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/anon_inodes.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.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;
157 * Creates the slab caches used by the aio routines, panic on
158 * failure as this is done early during the boot sequence.
160 static int __init aio_setup(void)
162 kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
163 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
165 pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
169 __initcall(aio_setup);
171 static void put_aio_ring_file(struct kioctx *ctx)
173 struct file *aio_ring_file = ctx->aio_ring_file;
175 truncate_setsize(aio_ring_file->f_inode, 0);
177 /* Prevent further access to the kioctx from migratepages */
178 spin_lock(&aio_ring_file->f_inode->i_mapping->private_lock);
179 aio_ring_file->f_inode->i_mapping->private_data = NULL;
180 ctx->aio_ring_file = NULL;
181 spin_unlock(&aio_ring_file->f_inode->i_mapping->private_lock);
187 static void aio_free_ring(struct kioctx *ctx)
191 for (i = 0; i < ctx->nr_pages; i++) {
192 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
193 page_count(ctx->ring_pages[i]));
194 put_page(ctx->ring_pages[i]);
197 put_aio_ring_file(ctx);
199 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
200 kfree(ctx->ring_pages);
201 ctx->ring_pages = NULL;
205 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
207 vma->vm_ops = &generic_file_vm_ops;
211 static const struct file_operations aio_ring_fops = {
212 .mmap = aio_ring_mmap,
215 static int aio_set_page_dirty(struct page *page)
220 #if IS_ENABLED(CONFIG_MIGRATION)
221 static int aio_migratepage(struct address_space *mapping, struct page *new,
222 struct page *old, enum migrate_mode mode)
228 /* Writeback must be complete */
229 BUG_ON(PageWriteback(old));
232 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
233 if (rc != MIGRATEPAGE_SUCCESS) {
240 /* We can potentially race against kioctx teardown here. Use the
241 * address_space's private data lock to protect the mapping's
244 spin_lock(&mapping->private_lock);
245 ctx = mapping->private_data;
248 spin_lock_irqsave(&ctx->completion_lock, flags);
249 migrate_page_copy(new, old);
251 if (idx < (pgoff_t)ctx->nr_pages)
252 ctx->ring_pages[idx] = new;
253 spin_unlock_irqrestore(&ctx->completion_lock, flags);
256 spin_unlock(&mapping->private_lock);
262 static const struct address_space_operations aio_ctx_aops = {
263 .set_page_dirty = aio_set_page_dirty,
264 #if IS_ENABLED(CONFIG_MIGRATION)
265 .migratepage = aio_migratepage,
269 static int aio_setup_ring(struct kioctx *ctx)
271 struct aio_ring *ring;
272 unsigned nr_events = ctx->max_reqs;
273 struct mm_struct *mm = current->mm;
274 unsigned long size, populate;
279 /* Compensate for the ring buffer's head/tail overlap entry */
280 nr_events += 2; /* 1 is required, 2 for good luck */
282 size = sizeof(struct aio_ring);
283 size += sizeof(struct io_event) * nr_events;
285 nr_pages = PFN_UP(size);
289 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
291 ctx->aio_ring_file = NULL;
295 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
296 file->f_inode->i_mapping->private_data = ctx;
297 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
299 for (i = 0; i < nr_pages; i++) {
301 page = find_or_create_page(file->f_inode->i_mapping,
302 i, GFP_HIGHUSER | __GFP_ZERO);
305 pr_debug("pid(%d) page[%d]->count=%d\n",
306 current->pid, i, page_count(page));
307 SetPageUptodate(page);
311 ctx->aio_ring_file = file;
312 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
313 / sizeof(struct io_event);
315 ctx->ring_pages = ctx->internal_pages;
316 if (nr_pages > AIO_RING_PAGES) {
317 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
319 if (!ctx->ring_pages)
323 ctx->mmap_size = nr_pages * PAGE_SIZE;
324 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
326 down_write(&mm->mmap_sem);
327 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
328 PROT_READ | PROT_WRITE,
329 MAP_SHARED | MAP_POPULATE, 0, &populate);
330 if (IS_ERR((void *)ctx->mmap_base)) {
331 up_write(&mm->mmap_sem);
337 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
339 /* We must do this while still holding mmap_sem for write, as we
340 * need to be protected against userspace attempting to mremap()
341 * or munmap() the ring buffer.
