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);
203 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
205 vma->vm_ops = &generic_file_vm_ops;
209 static const struct file_operations aio_ring_fops = {
210 .mmap = aio_ring_mmap,
213 static int aio_set_page_dirty(struct page *page)
218 #if IS_ENABLED(CONFIG_MIGRATION)
219 static int aio_migratepage(struct address_space *mapping, struct page *new,
220 struct page *old, enum migrate_mode mode)
226 /* Writeback must be complete */
227 BUG_ON(PageWriteback(old));
230 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode);
231 if (rc != MIGRATEPAGE_SUCCESS) {
238 /* We can potentially race against kioctx teardown here. Use the
239 * address_space's private data lock to protect the mapping's
242 spin_lock(&mapping->private_lock);
243 ctx = mapping->private_data;
246 spin_lock_irqsave(&ctx->completion_lock, flags);
247 migrate_page_copy(new, old);
249 if (idx < (pgoff_t)ctx->nr_pages)
250 ctx->ring_pages[idx] = new;
251 spin_unlock_irqrestore(&ctx->completion_lock, flags);
254 spin_unlock(&mapping->private_lock);
260 static const struct address_space_operations aio_ctx_aops = {
261 .set_page_dirty = aio_set_page_dirty,
262 #if IS_ENABLED(CONFIG_MIGRATION)
263 .migratepage = aio_migratepage,
267 static int aio_setup_ring(struct kioctx *ctx)
269 struct aio_ring *ring;
270 unsigned nr_events = ctx->max_reqs;
271 struct mm_struct *mm = current->mm;
272 unsigned long size, populate;
277 /* Compensate for the ring buffer's head/tail overlap entry */
278 nr_events += 2; /* 1 is required, 2 for good luck */
280 size = sizeof(struct aio_ring);
281 size += sizeof(struct io_event) * nr_events;
283 nr_pages = PFN_UP(size);
287 file = anon_inode_getfile_private("[aio]", &aio_ring_fops, ctx, O_RDWR);
289 ctx->aio_ring_file = NULL;
293 file->f_inode->i_mapping->a_ops = &aio_ctx_aops;
294 file->f_inode->i_mapping->private_data = ctx;
295 file->f_inode->i_size = PAGE_SIZE * (loff_t)nr_pages;
297 for (i = 0; i < nr_pages; i++) {
299 page = find_or_create_page(file->f_inode->i_mapping,
300 i, GFP_HIGHUSER | __GFP_ZERO);
303 pr_debug("pid(%d) page[%d]->count=%d\n",
304 current->pid, i, page_count(page));
305 SetPageUptodate(page);
309 ctx->aio_ring_file = file;
310 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
311 / sizeof(struct io_event);
313 ctx->ring_pages = ctx->internal_pages;
314 if (nr_pages > AIO_RING_PAGES) {
315 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
317 if (!ctx->ring_pages)
321 ctx->mmap_size = nr_pages * PAGE_SIZE;
322 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
324 down_write(&mm->mmap_sem);
325 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
326 PROT_READ | PROT_WRITE,
327 MAP_SHARED | MAP_POPULATE, 0, &populate);
328 if (IS_ERR((void *)ctx->mmap_base)) {
329 up_write(&mm->mmap_sem);
335 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
337 /* We must do this while still holding mmap_sem for write, as we
338 * need to be protected against userspace attempting to mremap()
339 * or munmap() the ring buffer.
341 ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
342 1, 0, ctx->ring_pages, NULL);
344 /* Dropping the reference here is safe as the page cache will hold
345 * onto the pages for us. It is also required so that page migration
346 * can unmap the pages and get the right reference count.
