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 */
59 unsigned compat_features;
60 unsigned incompat_features;
61 unsigned header_length; /* size of aio_ring */
64 struct io_event io_events[0];
65 }; /* 128 bytes + ring size */
67 #define AIO_RING_PAGES 8
72 struct kioctx *table[];
76 unsigned reqs_available;
80 struct percpu_ref users;
83 struct percpu_ref reqs;
85 unsigned long user_id;
87 struct __percpu kioctx_cpu *cpu;
90 * For percpu reqs_available, number of slots we move to/from global
95 * This is what userspace passed to io_setup(), it's not used for
96 * anything but counting against the global max_reqs quota.
98 * The real limit is nr_events - 1, which will be larger (see
103 /* Size of ringbuffer, in units of struct io_event */
106 unsigned long mmap_base;
107 unsigned long mmap_size;
109 struct page **ring_pages;
112 struct work_struct free_work;
116 * This counts the number of available slots in the ringbuffer,
117 * so we avoid overflowing it: it's decremented (if positive)
118 * when allocating a kiocb and incremented when the resulting
119 * io_event is pulled off the ringbuffer.
121 * We batch accesses to it with a percpu version.
123 atomic_t reqs_available;
124 } ____cacheline_aligned_in_smp;
128 struct list_head active_reqs; /* used for cancellation */
129 } ____cacheline_aligned_in_smp;
132 struct mutex ring_lock;
133 wait_queue_head_t wait;
134 } ____cacheline_aligned_in_smp;
138 spinlock_t completion_lock;
139 } ____cacheline_aligned_in_smp;
141 struct page *internal_pages[AIO_RING_PAGES];
142 struct file *aio_ring_file;
147 /*------ sysctl variables----*/
148 static DEFINE_SPINLOCK(aio_nr_lock);
149 unsigned long aio_nr; /* current system wide number of aio requests */
150 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
151 /*----end sysctl variables---*/
153 static struct kmem_cache *kiocb_cachep;
154 static struct kmem_cache *kioctx_cachep;
156 static struct vfsmount *aio_mnt;
158 static const struct file_operations aio_ring_fops;
159 static const struct address_space_operations aio_ctx_aops;
161 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
163 struct qstr this = QSTR_INIT("[aio]", 5);
166 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
168 return ERR_CAST(inode);
170 inode->i_mapping->a_ops = &aio_ctx_aops;
171 inode->i_mapping->private_data = ctx;
172 inode->i_size = PAGE_SIZE * nr_pages;
174 path.dentry = d_alloc_pseudo(aio_mnt->mnt_sb, &this);
177 return ERR_PTR(-ENOMEM);
179 path.mnt = mntget(aio_mnt);
181 d_instantiate(path.dentry, inode);
182 file = alloc_file(&path, FMODE_READ | FMODE_WRITE, &aio_ring_fops);
188 file->f_flags = O_RDWR;
189 file->private_data = ctx;
193 static struct dentry *aio_mount(struct file_system_type *fs_type,
194 int flags, const char *dev_name, void *data)
196 static const struct dentry_operations ops = {
197 .d_dname = simple_dname,
199 return mount_pseudo(fs_type, "aio:", NULL, &ops, 0xa10a10a1);
203 * Creates the slab caches used by the aio routines, panic on
204 * failure as this is done early during the boot sequence.