343 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
344 1, 0, ctx->ring_pages, NULL);
346 /* Dropping the reference here is safe as the page cache will hold
347 * onto the pages for us. It is also required so that page migration
348 * can unmap the pages and get the right reference count.
350 for (i = 0; i < ctx->nr_pages; i++)
351 put_page(ctx->ring_pages[i]);
353 up_write(&mm->mmap_sem);
355 if (unlikely(ctx->nr_pages != nr_pages)) {
360 ctx->user_id = ctx->mmap_base;
361 ctx->nr_events = nr_events; /* trusted copy */
363 ring = kmap_atomic(ctx->ring_pages[0]);
364 ring->nr = nr_events; /* user copy */
366 ring->head = ring->tail = 0;
367 ring->magic = AIO_RING_MAGIC;
368 ring->compat_features = AIO_RING_COMPAT_FEATURES;
369 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
370 ring->header_length = sizeof(struct aio_ring);
372 flush_dcache_page(ctx->ring_pages[0]);
377 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
378 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
379 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
381 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
383 struct kioctx *ctx = req->ki_ctx;
386 spin_lock_irqsave(&ctx->ctx_lock, flags);
388 if (!req->ki_list.next)
389 list_add(&req->ki_list, &ctx->active_reqs);
391 req->ki_cancel = cancel;
393 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
395 EXPORT_SYMBOL(kiocb_set_cancel_fn);
397 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
399 kiocb_cancel_fn *old, *cancel;
402 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
403 * actually has a cancel function, hence the cmpxchg()
406 cancel = ACCESS_ONCE(kiocb->ki_cancel);
408 if (!cancel || cancel == KIOCB_CANCELLED)
412 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
413 } while (cancel != old);
415 return cancel(kiocb);
418 static void free_ioctx(struct work_struct *work)
420 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
422 pr_debug("freeing %p\n", ctx);
425 free_percpu(ctx->cpu);
426 kmem_cache_free(kioctx_cachep, ctx);
429 static void free_ioctx_reqs(struct percpu_ref *ref)
431 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
433 INIT_WORK(&ctx->free_work, free_ioctx);
434 schedule_work(&ctx->free_work);
438 * When this function runs, the kioctx has been removed from the "hash table"
439 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
440 * now it's safe to cancel any that need to be.
442 static void free_ioctx_users(struct percpu_ref *ref)
444 struct kioctx *ctx = container_of(ref, struct kioctx, users);
447 spin_lock_irq(&ctx->ctx_lock);
449 while (!list_empty(&ctx->active_reqs)) {
450 req = list_first_entry(&ctx->active_reqs,
451 struct kiocb, ki_list);
453 list_del_init(&req->ki_list);
454 kiocb_cancel(ctx, req);
457 spin_unlock_irq(&ctx->ctx_lock);
459 percpu_ref_kill(&ctx->reqs);
460 percpu_ref_put(&ctx->reqs);
463 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
466 struct kioctx_table *table, *old;
467 struct aio_ring *ring;
469 spin_lock(&mm->ioctx_lock);
471 table = rcu_dereference(mm->ioctx_table);
475 for (i = 0; i < table->nr; i++)
476 if (!table->table[i]) {
478 table->table[i] = ctx;
480 spin_unlock(&mm->ioctx_lock);
482 ring = kmap_atomic(ctx->ring_pages[0]);
488 new_nr = (table ? table->nr : 1) * 4;
491 spin_unlock(&mm->ioctx_lock);
493 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
500 spin_lock(&mm->ioctx_lock);
502 old = rcu_dereference(mm->ioctx_table);
505 rcu_assign_pointer(mm->ioctx_table, table);
506 } else if (table->nr > old->nr) {
507 memcpy(table->table, old->table,
508 old->nr * sizeof(struct kioctx *));
510 rcu_assign_pointer(mm->ioctx_table, table);
519 static void aio_nr_sub(unsigned nr)
521 spin_lock(&aio_nr_lock);
522 if (WARN_ON(aio_nr - nr > aio_nr))
526 spin_unlock(&aio_nr_lock);
530 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
532 static struct kioctx *ioctx_alloc(unsigned nr_events)
534 struct mm_struct *mm = current->mm;
539 * We keep track of the number of available ringbuffer slots, to prevent
540 * overflow (reqs_available), and we also use percpu counters for this.
