 ring->head = ring->tail = 0;
501 ring->magic = AIO_RING_MAGIC;
502 ring->compat_features = AIO_RING_COMPAT_FEATURES;
503 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
504 ring->header_length = sizeof(struct aio_ring);
506 flush_dcache_page(ctx->ring_pages[0]);
511 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
512 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
513 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
515 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
517 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, common);
518 struct kioctx *ctx = req->ki_ctx;
521 spin_lock_irqsave(&ctx->ctx_lock, flags);
523 if (!req->ki_list.next)
524 list_add(&req->ki_list, &ctx->active_reqs);
526 req->ki_cancel = cancel;
528 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
530 EXPORT_SYMBOL(kiocb_set_cancel_fn);
532 static int kiocb_cancel(struct aio_kiocb *kiocb)
534 kiocb_cancel_fn *old, *cancel;
537 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
538 * actually has a cancel function, hence the cmpxchg()
541 cancel = ACCESS_ONCE(kiocb->ki_cancel);
543 if (!cancel || cancel == KIOCB_CANCELLED)
547 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
548 } while (cancel != old);
550 return cancel(&kiocb->common);
553 static void free_ioctx(struct work_struct *work)
555 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
557 pr_debug("freeing %p\n", ctx);
560 free_percpu(ctx->cpu);
561 percpu_ref_exit(&ctx->reqs);
562 percpu_ref_exit(&ctx->users);
563 kmem_cache_free(kioctx_cachep, ctx);
566 static void free_ioctx_reqs(struct percpu_ref *ref)
568 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
570 /* At this point we know that there are no any in-flight requests */
571 if (ctx->requests_done)
572 complete(ctx->requests_done);
574 INIT_WORK(&ctx->free_work, free_ioctx);
575 schedule_work(&ctx->free_work);
579 * When this function runs, the kioctx has been removed from the "hash table"
580 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
581 * now it's safe to cancel any that need to be.
583 static void free_ioctx_users(struct percpu_ref *ref)
585 struct kioctx *ctx = container_of(ref, struct kioctx, users);
586 struct aio_kiocb *req;
588 spin_lock_irq(&ctx->ctx_lock);
590 while (!list_empty(&ctx->active_reqs)) {
591 req = list_first_entry(&ctx->active_reqs,
592 struct aio_kiocb, ki_list);
594 list_del_init(&req->ki_list);
598 spin_unlock_irq(&ctx->ctx_lock);
600 percpu_ref_kill(&ctx->reqs);
601 percpu_ref_put(&ctx->reqs);
604 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
607 struct kioctx_table *table, *old;
608 struct aio_ring *ring;
610 spin_lock(&mm->ioctx_lock);
611 table = rcu_dereference_raw(mm->ioctx_table);
615 for (i = 0; i < table->nr; i++)
616 if (!table->table[i]) {
618 table->table[i] = ctx;
619 spin_unlock(&mm->ioctx_lock);
621 /* While kioctx setup is in progress,
622 * we are protected from page migration
623 * changes ring_pages by ->ring_lock.
625 ring = kmap_atomic(ctx->ring_pages[0]);
631 new_nr = (table ? table->nr : 1) * 4;
632 spin_unlock(&mm->ioctx_lock);
634 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
641 spin_lock(&mm->ioctx_lock);
642 old = rcu_dereference_raw(mm->ioctx_table);
645 rcu_assign_pointer(mm->ioctx_table, table);
646 } else if (table->nr > old->nr) {
647 memcpy(table->table, old->table,
648 old->nr * sizeof(struct kioctx *));
650 rcu_assign_pointer(mm->ioctx_table, table);
659 static void aio_nr_sub(unsigned nr)
661 spin_lock(&aio_nr_lock);
662 if (WARN_ON(aio_nr - nr > aio_nr))
666 spin_unlock(&aio_nr_lock);
670 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
672 static struct kioctx *ioctx_alloc(unsigned nr_events)
674 struct mm_struct *mm = current->mm;
679 * We keep track of the number of available ringbuffer slots, to prevent
680 * overflow (reqs_available), and we also use percpu counters for this.