206 static int __init aio_setup(void)
208 static struct file_system_type aio_fs = {
211 .kill_sb = kill_anon_super,
213 aio_mnt = kern_mount(&aio_fs);
215 panic("Failed to create aio fs mount.");
217 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
218 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
220 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
224 __initcall(aio_setup);
226 static void put_aio_ring_file(struct kioctx *ctx)
228 struct file *aio_ring_file = ctx->aio_ring_file;
230 truncate_setsize(aio_ring_file->f_inode, 0);
232 /* Prevent further access to the kioctx from migratepages */
233 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
234 aio_ring_file->f_inode->i_mapping->private_data = NULL;
235 ctx->aio_ring_file = NULL;
236 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
242 static void aio_free_ring(struct kioctx *ctx)
246 for (i = 0; i < ctx->nr_pages; i++) {
247 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
248 page_count(ctx->ring_pages[i]));
249 put_page(ctx->ring_pages[i]);
252 put_aio_ring_file(ctx);
254 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
255 kfree(ctx->ring_pages);
256 ctx->ring_pages = NULL;
260 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
262 vma->vm_ops = &generic_file_vm_ops;
266 static const struct file_operations aio_ring_fops = {
267 .mmap = aio_ring_mmap,
270 static int aio_set_page_dirty(struct page *page)
275 #if IS_ENABLED(CONFIG_MIGRATION)
276 static int aio_migratepage(struct address_space *mapping, struct page *new,
277 struct page *old, enum migrate_mode mode)
283 /* Writeback must be complete */
284 BUG_ON(PageWriteback(old));
287 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
288 if (rc != MIGRATEPAGE_SUCCESS) {
295 /* We can potentially race against kioctx teardown here. Use the
296 * address_space's private data lock to protect the mapping's
299 spin_lock(&mapping->private_lock);
300 ctx = mapping->private_data;
303 spin_lock_irqsave(&ctx->completion_lock, flags);
304 migrate_page_copy(new, old);
306 if (idx < (pgoff_t)ctx->nr_pages)
307 ctx->ring_pages[idx] = new;
308 spin_unlock_irqrestore(&ctx->completion_lock, flags);
311 spin_unlock(&mapping->private_lock);
317 static const struct address_space_operations aio_ctx_aops = {
318 .set_page_dirty = aio_set_page_dirty,
319 #if IS_ENABLED(CONFIG_MIGRATION)
320 .migratepage = aio_migratepage,
324 static int aio_setup_ring(struct kioctx *ctx)
326 struct aio_ring *ring;
327 unsigned nr_events = ctx->max_reqs;
328 struct mm_struct *mm = current->mm;
329 unsigned long size, populate;
334 /* Compensate for the ring buffer's head/tail overlap entry */
335 nr_events += 2; /* 1 is required, 2 for good luck */
337 size = sizeof(struct aio_ring);
338 size += sizeof(struct io_event) * nr_events;
340 nr_pages = PFN_UP(size);
344 file = aio_private_file(ctx, nr_pages);
346 ctx->aio_ring_file = NULL;
350 for (i = 0; i < nr_pages; i++) {
352 page = find_or_create_page(file->f_inode->i_mapping,
353 i, GFP_HIGHUSER | __GFP_ZERO);
356 pr_debug("pid(%d) page[%d]->count=%d\n",
357 current->pid, i, page_count(page));
358 SetPageUptodate(page);
362 ctx->aio_ring_file = file;
363 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
364 / sizeof(struct io_event);
366 ctx->ring_pages = ctx->internal_pages;
367 if (nr_pages > AIO_RING_PAGES) {
368 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
370 if (!ctx->ring_pages) {
371 put_aio_ring_file(ctx);
376 ctx->mmap_size = nr_pages * PAGE_SIZE;
377 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
379 down_write(&mm->mmap_sem);
380 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
381 PROT_READ | PROT_WRITE,
382 MAP_SHARED | MAP_POPULATE, 0, &populate);
383 if (IS_ERR((void *)ctx->mmap_base)) {
384 up_write(&mm->mmap_sem);
390 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
392 /* We must do this while still holding mmap_sem for write, as we
393 * need to be protected against userspace attempting to mremap()
394 * or munmap() the ring buffer.
396 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
397 1, 0, ctx->ring_pages, NULL);
399 /* Dropping the reference here is safe as the page cache will hold
400 * onto the pages for us. It is also required so that page migration
401 * can unmap the pages and get the right reference count.
