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.
8 * Copyright 2018 Christoph Hellwig.
10 * See ../COPYING for licensing terms.
12 #define pr_fmt(fmt) "%s: " fmt, __func__
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
25 #include <linux/sched/signal.h>
27 #include <linux/file.h>
29 #include <linux/mman.h>
30 #include <linux/mmu_context.h>
31 #include <linux/percpu.h>
32 #include <linux/slab.h>
33 #include <linux/timer.h>
34 #include <linux/aio.h>
35 #include <linux/highmem.h>
36 #include <linux/workqueue.h>
37 #include <linux/security.h>
38 #include <linux/eventfd.h>
39 #include <linux/blkdev.h>
40 #include <linux/compat.h>
41 #include <linux/migrate.h>
42 #include <linux/ramfs.h>
43 #include <linux/percpu-refcount.h>
44 #include <linux/mount.h>
46 #include <asm/kmap_types.h>
47 #include <linux/uaccess.h>
48 #include <linux/nospec.h>
54 #define AIO_RING_MAGIC 0xa10a10a1
55 #define AIO_RING_COMPAT_FEATURES 1
56 #define AIO_RING_INCOMPAT_FEATURES 0
58 unsigned id; /* kernel internal index number */
59 unsigned nr; /* number of io_events */
60 unsigned head; /* Written to by userland or under ring_lock
61 * mutex by aio_read_events_ring(). */
65 unsigned compat_features;
66 unsigned incompat_features;
67 unsigned header_length; /* size of aio_ring */
70 struct io_event io_events[0];
71 }; /* 128 bytes + ring size */
73 #define AIO_RING_PAGES 8
78 struct kioctx __rcu *table[];
82 unsigned reqs_available;
86 struct completion comp;
91 struct percpu_ref users;
94 struct percpu_ref reqs;
96 unsigned long user_id;
98 struct __percpu kioctx_cpu *cpu;
101 * For percpu reqs_available, number of slots we move to/from global
106 * This is what userspace passed to io_setup(), it's not used for
107 * anything but counting against the global max_reqs quota.
109 * The real limit is nr_events - 1, which will be larger (see
114 /* Size of ringbuffer, in units of struct io_event */
117 unsigned long mmap_base;
118 unsigned long mmap_size;
120 struct page **ring_pages;
123 struct rcu_work free_rwork; /* see free_ioctx() */
126 * signals when all in-flight requests are done
128 struct ctx_rq_wait *rq_wait;
132 * This counts the number of available slots in the ringbuffer,
133 * so we avoid overflowing it: it's decremented (if positive)
134 * when allocating a kiocb and incremented when the resulting
135 * io_event is pulled off the ringbuffer.
137 * We batch accesses to it with a percpu version.
139 atomic_t reqs_available;
140 } ____cacheline_aligned_in_smp;
144 struct list_head active_reqs; /* used for cancellation */
145 } ____cacheline_aligned_in_smp;
148 struct mutex ring_lock;
149 wait_queue_head_t wait;
150 } ____cacheline_aligned_in_smp;
154 unsigned completed_events;
155 spinlock_t completion_lock;
156 } ____cacheline_aligned_in_smp;
158 struct page *internal_pages[AIO_RING_PAGES];
159 struct file *aio_ring_file;
165 * First field must be the file pointer in all the
166 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
170 struct work_struct work;
177 struct wait_queue_head *head;
181 bool work_need_resched;
182 struct wait_queue_entry wait;
183 struct work_struct work;
187 * NOTE! Each of the iocb union members has the file pointer
188 * as the first entry in their struct definition. So you can
189 * access the file pointer through any of the sub-structs,
190 * or directly as just 'ki_filp' in this struct.
194 struct file *ki_filp;
196 struct fsync_iocb fsync;
197 struct poll_iocb poll;
200 struct kioctx *ki_ctx;
201 kiocb_cancel_fn *ki_cancel;
203 struct io_event ki_res;
205 struct list_head ki_list; /* the aio core uses this
206 * for cancellation */
207 refcount_t ki_refcnt;
210 * If the aio_resfd field of the userspace iocb is not zero,
211 * this is the underlying eventfd context to deliver events to.
213 struct eventfd_ctx *ki_eventfd;
216 /*------ sysctl variables----*/
217 static DEFINE_SPINLOCK(aio_nr_lock);
218 unsigned long aio_nr; /* current system wide number of aio requests */
219 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
220 /*----end sysctl variables---*/
222 static struct kmem_cache *kiocb_cachep;
223 static struct kmem_cache *kioctx_cachep;
225 static struct vfsmount *aio_mnt;
227 static const struct file_operations aio_ring_fops;
228 static const struct address_space_operations aio_ctx_aops;
230 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
233 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
235 return ERR_CAST(inode);
237 inode->i_mapping->a_ops = &aio_ctx_aops;
238 inode->i_mapping->private_data = ctx;
239 inode->i_size = PAGE_SIZE * nr_pages;
241 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
242 O_RDWR, &aio_ring_fops);
248 static struct dentry *aio_mount(struct file_system_type *fs_type,
249 int flags, const char *dev_name, void *data)
251 struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
255 root->d_sb->s_iflags |= SB_I_NOEXEC;
260 * Creates the slab caches used by the aio routines, panic on
261 * failure as this is done early during the boot sequence.
263 static int __init aio_setup(void)
265 static struct file_system_type aio_fs = {
268 .kill_sb = kill_anon_super,
270 aio_mnt = kern_mount(&aio_fs);
272 panic("Failed to create aio fs mount.");
274 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
275 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
278 __initcall(aio_setup);
280 static void put_aio_ring_file(struct kioctx *ctx)
282 struct file *aio_ring_file = ctx->aio_ring_file;
283 struct address_space *i_mapping;
286 truncate_setsize(file_inode(aio_ring_file), 0);
288 /* Prevent further access to the kioctx from migratepages */
289 i_mapping = aio_ring_file->f_mapping;
290 spin_lock(&i_mapping->private_lock);
291 i_mapping->private_data = NULL;
292 ctx->aio_ring_file = NULL;
293 spin_unlock(&i_mapping->private_lock);
299 static void aio_free_ring(struct kioctx *ctx)
303 /* Disconnect the kiotx from the ring file. This prevents future
304 * accesses to the kioctx from page migration.
