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>
45 #include <linux/pseudo_fs.h>
47 #include <asm/kmap_types.h>
48 #include <linux/uaccess.h>
49 #include <linux/nospec.h>
55 #define AIO_RING_MAGIC 0xa10a10a1
56 #define AIO_RING_COMPAT_FEATURES 1
57 #define AIO_RING_INCOMPAT_FEATURES 0
59 unsigned id; /* kernel internal index number */
60 unsigned nr; /* number of io_events */
61 unsigned head; /* Written to by userland or under ring_lock
62 * mutex by aio_read_events_ring(). */
66 unsigned compat_features;
67 unsigned incompat_features;
68 unsigned header_length; /* size of aio_ring */
71 struct io_event io_events[0];
72 }; /* 128 bytes + ring size */
75 * Plugging is meant to work with larger batches of IOs. If we don't
76 * have more than the below, then don't bother setting up a plug.
78 #define AIO_PLUG_THRESHOLD 2
80 #define AIO_RING_PAGES 8
85 struct kioctx __rcu *table[];
89 unsigned reqs_available;
93 struct completion comp;
98 struct percpu_ref users;
101 struct percpu_ref reqs;
103 unsigned long user_id;
105 struct __percpu kioctx_cpu *cpu;
108 * For percpu reqs_available, number of slots we move to/from global
113 * This is what userspace passed to io_setup(), it's not used for
114 * anything but counting against the global max_reqs quota.
116 * The real limit is nr_events - 1, which will be larger (see
121 /* Size of ringbuffer, in units of struct io_event */
124 unsigned long mmap_base;
125 unsigned long mmap_size;
127 struct page **ring_pages;
130 struct rcu_work free_rwork; /* see free_ioctx() */
133 * signals when all in-flight requests are done
135 struct ctx_rq_wait *rq_wait;
139 * This counts the number of available slots in the ringbuffer,
140 * so we avoid overflowing it: it's decremented (if positive)
141 * when allocating a kiocb and incremented when the resulting
142 * io_event is pulled off the ringbuffer.
144 * We batch accesses to it with a percpu version.
146 atomic_t reqs_available;
147 } ____cacheline_aligned_in_smp;
151 struct list_head active_reqs; /* used for cancellation */
152 } ____cacheline_aligned_in_smp;
155 struct mutex ring_lock;
156 wait_queue_head_t wait;
157 } ____cacheline_aligned_in_smp;
161 unsigned completed_events;
162 spinlock_t completion_lock;
163 } ____cacheline_aligned_in_smp;
165 struct page *internal_pages[AIO_RING_PAGES];
166 struct file *aio_ring_file;
172 * First field must be the file pointer in all the
173 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
177 struct work_struct work;
184 struct wait_queue_head *head;
188 bool work_need_resched;
189 struct wait_queue_entry wait;
190 struct work_struct work;
194 * NOTE! Each of the iocb union members has the file pointer
195 * as the first entry in their struct definition. So you can
196 * access the file pointer through any of the sub-structs,
197 * or directly as just 'ki_filp' in this struct.
201 struct file *ki_filp;
203 struct fsync_iocb fsync;
204 struct poll_iocb poll;
207 struct kioctx *ki_ctx;
208 kiocb_cancel_fn *ki_cancel;
210 struct io_event ki_res;
212 struct list_head ki_list; /* the aio core uses this
213 * for cancellation */
214 refcount_t ki_refcnt;
217 * If the aio_resfd field of the userspace iocb is not zero,
218 * this is the underlying eventfd context to deliver events to.
220 struct eventfd_ctx *ki_eventfd;
223 /*------ sysctl variables----*/
224 static DEFINE_SPINLOCK(aio_nr_lock);
225 unsigned long aio_nr; /* current system wide number of aio requests */
226 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
227 /*----end sysctl variables---*/
229 static struct kmem_cache *kiocb_cachep;
230 static struct kmem_cache *kioctx_cachep;
232 static struct vfsmount *aio_mnt;
234 static const struct file_operations aio_ring_fops;
235 static const struct address_space_operations aio_ctx_aops;
237 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
240 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
242 return ERR_CAST(inode);
244 inode->i_mapping->a_ops = &aio_ctx_aops;
245 inode->i_mapping->private_data = ctx;
246 inode->i_size = PAGE_SIZE * nr_pages;
248 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
249 O_RDWR, &aio_ring_fops);
255 static int aio_init_fs_context(struct fs_context *fc)
257 if (!init_pseudo(fc, AIO_RING_MAGIC))
259 fc->s_iflags |= SB_I_NOEXEC;
264 * Creates the slab caches used by the aio routines, panic on
265 * failure as this is done early during the boot sequence.