348 for (i = 0; i < ctx->nr_pages; i++)
349 put_page(ctx->ring_pages[i]);
351 up_write(&mm->mmap_sem);
353 if (unlikely(ctx->nr_pages != nr_pages)) {
358 ctx->user_id = ctx->mmap_base;
359 ctx->nr_events = nr_events; /* trusted copy */
361 ring = kmap_atomic(ctx->ring_pages[0]);
362 ring->nr = nr_events; /* user copy */
364 ring->head = ring->tail = 0;
365 ring->magic = AIO_RING_MAGIC;
366 ring->compat_features = AIO_RING_COMPAT_FEATURES;
367 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
368 ring->header_length = sizeof(struct aio_ring);
370 flush_dcache_page(ctx->ring_pages[0]);
375 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
376 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
377 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
379 void kiocb_set_cancel_fn(struct kiocb *req, kiocb_cancel_fn *cancel)
381 struct kioctx *ctx = req->ki_ctx;
384 spin_lock_irqsave(&ctx->ctx_lock, flags);
386 if (!req->ki_list.next)
387 list_add(&req->ki_list, &ctx->active_reqs);
389 req->ki_cancel = cancel;
391 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
393 EXPORT_SYMBOL(kiocb_set_cancel_fn);
395 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb)
397 kiocb_cancel_fn *old, *cancel;
400 * Don't want to set kiocb->ki_cancel = KIOCB_CANCELLED unless it
401 * actually has a cancel function, hence the cmpxchg()
404 cancel = ACCESS_ONCE(kiocb->ki_cancel);
406 if (!cancel || cancel == KIOCB_CANCELLED)
410 cancel = cmpxchg(&kiocb->ki_cancel, old, KIOCB_CANCELLED);
411 } while (cancel != old);
413 return cancel(kiocb);
416 static void free_ioctx(struct work_struct *work)
418 struct kioctx *ctx = container_of(work, struct kioctx, free_work);
420 pr_debug("freeing %p\n", ctx);
423 free_percpu(ctx->cpu);
424 kmem_cache_free(kioctx_cachep, ctx);
427 static void free_ioctx_reqs(struct percpu_ref *ref)
429 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
431 INIT_WORK(&ctx->free_work, free_ioctx);
432 schedule_work(&ctx->free_work);
436 * When this function runs, the kioctx has been removed from the "hash table"
437 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
438 * now it's safe to cancel any that need to be.
440 static void free_ioctx_users(struct percpu_ref *ref)
442 struct kioctx *ctx = container_of(ref, struct kioctx, users);
445 spin_lock_irq(&ctx->ctx_lock);
447 while (!list_empty(&ctx->active_reqs)) {
448 req = list_first_entry(&ctx->active_reqs,
449 struct kiocb, ki_list);
451 list_del_init(&req->ki_list);
452 kiocb_cancel(ctx, req);
455 spin_unlock_irq(&ctx->ctx_lock);
457 percpu_ref_kill(&ctx->reqs);
458 percpu_ref_put(&ctx->reqs);
461 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
464 struct kioctx_table *table, *old;
465 struct aio_ring *ring;
467 spin_lock(&mm->ioctx_lock);
469 table = rcu_dereference(mm->ioctx_table);
473 for (i = 0; i < table->nr; i++)
474 if (!table->table[i]) {
476 table->table[i] = ctx;
478 spin_unlock(&mm->ioctx_lock);
480 ring = kmap_atomic(ctx->ring_pages[0]);
486 new_nr = (table ? table->nr : 1) * 4;
489 spin_unlock(&mm->ioctx_lock);
491 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
498 spin_lock(&mm->ioctx_lock);
500 old = rcu_dereference(mm->ioctx_table);
503 rcu_assign_pointer(mm->ioctx_table, table);
504 } else if (table->nr > old->nr) {
505 memcpy(table->table, old->table,
506 old->nr * sizeof(struct kioctx *));
508 rcu_assign_pointer(mm->ioctx_table, table);
517 static void aio_nr_sub(unsigned nr)
519 spin_lock(&aio_nr_lock);
520 if (WARN_ON(aio_nr - nr > aio_nr))
524 spin_unlock(&aio_nr_lock);
528 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
530 static struct kioctx *ioctx_alloc(unsigned nr_events)
532 struct mm_struct *mm = current->mm;
537 * We keep track of the number of available ringbuffer slots, to prevent
538 * overflow (reqs_available), and we also use percpu counters for this.