682 * So since up to half the slots might be on other cpu's percpu counters
683 * and unavailable, double nr_events so userspace sees what they
684 * expected: additionally, we move req_batch slots to/from percpu
685 * counters at a time, so make sure that isn't 0:
687 nr_events = max(nr_events, num_possible_cpus() * 4);
690 /* Prevent overflows */
691 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
692 pr_debug("ENOMEM: nr_events too high\n");
693 return ERR_PTR(-EINVAL);
696 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
697 return ERR_PTR(-EAGAIN);
699 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
701 return ERR_PTR(-ENOMEM);
703 ctx->max_reqs = nr_events;
705 spin_lock_init(&ctx->ctx_lock);
706 spin_lock_init(&ctx->completion_lock);
707 mutex_init(&ctx->ring_lock);
708 /* Protect against page migration throughout kiotx setup by keeping
709 * the ring_lock mutex held until setup is complete. */
710 mutex_lock(&ctx->ring_lock);
711 init_waitqueue_head(&ctx->wait);
713 INIT_LIST_HEAD(&ctx->active_reqs);
715 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
718 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
721 ctx->cpu = alloc_percpu(struct kioctx_cpu);
725 err = aio_setup_ring(ctx);
729 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
730 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
731 if (ctx->req_batch < 1)
734 /* limit the number of system wide aios */
735 spin_lock(&aio_nr_lock);
736 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
737 aio_nr + nr_events < aio_nr) {
738 spin_unlock(&aio_nr_lock);
742 aio_nr += ctx->max_reqs;
743 spin_unlock(&aio_nr_lock);
745 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
746 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
748 err = ioctx_add_table(ctx, mm);
752 /* Release the ring_lock mutex now that all setup is complete. */
753 mutex_unlock(&ctx->ring_lock);
755 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
756 ctx, ctx->user_id, mm, ctx->nr_events);
760 aio_nr_sub(ctx->max_reqs);
764 mutex_unlock(&ctx->ring_lock);
765 free_percpu(ctx->cpu);
766 percpu_ref_exit(&ctx->reqs);
767 percpu_ref_exit(&ctx->users);
768 kmem_cache_free(kioctx_cachep, ctx);
769 pr_debug("error allocating ioctx %d\n", err);
774 * Cancels all outstanding aio requests on an aio context. Used
775 * when the processes owning a context have all exited to encourage
776 * the rapid destruction of the kioctx.
778 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
779 struct completion *requests_done)
781 struct kioctx_table *table;
783 if (atomic_xchg(&ctx->dead, 1))
787 spin_lock(&mm->ioctx_lock);
788 table = rcu_dereference_raw(mm->ioctx_table);
789 WARN_ON(ctx != table->table[ctx->id]);
790 table->table[ctx->id] = NULL;
791 spin_unlock(&mm->ioctx_lock);
793 /* percpu_ref_kill() will do the necessary call_rcu() */
794 wake_up_all(&ctx->wait);
797 * It'd be more correct to do this in free_ioctx(), after all
798 * the outstanding kiocbs have finished - but by then io_destroy
799 * has already returned, so io_setup() could potentially return
800 * -EAGAIN with no ioctxs actually in use (as far as userspace
803 aio_nr_sub(ctx->max_reqs);
806 vm_munmap(ctx->mmap_base, ctx->mmap_size);
808 ctx->requests_done = requests_done;
809 percpu_ref_kill(&ctx->users);
814 * exit_aio: called when the last user of mm goes away. At this point, there is
815 * no way for any new requests to be submited or any of the io_* syscalls to be
816 * called on the context.
818 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
821 void exit_aio(struct mm_struct *mm)
823 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
829 for (i = 0; i < table->nr; ++i) {
830 struct kioctx *ctx = table->table[i];
831 struct completion requests_done =
832 COMPLETION_INITIALIZER_ONSTACK(requests_done);
837 * We don't need to bother with munmap() here - exit_mmap(mm)
838 * is coming and it'll unmap everything. And we simply can't,
839 * this is not necessarily our ->mm.