267 static int __init aio_setup(void)
269 static struct file_system_type aio_fs = {
271 .init_fs_context = aio_init_fs_context,
272 .kill_sb = kill_anon_super,
274 aio_mnt = kern_mount(&aio_fs);
276 panic("Failed to create aio fs mount.");
278 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
279 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
282 __initcall(aio_setup);
284 static void put_aio_ring_file(struct kioctx *ctx)
286 struct file *aio_ring_file = ctx->aio_ring_file;
287 struct address_space *i_mapping;
290 truncate_setsize(file_inode(aio_ring_file), 0);
292 /* Prevent further access to the kioctx from migratepages */
293 i_mapping = aio_ring_file->f_mapping;
294 spin_lock(&i_mapping->private_lock);
295 i_mapping->private_data = NULL;
296 ctx->aio_ring_file = NULL;
297 spin_unlock(&i_mapping->private_lock);
303 static void aio_free_ring(struct kioctx *ctx)
307 /* Disconnect the kiotx from the ring file. This prevents future
308 * accesses to the kioctx from page migration.
310 put_aio_ring_file(ctx);
312 for (i = 0; i < ctx->nr_pages; i++) {
314 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
315 page_count(ctx->ring_pages[i]));
316 page = ctx->ring_pages[i];
319 ctx->ring_pages[i] = NULL;
323 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
324 kfree(ctx->ring_pages);
325 ctx->ring_pages = NULL;
329 static int aio_ring_mremap(struct vm_area_struct *vma)
331 struct file *file = vma->vm_file;
332 struct mm_struct *mm = vma->vm_mm;
333 struct kioctx_table *table;
334 int i, res = -EINVAL;
336 spin_lock(&mm->ioctx_lock);
338 table = rcu_dereference(mm->ioctx_table);
342 for (i = 0; i < table->nr; i++) {
345 ctx = rcu_dereference(table->table[i]);
346 if (ctx && ctx->aio_ring_file == file) {
347 if (!atomic_read(&ctx->dead)) {
348 ctx->user_id = ctx->mmap_base = vma->vm_start;
357 spin_unlock(&mm->ioctx_lock);
361 static const struct vm_operations_struct aio_ring_vm_ops = {
362 .mremap = aio_ring_mremap,
363 #if IS_ENABLED(CONFIG_MMU)
364 .fault = filemap_fault,
365 .map_pages = filemap_map_pages,
366 .page_mkwrite = filemap_page_mkwrite,
370 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
372 vma->vm_flags |= VM_DONTEXPAND;
373 vma->vm_ops = &aio_ring_vm_ops;
377 static const struct file_operations aio_ring_fops = {
378 .mmap = aio_ring_mmap,
381 #if IS_ENABLED(CONFIG_MIGRATION)
382 static int aio_migratepage(struct address_space *mapping, struct page *new,
383 struct page *old, enum migrate_mode mode)
391 * We cannot support the _NO_COPY case here, because copy needs to
392 * happen under the ctx->completion_lock. That does not work with the
393 * migration workflow of MIGRATE_SYNC_NO_COPY.
395 if (mode == MIGRATE_SYNC_NO_COPY)
400 /* mapping->private_lock here protects against the kioctx teardown. */
401 spin_lock(&mapping->private_lock);
402 ctx = mapping->private_data;
408 /* The ring_lock mutex. The prevents aio_read_events() from writing
409 * to the ring's head, and prevents page migration from mucking in
410 * a partially initialized kiotx.
412 if (!mutex_trylock(&ctx->ring_lock)) {
418 if (idx < (pgoff_t)ctx->nr_pages) {
419 /* Make sure the old page hasn't already been changed */
420 if (ctx->ring_pages[idx] != old)
428 /* Writeback must be complete */
429 BUG_ON(PageWriteback(old));
432 rc = migrate_page_move_mapping(mapping, new, old, 1);
433 if (rc != MIGRATEPAGE_SUCCESS) {
438 /* Take completion_lock to prevent other writes to the ring buffer
439 * while the old page is copied to the new. This prevents new
440 * events from being lost.