540 * So since up to half the slots might be on other cpu's percpu counters
541 * and unavailable, double nr_events so userspace sees what they
542 * expected: additionally, we move req_batch slots to/from percpu
543 * counters at a time, so make sure that isn't 0:
545 nr_events = max(nr_events, num_possible_cpus() * 4);
548 /* Prevent overflows */
549 if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
550 (nr_events > (0x10000000U / sizeof(struct kiocb)))) {
551 pr_debug("ENOMEM: nr_events too high\n");
552 return ERR_PTR(-EINVAL);
555 if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
556 return ERR_PTR(-EAGAIN);
558 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
560 return ERR_PTR(-ENOMEM);
562 ctx->max_reqs = nr_events;
564 if (percpu_ref_init(&ctx->users, free_ioctx_users))
567 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs))
570 spin_lock_init(&ctx->ctx_lock);
571 spin_lock_init(&ctx->completion_lock);
572 mutex_init(&ctx->ring_lock);
573 init_waitqueue_head(&ctx->wait);
575 INIT_LIST_HEAD(&ctx->active_reqs);
577 ctx->cpu = alloc_percpu(struct kioctx_cpu);
581 if (aio_setup_ring(ctx) < 0)
584 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
585 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
586 if (ctx->req_batch < 1)
589 /* limit the number of system wide aios */
590 spin_lock(&aio_nr_lock);
591 if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
592 aio_nr + nr_events < aio_nr) {
593 spin_unlock(&aio_nr_lock);
597 aio_nr += ctx->max_reqs;
598 spin_unlock(&aio_nr_lock);
600 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
602 err = ioctx_add_table(ctx, mm);
606 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
607 ctx, ctx->user_id, mm, ctx->nr_events);
611 aio_nr_sub(ctx->max_reqs);
614 free_percpu(ctx->cpu);
615 free_percpu(ctx->reqs.pcpu_count);
616 free_percpu(ctx->users.pcpu_count);
617 kmem_cache_free(kioctx_cachep, ctx);
618 pr_debug("error allocating ioctx %d\n", err);
623 * Cancels all outstanding aio requests on an aio context. Used
624 * when the processes owning a context have all exited to encourage
625 * the rapid destruction of the kioctx.
627 static void kill_ioctx(struct mm_struct *mm, struct kioctx *ctx)
629 if (!atomic_xchg(&ctx->dead, 1)) {
630 struct kioctx_table *table;
632 spin_lock(&mm->ioctx_lock);
634 table = rcu_dereference(mm->ioctx_table);
636 WARN_ON(ctx != table->table[ctx->id]);
637 table->table[ctx->id] = NULL;
639 spin_unlock(&mm->ioctx_lock);
641 /* percpu_ref_kill() will do the necessary call_rcu() */
642 wake_up_all(&ctx->wait);
645 * It'd be more correct to do this in free_ioctx(), after all
646 * the outstanding kiocbs have finished - but by then io_destroy
647 * has already returned, so io_setup() could potentially return
648 * -EAGAIN with no ioctxs actually in use (as far as userspace
651 aio_nr_sub(ctx->max_reqs);
654 vm_munmap(ctx->mmap_base, ctx->mmap_size);
656 percpu_ref_kill(&ctx->users);
660 /* wait_on_sync_kiocb:
661 * Waits on the given sync kiocb to complete.
663 ssize_t wait_on_sync_kiocb(struct kiocb *req)
665 while (!req->ki_ctx) {
666 set_current_state(TASK_UNINTERRUPTIBLE);
671 __set_current_state(TASK_RUNNING);
672 return req->ki_user_data;
674 EXPORT_SYMBOL(wait_on_sync_kiocb);
677 * exit_aio: called when the last user of mm goes away. At this point, there is
678 * no way for any new requests to be submited or any of the io_* syscalls to be
679 * called on the context.
681 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
684 void exit_aio(struct mm_struct *mm)
686 struct kioctx_table *table;
692 table = rcu_dereference(mm->ioctx_table);
695 if (!table || i >= table->nr) {
697 rcu_assign_pointer(mm->ioctx_table, NULL);
703 ctx = table->table[i++];
709 * We don't need to bother with munmap() here -
710 * exit_mmap(mm) is coming and it'll unmap everything.
711 * Since aio_free_ring() uses non-zero ->mmap_size
712 * as indicator that it needs to unmap the area,
713 * just set it to 0; aio_free_ring() is the only
714 * place that uses ->mmap_size, so it's safe.
722 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
724 struct kioctx_cpu *kcpu;
727 kcpu = this_cpu_ptr(ctx->cpu);
729 kcpu->reqs_available += nr;
730 while (kcpu->reqs_available >= ctx->req_batch * 2) {
731 kcpu->reqs_available -= ctx->req_batch;
732 atomic_add(ctx->req_batch, &ctx->reqs_available);
738 static bool get_reqs_available(struct kioctx *ctx)
740 struct kioctx_cpu *kcpu;
744 kcpu = this_cpu_ptr(ctx->cpu);
746 if (!kcpu->reqs_available) {
747 int old, avail = atomic_read(&ctx->reqs_available);
750 if (avail < ctx->req_batch)
754 avail = atomic_cmpxchg(&ctx->reqs_available,
755 avail, avail - ctx->req_batch);
756 } while (avail != old);
758 kcpu->reqs_available += ctx->req_batch;
762 kcpu->reqs_available--;
769 * Allocate a slot for an aio request.