306 put_aio_ring_file(ctx);
308 for (i = 0; i < ctx->nr_pages; i++) {
310 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
311 page_count(ctx->ring_pages[i]));
312 page = ctx->ring_pages[i];
315 ctx->ring_pages[i] = NULL;
319 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
320 kfree(ctx->ring_pages);
321 ctx->ring_pages = NULL;
325 static int aio_ring_mremap(struct vm_area_struct *vma)
327 struct file *file = vma->vm_file;
328 struct mm_struct *mm = vma->vm_mm;
329 struct kioctx_table *table;
330 int i, res = -EINVAL;
332 spin_lock(&mm->ioctx_lock);
334 table = rcu_dereference(mm->ioctx_table);
335 for (i = 0; i < table->nr; i++) {
338 ctx = rcu_dereference(table->table[i]);
339 if (ctx && ctx->aio_ring_file == file) {
340 if (!atomic_read(&ctx->dead)) {
341 ctx->user_id = ctx->mmap_base = vma->vm_start;
349 spin_unlock(&mm->ioctx_lock);
353 static const struct vm_operations_struct aio_ring_vm_ops = {
354 .mremap = aio_ring_mremap,
355 #if IS_ENABLED(CONFIG_MMU)
356 .fault = filemap_fault,
357 .map_pages = filemap_map_pages,
358 .page_mkwrite = filemap_page_mkwrite,
362 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
364 vma->vm_flags |= VM_DONTEXPAND;
365 vma->vm_ops = &aio_ring_vm_ops;
369 static const struct file_operations aio_ring_fops = {
370 .mmap = aio_ring_mmap,
373 #if IS_ENABLED(CONFIG_MIGRATION)
374 static int aio_migratepage(struct address_space *mapping, struct page *new,
375 struct page *old, enum migrate_mode mode)
383 * We cannot support the _NO_COPY case here, because copy needs to
384 * happen under the ctx->completion_lock. That does not work with the
385 * migration workflow of MIGRATE_SYNC_NO_COPY.
387 if (mode == MIGRATE_SYNC_NO_COPY)
392 /* mapping->private_lock here protects against the kioctx teardown. */
393 spin_lock(&mapping->private_lock);
394 ctx = mapping->private_data;
400 /* The ring_lock mutex. The prevents aio_read_events() from writing
401 * to the ring's head, and prevents page migration from mucking in
402 * a partially initialized kiotx.
404 if (!mutex_trylock(&ctx->ring_lock)) {
410 if (idx < (pgoff_t)ctx->nr_pages) {
411 /* Make sure the old page hasn't already been changed */
412 if (ctx->ring_pages[idx] != old)
420 /* Writeback must be complete */
421 BUG_ON(PageWriteback(old));
424 rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
425 if (rc != MIGRATEPAGE_SUCCESS) {
430 /* Take completion_lock to prevent other writes to the ring buffer
431 * while the old page is copied to the new. This prevents new
432 * events from being lost.
434 spin_lock_irqsave(&ctx->completion_lock, flags);
435 migrate_page_copy(new, old);
436 BUG_ON(ctx->ring_pages[idx] != old);
437 ctx->ring_pages[idx] = new;
438 spin_unlock_irqrestore(&ctx->completion_lock, flags);
440 /* The old page is no longer accessible. */
444 mutex_unlock(&ctx->ring_lock);
446 spin_unlock(&mapping->private_lock);
451 static const struct address_space_operations aio_ctx_aops = {
452 .set_page_dirty = __set_page_dirty_no_writeback,
453 #if IS_ENABLED(CONFIG_MIGRATION)
454 .migratepage = aio_migratepage,
458 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
460 struct aio_ring *ring;
461 struct mm_struct *mm = current->mm;
462 unsigned long size, unused;
467 /* Compensate for the ring buffer's head/tail overlap entry */
468 nr_events += 2; /* 1 is required, 2 for good luck */
470 size = sizeof(struct aio_ring);
471 size += sizeof(struct io_event) * nr_events;
473 nr_pages = PFN_UP(size);
477 file = aio_private_file(ctx, nr_pages);
479 ctx->aio_ring_file = NULL;
483 ctx->aio_ring_file = file;
484 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
485 / sizeof(struct io_event);
487 ctx->ring_pages = ctx->internal_pages;
488 if (nr_pages > AIO_RING_PAGES) {
489 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
491 if (!ctx->ring_pages) {
492 put_aio_ring_file(ctx);
497 for (i = 0; i < nr_pages; i++) {
499 page = find_or_create_page(file->f_mapping,
500 i, GFP_HIGHUSER | __GFP_ZERO);
503 pr_debug("pid(%d) page[%d]->count=%d\n",
504 current->pid, i, page_count(page));
505 SetPageUptodate(page);
508 ctx->ring_pages[i] = page;
512 if (unlikely(i != nr_pages)) {
517 ctx->mmap_size = nr_pages * PAGE_SIZE;
518 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
520 if (down_write_killable(&mm->mmap_sem)) {
526 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
527 PROT_READ | PROT_WRITE,
528 MAP_SHARED, 0, &unused, NULL);
529 up_write(&mm->mmap_sem);
530 if (IS_ERR((void *)ctx->mmap_base)) {
536 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
538 ctx->user_id = ctx->mmap_base;
539 ctx->nr_events = nr_events; /* trusted copy */
541 ring = kmap_atomic(ctx->ring_pages[0]);
542 ring->nr = nr_events; /* user copy */
544 ring->head = ring->tail = 0;
545 ring->magic = AIO_RING_MAGIC;
546 ring->compat_features = AIO_RING_COMPAT_FEATURES;
547 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
548 ring->header_length = sizeof(struct aio_ring);
550 flush_dcache_page(ctx->ring_pages[0]);
555 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
556 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
557 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
559 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
561 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
562 struct kioctx *ctx = req->ki_ctx;
565 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
568 spin_lock_irqsave(&ctx->ctx_lock, flags);
569 list_add_tail(&req->ki_list, &ctx->active_reqs);
570 req->ki_cancel = cancel;
571 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
573 EXPORT_SYMBOL(kiocb_set_cancel_fn);
576 * free_ioctx() should be RCU delayed to synchronize against the RCU
577 * protected lookup_ioctx() and also needs process context to call
578 * aio_free_ring(). Use rcu_work.
580 static void free_ioctx(struct work_struct *work)
582 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
584 pr_debug("freeing %p\n", ctx);
587 free_percpu(ctx->cpu);
588 percpu_ref_exit(&ctx->reqs);
589 percpu_ref_exit(&ctx->users);
590 kmem_cache_free(kioctx_cachep, ctx);
593 static void free_ioctx_reqs(struct percpu_ref *ref)
595 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
597 /* At this point we know that there are no any in-flight requests */
598 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
599 complete(&ctx->rq_wait->comp);
601 /* Synchronize against RCU protected table->table[] dereferences */
602 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
603 queue_rcu_work(system_wq, &ctx->free_rwork);
607 * When this function runs, the kioctx has been removed from the "hash table"
608 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
609 * now it's safe to cancel any that need to be.