403 for (i = 0; i < ctx->nr_pages; i++)
404 put_page(ctx->ring_pages[i]);
406 up_write(&mm->mmap_sem);
408 if (unlikely(ctx->nr_pages != nr_pages)) {
413 ctx->user_id = ctx->mmap_base;
414 ctx->nr_events = nr_events; /* trusted copy */
416 ring = kmap_atomic(ctx->ring_pages[0]);
417 ring->nr = nr_events; /* user copy */
419 ring->head = ring->tail = 0;
420 ring->magic = AIO_RING_MAGIC;
421 ring->compat_features = AIO_RING_COMPAT_FEATURES;
422 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
423 ring->header_length = sizeof(struct aio_ring);
425 flush_dcache_page(ctx->ring_pages[0]);
430 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
431 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
432 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
434 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
436 struct kioctx *ctx = req->ki_ctx;
439 spin_lock_irqsave(&ctx->ctx_lock, flags);
441 if (!req->ki_list.next)
442 list_add(&req->ki_list, &ctx->active_reqs);
444 req->ki_cancel = cancel;
446 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
448 EXPORT_SYMBOL(kiocb_set_cancel_fn);
450 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
452 kiocb_cancel_fn *old, *cancel;
455 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
456 * actually has a cancel function, hence the cmpxchg()
459 cancel = ACCESS_ONCE(kiocb->ki_cancel);
461 if (!cancel || cancel == KIOCB_CANCELLED)
465 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
466 } while (cancel != old);
468 return cancel(kiocb);
471 static void free_ioctx(struct work_struct *work)
473 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
475 pr_debug("freeing %p\n", ctx);
478 free_percpu(ctx->cpu);
479 kmem_cache_free(kioctx_cachep, ctx);
482 static void free_ioctx_reqs(struct percpu_ref *ref)
484 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
486 INIT_WORK(&ctx->free_work, free_ioctx);
487 schedule_work(&ctx->free_work);
491 * When this function runs, the kioctx has been removed from the "hash table"
492 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
493 * now it's safe to cancel any that need to be.
495 static void free_ioctx_users(struct percpu_ref *ref)
497 struct kioctx *ctx = container_of(ref, struct kioctx, users);
500 spin_lock_irq(&ctx->ctx_lock);
502 while (!list_empty(&ctx->active_reqs)) {
503 req = list_first_entry(&ctx->active_reqs,
504 struct kiocb, ki_list);
506 list_del_init(&req->ki_list);
507 kiocb_cancel(ctx, req);
510 spin_unlock_irq(&ctx->ctx_lock);
512 percpu_ref_kill(&ctx->reqs);
513 percpu_ref_put(&ctx->reqs);
516 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
519 struct kioctx_table *table, *old;
520 struct aio_ring *ring;
522 spin_lock(&mm->ioctx_lock);
524 table = rcu_dereference(mm->ioctx_table);
528 for (i = 0; i < table->nr; i++)
529 if (!table->table[i]) {
531 table->table[i] = ctx;
533 spin_unlock(&mm->ioctx_lock);
535 ring = kmap_atomic(ctx->ring_pages[0]);
541 new_nr = (table ? table->nr : 1) * 4;
544 spin_unlock(&mm->ioctx_lock);
546 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
553 spin_lock(&mm->ioctx_lock);
555 old = rcu_dereference(mm->ioctx_table);
558 rcu_assign_pointer(mm->ioctx_table, table);
559 } else if (table->nr > old->nr) {
560 memcpy(table->table, old->table,
561 old->nr * sizeof(struct kioctx *));
563 rcu_assign_pointer(mm->ioctx_table, table);
572 static void aio_nr_sub(unsigned nr)
574 spin_lock(&aio_nr_lock);
575 if (WARN_ON(aio_nr - nr > aio_nr))
579 spin_unlock(&aio_nr_lock);
583 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
585 static struct kioctx *ioctx_alloc(unsigned nr_events)
587 struct mm_struct *mm = current->mm;
592 * We keep track of the number of available ringbuffer slots, to prevent
593 * overflow (reqs_available), and we also use percpu counters for this.