770 * Returns NULL if no requests are free.
772 static inline struct kiocb *aio_get_req(struct kioctx *ctx)
776 if (!get_reqs_available(ctx))
779 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
783 percpu_ref_get(&ctx->reqs);
788 put_reqs_available(ctx, 1);
792 static void kiocb_free(struct kiocb *req)
796 if (req->ki_eventfd != NULL)
797 eventfd_ctx_put(req->ki_eventfd);
798 kmem_cache_free(kiocb_cachep, req);
801 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
803 struct aio_ring __user *ring = (void __user *)ctx_id;
804 struct mm_struct *mm = current->mm;
805 struct kioctx *ctx, *ret = NULL;
806 struct kioctx_table *table;
809 if (get_user(id, &ring->id))
813 table = rcu_dereference(mm->ioctx_table);
815 if (!table || id >= table->nr)
818 ctx = table->table[id];
819 if (ctx && ctx->user_id == ctx_id) {
820 percpu_ref_get(&ctx->users);
829 * Called when the io request on the given iocb is complete.
831 void aio_complete(struct kiocb *iocb, long res, long res2)
833 struct kioctx *ctx = iocb->ki_ctx;
834 struct aio_ring *ring;
835 struct io_event *ev_page, *event;
840 * Special case handling for sync iocbs:
841 * - events go directly into the iocb for fast handling
842 * - the sync task with the iocb in its stack holds the single iocb
843 * ref, no other paths have a way to get another ref
844 * - the sync task helpfully left a reference to itself in the iocb
846 if (is_sync_kiocb(iocb)) {
847 iocb->ki_user_data = res;
849 iocb->ki_ctx = ERR_PTR(-EXDEV);
850 wake_up_process(iocb->ki_obj.tsk);
854 if (iocb->ki_list.next) {
857 spin_lock_irqsave(&ctx->ctx_lock, flags);
858 list_del(&iocb->ki_list);
859 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
863 * Add a completion event to the ring buffer. Must be done holding
864 * ctx->completion_lock to prevent other code from messing with the tail
865 * pointer since we might be called from irq context.
867 spin_lock_irqsave(&ctx->completion_lock, flags);
870 pos = tail + AIO_EVENTS_OFFSET;
872 if (++tail >= ctx->nr_events)
875 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
876 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
878 event->obj = (u64)(unsigned long)iocb->ki_obj.user;
879 event->data = iocb->ki_user_data;
883 kunmap_atomic(ev_page);
884 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
886 pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
887 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
890 /* after flagging the request as done, we
891 * must never even look at it again
893 smp_wmb(); /* make event visible before updating tail */
897 ring = kmap_atomic(ctx->ring_pages[0]);
900 flush_dcache_page(ctx->ring_pages[0]);
902 spin_unlock_irqrestore(&ctx->completion_lock, flags);
904 pr_debug("added to ring %p at [%u]\n", iocb, tail);
907 * Check if the user asked us to deliver the result through an
908 * eventfd. The eventfd_signal() function is safe to be called
911 if (iocb->ki_eventfd != NULL)
912 eventfd_signal(iocb->ki_eventfd, 1);
914 /* everything turned out well, dispose of the aiocb. */
918 * We have to order our ring_info tail store above and test
919 * of the wait list below outside the wait lock. This is
920 * like in wake_up_bit() where clearing a bit has to be
921 * ordered with the unlocked test.