611 static void free_ioctx_users(struct percpu_ref *ref)
613 struct kioctx *ctx = container_of(ref, struct kioctx, users);
614 struct aio_kiocb *req;
616 spin_lock_irq(&ctx->ctx_lock);
618 while (!list_empty(&ctx->active_reqs)) {
619 req = list_first_entry(&ctx->active_reqs,
620 struct aio_kiocb, ki_list);
621 req->ki_cancel(&req->rw);
622 list_del_init(&req->ki_list);
625 spin_unlock_irq(&ctx->ctx_lock);
627 percpu_ref_kill(&ctx->reqs);
628 percpu_ref_put(&ctx->reqs);
631 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
634 struct kioctx_table *table, *old;
635 struct aio_ring *ring;
637 spin_lock(&mm->ioctx_lock);
638 table = rcu_dereference_raw(mm->ioctx_table);
642 for (i = 0; i < table->nr; i++)
643 if (!rcu_access_pointer(table->table[i])) {
645 rcu_assign_pointer(table->table[i], ctx);
646 spin_unlock(&mm->ioctx_lock);
648 /* While kioctx setup is in progress,
649 * we are protected from page migration
650 * changes ring_pages by ->ring_lock.
652 ring = kmap_atomic(ctx->ring_pages[0]);
658 new_nr = (table ? table->nr : 1) * 4;
659 spin_unlock(&mm->ioctx_lock);
661 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
668 spin_lock(&mm->ioctx_lock);
669 old = rcu_dereference_raw(mm->ioctx_table);
672 rcu_assign_pointer(mm->ioctx_table, table);
673 } else if (table->nr > old->nr) {
674 memcpy(table->table, old->table,
675 old->nr * sizeof(struct kioctx *));
677 rcu_assign_pointer(mm->ioctx_table, table);
686 static void aio_nr_sub(unsigned nr)
688 spin_lock(&aio_nr_lock);
689 if (WARN_ON(aio_nr - nr > aio_nr))
693 spin_unlock(&aio_nr_lock);
697 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
699 static struct kioctx *ioctx_alloc(unsigned nr_events)
701 struct mm_struct *mm = current->mm;
706 * Store the original nr_events -- what userspace passed to io_setup(),
707 * for counting against the global limit -- before it changes.
709 unsigned int max_reqs = nr_events;
712 * We keep track of the number of available ringbuffer slots, to prevent
713 * overflow (reqs_available), and we also use percpu counters for this.
715 * So since up to half the slots might be on other cpu's percpu counters
716 * and unavailable, double nr_events so userspace sees what they
717 * expected: additionally, we move req_batch slots to/from percpu
718 * counters at a time, so make sure that isn't 0:
720 nr_events = max(nr_events, num_possible_cpus() * 4);
723 /* Prevent overflows */
724 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
725 pr_debug("ENOMEM: nr_events too high\n");
726 return ERR_PTR(-EINVAL);
729 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
730 return ERR_PTR(-EAGAIN);
732 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
734 return ERR_PTR(-ENOMEM);
736 ctx->max_reqs = max_reqs;
738 spin_lock_init(&ctx->ctx_lock);
739 spin_lock_init(&ctx->completion_lock);
740 mutex_init(&ctx->ring_lock);
741 /* Protect against page migration throughout kiotx setup by keeping
742 * the ring_lock mutex held until setup is complete. */
743 mutex_lock(&ctx->ring_lock);
744 init_waitqueue_head(&ctx->wait);
746 INIT_LIST_HEAD(&ctx->active_reqs);
748 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
751 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
754 ctx->cpu = alloc_percpu(struct kioctx_cpu);
758 err = aio_setup_ring(ctx, nr_events);
762 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
763 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
764 if (ctx->req_batch < 1)
767 /* limit the number of system wide aios */
768 spin_lock(&aio_nr_lock);
769 if (aio_nr + ctx->max_reqs > aio_max_nr ||
770 aio_nr + ctx->max_reqs < aio_nr) {
771 spin_unlock(&aio_nr_lock);
775 aio_nr += ctx->max_reqs;
776 spin_unlock(&aio_nr_lock);
778 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
779 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
781 err = ioctx_add_table(ctx, mm);
785 /* Release the ring_lock mutex now that all setup is complete. */
786 mutex_unlock(&ctx->ring_lock);
788 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
789 ctx, ctx->user_id, mm, ctx->nr_events);
793 aio_nr_sub(ctx->max_reqs);
795 atomic_set(&ctx->dead, 1);
797 vm_munmap(ctx->mmap_base, ctx->mmap_size);
800 mutex_unlock(&ctx->ring_lock);
801 free_percpu(ctx->cpu);
802 percpu_ref_exit(&ctx->reqs);
803 percpu_ref_exit(&ctx->users);
804 kmem_cache_free(kioctx_cachep, ctx);
805 pr_debug("error allocating ioctx %d\n", err);
810 * Cancels all outstanding aio requests on an aio context. Used
811 * when the processes owning a context have all exited to encourage
812 * the rapid destruction of the kioctx.
814 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
815 struct ctx_rq_wait *wait)
817 struct kioctx_table *table;
819 spin_lock(&mm->ioctx_lock);
820 if (atomic_xchg(&ctx->dead, 1)) {
821 spin_unlock(&mm->ioctx_lock);
825 table = rcu_dereference_raw(mm->ioctx_table);
826 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
827 RCU_INIT_POINTER(table->table[ctx->id], NULL);
828 spin_unlock(&mm->ioctx_lock);
830 /* free_ioctx_reqs() will do the necessary RCU synchronization */
831 wake_up_all(&ctx->wait);
834 * It'd be more correct to do this in free_ioctx(), after all
835 * the outstanding kiocbs have finished - but by then io_destroy
836 * has already returned, so io_setup() could potentially return
837 * -EAGAIN with no ioctxs actually in use (as far as userspace
840 aio_nr_sub(ctx->max_reqs);
843 vm_munmap(ctx->mmap_base, ctx->mmap_size);
846 percpu_ref_kill(&ctx->users);
851 * exit_aio: called when the last user of mm goes away. At this point, there is
852 * no way for any new requests to be submited or any of the io_* syscalls to be
853 * called on the context.
855 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
858 void exit_aio(struct mm_struct *mm)
860 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
861 struct ctx_rq_wait wait;
867 atomic_set(&wait.count, table->nr);
868 init_completion(&wait.comp);
871 for (i = 0; i < table->nr; ++i) {
873 rcu_dereference_protected(table->table[i], true);
881 * We don't need to bother with munmap() here - exit_mmap(mm)
882 * is coming and it'll unmap everything. And we simply can't,
883 * this is not necessarily our ->mm.