442 spin_lock_irqsave(&ctx->completion_lock, flags);
443 migrate_page_copy(new, old);
444 BUG_ON(ctx->ring_pages[idx] != old);
445 ctx->ring_pages[idx] = new;
446 spin_unlock_irqrestore(&ctx->completion_lock, flags);
448 /* The old page is no longer accessible. */
452 mutex_unlock(&ctx->ring_lock);
454 spin_unlock(&mapping->private_lock);
459 static const struct address_space_operations aio_ctx_aops = {
460 .set_page_dirty = __set_page_dirty_no_writeback,
461 #if IS_ENABLED(CONFIG_MIGRATION)
462 .migratepage = aio_migratepage,
466 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
468 struct aio_ring *ring;
469 struct mm_struct *mm = current->mm;
470 unsigned long size, unused;
475 /* Compensate for the ring buffer's head/tail overlap entry */
476 nr_events += 2; /* 1 is required, 2 for good luck */
478 size = sizeof(struct aio_ring);
479 size += sizeof(struct io_event) * nr_events;
481 nr_pages = PFN_UP(size);
485 file = aio_private_file(ctx, nr_pages);
487 ctx->aio_ring_file = NULL;
491 ctx->aio_ring_file = file;
492 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
493 / sizeof(struct io_event);
495 ctx->ring_pages = ctx->internal_pages;
496 if (nr_pages > AIO_RING_PAGES) {
497 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
499 if (!ctx->ring_pages) {
500 put_aio_ring_file(ctx);
505 for (i = 0; i < nr_pages; i++) {
507 page = find_or_create_page(file->f_mapping,
508 i, GFP_HIGHUSER | __GFP_ZERO);
511 pr_debug("pid(%d) page[%d]->count=%d\n",
512 current->pid, i, page_count(page));
513 SetPageUptodate(page);
516 ctx->ring_pages[i] = page;
520 if (unlikely(i != nr_pages)) {
525 ctx->mmap_size = nr_pages * PAGE_SIZE;
526 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
528 if (down_write_killable(&mm->mmap_sem)) {
534 ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
535 PROT_READ | PROT_WRITE,
536 MAP_SHARED, 0, &unused, NULL);
537 up_write(&mm->mmap_sem);
538 if (IS_ERR((void *)ctx->mmap_base)) {
544 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
546 ctx->user_id = ctx->mmap_base;
547 ctx->nr_events = nr_events; /* trusted copy */
549 ring = kmap_atomic(ctx->ring_pages[0]);
550 ring->nr = nr_events; /* user copy */
552 ring->head = ring->tail = 0;
553 ring->magic = AIO_RING_MAGIC;
554 ring->compat_features = AIO_RING_COMPAT_FEATURES;
555 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
556 ring->header_length = sizeof(struct aio_ring);
558 flush_dcache_page(ctx->ring_pages[0]);
563 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
564 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
565 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
567 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
569 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
570 struct kioctx *ctx = req->ki_ctx;
573 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
576 spin_lock_irqsave(&ctx->ctx_lock, flags);
577 list_add_tail(&req->ki_list, &ctx->active_reqs);
578 req->ki_cancel = cancel;
579 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
581 EXPORT_SYMBOL(kiocb_set_cancel_fn);
584 * free_ioctx() should be RCU delayed to synchronize against the RCU
585 * protected lookup_ioctx() and also needs process context to call
586 * aio_free_ring(). Use rcu_work.
588 static void free_ioctx(struct work_struct *work)
590 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
592 pr_debug("freeing %p\n", ctx);
595 free_percpu(ctx->cpu);
596 percpu_ref_exit(&ctx->reqs);
597 percpu_ref_exit(&ctx->users);
598 kmem_cache_free(kioctx_cachep, ctx);
601 static void free_ioctx_reqs(struct percpu_ref *ref)
603 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
605 /* At this point we know that there are no any in-flight requests */
606 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
607 complete(&ctx->rq_wait->comp);
609 /* Synchronize against RCU protected table->table[] dereferences */
610 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
611 queue_rcu_work(system_wq, &ctx->free_rwork);
615 * When this function runs, the kioctx has been removed from the "hash table"
616 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
617 * now it's safe to cancel any that need to be.
619 static void free_ioctx_users(struct percpu_ref *ref)
621 struct kioctx *ctx = container_of(ref, struct kioctx, users);
622 struct aio_kiocb *req;
624 spin_lock_irq(&ctx->ctx_lock);
626 while (!list_empty(&ctx->active_reqs)) {
627 req = list_first_entry(&ctx->active_reqs,
628 struct aio_kiocb, ki_list);
629 req->ki_cancel(&req->rw);
630 list_del_init(&req->ki_list);
633 spin_unlock_irq(&ctx->ctx_lock);
635 percpu_ref_kill(&ctx->reqs);
636 percpu_ref_put(&ctx->reqs);
639 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
642 struct kioctx_table *table, *old;
643 struct aio_ring *ring;
645 spin_lock(&mm->ioctx_lock);
646 table = rcu_dereference_raw(mm->ioctx_table);
650 for (i = 0; i < table->nr; i++)
651 if (!rcu_access_pointer(table->table[i])) {
653 rcu_assign_pointer(table->table[i], ctx);
654 spin_unlock(&mm->ioctx_lock);
656 /* While kioctx setup is in progress,
657 * we are protected from page migration
658 * changes ring_pages by ->ring_lock.
660 ring = kmap_atomic(ctx->ring_pages[0]);
666 new_nr = (table ? table->nr : 1) * 4;
667 spin_unlock(&mm->ioctx_lock);
669 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
676 spin_lock(&mm->ioctx_lock);
677 old = rcu_dereference_raw(mm->ioctx_table);
680 rcu_assign_pointer(mm->ioctx_table, table);
681 } else if (table->nr > old->nr) {
682 memcpy(table->table, old->table,
683 old->nr * sizeof(struct kioctx *));
685 rcu_assign_pointer(mm->ioctx_table, table);
694 static void aio_nr_sub(unsigned nr)
696 spin_lock(&aio_nr_lock);
697 if (WARN_ON(aio_nr - nr > aio_nr))
701 spin_unlock(&aio_nr_lock);
705 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
707 static struct kioctx *ioctx_alloc(unsigned nr_events)
709 struct mm_struct *mm = current->mm;
714 * Store the original nr_events -- what userspace passed to io_setup(),
715 * for counting against the global limit -- before it changes.