542 * So since up to half the slots might be on other cpu's percpu counters
543 * and unavailable, double nr_events so userspace sees what they
544 * expected: additionally, we move req_batch slots to/from percpu
545 * counters at a time, so make sure that isn't 0:
547 nr_events = max(nr_events, num_possible_cpus() * 4);
550 /* Prevent overflows */
551 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
552 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
553 pr_debug("ENOMEM: nr_events too high\n");
554 return ERR_PTR(-EINVAL);
557 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
558 return ERR_PTR(-EAGAIN);
560 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
562 return ERR_PTR(-ENOMEM);
564 ctx->max_reqs = nr_events;
566 if (percpu_ref_init(&ctx->users, free_ioctx_users))
569 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
572 spin_lock_init(&ctx->ctx_lock);
573 spin_lock_init(&ctx->completion_lock);
574 mutex_init(&ctx->ring_lock);
575 init_waitqueue_head(&ctx->wait);
577 INIT_LIST_HEAD(&ctx->active_reqs);
579 ctx->cpu = alloc_percpu(struct kioctx_cpu);
583 if (aio_setup_ring(ctx) < 0)
586 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
587 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
588 if (ctx->req_batch < 1)
591 /* limit the number of system wide aios */
592 spin_lock(&aio_nr_lock);
593 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
594 aio_nr + nr_events < aio_nr) {
595 spin_unlock(&aio_nr_lock);
599 aio_nr += ctx->max_reqs;
600 spin_unlock(&aio_nr_lock);
602 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
604 err = ioctx_add_table(ctx, mm);
608 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
609 ctx, ctx->user_id, mm, ctx->nr_events);
613 aio_nr_sub(ctx->max_reqs);
615 free_percpu(ctx->cpu);
616 free_percpu(ctx->reqs.pcpu_count);
617 free_percpu(ctx->users.pcpu_count);
618 kmem_cache_free(kioctx_cachep, ctx);
619 pr_debug("error allocating ioctx %d\n", err);
624 * Cancels all outstanding aio requests on an aio context. Used
625 * when the processes owning a context have all exited to encourage
626 * the rapid destruction of the kioctx.
628 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
630 if (!atomic_xchg(&ctx->dead, 1)) {
631 struct kioctx_table *table;
633 spin_lock(&mm->ioctx_lock);
635 table = rcu_dereference(mm->ioctx_table);
637 WARN_ON(ctx != table->table[ctx->id]);
638 table->table[ctx->id] = NULL;
640 spin_unlock(&mm->ioctx_lock);
642 /* percpu_ref_kill() will do the necessary call_rcu() */
643 wake_up_all(&ctx->wait);
646 * It'd be more correct to do this in free_ioctx(), after all
647 * the outstanding kiocbs have finished - but by then io_destroy
648 * has already returned, so io_setup() could potentially return
649 * -EAGAIN with no ioctxs actually in use (as far as userspace
652 aio_nr_sub(ctx->max_reqs);
655 vm_munmap(ctx->mmap_base, ctx->mmap_size);
657 percpu_ref_kill(&ctx->users);
661 /* wait_on_sync_kiocb:
662 * Waits on the given sync kiocb to complete.
664 ssize_t wait_on_sync_kiocb(struct kiocb *req)
666 while (!req->ki_ctx) {
667 set_current_state(TASK_UNINTERRUPTIBLE);
672 __set_current_state(TASK_RUNNING);
673 return req->ki_user_data;
675 EXPORT_SYMBOL(wait_on_sync_kiocb);
678 * exit_aio: called when the last user of mm goes away. At this point, there is
679 * no way for any new requests to be submited or any of the io_* syscalls to be
680 * called on the context.
682 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
685 void exit_aio(struct mm_struct *mm)
687 struct kioctx_table *table;
693 table = rcu_dereference(mm->ioctx_table);
696 if (!table || i >= table->nr) {
698 rcu_assign_pointer(mm->ioctx_table, NULL);
704 ctx = table->table[i++];
710 * We don't need to bother with munmap() here -
711 * exit_mmap(mm) is coming and it'll unmap everything.