884 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
885 * that it needs to unmap the area, just set it to 0.
888 kill_ioctx(mm, ctx, &wait);
891 if (!atomic_sub_and_test(skipped, &wait.count)) {
892 /* Wait until all IO for the context are done. */
893 wait_for_completion(&wait.comp);
896 RCU_INIT_POINTER(mm->ioctx_table, NULL);
900 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
902 struct kioctx_cpu *kcpu;
905 local_irq_save(flags);
906 kcpu = this_cpu_ptr(ctx->cpu);
907 kcpu->reqs_available += nr;
909 while (kcpu->reqs_available >= ctx->req_batch * 2) {
910 kcpu->reqs_available -= ctx->req_batch;
911 atomic_add(ctx->req_batch, &ctx->reqs_available);
914 local_irq_restore(flags);
917 static bool __get_reqs_available(struct kioctx *ctx)
919 struct kioctx_cpu *kcpu;
923 local_irq_save(flags);
924 kcpu = this_cpu_ptr(ctx->cpu);
925 if (!kcpu->reqs_available) {
926 int old, avail = atomic_read(&ctx->reqs_available);
929 if (avail < ctx->req_batch)
933 avail = atomic_cmpxchg(&ctx->reqs_available,
934 avail, avail - ctx->req_batch);
935 } while (avail != old);
937 kcpu->reqs_available += ctx->req_batch;
941 kcpu->reqs_available--;
943 local_irq_restore(flags);
947 /* refill_reqs_available
948 * Updates the reqs_available reference counts used for tracking the
949 * number of free slots in the completion ring. This can be called
950 * from aio_complete() (to optimistically update reqs_available) or
951 * from aio_get_req() (the we're out of events case). It must be
952 * called holding ctx->completion_lock.
954 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
957 unsigned events_in_ring, completed;
959 /* Clamp head since userland can write to it. */
960 head %= ctx->nr_events;
962 events_in_ring = tail - head;
964 events_in_ring = ctx->nr_events - (head - tail);
966 completed = ctx->completed_events;
967 if (events_in_ring < completed)
968 completed -= events_in_ring;
975 ctx->completed_events -= completed;
976 put_reqs_available(ctx, completed);
979 /* user_refill_reqs_available
980 * Called to refill reqs_available when aio_get_req() encounters an
981 * out of space in the completion ring.
983 static void user_refill_reqs_available(struct kioctx *ctx)
985 spin_lock_irq(&ctx->completion_lock);
986 if (ctx->completed_events) {
987 struct aio_ring *ring;
990 /* Access of ring->head may race with aio_read_events_ring()
991 * here, but that's okay since whether we read the old version
992 * or the new version, and either will be valid. The important
993 * part is that head cannot pass tail since we prevent
994 * aio_complete() from updating tail by holding
995 * ctx->completion_lock. Even if head is invalid, the check
996 * against ctx->completed_events below will make sure we do the
999 ring = kmap_atomic(ctx->ring_pages[0]);
1001 kunmap_atomic(ring);
1003 refill_reqs_available(ctx, head, ctx->tail);
1006 spin_unlock_irq(&ctx->completion_lock);
1009 static bool get_reqs_available(struct kioctx *ctx)
1011 if (__get_reqs_available(ctx))
1013 user_refill_reqs_available(ctx);
1014 return __get_reqs_available(ctx);
1018 * Allocate a slot for an aio request.
1019 * Returns NULL if no requests are free.
1021 * The refcount is initialized to 2 - one for the async op completion,
1022 * one for the synchronous code that does this.
1024 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1026 struct aio_kiocb *req;
1028 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1032 percpu_ref_get(&ctx->reqs);
1034 INIT_LIST_HEAD(&req->ki_list);
1035 refcount_set(&req->ki_refcnt, 2);
1036 req->ki_eventfd = NULL;
1040 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1042 struct aio_ring __user *ring = (void __user *)ctx_id;
1043 struct mm_struct *mm = current->mm;
1044 struct kioctx *ctx, *ret = NULL;
1045 struct kioctx_table *table;
1048 if (get_user(id, &ring->id))
1052 table = rcu_dereference(mm->ioctx_table);
1054 if (!table || id >= table->nr)
1057 id = array_index_nospec(id, table->nr);
1058 ctx = rcu_dereference(table->table[id]);
1059 if (ctx && ctx->user_id == ctx_id) {
1060 if (percpu_ref_tryget_live(&ctx->users))
1068 static inline void iocb_destroy(struct aio_kiocb *iocb)
1071 fput(iocb->ki_filp);
1072 percpu_ref_put(&iocb->ki_ctx->reqs);
1073 kmem_cache_free(kiocb_cachep, iocb);
1077 * Called when the io request on the given iocb is complete.
1079 static void aio_complete(struct aio_kiocb *iocb)
1081 struct kioctx *ctx = iocb->ki_ctx;
1082 struct aio_ring *ring;
1083 struct io_event *ev_page, *event;
1084 unsigned tail, pos, head;
1085 unsigned long flags;
1088 * Add a completion event to the ring buffer. Must be done holding
1089 * ctx->completion_lock to prevent other code from messing with the tail
1090 * pointer since we might be called from irq context.
1092 spin_lock_irqsave(&ctx->completion_lock, flags);
1095 pos = tail + AIO_EVENTS_OFFSET;
1097 if (++tail >= ctx->nr_events)
1100 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1101 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1103 *event = iocb->ki_res;
1105 kunmap_atomic(ev_page);
1106 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1108 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1109 (void __user *)(unsigned long)iocb->ki_res.obj,
1110 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1112 /* after flagging the request as done, we
1113 * must never even look at it again
1115 smp_wmb(); /* make event visible before updating tail */
1119 ring = kmap_atomic(ctx->ring_pages[0]);
1122 kunmap_atomic(ring);
1123 flush_dcache_page(ctx->ring_pages[0]);
1125 ctx->completed_events++;
1126 if (ctx->completed_events > 1)
1127 refill_reqs_available(ctx, head, tail);
1128 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1130 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1133 * Check if the user asked us to deliver the result through an
1134 * eventfd. The eventfd_signal() function is safe to be called
1137 if (iocb->ki_eventfd) {
1138 eventfd_signal(iocb->ki_eventfd, 1);
1139 eventfd_ctx_put(iocb->ki_eventfd);
1143 * We have to order our ring_info tail store above and test
1144 * of the wait list below outside the wait lock. This is
1145 * like in wake_up_bit() where clearing a bit has to be
1146 * ordered with the unlocked test.