717 unsigned int max_reqs = nr_events;
720 * We keep track of the number of available ringbuffer slots, to prevent
721 * overflow (reqs_available), and we also use percpu counters for this.
723 * So since up to half the slots might be on other cpu's percpu counters
724 * and unavailable, double nr_events so userspace sees what they
725 * expected: additionally, we move req_batch slots to/from percpu
726 * counters at a time, so make sure that isn't 0:
728 nr_events = max(nr_events, num_possible_cpus() * 4);
731 /* Prevent overflows */
732 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
733 pr_debug("ENOMEM: nr_events too high\n");
734 return ERR_PTR(-EINVAL);
737 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
738 return ERR_PTR(-EAGAIN);
740 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
742 return ERR_PTR(-ENOMEM);
744 ctx->max_reqs = max_reqs;
746 spin_lock_init(&ctx->ctx_lock);
747 spin_lock_init(&ctx->completion_lock);
748 mutex_init(&ctx->ring_lock);
749 /* Protect against page migration throughout kiotx setup by keeping
750 * the ring_lock mutex held until setup is complete. */
751 mutex_lock(&ctx->ring_lock);
752 init_waitqueue_head(&ctx->wait);
754 INIT_LIST_HEAD(&ctx->active_reqs);
756 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
759 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
762 ctx->cpu = alloc_percpu(struct kioctx_cpu);
766 err = aio_setup_ring(ctx, nr_events);
770 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
771 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
772 if (ctx->req_batch < 1)
775 /* limit the number of system wide aios */
776 spin_lock(&aio_nr_lock);
777 if (aio_nr + ctx->max_reqs > aio_max_nr ||
778 aio_nr + ctx->max_reqs < aio_nr) {
779 spin_unlock(&aio_nr_lock);
783 aio_nr += ctx->max_reqs;
784 spin_unlock(&aio_nr_lock);
786 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
787 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
789 err = ioctx_add_table(ctx, mm);
793 /* Release the ring_lock mutex now that all setup is complete. */
794 mutex_unlock(&ctx->ring_lock);
796 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
797 ctx, ctx->user_id, mm, ctx->nr_events);
801 aio_nr_sub(ctx->max_reqs);
803 atomic_set(&ctx->dead, 1);
805 vm_munmap(ctx->mmap_base, ctx->mmap_size);
808 mutex_unlock(&ctx->ring_lock);
809 free_percpu(ctx->cpu);
810 percpu_ref_exit(&ctx->reqs);
811 percpu_ref_exit(&ctx->users);
812 kmem_cache_free(kioctx_cachep, ctx);
813 pr_debug("error allocating ioctx %d\n", err);
818 * Cancels all outstanding aio requests on an aio context. Used
819 * when the processes owning a context have all exited to encourage
820 * the rapid destruction of the kioctx.
822 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
823 struct ctx_rq_wait *wait)
825 struct kioctx_table *table;
827 spin_lock(&mm->ioctx_lock);
828 if (atomic_xchg(&ctx->dead, 1)) {
829 spin_unlock(&mm->ioctx_lock);
833 table = rcu_dereference_raw(mm->ioctx_table);
834 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
835 RCU_INIT_POINTER(table->table[ctx->id], NULL);
836 spin_unlock(&mm->ioctx_lock);
838 /* free_ioctx_reqs() will do the necessary RCU synchronization */
839 wake_up_all(&ctx->wait);
842 * It'd be more correct to do this in free_ioctx(), after all
843 * the outstanding kiocbs have finished - but by then io_destroy
844 * has already returned, so io_setup() could potentially return
845 * -EAGAIN with no ioctxs actually in use (as far as userspace
848 aio_nr_sub(ctx->max_reqs);
851 vm_munmap(ctx->mmap_base, ctx->mmap_size);
854 percpu_ref_kill(&ctx->users);
859 * exit_aio: called when the last user of mm goes away. At this point, there is
860 * no way for any new requests to be submited or any of the io_* syscalls to be
861 * called on the context.
863 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
866 void exit_aio(struct mm_struct *mm)
868 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
869 struct ctx_rq_wait wait;
875 atomic_set(&wait.count, table->nr);
876 init_completion(&wait.comp);
879 for (i = 0; i < table->nr; ++i) {
881 rcu_dereference_protected(table->table[i], true);
889 * We don't need to bother with munmap() here - exit_mmap(mm)
890 * is coming and it'll unmap everything. And we simply can't,
891 * this is not necessarily our ->mm.
892 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
893 * that it needs to unmap the area, just set it to 0.