840 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
841 * that it needs to unmap the area, just set it to 0.
844 kill_ioctx(mm, ctx, &requests_done);
846 /* Wait until all IO for the context are done. */
847 wait_for_completion(&requests_done);
850 RCU_INIT_POINTER(mm->ioctx_table, NULL);
854 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
856 struct kioctx_cpu *kcpu;
859 local_irq_save(flags);
860 kcpu = this_cpu_ptr(ctx->cpu);
861 kcpu->reqs_available += nr;
863 while (kcpu->reqs_available >= ctx->req_batch * 2) {
864 kcpu->reqs_available -= ctx->req_batch;
865 atomic_add(ctx->req_batch, &ctx->reqs_available);
868 local_irq_restore(flags);
871 static bool get_reqs_available(struct kioctx *ctx)
873 struct kioctx_cpu *kcpu;
877 local_irq_save(flags);
878 kcpu = this_cpu_ptr(ctx->cpu);
879 if (!kcpu->reqs_available) {
880 int old, avail = atomic_read(&ctx->reqs_available);
883 if (avail < ctx->req_batch)
887 avail = atomic_cmpxchg(&ctx->reqs_available,
888 avail, avail - ctx->req_batch);
889 } while (avail != old);
891 kcpu->reqs_available += ctx->req_batch;
895 kcpu->reqs_available--;
897 local_irq_restore(flags);
901 /* refill_reqs_available
902 * Updates the reqs_available reference counts used for tracking the
903 * number of free slots in the completion ring. This can be called
904 * from aio_complete() (to optimistically update reqs_available) or
905 * from aio_get_req() (the we're out of events case). It must be
906 * called holding ctx->completion_lock.
908 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
911 unsigned events_in_ring, completed;
913 /* Clamp head since userland can write to it. */
914 head %= ctx->nr_events;
916 events_in_ring = tail - head;
918 events_in_ring = ctx->nr_events - (head - tail);
920 completed = ctx->completed_events;
921 if (events_in_ring < completed)
922 completed -= events_in_ring;
929 ctx->completed_events -= completed;
930 put_reqs_available(ctx, completed);
933 /* user_refill_reqs_available
934 * Called to refill reqs_available when aio_get_req() encounters an
935 * out of space in the completion ring.
937 static void user_refill_reqs_available(struct kioctx *ctx)
939 spin_lock_irq(&ctx->completion_lock);
940 if (ctx->completed_events) {
941 struct aio_ring *ring;
944 /* Access of ring->head may race with aio_read_events_ring()
945 * here, but that's okay since whether we read the old version
946 * or the new version, and either will be valid. The important
947 * part is that head cannot pass tail since we prevent
948 * aio_complete() from updating tail by holding
949 * ctx->completion_lock. Even if head is invalid, the check
950 * against ctx->completed_events below will make sure we do the
953 ring = kmap_atomic(ctx->ring_pages[0]);
957 refill_reqs_available(ctx, head, ctx->tail);
960 spin_unlock_irq(&ctx->completion_lock);
964 * Allocate a slot for an aio request.
965 * Returns NULL if no requests are free.
967 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
969 struct aio_kiocb *req;
971 if (!get_reqs_available(ctx)) {
972 user_refill_reqs_available(ctx);
973 if (!get_reqs_available(ctx))
977 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
981 percpu_ref_get(&ctx->reqs);
986 put_reqs_available(ctx, 1);
990 static void kiocb_free(struct aio_kiocb *req)
992 if (req->common.ki_filp)
993 fput(req->common.ki_filp);
994 if (req->ki_eventfd != NULL)
995 eventfd_ctx_put(req->ki_eventfd);
996 kmem_cache_free(kiocb_cachep, req);
999 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1001 struct aio_ring __user *ring = (void __user *)ctx_id;
1002 struct mm_struct *mm = current->mm;
1003 struct kioctx *ctx, *ret = NULL;
1004 struct kioctx_table *table;
1007 if (get_user(id, &ring->id))
1011 table = rcu_dereference(mm->ioctx_table);
1013 if (!table || id >= table->nr)
1016 ctx = table->table[id];
1017 if (ctx && ctx->user_id == ctx_id) {
1018 percpu_ref_get(&ctx->users);
1027 * Called when the io request on the given iocb is complete.