1150 if (waitqueue_active(&ctx->wait))
1151 wake_up(&ctx->wait);
1154 static inline void iocb_put(struct aio_kiocb *iocb)
1156 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1162 /* aio_read_events_ring
1163 * Pull an event off of the ioctx's event ring. Returns the number of
1166 static long aio_read_events_ring(struct kioctx *ctx,
1167 struct io_event __user *event, long nr)
1169 struct aio_ring *ring;
1170 unsigned head, tail, pos;
1175 * The mutex can block and wake us up and that will cause
1176 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1177 * and repeat. This should be rare enough that it doesn't cause
1178 * peformance issues. See the comment in read_events() for more detail.
1180 sched_annotate_sleep();
1181 mutex_lock(&ctx->ring_lock);
1183 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1184 ring = kmap_atomic(ctx->ring_pages[0]);
1187 kunmap_atomic(ring);
1190 * Ensure that once we've read the current tail pointer, that
1191 * we also see the events that were stored up to the tail.
1195 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1200 head %= ctx->nr_events;
1201 tail %= ctx->nr_events;
1205 struct io_event *ev;
1208 avail = (head <= tail ? tail : ctx->nr_events) - head;
1212 pos = head + AIO_EVENTS_OFFSET;
1213 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1214 pos %= AIO_EVENTS_PER_PAGE;
1216 avail = min(avail, nr - ret);
1217 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1220 copy_ret = copy_to_user(event + ret, ev + pos,
1221 sizeof(*ev) * avail);
1224 if (unlikely(copy_ret)) {
1231 head %= ctx->nr_events;
1234 ring = kmap_atomic(ctx->ring_pages[0]);
1236 kunmap_atomic(ring);
1237 flush_dcache_page(ctx->ring_pages[0]);
1239 pr_debug("%li h%u t%u\n", ret, head, tail);
1241 mutex_unlock(&ctx->ring_lock);
1246 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1247 struct io_event __user *event, long *i)
1249 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1254 if (unlikely(atomic_read(&ctx->dead)))
1260 return ret < 0 || *i >= min_nr;
1263 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1264 struct io_event __user *event,
1270 * Note that aio_read_events() is being called as the conditional - i.e.
1271 * we're calling it after prepare_to_wait() has set task state to
1272 * TASK_INTERRUPTIBLE.
1274 * But aio_read_events() can block, and if it blocks it's going to flip
1275 * the task state back to TASK_RUNNING.
1277 * This should be ok, provided it doesn't flip the state back to
1278 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1279 * will only happen if the mutex_lock() call blocks, and we then find
1280 * the ringbuffer empty. So in practice we should be ok, but it's
1281 * something to be aware of when touching this code.
1284 aio_read_events(ctx, min_nr, nr, event, &ret);
1286 wait_event_interruptible_hrtimeout(ctx->wait,
1287 aio_read_events(ctx, min_nr, nr, event, &ret),
1293 * Create an aio_context capable of receiving at least nr_events.
1294 * ctxp must not point to an aio_context that already exists, and
1295 * must be initialized to 0 prior to the call. On successful
1296 * creation of the aio_context, *ctxp is filled in with the resulting
1297 * handle. May fail with -EINVAL if *ctxp is not initialized,
1298 * if the specified nr_events exceeds internal limits. May fail
1299 * with -EAGAIN if the specified nr_events exceeds the user's limit
1300 * of available events. May fail with -ENOMEM if insufficient kernel
1301 * resources are available. May fail with -EFAULT if an invalid
1302 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1305 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1307 struct kioctx *ioctx = NULL;
1311 ret = get_user(ctx, ctxp);
1316 if (unlikely(ctx || nr_events == 0)) {
1317 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1322 ioctx = ioctx_alloc(nr_events);
1323 ret = PTR_ERR(ioctx);
1324 if (!IS_ERR(ioctx)) {
1325 ret = put_user(ioctx->user_id, ctxp);
1327 kill_ioctx(current->mm, ioctx, NULL);
1328 percpu_ref_put(&ioctx->users);
1335 #ifdef CONFIG_COMPAT
1336 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1338 struct kioctx *ioctx = NULL;
1342 ret = get_user(ctx, ctx32p);
1347 if (unlikely(ctx || nr_events == 0)) {
1348 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1353 ioctx = ioctx_alloc(nr_events);
1354 ret = PTR_ERR(ioctx);
1355 if (!IS_ERR(ioctx)) {
1356 /* truncating is ok because it's a user address */
1357 ret = put_user((u32)ioctx->user_id, ctx32p);
1359 kill_ioctx(current->mm, ioctx, NULL);
1360 percpu_ref_put(&ioctx->users);
1369 * Destroy the aio_context specified. May cancel any outstanding
1370 * AIOs and block on completion. Will fail with -ENOSYS if not
1371 * implemented. May fail with -EINVAL if the context pointed to
1374 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1376 struct kioctx *ioctx = lookup_ioctx(ctx);
1377 if (likely(NULL != ioctx)) {
1378 struct ctx_rq_wait wait;
1381 init_completion(&wait.comp);
1382 atomic_set(&wait.count, 1);
1384 /* Pass requests_done to kill_ioctx() where it can be set
1385 * in a thread-safe way. If we try to set it here then we have
1386 * a race condition if two io_destroy() called simultaneously.
1388 ret = kill_ioctx(current->mm, ioctx, &wait);
1389 percpu_ref_put(&ioctx->users);
1391 /* Wait until all IO for the context are done. Otherwise kernel
1392 * keep using user-space buffers even if user thinks the context
1396 wait_for_completion(&wait.comp);
1400 pr_debug("EINVAL: invalid context id\n");
1404 static void aio_remove_iocb(struct aio_kiocb *iocb)
1406 struct kioctx *ctx = iocb->ki_ctx;
1407 unsigned long flags;
1409 spin_lock_irqsave(&ctx->ctx_lock, flags);
1410 list_del(&iocb->ki_list);
1411 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1414 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1416 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1418 if (!list_empty_careful(&iocb->ki_list))
1419 aio_remove_iocb(iocb);
1421 if (kiocb->ki_flags & IOCB_WRITE) {
1422 struct inode *inode = file_inode(kiocb->ki_filp);
1425 * Tell lockdep we inherited freeze protection from submission
1428 if (S_ISREG(inode->i_mode))
1429 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1430 file_end_write(kiocb->ki_filp);
1433 iocb->ki_res.res = res;
1434 iocb->ki_res.res2 = res2;
1438 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1442 req->ki_complete = aio_complete_rw;
1443 req->private = NULL;
1444 req->ki_pos = iocb->aio_offset;
1445 req->ki_flags = iocb_flags(req->ki_filp);
1446 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1447 req->ki_flags |= IOCB_EVENTFD;
1448 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1449 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1451 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1452 * aio_reqprio is interpreted as an I/O scheduling
1453 * class and priority.