896 kill_ioctx(mm, ctx, &wait);
899 if (!atomic_sub_and_test(skipped, &wait.count)) {
900 /* Wait until all IO for the context are done. */
901 wait_for_completion(&wait.comp);
904 RCU_INIT_POINTER(mm->ioctx_table, NULL);
908 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
910 struct kioctx_cpu *kcpu;
913 local_irq_save(flags);
914 kcpu = this_cpu_ptr(ctx->cpu);
915 kcpu->reqs_available += nr;
917 while (kcpu->reqs_available >= ctx->req_batch * 2) {
918 kcpu->reqs_available -= ctx->req_batch;
919 atomic_add(ctx->req_batch, &ctx->reqs_available);
922 local_irq_restore(flags);
925 static bool __get_reqs_available(struct kioctx *ctx)
927 struct kioctx_cpu *kcpu;
931 local_irq_save(flags);
932 kcpu = this_cpu_ptr(ctx->cpu);
933 if (!kcpu->reqs_available) {
934 int old, avail = atomic_read(&ctx->reqs_available);
937 if (avail < ctx->req_batch)
941 avail = atomic_cmpxchg(&ctx->reqs_available,
942 avail, avail - ctx->req_batch);
943 } while (avail != old);
945 kcpu->reqs_available += ctx->req_batch;
949 kcpu->reqs_available--;
951 local_irq_restore(flags);
955 /* refill_reqs_available
956 * Updates the reqs_available reference counts used for tracking the
957 * number of free slots in the completion ring. This can be called
958 * from aio_complete() (to optimistically update reqs_available) or
959 * from aio_get_req() (the we're out of events case). It must be
960 * called holding ctx->completion_lock.
962 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
965 unsigned events_in_ring, completed;
967 /* Clamp head since userland can write to it. */
968 head %= ctx->nr_events;
970 events_in_ring = tail - head;
972 events_in_ring = ctx->nr_events - (head - tail);
974 completed = ctx->completed_events;
975 if (events_in_ring < completed)
976 completed -= events_in_ring;
983 ctx->completed_events -= completed;
984 put_reqs_available(ctx, completed);
987 /* user_refill_reqs_available
988 * Called to refill reqs_available when aio_get_req() encounters an
989 * out of space in the completion ring.
991 static void user_refill_reqs_available(struct kioctx *ctx)
993 spin_lock_irq(&ctx->completion_lock);
994 if (ctx->completed_events) {
995 struct aio_ring *ring;
998 /* Access of ring->head may race with aio_read_events_ring()
999 * here, but that's okay since whether we read the old version
1000 * or the new version, and either will be valid. The important
1001 * part is that head cannot pass tail since we prevent
1002 * aio_complete() from updating tail by holding
1003 * ctx->completion_lock. Even if head is invalid, the check
1004 * against ctx->completed_events below will make sure we do the
1007 ring = kmap_atomic(ctx->ring_pages[0]);
1009 kunmap_atomic(ring);
1011 refill_reqs_available(ctx, head, ctx->tail);
1014 spin_unlock_irq(&ctx->completion_lock);
1017 static bool get_reqs_available(struct kioctx *ctx)
1019 if (__get_reqs_available(ctx))
1021 user_refill_reqs_available(ctx);
1022 return __get_reqs_available(ctx);
1026 * Allocate a slot for an aio request.
1027 * Returns NULL if no requests are free.
1029 * The refcount is initialized to 2 - one for the async op completion,
1030 * one for the synchronous code that does this.
1032 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1034 struct aio_kiocb *req;
1036 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1040 if (unlikely(!get_reqs_available(ctx))) {
1041 kmem_cache_free(kiocb_cachep, req);
1045 percpu_ref_get(&ctx->reqs);
1047 INIT_LIST_HEAD(&req->ki_list);
1048 refcount_set(&req->ki_refcnt, 2);
1049 req->ki_eventfd = NULL;
1053 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1055 struct aio_ring __user *ring = (void __user *)ctx_id;
1056 struct mm_struct *mm = current->mm;
1057 struct kioctx *ctx, *ret = NULL;
1058 struct kioctx_table *table;
1061 if (get_user(id, &ring->id))
1065 table = rcu_dereference(mm->ioctx_table);
1067 if (!table || id >= table->nr)
1070 id = array_index_nospec(id, table->nr);
1071 ctx = rcu_dereference(table->table[id]);
1072 if (ctx && ctx->user_id == ctx_id) {
1073 if (percpu_ref_tryget_live(&ctx->users))
1081 static inline void iocb_destroy(struct aio_kiocb *iocb)
1083 if (iocb->ki_eventfd)
1084 eventfd_ctx_put(iocb->ki_eventfd);
1086 fput(iocb->ki_filp);
1087 percpu_ref_put(&iocb->ki_ctx->reqs);
1088 kmem_cache_free(kiocb_cachep, iocb);
1092 * Called when the io request on the given iocb is complete.