712 * Since aio_free_ring() uses non-zero ->mmap_size
713 * as indicator that it needs to unmap the area,
714 * just set it to 0; aio_free_ring() is the only
715 * place that uses ->mmap_size, so it's safe.
723 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
725 struct kioctx_cpu *kcpu;
728 kcpu = this_cpu_ptr(ctx->cpu);
730 kcpu->reqs_available += nr;
731 while (kcpu->reqs_available >= ctx->req_batch * 2) {
732 kcpu->reqs_available -= ctx->req_batch;
733 atomic_add(ctx->req_batch, &ctx->reqs_available);
739 static bool get_reqs_available(struct kioctx *ctx)
741 struct kioctx_cpu *kcpu;
745 kcpu = this_cpu_ptr(ctx->cpu);
747 if (!kcpu->reqs_available) {
748 int old, avail = atomic_read(&ctx->reqs_available);
751 if (avail < ctx->req_batch)
755 avail = atomic_cmpxchg(&ctx->reqs_available,
756 avail, avail - ctx->req_batch);
757 } while (avail != old);
759 kcpu->reqs_available += ctx->req_batch;
763 kcpu->reqs_available--;
770 * Allocate a slot for an aio request.
771 * Returns NULL if no requests are free.
773 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
777 if (!get_reqs_available(ctx))
780 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
784 percpu_ref_get(&ctx->reqs);
789 put_reqs_available(ctx, 1);
793 static void kiocb_free(struct kiocb *req)
797 if (req->ki_eventfd != NULL)
798 eventfd_ctx_put(req->ki_eventfd);
799 kmem_cache_free(kiocb_cachep, req);
802 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
804 struct aio_ring __user *ring = (void __user *)ctx_id;
805 struct mm_struct *mm = current->mm;
806 struct kioctx *ctx, *ret = NULL;
807 struct kioctx_table *table;
810 if (get_user(id, &ring->id))
814 table = rcu_dereference(mm->ioctx_table);
816 if (!table || id >= table->nr)
819 ctx = table->table[id];
820 if (ctx && ctx->user_id == ctx_id) {
821 percpu_ref_get(&ctx->users);
830 * Called when the io request on the given iocb is complete.
832 void aio_complete(struct kiocb *iocb, long res, long res2)
834 struct kioctx *ctx = iocb->ki_ctx;
835 struct aio_ring *ring;
836 struct io_event *ev_page, *event;
841 * Special case handling for sync iocbs:
842 * - events go directly into the iocb for fast handling
843 * - the sync task with the iocb in its stack holds the single iocb
844 * ref, no other paths have a way to get another ref
845 * - the sync task helpfully left a reference to itself in the iocb
847 if (is_sync_kiocb(iocb)) {
848 iocb->ki_user_data = res;
850 iocb->ki_ctx = ERR_PTR(-EXDEV);
851 wake_up_process(iocb->ki_obj.tsk);
855 if (iocb->ki_list.next) {
858 spin_lock_irqsave(&ctx->ctx_lock, flags);
859 list_del(&iocb->ki_list);
860 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
864 * Add a completion event to the ring buffer. Must be done holding
865 * ctx->completion_lock to prevent other code from messing with the tail
866 * pointer since we might be called from irq context.
868 spin_lock_irqsave(&ctx->completion_lock, flags);
871 pos = tail + AIO_EVENTS_OFFSET;
873 if (++tail >= ctx->nr_events)
876 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
877 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
879 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
880 event->data = iocb->ki_user_data;
884 kunmap_atomic(ev_page);
885 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
887 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
888 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
891 /* after flagging the request as done, we
892 * must never even look at it again
894 smp_wmb(); /* make event visible before updating tail */
898 ring = kmap_atomic(ctx->ring_pages[0]);
901 flush_dcache_page(ctx->ring_pages[0]);
903 spin_unlock_irqrestore(&ctx->completion_lock, flags);
905 pr_debug("added to ring %p at [%u]\n", iocb, tail);
908 * Check if the user asked us to deliver the result through an
909 * eventfd. The eventfd_signal() function is safe to be called
912 if (iocb->ki_eventfd != NULL)
913 eventfd_signal(iocb->ki_eventfd, 1);
915 /* everything turned out well, dispose of the aiocb. */
919 * We have to order our ring_info tail store above and test
920 * of the wait list below outside the wait lock. This is
921 * like in wake_up_bit() where clearing a bit has to be
922 * ordered with the unlocked test.