1455 ret = ioprio_check_cap(iocb->aio_reqprio);
1457 pr_debug("aio ioprio check cap error: %d\n", ret);
1461 req->ki_ioprio = iocb->aio_reqprio;
1463 req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0);
1465 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1469 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1473 static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
1474 bool vectored, bool compat, struct iov_iter *iter)
1476 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1477 size_t len = iocb->aio_nbytes;
1480 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1484 #ifdef CONFIG_COMPAT
1486 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1489 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1492 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1498 case -ERESTARTNOINTR:
1499 case -ERESTARTNOHAND:
1500 case -ERESTART_RESTARTBLOCK:
1502 * There's no easy way to restart the syscall since other AIO's
1503 * may be already running. Just fail this IO with EINTR.
1508 req->ki_complete(req, ret, 0);
1512 static ssize_t aio_read(struct kiocb *req, const struct iocb *iocb,
1513 bool vectored, bool compat)
1515 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1516 struct iov_iter iter;
1520 ret = aio_prep_rw(req, iocb);
1523 file = req->ki_filp;
1524 if (unlikely(!(file->f_mode & FMODE_READ)))
1527 if (unlikely(!file->f_op->read_iter))
1530 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1533 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1535 aio_rw_done(req, call_read_iter(file, req, &iter));
1540 static ssize_t aio_write(struct kiocb *req, const struct iocb *iocb,
1541 bool vectored, bool compat)
1543 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1544 struct iov_iter iter;
1548 ret = aio_prep_rw(req, iocb);
1551 file = req->ki_filp;
1553 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1555 if (unlikely(!file->f_op->write_iter))
1558 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1561 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1564 * Open-code file_start_write here to grab freeze protection,
1565 * which will be released by another thread in
1566 * aio_complete_rw(). Fool lockdep by telling it the lock got
1567 * released so that it doesn't complain about the held lock when
1568 * we return to userspace.
1570 if (S_ISREG(file_inode(file)->i_mode)) {
1571 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1572 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1574 req->ki_flags |= IOCB_WRITE;
1575 aio_rw_done(req, call_write_iter(file, req, &iter));
1581 static void aio_fsync_work(struct work_struct *work)
1583 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1584 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1586 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1587 revert_creds(old_cred);
1588 put_cred(iocb->fsync.creds);
1592 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1595 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1596 iocb->aio_rw_flags))
1599 if (unlikely(!req->file->f_op->fsync))
1602 req->creds = prepare_creds();
1606 req->datasync = datasync;
1607 INIT_WORK(&req->work, aio_fsync_work);
1608 schedule_work(&req->work);
1612 static void aio_poll_put_work(struct work_struct *work)
1614 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1615 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1621 * Safely lock the waitqueue which the request is on, synchronizing with the
1622 * case where the ->poll() provider decides to free its waitqueue early.
1624 * Returns true on success, meaning that req->head->lock was locked, req->wait
1625 * is on req->head, and an RCU read lock was taken. Returns false if the
1626 * request was already removed from its waitqueue (which might no longer exist).
1628 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1630 wait_queue_head_t *head;
1633 * While we hold the waitqueue lock and the waitqueue is nonempty,
1634 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1635 * lock in the first place can race with the waitqueue being freed.
1637 * We solve this as eventpoll does: by taking advantage of the fact that
1638 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1639 * we enter rcu_read_lock() and see that the pointer to the queue is
1640 * non-NULL, we can then lock it without the memory being freed out from
1641 * under us, then check whether the request is still on the queue.
1643 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1644 * case the caller deletes the entry from the queue, leaving it empty.
1645 * In that case, only RCU prevents the queue memory from being freed.
1648 head = smp_load_acquire(&req->head);
1650 spin_lock(&head->lock);
1651 if (!list_empty(&req->wait.entry))
1653 spin_unlock(&head->lock);
1659 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1661 spin_unlock(&req->head->lock);
1665 static void aio_poll_complete_work(struct work_struct *work)
1667 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1668 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1669 struct poll_table_struct pt = { ._key = req->events };
1670 struct kioctx *ctx = iocb->ki_ctx;
1673 if (!READ_ONCE(req->cancelled))
1674 mask = vfs_poll(req->file, &pt) & req->events;
1677 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1678 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1679 * synchronize with them. In the cancellation case the list_del_init
1680 * itself is not actually needed, but harmless so we keep it in to
1681 * avoid further branches in the fast path.
1683 spin_lock_irq(&ctx->ctx_lock);
1684 if (poll_iocb_lock_wq(req)) {
1685 if (!mask && !READ_ONCE(req->cancelled)) {
1687 * The request isn't actually ready to be completed yet.
1688 * Reschedule completion if another wakeup came in.
1690 if (req->work_need_resched) {
1691 schedule_work(&req->work);
1692 req->work_need_resched = false;
1694 req->work_scheduled = false;
1696 poll_iocb_unlock_wq(req);
1697 spin_unlock_irq(&ctx->ctx_lock);
1700 list_del_init(&req->wait.entry);
1701 poll_iocb_unlock_wq(req);
1702 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1703 list_del_init(&iocb->ki_list);
1704 iocb->ki_res.res = mangle_poll(mask);
1705 spin_unlock_irq(&ctx->ctx_lock);
1710 /* assumes we are called with irqs disabled */
1711 static int aio_poll_cancel(struct kiocb *iocb)
1713 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1714 struct poll_iocb *req = &aiocb->poll;
1716 if (poll_iocb_lock_wq(req)) {
1717 WRITE_ONCE(req->cancelled, true);
1718 if (!req->work_scheduled) {
1719 schedule_work(&aiocb->poll.work);
1720 req->work_scheduled = true;
1722 poll_iocb_unlock_wq(req);
1723 } /* else, the request was force-cancelled by POLLFREE already */
1728 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1731 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1732 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1733 __poll_t mask = key_to_poll(key);
1734 unsigned long flags;
1736 /* for instances that support it check for an event match first: */
1737 if (mask && !(mask & req->events))
1741 * Complete the request inline if possible. This requires that three
1742 * conditions be met:
1743 * 1. An event mask must have been passed. If a plain wakeup was done
1744 * instead, then mask == 0 and we have to call vfs_poll() to get
1745 * the events, so inline completion isn't possible.
1746 * 2. The completion work must not have already been scheduled.
1747 * 3. ctx_lock must not be busy. We have to use trylock because we
1748 * already hold the waitqueue lock, so this inverts the normal
1749 * locking order. Use irqsave/irqrestore because not all
1750 * filesystems (e.g. fuse) call this function with IRQs disabled,
1751 * yet IRQs have to be disabled before ctx_lock is obtained.