1094 static void aio_complete(struct aio_kiocb *iocb)
1096 struct kioctx *ctx = iocb->ki_ctx;
1097 struct aio_ring *ring;
1098 struct io_event *ev_page, *event;
1099 unsigned tail, pos, head;
1100 unsigned long flags;
1103 * Add a completion event to the ring buffer. Must be done holding
1104 * ctx->completion_lock to prevent other code from messing with the tail
1105 * pointer since we might be called from irq context.
1107 spin_lock_irqsave(&ctx->completion_lock, flags);
1110 pos = tail + AIO_EVENTS_OFFSET;
1112 if (++tail >= ctx->nr_events)
1115 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1116 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1118 *event = iocb->ki_res;
1120 kunmap_atomic(ev_page);
1121 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1123 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1124 (void __user *)(unsigned long)iocb->ki_res.obj,
1125 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1127 /* after flagging the request as done, we
1128 * must never even look at it again
1130 smp_wmb(); /* make event visible before updating tail */
1134 ring = kmap_atomic(ctx->ring_pages[0]);
1137 kunmap_atomic(ring);
1138 flush_dcache_page(ctx->ring_pages[0]);
1140 ctx->completed_events++;
1141 if (ctx->completed_events > 1)
1142 refill_reqs_available(ctx, head, tail);
1143 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1145 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1148 * Check if the user asked us to deliver the result through an
1149 * eventfd. The eventfd_signal() function is safe to be called
1152 if (iocb->ki_eventfd)
1153 eventfd_signal(iocb->ki_eventfd, 1);
1156 * We have to order our ring_info tail store above and test
1157 * of the wait list below outside the wait lock. This is
1158 * like in wake_up_bit() where clearing a bit has to be
1159 * ordered with the unlocked test.
1163 if (waitqueue_active(&ctx->wait))
1164 wake_up(&ctx->wait);
1167 static inline void iocb_put(struct aio_kiocb *iocb)
1169 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1175 /* aio_read_events_ring
1176 * Pull an event off of the ioctx's event ring. Returns the number of
1179 static long aio_read_events_ring(struct kioctx *ctx,
1180 struct io_event __user *event, long nr)
1182 struct aio_ring *ring;
1183 unsigned head, tail, pos;
1188 * The mutex can block and wake us up and that will cause
1189 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1190 * and repeat. This should be rare enough that it doesn't cause
1191 * peformance issues. See the comment in read_events() for more detail.
1193 sched_annotate_sleep();
1194 mutex_lock(&ctx->ring_lock);
1196 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1197 ring = kmap_atomic(ctx->ring_pages[0]);
1200 kunmap_atomic(ring);
1203 * Ensure that once we've read the current tail pointer, that
1204 * we also see the events that were stored up to the tail.
1208 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1213 head %= ctx->nr_events;
1214 tail %= ctx->nr_events;
1218 struct io_event *ev;
1221 avail = (head <= tail ? tail : ctx->nr_events) - head;
1225 pos = head + AIO_EVENTS_OFFSET;
1226 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1227 pos %= AIO_EVENTS_PER_PAGE;
1229 avail = min(avail, nr - ret);
1230 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1233 copy_ret = copy_to_user(event + ret, ev + pos,
1234 sizeof(*ev) * avail);
1237 if (unlikely(copy_ret)) {
1244 head %= ctx->nr_events;
1247 ring = kmap_atomic(ctx->ring_pages[0]);
1249 kunmap_atomic(ring);
1250 flush_dcache_page(ctx->ring_pages[0]);
1252 pr_debug("%li h%u t%u\n", ret, head, tail);
1254 mutex_unlock(&ctx->ring_lock);
1259 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1260 struct io_event __user *event, long *i)
1262 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1267 if (unlikely(atomic_read(&ctx->dead)))
1273 return ret < 0 || *i >= min_nr;
1276 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1277 struct io_event __user *event,
1283 * Note that aio_read_events() is being called as the conditional - i.e.
1284 * we're calling it after prepare_to_wait() has set task state to
1285 * TASK_INTERRUPTIBLE.
1287 * But aio_read_events() can block, and if it blocks it's going to flip
1288 * the task state back to TASK_RUNNING.
1290 * This should be ok, provided it doesn't flip the state back to
1291 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1292 * will only happen if the mutex_lock() call blocks, and we then find
1293 * the ringbuffer empty. So in practice we should be ok, but it's
1294 * something to be aware of when touching this code.
1297 aio_read_events(ctx, min_nr, nr, event, &ret);
1299 wait_event_interruptible_hrtimeout(ctx->wait,
1300 aio_read_events(ctx, min_nr, nr, event, &ret),
1306 * Create an aio_context capable of receiving at least nr_events.