926 if (waitqueue_active(&ctx->wait))
929 percpu_ref_put(&ctx->reqs);
931 EXPORT_SYMBOL(aio_complete);
934 * Pull an event off of the ioctx's event ring. Returns the number of
937 static long aio_read_events_ring(struct kioctx *ctx,
938 struct io_event __user *event, long nr)
940 struct aio_ring *ring;
941 unsigned head, tail, pos;
945 mutex_lock(&ctx->ring_lock);
947 ring = kmap_atomic(ctx->ring_pages[0]);
952 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
962 avail = (head <= tail ? tail : ctx->nr_events) - head;
966 avail = min(avail, nr - ret);
967 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
968 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
970 pos = head + AIO_EVENTS_OFFSET;
971 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
972 pos %= AIO_EVENTS_PER_PAGE;
975 copy_ret = copy_to_user(event + ret, ev + pos,
976 sizeof(*ev) * avail);
979 if (unlikely(copy_ret)) {
986 head %= ctx->nr_events;
989 ring = kmap_atomic(ctx->ring_pages[0]);
992 flush_dcache_page(ctx->ring_pages[0]);
994 pr_debug("%li h%u t%u\n", ret, head, tail);
996 put_reqs_available(ctx, ret);
998 mutex_unlock(&ctx->ring_lock);
1003 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1004 struct io_event __user *event, long *i)
1006 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1011 if (unlikely(atomic_read(&ctx->dead)))
1017 return ret < 0 || *i >= min_nr;
1020 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1021 struct io_event __user *event,
1022 struct timespec __user *timeout)
1024 ktime_t until = { .tv64 = KTIME_MAX };
1030 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1033 until = timespec_to_ktime(ts);
1037 * Note that aio_read_events() is being called as the conditional - i.e.
1038 * we're calling it after prepare_to_wait() has set task state to
1039 * TASK_INTERRUPTIBLE.
1041 * But aio_read_events() can block, and if it blocks it's going to flip
1042 * the task state back to TASK_RUNNING.
1044 * This should be ok, provided it doesn't flip the state back to
1045 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1046 * will only happen if the mutex_lock() call blocks, and we then find
1047 * the ringbuffer empty. So in practice we should be ok, but it's
1048 * something to be aware of when touching this code.
1050 wait_event_interruptible_hrtimeout(ctx->wait,
1051 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1053 if (!ret && signal_pending(current))
1060 * Create an aio_context capable of receiving at least nr_events.
1061 * ctxp must not point to an aio_context that already exists, and
1062 * must be initialized to 0 prior to the call. On successful
1063 * creation of the aio_context, *ctxp is filled in with the resulting
1064 * handle. May fail with -EINVAL if *ctxp is not initialized,
1065 * if the specified nr_events exceeds internal limits. May fail
1066 * with -EAGAIN if the specified nr_events exceeds the user's limit
1067 * of available events. May fail with -ENOMEM if insufficient kernel
1068 * resources are available. May fail with -EFAULT if an invalid
1069 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1072 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1074 struct kioctx *ioctx = NULL;
1078 ret = get_user(ctx, ctxp);
1083 if (unlikely(ctx || nr_events == 0)) {
1084 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1089 ioctx = ioctx_alloc(nr_events);
1090 ret = PTR_ERR(ioctx);
1091 if (!IS_ERR(ioctx)) {
1092 ret = put_user(ioctx->user_id, ctxp);
1094 kill_ioctx(current->mm, ioctx);
1095 percpu_ref_put(&ioctx->users);
1103 * Destroy the aio_context specified. May cancel any outstanding
1104 * AIOs and block on completion. Will fail with -ENOSYS if not
1105 * implemented. May fail with -EINVAL if the context pointed to
1108 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1110 struct kioctx *ioctx = lookup_ioctx(ctx);
1111 if (likely(NULL != ioctx)) {
1112 kill_ioctx(current->mm, ioctx);
1113 percpu_ref_put(&ioctx->users);
1116 pr_debug("EINVAL: io_destroy: invalid context id\n");
1120 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1121 unsigned long, loff_t);
1123 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1124 int rw, char __user *buf,
1125 unsigned long *nr_segs,
1126 struct iovec **iovec,
1131 *nr_segs = kiocb->ki_nbytes;
1133 #ifdef CONFIG_COMPAT
1135 ret = compat_rw_copy_check_uvector(rw,
1136 (struct compat_iovec __user *)buf,
1137 *nr_segs, 1, *iovec, iovec);
1140 ret = rw_copy_check_uvector(rw,
1141 (struct iovec __user *)buf,
1142 *nr_segs, 1, *iovec, iovec);
1146 /* ki_nbytes now reflect bytes instead of segs */
1147 kiocb->ki_nbytes = ret;
1151 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1152 int rw, char __user *buf,
1153 unsigned long *nr_segs,
1154 struct iovec *iovec)
1156 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1159 iovec->iov_base = buf;
1160 iovec->iov_len = kiocb->ki_nbytes;
1167 * Performs the initial checks and aio retry method
1168 * setup for the kiocb at the time of io submission.