925 if (waitqueue_active(&ctx->wait))
928 percpu_ref_put(&ctx->reqs);
930 EXPORT_SYMBOL(aio_complete);
933 * Pull an event off of the ioctx's event ring. Returns the number of
936 static long aio_read_events_ring(struct kioctx *ctx,
937 struct io_event __user *event, long nr)
939 struct aio_ring *ring;
940 unsigned head, tail, pos;
944 mutex_lock(&ctx->ring_lock);
946 ring = kmap_atomic(ctx->ring_pages[0]);
951 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
961 avail = (head <= tail ? tail : ctx->nr_events) - head;
965 avail = min(avail, nr - ret);
966 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
967 ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
969 pos = head + AIO_EVENTS_OFFSET;
970 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
971 pos %= AIO_EVENTS_PER_PAGE;
974 copy_ret = copy_to_user(event + ret, ev + pos,
975 sizeof(*ev) * avail);
978 if (unlikely(copy_ret)) {
985 head %= ctx->nr_events;
988 ring = kmap_atomic(ctx->ring_pages[0]);
991 flush_dcache_page(ctx->ring_pages[0]);
993 pr_debug("%li h%u t%u\n", ret, head, tail);
995 put_reqs_available(ctx, ret);
997 mutex_unlock(&ctx->ring_lock);
1002 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1003 struct io_event __user *event, long *i)
1005 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1010 if (unlikely(atomic_read(&ctx->dead)))
1016 return ret < 0 || *i >= min_nr;
1019 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1020 struct io_event __user *event,
1021 struct timespec __user *timeout)
1023 ktime_t until = { .tv64 = KTIME_MAX };
1029 if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
1032 until = timespec_to_ktime(ts);
1036 * Note that aio_read_events() is being called as the conditional - i.e.
1037 * we're calling it after prepare_to_wait() has set task state to
1038 * TASK_INTERRUPTIBLE.
1040 * But aio_read_events() can block, and if it blocks it's going to flip
1041 * the task state back to TASK_RUNNING.
1043 * This should be ok, provided it doesn't flip the state back to
1044 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1045 * will only happen if the mutex_lock() call blocks, and we then find
1046 * the ringbuffer empty. So in practice we should be ok, but it's
1047 * something to be aware of when touching this code.
1049 wait_event_interruptible_hrtimeout(ctx->wait,
1050 aio_read_events(ctx, min_nr, nr, event, &ret), until);
1052 if (!ret && signal_pending(current))
1059 * Create an aio_context capable of receiving at least nr_events.
1060 * ctxp must not point to an aio_context that already exists, and
1061 * must be initialized to 0 prior to the call. On successful
1062 * creation of the aio_context, *ctxp is filled in with the resulting
1063 * handle. May fail with -EINVAL if *ctxp is not initialized,
1064 * if the specified nr_events exceeds internal limits. May fail
1065 * with -EAGAIN if the specified nr_events exceeds the user's limit
1066 * of available events. May fail with -ENOMEM if insufficient kernel
1067 * resources are available. May fail with -EFAULT if an invalid
1068 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1071 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1073 struct kioctx *ioctx = NULL;
1077 ret = get_user(ctx, ctxp);
1082 if (unlikely(ctx || nr_events == 0)) {
1083 pr_debug("EINVAL: io_setup: ctx %lu nr_events %u\n",
1088 ioctx = ioctx_alloc(nr_events);
1089 ret = PTR_ERR(ioctx);
1090 if (!IS_ERR(ioctx)) {
1091 ret = put_user(ioctx->user_id, ctxp);
1093 kill_ioctx(current->mm, ioctx);
1094 percpu_ref_put(&ioctx->users);
1102 * Destroy the aio_context specified. May cancel any outstanding
1103 * AIOs and block on completion. Will fail with -ENOSYS if not
1104 * implemented. May fail with -EINVAL if the context pointed to
1107 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1109 struct kioctx *ioctx = lookup_ioctx(ctx);
1110 if (likely(NULL != ioctx)) {
1111 kill_ioctx(current->mm, ioctx);
1112 percpu_ref_put(&ioctx->users);
1115 pr_debug("EINVAL: io_destroy: invalid context id\n");
1119 typedef ssize_t (aio_rw_op)(struct kiocb *, const struct iovec *,
1120 unsigned long, loff_t);
1122 static ssize_t aio_setup_vectored_rw(struct kiocb *kiocb,
1123 int rw, char __user *buf,
1124 unsigned long *nr_segs,
1125 struct iovec **iovec,
1130 *nr_segs = kiocb->ki_nbytes;
1132 #ifdef CONFIG_COMPAT
1134 ret = compat_rw_copy_check_uvector(rw,
1135 (struct compat_iovec __user *)buf,
1136 *nr_segs, 1, *iovec, iovec);
1139 ret = rw_copy_check_uvector(rw,
1140 (struct iovec __user *)buf,
1141 *nr_segs, 1, *iovec, iovec);
1145 /* ki_nbytes now reflect bytes instead of segs */
1146 kiocb->ki_nbytes = ret;
1150 static ssize_t aio_setup_single_vector(struct kiocb *kiocb,
1151 int rw, char __user *buf,
1152 unsigned long *nr_segs,
1153 struct iovec *iovec)
1155 if (unlikely(!access_ok(!rw, buf, kiocb->ki_nbytes)))
1158 iovec->iov_base = buf;
1159 iovec->iov_len = kiocb->ki_nbytes;
1166 * Performs the initial checks and aio retry method
1167 * setup for the kiocb at the time of io submission.