1307 * ctxp must not point to an aio_context that already exists, and
1308 * must be initialized to 0 prior to the call. On successful
1309 * creation of the aio_context, *ctxp is filled in with the resulting
1310 * handle. May fail with -EINVAL if *ctxp is not initialized,
1311 * if the specified nr_events exceeds internal limits. May fail
1312 * with -EAGAIN if the specified nr_events exceeds the user's limit
1313 * of available events. May fail with -ENOMEM if insufficient kernel
1314 * resources are available. May fail with -EFAULT if an invalid
1315 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1318 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1320 struct kioctx *ioctx = NULL;
1324 ret = get_user(ctx, ctxp);
1329 if (unlikely(ctx || nr_events == 0)) {
1330 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1335 ioctx = ioctx_alloc(nr_events);
1336 ret = PTR_ERR(ioctx);
1337 if (!IS_ERR(ioctx)) {
1338 ret = put_user(ioctx->user_id, ctxp);
1340 kill_ioctx(current->mm, ioctx, NULL);
1341 percpu_ref_put(&ioctx->users);
1348 #ifdef CONFIG_COMPAT
1349 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1351 struct kioctx *ioctx = NULL;
1355 ret = get_user(ctx, ctx32p);
1360 if (unlikely(ctx || nr_events == 0)) {
1361 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1366 ioctx = ioctx_alloc(nr_events);
1367 ret = PTR_ERR(ioctx);
1368 if (!IS_ERR(ioctx)) {
1369 /* truncating is ok because it's a user address */
1370 ret = put_user((u32)ioctx->user_id, ctx32p);
1372 kill_ioctx(current->mm, ioctx, NULL);
1373 percpu_ref_put(&ioctx->users);
1382 * Destroy the aio_context specified. May cancel any outstanding
1383 * AIOs and block on completion. Will fail with -ENOSYS if not
1384 * implemented. May fail with -EINVAL if the context pointed to
1387 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1389 struct kioctx *ioctx = lookup_ioctx(ctx);
1390 if (likely(NULL != ioctx)) {
1391 struct ctx_rq_wait wait;
1394 init_completion(&wait.comp);
1395 atomic_set(&wait.count, 1);
1397 /* Pass requests_done to kill_ioctx() where it can be set
1398 * in a thread-safe way. If we try to set it here then we have
1399 * a race condition if two io_destroy() called simultaneously.
1401 ret = kill_ioctx(current->mm, ioctx, &wait);
1402 percpu_ref_put(&ioctx->users);
1404 /* Wait until all IO for the context are done. Otherwise kernel
1405 * keep using user-space buffers even if user thinks the context
1409 wait_for_completion(&wait.comp);
1413 pr_debug("EINVAL: invalid context id\n");
1417 static void aio_remove_iocb(struct aio_kiocb *iocb)
1419 struct kioctx *ctx = iocb->ki_ctx;
1420 unsigned long flags;
1422 spin_lock_irqsave(&ctx->ctx_lock, flags);
1423 list_del(&iocb->ki_list);
1424 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1427 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1429 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1431 if (!list_empty_careful(&iocb->ki_list))
1432 aio_remove_iocb(iocb);
1434 if (kiocb->ki_flags & IOCB_WRITE) {
1435 struct inode *inode = file_inode(kiocb->ki_filp);
1438 * Tell lockdep we inherited freeze protection from submission
1441 if (S_ISREG(inode->i_mode))
1442 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1443 file_end_write(kiocb->ki_filp);
1446 iocb->ki_res.res = res;
1447 iocb->ki_res.res2 = res2;
1451 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1455 req->ki_complete = aio_complete_rw;
1456 req->private = NULL;
1457 req->ki_pos = iocb->aio_offset;
1458 req->ki_flags = iocb_flags(req->ki_filp);
1459 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1460 req->ki_flags |= IOCB_EVENTFD;
1461 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1462 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1464 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1465 * aio_reqprio is interpreted as an I/O scheduling
1466 * class and priority.
1468 ret = ioprio_check_cap(iocb->aio_reqprio);
1470 pr_debug("aio ioprio check cap error: %d\n", ret);
1474 req->ki_ioprio = iocb->aio_reqprio;
1476 req->ki_ioprio = get_current_ioprio();
1478 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1482 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1486 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1487 struct iovec **iovec, bool vectored, bool compat,
1488 struct iov_iter *iter)
1490 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1491 size_t len = iocb->aio_nbytes;
1494 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1498 #ifdef CONFIG_COMPAT
1500 return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1503 return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1506 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1512 case -ERESTARTNOINTR:
1513 case -ERESTARTNOHAND:
1514 case -ERESTART_RESTARTBLOCK:
1516 * There's no easy way to restart the syscall since other AIO's
1517 * may be already running. Just fail this IO with EINTR.
1522 req->ki_complete(req, ret, 0);
1526 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1527 bool vectored, bool compat)
1529 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1530 struct iov_iter iter;
1534 ret = aio_prep_rw(req, iocb);
1537 file = req->ki_filp;
1538 if (unlikely(!(file->f_mode & FMODE_READ)))
1541 if (unlikely(!file->f_op->read_iter))
1544 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1547 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1549 aio_rw_done(req, call_read_iter(file, req, &iter));
1554 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1555 bool vectored, bool compat)
1557 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1558 struct iov_iter iter;
1562 ret = aio_prep_rw(req, iocb);
1565 file = req->ki_filp;
1567 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1569 if (unlikely(!file->f_op->write_iter))
1572 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1575 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1578 * Open-code file_start_write here to grab freeze protection,
1579 * which will be released by another thread in
1580 * aio_complete_rw(). Fool lockdep by telling it the lock got
1581 * released so that it doesn't complain about the held lock when
1582 * we return to userspace.