1170 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1171 char __user *buf, bool compat)
1173 struct file *file = req->ki_filp;
1175 unsigned long nr_segs;
1179 struct iovec inline_vec, *iovec = &inline_vec;
1182 case IOCB_CMD_PREAD:
1183 case IOCB_CMD_PREADV:
1186 rw_op = file->f_op->aio_read;
1189 case IOCB_CMD_PWRITE:
1190 case IOCB_CMD_PWRITEV:
1193 rw_op = file->f_op->aio_write;
1196 if (unlikely(!(file->f_mode & mode)))
1202 ret = (opcode == IOCB_CMD_PREADV ||
1203 opcode == IOCB_CMD_PWRITEV)
1204 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1206 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1211 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1213 if (iovec != &inline_vec)
1218 req->ki_nbytes = ret;
1220 /* XXX: move/kill - rw_verify_area()? */
1221 /* This matches the pread()/pwrite() logic */
1222 if (req->ki_pos < 0) {
1228 file_start_write(file);
1230 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1233 file_end_write(file);
1236 case IOCB_CMD_FDSYNC:
1237 if (!file->f_op->aio_fsync)
1240 ret = file->f_op->aio_fsync(req, 1);
1243 case IOCB_CMD_FSYNC:
1244 if (!file->f_op->aio_fsync)
1247 ret = file->f_op->aio_fsync(req, 0);
1251 pr_debug("EINVAL: no operation provided\n");
1255 if (iovec != &inline_vec)
1258 if (ret != -EIOCBQUEUED) {
1260 * There's no easy way to restart the syscall since other AIO's
1261 * may be already running. Just fail this IO with EINTR.
1263 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1264 ret == -ERESTARTNOHAND ||
1265 ret == -ERESTART_RESTARTBLOCK))
1267 aio_complete(req, ret, 0);
1273 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1274 struct iocb *iocb, bool compat)
1279 /* enforce forwards compatibility on users */
1280 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1281 pr_debug("EINVAL: reserve field set\n");
1285 /* prevent overflows */
1287 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1288 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1289 ((ssize_t)iocb->aio_nbytes < 0)
1291 pr_debug("EINVAL: io_submit: overflow check\n");
1295 req = aio_get_req(ctx);
1299 req->ki_filp = fget(iocb->aio_fildes);
1300 if (unlikely(!req->ki_filp)) {
1305 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1307 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1308 * instance of the file* now. The file descriptor must be
1309 * an eventfd() fd, and will be signaled for each completed
1310 * event using the eventfd_signal() function.