1029 static void aio_complete(struct kiocb *kiocb, long res, long res2)
1031 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, common);
1032 struct kioctx *ctx = iocb->ki_ctx;
1033 struct aio_ring *ring;
1034 struct io_event *ev_page, *event;
1035 unsigned tail, pos, head;
1036 unsigned long flags;
1039 * Special case handling for sync iocbs:
1040 * - events go directly into the iocb for fast handling
1041 * - the sync task with the iocb in its stack holds the single iocb
1042 * ref, no other paths have a way to get another ref
1043 * - the sync task helpfully left a reference to itself in the iocb
1045 BUG_ON(is_sync_kiocb(kiocb));
1047 if (iocb->ki_list.next) {
1048 unsigned long flags;
1050 spin_lock_irqsave(&ctx->ctx_lock, flags);
1051 list_del(&iocb->ki_list);
1052 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1056 * Add a completion event to the ring buffer. Must be done holding
1057 * ctx->completion_lock to prevent other code from messing with the tail
1058 * pointer since we might be called from irq context.
1060 spin_lock_irqsave(&ctx->completion_lock, flags);
1063 pos = tail + AIO_EVENTS_OFFSET;
1065 if (++tail >= ctx->nr_events)
1068 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1069 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1071 event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1072 event->data = iocb->ki_user_data;
1076 kunmap_atomic(ev_page);
1077 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1079 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1080 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1083 /* after flagging the request as done, we
1084 * must never even look at it again
1086 smp_wmb(); /* make event visible before updating tail */
1090 ring = kmap_atomic(ctx->ring_pages[0]);
1093 kunmap_atomic(ring);
1094 flush_dcache_page(ctx->ring_pages[0]);
1096 ctx->completed_events++;
1097 if (ctx->completed_events > 1)
1098 refill_reqs_available(ctx, head, tail);
1099 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1101 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1104 * Check if the user asked us to deliver the result through an
1105 * eventfd. The eventfd_signal() function is safe to be called
1108 if (iocb->ki_eventfd != NULL)
1109 eventfd_signal(iocb->ki_eventfd, 1);
1111 /* everything turned out well, dispose of the aiocb. */
1115 * We have to order our ring_info tail store above and test
1116 * of the wait list below outside the wait lock. This is
1117 * like in wake_up_bit() where clearing a bit has to be
1118 * ordered with the unlocked test.
1122 if (waitqueue_active(&ctx->wait))
1123 wake_up(&ctx->wait);
1125 percpu_ref_put(&ctx->reqs);
1128 /* aio_read_events_ring
1129 * Pull an event off of the ioctx's event ring. Returns the number of
1132 static long aio_read_events_ring(struct kioctx *ctx,
1133 struct io_event __user *event, long nr)
1135 struct aio_ring *ring;
1136 unsigned head, tail, pos;
1141 * The mutex can block and wake us up and that will cause
1142 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1143 * and repeat. This should be rare enough that it doesn't cause
1144 * peformance issues. See the comment in read_events() for more detail.
1146 sched_annotate_sleep();
1147 mutex_lock(&ctx->ring_lock);
1149 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1150 ring = kmap_atomic(ctx->ring_pages[0]);
1153 kunmap_atomic(ring);
1156 * Ensure that once we've read the current tail pointer, that
1157 * we also see the events that were stored up to the tail.