1169 static ssize_t aio_run_iocb(struct kiocb *req, unsigned opcode,
1170 char __user *buf, bool compat)
1172 struct file *file = req->ki_filp;
1174 unsigned long nr_segs;
1178 struct iovec inline_vec, *iovec = &inline_vec;
1181 case IOCB_CMD_PREAD:
1182 case IOCB_CMD_PREADV:
1185 rw_op = file->f_op->aio_read;
1188 case IOCB_CMD_PWRITE:
1189 case IOCB_CMD_PWRITEV:
1192 rw_op = file->f_op->aio_write;
1195 if (unlikely(!(file->f_mode & mode)))
1201 ret = (opcode == IOCB_CMD_PREADV ||
1202 opcode == IOCB_CMD_PWRITEV)
1203 ? aio_setup_vectored_rw(req, rw, buf, &nr_segs,
1205 : aio_setup_single_vector(req, rw, buf, &nr_segs,
1210 ret = rw_verify_area(rw, file, &req->ki_pos, req->ki_nbytes);
1212 if (iovec != &inline_vec)
1217 req->ki_nbytes = ret;
1219 /* XXX: move/kill - rw_verify_area()? */
1220 /* This matches the pread()/pwrite() logic */
1221 if (req->ki_pos < 0) {
1227 file_start_write(file);
1229 ret = rw_op(req, iovec, nr_segs, req->ki_pos);
1232 file_end_write(file);
1235 case IOCB_CMD_FDSYNC:
1236 if (!file->f_op->aio_fsync)
1239 ret = file->f_op->aio_fsync(req, 1);
1242 case IOCB_CMD_FSYNC:
1243 if (!file->f_op->aio_fsync)
1246 ret = file->f_op->aio_fsync(req, 0);
1250 pr_debug("EINVAL: no operation provided\n");
1254 if (iovec != &inline_vec)
1257 if (ret != -EIOCBQUEUED) {
1259 * There's no easy way to restart the syscall since other AIO's
1260 * may be already running. Just fail this IO with EINTR.
1262 if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
1263 ret == -ERESTARTNOHAND ||
1264 ret == -ERESTART_RESTARTBLOCK))
1266 aio_complete(req, ret, 0);
1272 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1273 struct iocb *iocb, bool compat)
1278 /* enforce forwards compatibility on users */
1279 if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
1280 pr_debug("EINVAL: reserve field set\n");
1284 /* prevent overflows */
1286 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1287 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1288 ((ssize_t)iocb->aio_nbytes < 0)
1290 pr_debug("EINVAL: io_submit: overflow check\n");
1294 req = aio_get_req(ctx);
1298 req->ki_filp = fget(iocb->aio_fildes);
1299 if (unlikely(!req->ki_filp)) {
1304 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1306 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1307 * instance of the file* now. The file descriptor must be
1308 * an eventfd() fd, and will be signaled for each completed
1309 * event using the eventfd_signal() function.
1311 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1312 if (IS_ERR(req->ki_eventfd)) {
1313 ret = PTR_ERR(req->ki_eventfd);
1314 req->ki_eventfd = NULL;
1319 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1320 if (unlikely(ret)) {
1321 pr_debug("EFAULT: aio_key\n");
1325 req->ki_obj.user = user_iocb;
1326 req->ki_user_data = iocb->aio_data;
1327 req->ki_pos = iocb->aio_offset;
1328 req->ki_nbytes = iocb->aio_nbytes;
1330 ret = aio_run_iocb(req, iocb->aio_lio_opcode,
1331 (char __user *)(unsigned long)iocb->aio_buf,
1338 put_reqs_available(ctx, 1);
1339 percpu_ref_put(&ctx->reqs);
1344 long do_io_submit(aio_context_t ctx_id, long nr,
1345 struct iocb __user *__user *iocbpp, bool compat)
1350 struct blk_plug plug;
1352 if (unlikely(nr < 0))
1355 if (unlikely(nr > LONG_MAX/sizeof(*iocbpp)))
1356 nr = LONG_MAX/sizeof(*iocbpp);
1358 if (unlikely(!access_ok(VERIFY_READ, iocbpp, (nr*sizeof(*iocbpp)))))
1361 ctx = lookup_ioctx(ctx_id);
1362 if (unlikely(!ctx)) {
1363 pr_debug("EINVAL: invalid context id\n");
1367 blk_start_plug(&plug);
1370 * AKPM: should this return a partial result if some of the IOs were
1371 * successfully submitted?