1312 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1313 if (IS_ERR(req->ki_eventfd)) {
1314 ret = PTR_ERR(req->ki_eventfd);
1315 req->ki_eventfd = NULL;
1320 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1321 if (unlikely(ret)) {
1322 pr_debug("EFAULT: aio_key\n");
1326 req->ki_obj.user = user_iocb;
1327 req->ki_user_data = iocb->aio_data;
1328 req->ki_pos = iocb->aio_offset;
1329 req->ki_nbytes = iocb->aio_nbytes;
1331 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1332 (char __user *)(unsigned long)iocb->aio_buf,
1339 put_reqs_available(ctx, 1);
1340 percpu_ref_put(&ctx->reqs);
1345 long do_io_submit(aio_context_t ctx_id, long nr,
1346 struct iocb __user *__user *iocbpp, bool compat)
1351 struct blk_plug plug;
1353 if (unlikely(nr < 0))
1356 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1357 nr = LONG_MAX/sizeof(*iocbpp);
1359 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1362 ctx = lookup_ioctx(ctx_id);
1363 if (unlikely(!ctx)) {
1364 pr_debug("EINVAL: invalid context id\n");
1368 blk_start_plug(&plug);
1371 * AKPM: should this return a partial result if some of the IOs were
1372 * successfully submitted?
1374 for (i=0; i<nr; i++) {
1375 struct iocb __user *user_iocb;
1378 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1383 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1388 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1392 blk_finish_plug(&plug);
1394 percpu_ref_put(&ctx->users);
1399 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1400 * the number of iocbs queued. May return -EINVAL if the aio_context
1401 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1402 * *iocbpp[0] is not properly initialized, if the operation specified
1403 * is invalid for the file descriptor in the iocb. May fail with
1404 * -EFAULT if any of the data structures point to invalid data. May
1405 * fail with -EBADF if the file descriptor specified in the first
1406 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1407 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1408 * fail with -ENOSYS if not implemented.
1410 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1411 struct iocb __user * __user *, iocbpp)
1413 return do_io_submit(ctx_id, nr, iocbpp, 0);
1417 * Finds a given iocb for cancellation.
1419 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1422 struct list_head *pos;
1424 assert_spin_locked(&ctx->ctx_lock);
1426 if (key != KIOCB_KEY)
1429 /* TODO: use a hash or array, this sucks. */
1430 list_for_each(pos, &ctx->active_reqs) {
1431 struct kiocb *kiocb = list_kiocb(pos);
1432 if (kiocb->ki_obj.user == iocb)
1439 * Attempts to cancel an iocb previously passed to io_submit. If
1440 * the operation is successfully cancelled, the resulting event is
1441 * copied into the memory pointed to by result without being placed
1442 * into the completion queue and 0 is returned. May fail with
1443 * -EFAULT if any of the data structures pointed to are invalid.
1444 * May fail with -EINVAL if aio_context specified by ctx_id is
1445 * invalid. May fail with -EAGAIN if the iocb specified was not
1446 * cancelled. Will fail with -ENOSYS if not implemented.
1448 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1449 struct io_event __user *, result)
1452 struct kiocb *kiocb;
1456 ret = get_user(key, &iocb->aio_key);
1460 ctx = lookup_ioctx(ctx_id);
1464 spin_lock_irq(&ctx->ctx_lock);
1466 kiocb = lookup_kiocb(ctx, iocb, key);
1468 ret = kiocb_cancel(ctx, kiocb);
1472 spin_unlock_irq(&ctx->ctx_lock);
1476 * The result argument is no longer used - the io_event is
1477 * always delivered via the ring buffer. -EINPROGRESS indicates
1478 * cancellation is progress:
1483 percpu_ref_put(&ctx->users);
1489 * Attempts to read at least min_nr events and up to nr events from
1490 * the completion queue for the aio_context specified by ctx_id. If
1491 * it succeeds, the number of read events is returned. May fail with
1492 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1493 * out of range, if timeout is out of range. May fail with -EFAULT
1494 * if any of the memory specified is invalid. May return 0 or
1495 * < min_nr if the timeout specified by timeout has elapsed
1496 * before sufficient events are available, where timeout == NULL
1497 * specifies an infinite timeout. Note that the timeout pointed to by
1498 * timeout is relative. Will fail with -ENOSYS if not implemented.
1500 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1503 struct io_event __user *, events,
1504 struct timespec __user *, timeout)
1506 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1509 if (likely(ioctx)) {
1510 if (likely(min_nr <= nr && min_nr >= 0))
1511 ret = read_events(ioctx, min_nr, nr, events, timeout);
1512 percpu_ref_put(&ioctx->users);