595 * So since up to half the slots might be on other cpu's percpu counters
596 * and unavailable, double nr_events so userspace sees what they
597 * expected: additionally, we move req_batch slots to/from percpu
598 * counters at a time, so make sure that isn't 0:
600 nr_events = max(nr_events, num_possible_cpus() * 4);
603 /* Prevent overflows */
604 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
605 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
606 pr_debug("ENOMEM: nr_events too high\n");
607 return ERR_PTR(-EINVAL);
610 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
611 return ERR_PTR(-EAGAIN);
613 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
615 return ERR_PTR(-ENOMEM);
617 ctx->max_reqs = nr_events;
619 if (percpu_ref_init(&ctx->users, free_ioctx_users))
622 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
625 spin_lock_init(&ctx->ctx_lock);
626 spin_lock_init(&ctx->completion_lock);
627 mutex_init(&ctx->ring_lock);
628 init_waitqueue_head(&ctx->wait);
630 INIT_LIST_HEAD(&ctx->active_reqs);
632 ctx->cpu = alloc_percpu(struct kioctx_cpu);
636 if (aio_setup_ring(ctx) < 0)
639 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
640 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
641 if (ctx->req_batch < 1)
644 /* limit the number of system wide aios */
645 spin_lock(&aio_nr_lock);
646 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
647 aio_nr + nr_events < aio_nr) {
648 spin_unlock(&aio_nr_lock);
652 aio_nr += ctx->max_reqs;
653 spin_unlock(&aio_nr_lock);
655 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
657 err = ioctx_add_table(ctx, mm);
661 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
662 ctx, ctx->user_id, mm, ctx->nr_events);
666 aio_nr_sub(ctx->max_reqs);
670 free_percpu(ctx->cpu);
671 free_percpu(ctx->reqs.pcpu_count);
672 free_percpu(ctx->users.pcpu_count);
673 kmem_cache_free(kioctx_cachep, ctx);
674 pr_debug("error allocating ioctx %d\n", err);
679 * Cancels all outstanding aio requests on an aio context. Used
680 * when the processes owning a context have all exited to encourage
681 * the rapid destruction of the kioctx.
683 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
685 if (!atomic_xchg(&ctx->dead, 1)) {
686 struct kioctx_table *table;
688 spin_lock(&mm->ioctx_lock);
690 table = rcu_dereference(mm->ioctx_table);
692 WARN_ON(ctx != table->table[ctx->id]);
693 table->table[ctx->id] = NULL;
695 spin_unlock(&mm->ioctx_lock);
697 /* percpu_ref_kill() will do the necessary call_rcu() */
698 wake_up_all(&ctx->wait);
701 * It'd be more correct to do this in free_ioctx(), after all
702 * the outstanding kiocbs have finished - but by then io_destroy
703 * has already returned, so io_setup() could potentially return
704 * -EAGAIN with no ioctxs actually in use (as far as userspace
707 aio_nr_sub(ctx->max_reqs);
710 vm_munmap(ctx->mmap_base, ctx->mmap_size);
712 percpu_ref_kill(&ctx->users);
716 /* wait_on_sync_kiocb:
717 * Waits on the given sync kiocb to complete.
719 ssize_t wait_on_sync_kiocb(struct kiocb *req)
721 while (!req->ki_ctx) {
722 set_current_state(TASK_UNINTERRUPTIBLE);
727 __set_current_state(TASK_RUNNING);
728 return req->ki_user_data;
730 EXPORT_SYMBOL(wait_on_sync_kiocb);
733 * exit_aio: called when the last user of mm goes away. At this point, there is
734 * no way for any new requests to be submited or any of the io_* syscalls to be
735 * called on the context.
737 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
740 void exit_aio(struct mm_struct *mm)
742 struct kioctx_table *table;
748 table = rcu_dereference(mm->ioctx_table);
751 if (!table || i >= table->nr) {
753 rcu_assign_pointer(mm->ioctx_table, NULL);
759 ctx = table->table[i++];
765 * We don't need to bother with munmap() here -
766 * exit_mmap(mm) is coming and it'll unmap everything.
767 * Since aio_free_ring() uses non-zero ->mmap_size
768 * as indicator that it needs to unmap the area,
769 * just set it to 0; aio_free_ring() is the only
770 * place that uses ->mmap_size, so it's safe.