1161 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1166 head %= ctx->nr_events;
1167 tail %= ctx->nr_events;
1171 struct io_event *ev;
1174 avail = (head <= tail ? tail : ctx->nr_events) - head;
1178 avail = min(avail, nr - ret);
1179 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1180 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1182 pos = head + AIO_EVENTS_OFFSET;
1183 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1184 pos %= AIO_EVENTS_PER_PAGE;
1187 copy_ret = copy_to_user(event + ret, ev + pos,
1188 sizeof(*ev) * avail);
1191 if (unlikely(copy_ret)) {
1198 head %= ctx->nr_events;
1201 ring = kmap_atomic(ctx->ring_pages[0]);
1203 kunmap_atomic(ring);
1204 flush_dcache_page(ctx->ring_pages[0]);
1206 pr_debug("%li h%u t%u\n", ret, head, tail);
1208 mutex_unlock(&ctx->ring_lock);
1213 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1214 struct io_event __user *event, long *i)
1216 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1221 if (unlikely(atomic_read(&ctx->dead)))
1227 return ret < 0 || *i >= min_nr;
1230 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1231 struct io_event __user *event,
1232 struct timespec __user *timeout)
1234 ktime_t until = { .tv64 = KTIME_MAX };
1240 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1243 until = timespec_to_ktime(ts);
1247 * Note that aio_read_events() is being called as the conditional - i.e.
1248 * we're calling it after prepare_to_wait() has set task state to
1249 * TASK_INTERRUPTIBLE.
1251 * But aio_read_events() can block, and if it blocks it's going to flip
1252 * the task state back to TASK_RUNNING.
1254 * This should be ok, provided it doesn't flip the state back to
1255 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1256 * will only happen if the mutex_lock() call blocks, and we then find
1257 * the ringbuffer empty. So in practice we should be ok, but it's
1258 * something to be aware of when touching this code.
1260 if (until.tv64 == 0)
1261 aio_read_events(ctx, min_nr, nr, event, &ret);
1263 wait_event_interruptible_hrtimeout(ctx->wait,
1264 aio_read_events(ctx, min_nr, nr, event, &ret),
1267 if (!ret && signal_pending(current))
1274 * Create an aio_context capable of receiving at least nr_events.
1275 * ctxp must not point to an aio_context that already exists, and
1276 * must be initialized to 0 prior to the call. On successful
1277 * creation of the aio_context, *ctxp is filled in with the resulting
1278 * handle. May fail with -EINVAL if *ctxp is not initialized,
1279 * if the specified nr_events exceeds internal limits. May fail
1280 * with -EAGAIN if the specified nr_events exceeds the user's limit
1281 * of available events. May fail with -ENOMEM if insufficient kernel
1282 * resources are available. May fail with -EFAULT if an invalid
1283 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1286 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1288 struct kioctx *ioctx = NULL;
1292 ret = get_user(ctx, ctxp);
1297 if (unlikely(ctx || nr_events == 0)) {
1298 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1303 ioctx = ioctx_alloc(nr_events);
1304 ret = PTR_ERR(ioctx);
1305 if (!IS_ERR(ioctx)) {
1306 ret = put_user(ioctx->user_id, ctxp);
1308 kill_ioctx(current->mm, ioctx, NULL);
1309 percpu_ref_put(&ioctx->users);
1317 * Destroy the aio_context specified. May cancel any outstanding
1318 * AIOs and block on completion. Will fail with -ENOSYS if not
1319 * implemented. May fail with -EINVAL if the context pointed to
1322 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1324 struct kioctx *ioctx = lookup_ioctx(ctx);
1325 if (likely(NULL != ioctx)) {
1326 struct completion requests_done =
1327 COMPLETION_INITIALIZER_ONSTACK(requests_done);
1330 /* Pass requests_done to kill_ioctx() where it can be set
1331 * in a thread-safe way. If we try to set it here then we have
1332 * a race condition if two io_destroy() called simultaneously.