1753 if (mask && !req->work_scheduled &&
1754 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1755 struct kioctx *ctx = iocb->ki_ctx;
1757 list_del_init(&req->wait.entry);
1758 list_del(&iocb->ki_list);
1759 iocb->ki_res.res = mangle_poll(mask);
1760 if (iocb->ki_eventfd && eventfd_signal_count()) {
1762 INIT_WORK(&req->work, aio_poll_put_work);
1763 schedule_work(&req->work);
1765 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1770 * Schedule the completion work if needed. If it was already
1771 * scheduled, record that another wakeup came in.
1773 * Don't remove the request from the waitqueue here, as it might
1774 * not actually be complete yet (we won't know until vfs_poll()
1775 * is called), and we must not miss any wakeups. POLLFREE is an
1776 * exception to this; see below.
1778 if (req->work_scheduled) {
1779 req->work_need_resched = true;
1781 schedule_work(&req->work);
1782 req->work_scheduled = true;
1786 * If the waitqueue is being freed early but we can't complete
1787 * the request inline, we have to tear down the request as best
1788 * we can. That means immediately removing the request from its
1789 * waitqueue and preventing all further accesses to the
1790 * waitqueue via the request. We also need to schedule the
1791 * completion work (done above). Also mark the request as
1792 * cancelled, to potentially skip an unneeded call to ->poll().
1794 if (mask & POLLFREE) {
1795 WRITE_ONCE(req->cancelled, true);
1796 list_del_init(&req->wait.entry);
1799 * Careful: this *must* be the last step, since as soon
1800 * as req->head is NULL'ed out, the request can be
1801 * completed and freed, since aio_poll_complete_work()
1802 * will no longer need to take the waitqueue lock.
1804 smp_store_release(&req->head, NULL);
1810 struct aio_poll_table {
1811 struct poll_table_struct pt;
1812 struct aio_kiocb *iocb;
1818 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1819 struct poll_table_struct *p)
1821 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1823 /* multiple wait queues per file are not supported */
1824 if (unlikely(pt->queued)) {
1825 pt->error = -EINVAL;
1831 pt->iocb->poll.head = head;
1832 add_wait_queue(head, &pt->iocb->poll.wait);
1835 static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1837 struct kioctx *ctx = aiocb->ki_ctx;
1838 struct poll_iocb *req = &aiocb->poll;
1839 struct aio_poll_table apt;
1840 bool cancel = false;
1843 /* reject any unknown events outside the normal event mask. */
1844 if ((u16)iocb->aio_buf != iocb->aio_buf)
1846 /* reject fields that are not defined for poll */
1847 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1850 INIT_WORK(&req->work, aio_poll_complete_work);
1851 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1854 req->cancelled = false;
1855 req->work_scheduled = false;
1856 req->work_need_resched = false;
1858 apt.pt._qproc = aio_poll_queue_proc;
1859 apt.pt._key = req->events;
1862 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1864 /* initialized the list so that we can do list_empty checks */
1865 INIT_LIST_HEAD(&req->wait.entry);
1866 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1868 mask = vfs_poll(req->file, &apt.pt) & req->events;
1869 spin_lock_irq(&ctx->ctx_lock);
1870 if (likely(apt.queued)) {
1871 bool on_queue = poll_iocb_lock_wq(req);
1873 if (!on_queue || req->work_scheduled) {
1875 * aio_poll_wake() already either scheduled the async
1876 * completion work, or completed the request inline.
1878 if (apt.error) /* unsupported case: multiple queues */
1883 if (mask || apt.error) {
1884 /* Steal to complete synchronously. */
1885 list_del_init(&req->wait.entry);
1886 } else if (cancel) {
1887 /* Cancel if possible (may be too late though). */
1888 WRITE_ONCE(req->cancelled, true);
1889 } else if (on_queue) {
1891 * Actually waiting for an event, so add the request to
1892 * active_reqs so that it can be cancelled if needed.
1894 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1895 aiocb->ki_cancel = aio_poll_cancel;
1898 poll_iocb_unlock_wq(req);
1900 if (mask) { /* no async, we'd stolen it */
1901 aiocb->ki_res.res = mangle_poll(mask);
1904 spin_unlock_irq(&ctx->ctx_lock);
1910 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1911 struct iocb __user *user_iocb, bool compat)
1913 struct aio_kiocb *req;
1916 /* enforce forwards compatibility on users */
1917 if (unlikely(iocb->aio_reserved2)) {
1918 pr_debug("EINVAL: reserve field set\n");
1922 /* prevent overflows */
1924 (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
1925 (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
1926 ((ssize_t)iocb->aio_nbytes < 0)
1928 pr_debug("EINVAL: overflow check\n");
1932 if (!get_reqs_available(ctx))
1936 req = aio_get_req(ctx);
1938 goto out_put_reqs_available;
1940 req->ki_filp = fget(iocb->aio_fildes);
1942 if (unlikely(!req->ki_filp))
1945 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1947 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1948 * instance of the file* now. The file descriptor must be
1949 * an eventfd() fd, and will be signaled for each completed
1950 * event using the eventfd_signal() function.
1952 req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
1953 if (IS_ERR(req->ki_eventfd)) {
1954 ret = PTR_ERR(req->ki_eventfd);
1955 req->ki_eventfd = NULL;
1960 ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1961 if (unlikely(ret)) {
1962 pr_debug("EFAULT: aio_key\n");
1966 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1967 req->ki_res.data = iocb->aio_data;
1968 req->ki_res.res = 0;
1969 req->ki_res.res2 = 0;
1971 switch (iocb->aio_lio_opcode) {
1972 case IOCB_CMD_PREAD:
1973 ret = aio_read(&req->rw, iocb, false, compat);
1975 case IOCB_CMD_PWRITE:
1976 ret = aio_write(&req->rw, iocb, false, compat);
1978 case IOCB_CMD_PREADV:
1979 ret = aio_read(&req->rw, iocb, true, compat);
1981 case IOCB_CMD_PWRITEV:
1982 ret = aio_write(&req->rw, iocb, true, compat);
1984 case IOCB_CMD_FSYNC:
1985 ret = aio_fsync(&req->fsync, iocb, false);
1987 case IOCB_CMD_FDSYNC:
1988 ret = aio_fsync(&req->fsync, iocb, true);
1991 ret = aio_poll(req, iocb);
1994 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1999 /* Done with the synchronous reference */
2003 * If ret is 0, we'd either done aio_complete() ourselves or have
2004 * arranged for that to be done asynchronously. Anything non-zero
2005 * means that we need to destroy req ourselves.