778 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
780 struct kioctx_cpu *kcpu;
783 kcpu = this_cpu_ptr(ctx->cpu);
785 kcpu->reqs_available += nr;
786 while (kcpu->reqs_available >= ctx->req_batch * 2) {
787 kcpu->reqs_available -= ctx->req_batch;
788 atomic_add(ctx->req_batch, &ctx->reqs_available);
794 static bool get_reqs_available(struct kioctx *ctx)
796 struct kioctx_cpu *kcpu;
800 kcpu = this_cpu_ptr(ctx->cpu);
802 if (!kcpu->reqs_available) {
803 int old, avail = atomic_read(&ctx->reqs_available);
806 if (avail < ctx->req_batch)
810 avail = atomic_cmpxchg(&ctx->reqs_available,
811 avail, avail - ctx->req_batch);
812 } while (avail != old);
814 kcpu->reqs_available += ctx->req_batch;
818 kcpu->reqs_available--;
825 * Allocate a slot for an aio request.
826 * Returns NULL if no requests are free.
828 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
832 if (!get_reqs_available(ctx))
835 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
839 percpu_ref_get(&ctx->reqs);
844 put_reqs_available(ctx, 1);
848 static void kiocb_free(struct kiocb *req)
852 if (req->ki_eventfd != NULL)
853 eventfd_ctx_put(req->ki_eventfd);
854 kmem_cache_free(kiocb_cachep, req);
857 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
859 struct aio_ring __user *ring = (void __user *)ctx_id;
860 struct mm_struct *mm = current->mm;
861 struct kioctx *ctx, *ret = NULL;
862 struct kioctx_table *table;
865 if (get_user(id, &ring->id))
869 table = rcu_dereference(mm->ioctx_table);
871 if (!table || id >= table->nr)
874 ctx = table->table[id];
875 if (ctx && ctx->user_id == ctx_id) {
876 percpu_ref_get(&ctx->users);
885 * Called when the io request on the given iocb is complete.
887 void aio_complete(struct kiocb *iocb, long res, long res2)
889 struct kioctx *ctx = iocb->ki_ctx;
890 struct aio_ring *ring;
891 struct io_event *ev_page, *event;
896 * Special case handling for sync iocbs:
897 * - events go directly into the iocb for fast handling
898 * - the sync task with the iocb in its stack holds the single iocb
899 * ref, no other paths have a way to get another ref
900 * - the sync task helpfully left a reference to itself in the iocb
902 if (is_sync_kiocb(iocb)) {
903 iocb->ki_user_data = res;
905 iocb->ki_ctx = ERR_PTR(-EXDEV);
906 wake_up_process(iocb->ki_obj.tsk);
910 if (iocb->ki_list.next) {
913 spin_lock_irqsave(&ctx->ctx_lock, flags);
914 list_del(&iocb->ki_list);
915 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
919 * Add a completion event to the ring buffer. Must be done holding
920 * ctx->completion_lock to prevent other code from messing with the tail
921 * pointer since we might be called from irq context.
923 spin_lock_irqsave(&ctx->completion_lock, flags);
926 pos = tail + AIO_EVENTS_OFFSET;
928 if (++tail >= ctx->nr_events)
931 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
932 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
934 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
935 event->data = iocb->ki_user_data;
939 kunmap_atomic(ev_page);
940 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
942 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
943 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
946 /* after flagging the request as done, we
947 * must never even look at it again
949 smp_wmb(); /* make event visible before updating tail */
953 ring = kmap_atomic(ctx->ring_pages[0]);
956 flush_dcache_page(ctx->ring_pages[0]);
958 spin_unlock_irqrestore(&ctx->completion_lock, flags);
960 pr_debug("added to ring %p at [%u]\n", iocb, tail);
963 * Check if the user asked us to deliver the result through an
964 * eventfd. The eventfd_signal() function is safe to be called
967 if (iocb->ki_eventfd != NULL)
968 eventfd_signal(iocb->ki_eventfd, 1);
970 /* everything turned out well, dispose of the aiocb. */
974 * We have to order our ring_info tail store above and test
975 * of the wait list below outside the wait lock. This is
976 * like in wake_up_bit() where clearing a bit has to be
977 * ordered with the unlocked test.