1334 ret = kill_ioctx(current->mm, ioctx, &requests_done);
1335 percpu_ref_put(&ioctx->users);
1337 /* Wait until all IO for the context are done. Otherwise kernel
1338 * keep using user-space buffers even if user thinks the context
1342 wait_for_completion(&requests_done);
1346 pr_debug("EINVAL: invalid context id\n");
1350 typedef ssize_t (rw_iter_op)(struct kiocb *, struct iov_iter *);
1352 static int aio_setup_vectored_rw(int rw, char __user *buf, size_t len,
1353 struct iovec **iovec,
1355 struct iov_iter *iter)
1357 #ifdef CONFIG_COMPAT
1359 return compat_import_iovec(rw,
1360 (struct compat_iovec __user *)buf,
1361 len, UIO_FASTIOV, iovec, iter);
1363 return import_iovec(rw, (struct iovec __user *)buf,
1364 len, UIO_FASTIOV, iovec, iter);
1369 * Performs the initial checks and io submission.
1371 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1372 char __user *buf, size_t len, bool compat)
1374 struct file *file = req->ki_filp;
1378 rw_iter_op *iter_op;
1379 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1380 struct iov_iter iter;
1383 case IOCB_CMD_PREAD:
1384 case IOCB_CMD_PREADV:
1387 iter_op = file->f_op->read_iter;
1390 case IOCB_CMD_PWRITE:
1391 case IOCB_CMD_PWRITEV:
1394 iter_op = file->f_op->write_iter;
1397 if (unlikely(!(file->f_mode & mode)))
1403 if (opcode == IOCB_CMD_PREADV || opcode == IOCB_CMD_PWRITEV)
1404 ret = aio_setup_vectored_rw(rw, buf, len,
1405 &iovec, compat, &iter);
1407 ret = import_single_range(rw, buf, len, iovec, &iter);
1411 ret = rw_verify_area(rw, file, &req->ki_pos,
1412 iov_iter_count(&iter));
1421 file_start_write(file);
1423 ret = iter_op(req, &iter);
1426 file_end_write(file);
1430 case IOCB_CMD_FDSYNC:
1431 if (!file->f_op->aio_fsync)
1434 ret = file->f_op->aio_fsync(req, 1);
1437 case IOCB_CMD_FSYNC:
1438 if (!file->f_op->aio_fsync)
1441 ret = file->f_op->aio_fsync(req, 0);
1445 pr_debug("EINVAL: no operation provided\n");
1449 if (ret != -EIOCBQUEUED) {
1451 * There's no easy way to restart the syscall since other AIO's
1452 * may be already running. Just fail this IO with EINTR.
1454 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1455 ret == -ERESTARTNOHAND ||
1456 ret == -ERESTART_RESTARTBLOCK))
1458 aio_complete(req, ret, 0);
1464 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1465 struct iocb *iocb, bool compat)
1467 struct aio_kiocb *req;
1470 /* enforce forwards compatibility on users */
1471 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1472 pr_debug("EINVAL: reserve field set\n");
1476 /* prevent overflows */
1478 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1479 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1480 ((ssize_t)iocb->aio_nbytes < 0)
1482 pr_debug("EINVAL: overflow check\n");
1486 req = aio_get_req(ctx);
1490 req->common.ki_filp = fget(iocb->aio_fildes);
1491 if (unlikely(!req->common.ki_filp)) {
1495 req->common.ki_pos = iocb->aio_offset;
1496 req->common.ki_complete = aio_complete;
1497 req->common.ki_flags = 0;
1499 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1501 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1502 * instance of the file* now. The file descriptor must be
1503 * an eventfd() fd, and will be signaled for each completed
1504 * event using the eventfd_signal() function.