2011 if (req->ki_eventfd)
2012 eventfd_ctx_put(req->ki_eventfd);
2014 out_put_reqs_available:
2015 put_reqs_available(ctx, 1);
2019 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2024 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2027 return __io_submit_one(ctx, &iocb, user_iocb, compat);
2031 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2032 * the number of iocbs queued. May return -EINVAL if the aio_context
2033 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2034 * *iocbpp[0] is not properly initialized, if the operation specified
2035 * is invalid for the file descriptor in the iocb. May fail with
2036 * -EFAULT if any of the data structures point to invalid data. May
2037 * fail with -EBADF if the file descriptor specified in the first
2038 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2039 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2040 * fail with -ENOSYS if not implemented.
2042 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2043 struct iocb __user * __user *, iocbpp)
2048 struct blk_plug plug;
2050 if (unlikely(nr < 0))
2053 ctx = lookup_ioctx(ctx_id);
2054 if (unlikely(!ctx)) {
2055 pr_debug("EINVAL: invalid context id\n");
2059 if (nr > ctx->nr_events)
2060 nr = ctx->nr_events;
2062 blk_start_plug(&plug);
2063 for (i = 0; i < nr; i++) {
2064 struct iocb __user *user_iocb;
2066 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2071 ret = io_submit_one(ctx, user_iocb, false);
2075 blk_finish_plug(&plug);
2077 percpu_ref_put(&ctx->users);
2081 #ifdef CONFIG_COMPAT
2082 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2083 int, nr, compat_uptr_t __user *, iocbpp)
2088 struct blk_plug plug;
2090 if (unlikely(nr < 0))
2093 ctx = lookup_ioctx(ctx_id);
2094 if (unlikely(!ctx)) {
2095 pr_debug("EINVAL: invalid context id\n");
2099 if (nr > ctx->nr_events)
2100 nr = ctx->nr_events;
2102 blk_start_plug(&plug);
2103 for (i = 0; i < nr; i++) {
2104 compat_uptr_t user_iocb;
2106 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2111 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2115 blk_finish_plug(&plug);
2117 percpu_ref_put(&ctx->users);
2123 * Attempts to cancel an iocb previously passed to io_submit. If
2124 * the operation is successfully cancelled, the resulting event is
2125 * copied into the memory pointed to by result without being placed
2126 * into the completion queue and 0 is returned. May fail with
2127 * -EFAULT if any of the data structures pointed to are invalid.
2128 * May fail with -EINVAL if aio_context specified by ctx_id is
2129 * invalid. May fail with -EAGAIN if the iocb specified was not
2130 * cancelled. Will fail with -ENOSYS if not implemented.
2132 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2133 struct io_event __user *, result)
2136 struct aio_kiocb *kiocb;
2139 u64 obj = (u64)(unsigned long)iocb;
2141 if (unlikely(get_user(key, &iocb->aio_key)))
2143 if (unlikely(key != KIOCB_KEY))
2146 ctx = lookup_ioctx(ctx_id);
2150 spin_lock_irq(&ctx->ctx_lock);
2151 /* TODO: use a hash or array, this sucks. */
2152 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2153 if (kiocb->ki_res.obj == obj) {
2154 ret = kiocb->ki_cancel(&kiocb->rw);
2155 list_del_init(&kiocb->ki_list);
2159 spin_unlock_irq(&ctx->ctx_lock);
2163 * The result argument is no longer used - the io_event is
2164 * always delivered via the ring buffer. -EINPROGRESS indicates
2165 * cancellation is progress:
2170 percpu_ref_put(&ctx->users);
2175 static long do_io_getevents(aio_context_t ctx_id,
2178 struct io_event __user *events,
2179 struct timespec64 *ts)
2181 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2182 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2185 if (likely(ioctx)) {
2186 if (likely(min_nr <= nr && min_nr >= 0))
2187 ret = read_events(ioctx, min_nr, nr, events, until);
2188 percpu_ref_put(&ioctx->users);
2195 * Attempts to read at least min_nr events and up to nr events from
2196 * the completion queue for the aio_context specified by ctx_id. If
2197 * it succeeds, the number of read events is returned. May fail with
2198 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2199 * out of range, if timeout is out of range. May fail with -EFAULT
2200 * if any of the memory specified is invalid. May return 0 or
2201 * < min_nr if the timeout specified by timeout has elapsed
2202 * before sufficient events are available, where timeout == NULL
2203 * specifies an infinite timeout. Note that the timeout pointed to by
2204 * timeout is relative. Will fail with -ENOSYS if not implemented.
2206 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2209 struct io_event __user *, events,
2210 struct timespec __user *, timeout)
2212 struct timespec64 ts;
2215 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2218 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2219 if (!ret && signal_pending(current))
2224 struct __aio_sigset {
2225 const sigset_t __user *sigmask;
2229 SYSCALL_DEFINE6(io_pgetevents,
2230 aio_context_t, ctx_id,
2233 struct io_event __user *, events,
2234 struct timespec __user *, timeout,
2235 const struct __aio_sigset __user *, usig)
2237 struct __aio_sigset ksig = { NULL, };
2238 sigset_t ksigmask, sigsaved;
2239 struct timespec64 ts;
2242 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2245 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2249 if (ksig.sigsetsize != sizeof(sigset_t))
2251 if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask)))
2253 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2254 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2257 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2258 if (signal_pending(current)) {
2260 current->saved_sigmask = sigsaved;
2261 set_restore_sigmask();
2265 ret = -ERESTARTNOHAND;
2268 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2274 #ifdef CONFIG_COMPAT
2275 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
2276 compat_long_t, min_nr,
2278 struct io_event __user *, events,
2279 struct compat_timespec __user *, timeout)
2281 struct timespec64 t;
2284 if (timeout && compat_get_timespec64(&t, timeout))
2287 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2288 if (!ret && signal_pending(current))
2294 struct __compat_aio_sigset {
2295 compat_sigset_t __user *sigmask;
2296 compat_size_t sigsetsize;
2299 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2300 compat_aio_context_t, ctx_id,
2301 compat_long_t, min_nr,
2303 struct io_event __user *, events,
2304 struct compat_timespec __user *, timeout,
2305 const struct __compat_aio_sigset __user *, usig)
2307 struct __compat_aio_sigset ksig = { NULL, };
2308 sigset_t ksigmask, sigsaved;
2309 struct timespec64 t;
2312 if (timeout && compat_get_timespec64(&t, timeout))
2315 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2319 if (ksig.sigsetsize != sizeof(compat_sigset_t))
2321 if (get_compat_sigset(&ksigmask, ksig.sigmask))
2323 sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2324 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2327 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2328 if (signal_pending(current)) {
2330 current->saved_sigmask = sigsaved;
2331 set_restore_sigmask();
2334 ret = -ERESTARTNOHAND;
2337 sigprocmask(SIG_SETMASK, &sigsaved, NULL);