981 if (waitqueue_active(&ctx->wait))
984 percpu_ref_put(&ctx->reqs);
986 EXPORT_SYMBOL(aio_complete);
989 * Pull an event off of the ioctx's event ring. Returns the number of
992 static long aio_read_events_ring(struct kioctx *ctx,
993 struct io_event __user *event, long nr)
995 struct aio_ring *ring;
996 unsigned head, tail, pos;
1000 mutex_lock(&ctx->ring_lock);
1002 ring = kmap_atomic(ctx->ring_pages[0]);
1005 kunmap_atomic(ring);
1007 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1014 struct io_event *ev;
1017 avail = (head <= tail ? tail : ctx->nr_events) - head;
1021 avail = min(avail, nr - ret);
1022 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
1023 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
1025 pos = head + AIO_EVENTS_OFFSET;
1026 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1027 pos %= AIO_EVENTS_PER_PAGE;
1030 copy_ret = copy_to_user(event + ret, ev + pos,
1031 sizeof(*ev) * avail);
1034 if (unlikely(copy_ret)) {
1041 head %= ctx->nr_events;
1044 ring = kmap_atomic(ctx->ring_pages[0]);
1046 kunmap_atomic(ring);
1047 flush_dcache_page(ctx->ring_pages[0]);
1049 pr_debug("%li h%u t%u\n", ret, head, tail);
1051 put_reqs_available(ctx, ret);
1053 mutex_unlock(&ctx->ring_lock);
1058 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1059 struct io_event __user *event, long *i)
1061 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1066 if (unlikely(atomic_read(&ctx->dead)))
1072 return ret < 0 || *i >= min_nr;
1075 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1076 struct io_event __user *event,
1077 struct timespec __user *timeout)
1079 ktime_t until = { .tv64 = KTIME_MAX };
1085 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1088 until = timespec_to_ktime(ts);
1092 * Note that aio_read_events() is being called as the conditional - i.e.
1093 * we're calling it after prepare_to_wait() has set task state to
1094 * TASK_INTERRUPTIBLE.
1096 * But aio_read_events() can block, and if it blocks it's going to flip
1097 * the task state back to TASK_RUNNING.
1099 * This should be ok, provided it doesn't flip the state back to
1100 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1101 * will only happen if the mutex_lock() call blocks, and we then find
1102 * the ringbuffer empty. So in practice we should be ok, but it's
1103 * something to be aware of when touching this code.
1105 wait_event_interruptible_hrtimeout(ctx->wait,
1106 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1108 if (!ret && signal_pending(current))
1115 * Create an aio_context capable of receiving at least nr_events.
1116 * ctxp must not point to an aio_context that already exists, and
1117 * must be initialized to 0 prior to the call. On successful
1118 * creation of the aio_context, *ctxp is filled in with the resulting
1119 * handle. May fail with -EINVAL if *ctxp is not initialized,
1120 * if the specified nr_events exceeds internal limits. May fail
1121 * with -EAGAIN if the specified nr_events exceeds the user's limit
1122 * of available events. May fail with -ENOMEM if insufficient kernel
1123 * resources are available. May fail with -EFAULT if an invalid
1124 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1127 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1129 struct kioctx *ioctx = NULL;
1133 ret = get_user(ctx, ctxp);
1138 if (unlikely(ctx || nr_events == 0)) {
1139 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1144 ioctx = ioctx_alloc(nr_events);
1145 ret = PTR_ERR(ioctx);
1146 if (!IS_ERR(ioctx)) {
1147 ret = put_user(ioctx->user_id, ctxp);
1149 kill_ioctx(current->mm, ioctx);
1150 percpu_ref_put(&ioctx->users);
1158 * Destroy the aio_context specified. May cancel any outstanding
1159 * AIOs and block on completion. Will fail with -ENOSYS if not
1160 * implemented. May fail with -EINVAL if the context pointed to
1163 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1165 struct kioctx *ioctx = lookup_ioctx(ctx);
1166 if (likely(NULL != ioctx)) {
1167 kill_ioctx(current->mm, ioctx);
1168 percpu_ref_put(&ioctx->users);
1171 pr_debug("EINVAL: io_destroy: invalid context id\n");
1175 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1176 unsigned long, loff_t);
1178 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1179 int rw, char __user *buf,
1180 unsigned long *nr_segs,
1181 struct iovec **iovec,
1186 *nr_segs = kiocb->ki_nbytes;
1188 #ifdef CONFIG_COMPAT
1190 ret = compat_rw_copy_check_uvector(rw,
1191 (struct compat_iovec __user *)buf,
1192 *nr_segs, 1, *iovec, iovec);
1195 ret = rw_copy_check_uvector(rw,
1196 (struct iovec __user *)buf,
1197 *nr_segs, 1, *iovec, iovec);
1201 /* ki_nbytes now reflect bytes instead of segs */
1202 kiocb->ki_nbytes = ret;
1206 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1207 int rw, char __user *buf,
1208 unsigned long *nr_segs,
1209 struct iovec *iovec)
1211 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1214 iovec->iov_base = buf;
1215 iovec->iov_len = kiocb->ki_nbytes;
1222 * Performs the initial checks and aio retry method
1223 * setup for the kiocb at the time of io submission.