1506 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1507 if (IS_ERR(req->ki_eventfd)) {
1508 ret = PTR_ERR(req->ki_eventfd);
1509 req->ki_eventfd = NULL;
1513 req->common.ki_flags |= IOCB_EVENTFD;
1516 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1517 if (unlikely(ret)) {
1518 pr_debug("EFAULT: aio_key\n");
1522 req->ki_user_iocb = user_iocb;
1523 req->ki_user_data = iocb->aio_data;
1525 ret = aio_run_iocb(&req->common, iocb->aio_lio_opcode,
1526 (char __user *)(unsigned long)iocb->aio_buf,
1534 put_reqs_available(ctx, 1);
1535 percpu_ref_put(&ctx->reqs);
1540 long do_io_submit(aio_context_t ctx_id, long nr,
1541 struct iocb __user *__user *iocbpp, bool compat)
1546 struct blk_plug plug;
1548 if (unlikely(nr < 0))
1551 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1552 nr = LONG_MAX/sizeof(*iocbpp);
1554 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1557 ctx = lookup_ioctx(ctx_id);
1558 if (unlikely(!ctx)) {
1559 pr_debug("EINVAL: invalid context id\n");
1563 blk_start_plug(&plug);
1566 * AKPM: should this return a partial result if some of the IOs were
1567 * successfully submitted?
1569 for (i=0; i<nr; i++) {
1570 struct iocb __user *user_iocb;
1573 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1578 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1583 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1587 blk_finish_plug(&plug);
1589 percpu_ref_put(&ctx->users);
1594 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1595 * the number of iocbs queued. May return -EINVAL if the aio_context
1596 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1597 * *iocbpp[0] is not properly initialized, if the operation specified
1598 * is invalid for the file descriptor in the iocb. May fail with
1599 * -EFAULT if any of the data structures point to invalid data. May
1600 * fail with -EBADF if the file descriptor specified in the first
1601 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1602 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1603 * fail with -ENOSYS if not implemented.
1605 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1606 struct iocb __user * __user *, iocbpp)
1608 return do_io_submit(ctx_id, nr, iocbpp, 0);
1612 * Finds a given iocb for cancellation.
1614 static struct aio_kiocb *
1615 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb, u32 key)
1617 struct aio_kiocb *kiocb;
1619 assert_spin_locked(&ctx->ctx_lock);
1621 if (key != KIOCB_KEY)
1624 /* TODO: use a hash or array, this sucks. */
1625 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1626 if (kiocb->ki_user_iocb == iocb)
1633 * Attempts to cancel an iocb previously passed to io_submit. If
1634 * the operation is successfully cancelled, the resulting event is
1635 * copied into the memory pointed to by result without being placed
1636 * into the completion queue and 0 is returned. May fail with
1637 * -EFAULT if any of the data structures pointed to are invalid.
1638 * May fail with -EINVAL if aio_context specified by ctx_id is
1639 * invalid. May fail with -EAGAIN if the iocb specified was not
1640 * cancelled. Will fail with -ENOSYS if not implemented.
1642 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1643 struct io_event __user *, result)
1646 struct aio_kiocb *kiocb;
1650 ret = get_user(key, &iocb->aio_key);
1654 ctx = lookup_ioctx(ctx_id);
1658 spin_lock_irq(&ctx->ctx_lock);
1660 kiocb = lookup_kiocb(ctx, iocb, key);
1662 ret = kiocb_cancel(kiocb);
1666 spin_unlock_irq(&ctx->ctx_lock);
1670 * The result argument is no longer used - the io_event is
1671 * always delivered via the ring buffer. -EINPROGRESS indicates
1672 * cancellation is progress:
1677 percpu_ref_put(&ctx->users);
1683 * Attempts to read at least min_nr events and up to nr events from
1684 * the completion queue for the aio_context specified by ctx_id. If
1685 * it succeeds, the number of read events is returned. May fail with
1686 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1687 * out of range, if timeout is out of range. May fail with -EFAULT
1688 * if any of the memory specified is invalid. May return 0 or
1689 * < min_nr if the timeout specified by timeout has elapsed
1690 * before sufficient events are available, where timeout == NULL
1691 * specifies an infinite timeout. Note that the timeout pointed to by
1692 * timeout is relative. Will fail with -ENOSYS if not implemented.
1694 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1697 struct io_event __user *, events,
1698 struct timespec __user *, timeout)
1700 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1703 if (likely(ioctx)) {
1704 if (likely(min_nr <= nr && min_nr >= 0))
1705 ret = read_events(ioctx, min_nr, nr, events, timeout);
1706 percpu_ref_put(&ioctx->users);