1373 for (i=0; i<nr; i++) {
1374 struct iocb __user *user_iocb;
1377 if (unlikely(__get_user(user_iocb, iocbpp + i))) {
1382 if (unlikely(copy_from_user(&tmp, user_iocb, sizeof(tmp)))) {
1387 ret = io_submit_one(ctx, user_iocb, &tmp, compat);
1391 blk_finish_plug(&plug);
1393 percpu_ref_put(&ctx->users);
1398 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1399 * the number of iocbs queued. May return -EINVAL if the aio_context
1400 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1401 * *iocbpp[0] is not properly initialized, if the operation specified
1402 * is invalid for the file descriptor in the iocb. May fail with
1403 * -EFAULT if any of the data structures point to invalid data. May
1404 * fail with -EBADF if the file descriptor specified in the first
1405 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1406 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1407 * fail with -ENOSYS if not implemented.
1409 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1410 struct iocb __user * __user *, iocbpp)
1412 return do_io_submit(ctx_id, nr, iocbpp, 0);
1416 * Finds a given iocb for cancellation.
1418 static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
1421 struct list_head *pos;
1423 assert_spin_locked(&ctx->ctx_lock);
1425 if (key != KIOCB_KEY)
1428 /* TODO: use a hash or array, this sucks. */
1429 list_for_each(pos, &ctx->active_reqs) {
1430 struct kiocb *kiocb = list_kiocb(pos);
1431 if (kiocb->ki_obj.user == iocb)
1438 * Attempts to cancel an iocb previously passed to io_submit. If
1439 * the operation is successfully cancelled, the resulting event is
1440 * copied into the memory pointed to by result without being placed
1441 * into the completion queue and 0 is returned. May fail with
1442 * -EFAULT if any of the data structures pointed to are invalid.
1443 * May fail with -EINVAL if aio_context specified by ctx_id is
1444 * invalid. May fail with -EAGAIN if the iocb specified was not
1445 * cancelled. Will fail with -ENOSYS if not implemented.
1447 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1448 struct io_event __user *, result)
1451 struct kiocb *kiocb;
1455 ret = get_user(key, &iocb->aio_key);
1459 ctx = lookup_ioctx(ctx_id);
1463 spin_lock_irq(&ctx->ctx_lock);
1465 kiocb = lookup_kiocb(ctx, iocb, key);
1467 ret = kiocb_cancel(ctx, kiocb);
1471 spin_unlock_irq(&ctx->ctx_lock);
1475 * The result argument is no longer used - the io_event is
1476 * always delivered via the ring buffer. -EINPROGRESS indicates
1477 * cancellation is progress:
1482 percpu_ref_put(&ctx->users);
1488 * Attempts to read at least min_nr events and up to nr events from
1489 * the completion queue for the aio_context specified by ctx_id. If
1490 * it succeeds, the number of read events is returned. May fail with
1491 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
1492 * out of range, if timeout is out of range. May fail with -EFAULT
1493 * if any of the memory specified is invalid. May return 0 or
1494 * < min_nr if the timeout specified by timeout has elapsed
1495 * before sufficient events are available, where timeout == NULL
1496 * specifies an infinite timeout. Note that the timeout pointed to by
1497 * timeout is relative. Will fail with -ENOSYS if not implemented.
1499 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
1502 struct io_event __user *, events,
1503 struct timespec __user *, timeout)
1505 struct kioctx *ioctx = lookup_ioctx(ctx_id);
1508 if (likely(ioctx)) {
1509 if (likely(min_nr <= nr && min_nr >= 0))
1510 ret = read_events(ioctx, min_nr, nr, events, timeout);
1511 percpu_ref_put(&ioctx->users);