1225 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1226 char __user *buf, bool compat)
1228 struct file *file = req->ki_filp;
1230 unsigned long nr_segs;
1234 struct iovec inline_vec, *iovec = &inline_vec;
1237 case IOCB_CMD_PREAD:
1238 case IOCB_CMD_PREADV:
1241 rw_op = file->f_op->aio_read;
1244 case IOCB_CMD_PWRITE:
1245 case IOCB_CMD_PWRITEV:
1248 rw_op = file->f_op->aio_write;
1251 if (unlikely(!(file->f_mode & mode)))
1257 ret = (opcode == IOCB_CMD_PREADV ||
1258 opcode == IOCB_CMD_PWRITEV)
1259 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1261 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1266 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1268 if (iovec != &inline_vec)
1273 req->ki_nbytes = ret;
1275 /* XXX: move/kill - rw_verify_area()? */
1276 /* This matches the pread()/pwrite() logic */
1277 if (req->ki_pos < 0) {
1283 file_start_write(file);
1285 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1288 file_end_write(file);
1291 case IOCB_CMD_FDSYNC:
1292 if (!file->f_op->aio_fsync)
1295 ret = file->f_op->aio_fsync(req, 1);
1298 case IOCB_CMD_FSYNC:
1299 if (!file->f_op->aio_fsync)
1302 ret = file->f_op->aio_fsync(req, 0);
1306 pr_debug("EINVAL: no operation provided\n");
1310 if (iovec != &inline_vec)
1313 if (ret != -EIOCBQUEUED) {
1315 * There's no easy way to restart the syscall since other AIO's
1316 * may be already running. Just fail this IO with EINTR.
1318 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1319 ret == -ERESTARTNOHAND ||
1320 ret == -ERESTART_RESTARTBLOCK))
1322 aio_complete(req, ret, 0);
1328 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1329 struct iocb *iocb, bool compat)
1334 /* enforce forwards compatibility on users */
1335 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1336 pr_debug("EINVAL: reserve field set\n");
1340 /* prevent overflows */
1342 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1343 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1344 ((ssize_t)iocb->aio_nbytes < 0)
1346 pr_debug("EINVAL: io_submit: overflow check\n");
1350 req = aio_get_req(ctx);
1354 req->ki_filp = fget(iocb->aio_fildes);
1355 if (unlikely(!req->ki_filp)) {
1360 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1362 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1363 * instance of the file* now. The file descriptor must be
1364 * an eventfd() fd, and will be signaled for each completed
1365 * event using the eventfd_signal() function.
1367 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1368 if (IS_ERR(req->ki_eventfd)) {
1369 ret = PTR_ERR(req->ki_eventfd);
1370 req->ki_eventfd = NULL;
1375 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1376 if (unlikely(ret)) {
1377 pr_debug("EFAULT: aio_key\n");
1381 req->ki_obj.user = user_iocb;
1382 req->ki_user_data = iocb->aio_data;
1383 req->ki_pos = iocb->aio_offset;
1384 req->ki_nbytes = iocb->aio_nbytes;
1386 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1387 (char __user *)(unsigned long)iocb->aio_buf,
1394 put_reqs_available(ctx, 1);
1395 percpu_ref_put(&ctx->reqs);
1400 long do_io_submit(aio_context_t ctx_id, long nr,
1401 struct iocb __user *__user *iocbpp, bool compat)
1406 struct blk_plug plug;
1408 if (unlikely(nr < 0))
1411 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1412 nr = LONG_MAX/sizeof(*iocbpp);
1414 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1417 ctx = lookup_ioctx(ctx_id);
1418 if (unlikely(!ctx)) {
1419 pr_debug("EINVAL: invalid context id\n");
1423 blk_start_plug(&plug);
1426 * AKPM: should this return a partial result if some of the IOs were
1427 * successfully submitted?
1429 for (i=0; i<nr; i++) {
1430 struct iocb __user *user_iocb;
1433 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1438 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1443 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1447 blk_finish_plug(&plug);
1449 percpu_ref_put(&ctx->users);
1454 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1455 * the number of iocbs queued. May return -EINVAL if the aio_context
1456 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1457 * *iocbpp[0] is not properly initialized, if the operation specified
1458 * is invalid for the file descriptor in the iocb. May fail with
1459 * -EFAULT if any of the data structures point to invalid data. May
1460 * fail with -EBADF if the file descriptor specified in the first
1461 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1462 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1463 * fail with -ENOSYS if not implemented.
1465 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1466 struct iocb __user * __user *, iocbpp)
1468 return do_io_submit(ctx_id, nr, iocbpp, 0);
1472 * Finds a given iocb for cancellation.
1474 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1477 struct list_head *pos;
1479 assert_spin_locked(&ctx->ctx_lock);
1481 if (key != KIOCB_KEY)
1484 /* TODO: use a hash or array, this sucks. */
1485 list_for_each(pos, &ctx->active_reqs) {
1486 struct kiocb *kiocb = list_kiocb(pos);
1487 if (kiocb->ki_obj.user == iocb)
1494 * Attempts to cancel an iocb previously passed to io_submit. If
1495 * the operation is successfully cancelled, the resulting event is
1496 * copied into the memory pointed to by result without being placed
1497 * into the completion queue and 0 is returned. May fail with
1498 * -EFAULT if any of the data structures pointed to are invalid.
1499 * May fail with -EINVAL if aio_context specified by ctx_id is
1500 * invalid. May fail with -EAGAIN if the iocb specified was not
1501 * cancelled. Will fail with -ENOSYS if not implemented.
1503 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1504 struct io_event __user *, result)
1507 struct kiocb *kiocb;
1511 ret = get_user(key, &iocb->aio_key);
1515 ctx = lookup_ioctx(ctx_id);
1519 spin_lock_irq(&ctx->ctx_lock);
1521 kiocb = lookup_kiocb(ctx, iocb, key);
1523 ret = kiocb_cancel(ctx, kiocb);
1527 spin_unlock_irq(&ctx->ctx_lock);
1531 * The result argument is no longer used - the io_event is
1532 * always delivered via the ring buffer. -EINPROGRESS indicates
1533 * cancellation is progress:
1538 percpu_ref_put(&ctx->users);
1544 * Attempts to read at least min_nr events and up to nr events from
1545 * the completion queue for the aio_context specified by ctx_id. If
1546 * it succeeds, the number of read events is returned. May fail with
1547 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1548 * out of range, if timeout is out of range. May fail with -EFAULT
1549 * if any of the memory specified is invalid. May return 0 or
1550 * < min_nr if the timeout specified by timeout has elapsed
1551 * before sufficient events are available, where timeout == NULL
1552 * specifies an infinite timeout. Note that the timeout pointed to by
1553 * timeout is relative. Will fail with -ENOSYS if not implemented.
1555 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1558 struct io_event __user *, events,
1559 struct timespec __user *, timeout)
1561 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1564 if (likely(ioctx)) {
1565 if (likely(min_nr <= nr && min_nr >= 0))
1566 ret = read_events(ioctx, min_nr, nr, events, timeout);
1567 percpu_ref_put(&ioctx->users);