1584 if (S_ISREG(file_inode(file)->i_mode)) {
1585 __sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1586 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1588 req->ki_flags |= IOCB_WRITE;
1589 aio_rw_done(req, call_write_iter(file, req, &iter));
1595 static void aio_fsync_work(struct work_struct *work)
1597 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1598 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1600 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1601 revert_creds(old_cred);
1602 put_cred(iocb->fsync.creds);
1606 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1609 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1610 iocb->aio_rw_flags))
1613 if (unlikely(!req->file->f_op->fsync))
1616 req->creds = prepare_creds();
1620 req->datasync = datasync;
1621 INIT_WORK(&req->work, aio_fsync_work);
1622 schedule_work(&req->work);
1626 static void aio_poll_put_work(struct work_struct *work)
1628 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1629 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1635 * Safely lock the waitqueue which the request is on, synchronizing with the
1636 * case where the ->poll() provider decides to free its waitqueue early.
1638 * Returns true on success, meaning that req->head->lock was locked, req->wait
1639 * is on req->head, and an RCU read lock was taken. Returns false if the
1640 * request was already removed from its waitqueue (which might no longer exist).
1642 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1644 wait_queue_head_t *head;
1647 * While we hold the waitqueue lock and the waitqueue is nonempty,
1648 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1649 * lock in the first place can race with the waitqueue being freed.
1651 * We solve this as eventpoll does: by taking advantage of the fact that
1652 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1653 * we enter rcu_read_lock() and see that the pointer to the queue is
1654 * non-NULL, we can then lock it without the memory being freed out from
1655 * under us, then check whether the request is still on the queue.
1657 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1658 * case the caller deletes the entry from the queue, leaving it empty.
1659 * In that case, only RCU prevents the queue memory from being freed.
1662 head = smp_load_acquire(&req->head);
1664 spin_lock(&head->lock);
1665 if (!list_empty(&req->wait.entry))
1667 spin_unlock(&head->lock);
1673 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1675 spin_unlock(&req->head->lock);
1679 static void aio_poll_complete_work(struct work_struct *work)
1681 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1682 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1683 struct poll_table_struct pt = { ._key = req->events };
1684 struct kioctx *ctx = iocb->ki_ctx;
1687 if (!READ_ONCE(req->cancelled))
1688 mask = vfs_poll(req->file, &pt) & req->events;
1691 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1692 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1693 * synchronize with them. In the cancellation case the list_del_init
1694 * itself is not actually needed, but harmless so we keep it in to
1695 * avoid further branches in the fast path.
1697 spin_lock_irq(&ctx->ctx_lock);
1698 if (poll_iocb_lock_wq(req)) {
1699 if (!mask && !READ_ONCE(req->cancelled)) {
1701 * The request isn't actually ready to be completed yet.
1702 * Reschedule completion if another wakeup came in.
1704 if (req->work_need_resched) {
1705 schedule_work(&req->work);
1706 req->work_need_resched = false;
1708 req->work_scheduled = false;
1710 poll_iocb_unlock_wq(req);
1711 spin_unlock_irq(&ctx->ctx_lock);
1714 list_del_init(&req->wait.entry);
1715 poll_iocb_unlock_wq(req);
1716 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1717 list_del_init(&iocb->ki_list);
1718 iocb->ki_res.res = mangle_poll(mask);
1719 spin_unlock_irq(&ctx->ctx_lock);
1724 /* assumes we are called with irqs disabled */
1725 static int aio_poll_cancel(struct kiocb *iocb)
1727 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1728 struct poll_iocb *req = &aiocb->poll;
1730 if (poll_iocb_lock_wq(req)) {
1731 WRITE_ONCE(req->cancelled, true);
1732 if (!req->work_scheduled) {
1733 schedule_work(&aiocb->poll.work);
1734 req->work_scheduled = true;
1736 poll_iocb_unlock_wq(req);
1737 } /* else, the request was force-cancelled by POLLFREE already */
1742 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1745 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1746 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1747 __poll_t mask = key_to_poll(key);
1748 unsigned long flags;
1750 /* for instances that support it check for an event match first: */
1751 if (mask && !(mask & req->events))
1755 * Complete the request inline if possible. This requires that three
1756 * conditions be met:
1757 * 1. An event mask must have been passed. If a plain wakeup was done
1758 * instead, then mask == 0 and we have to call vfs_poll() to get
1759 * the events, so inline completion isn't possible.
1760 * 2. The completion work must not have already been scheduled.
1761 * 3. ctx_lock must not be busy. We have to use trylock because we
1762 * already hold the waitqueue lock, so this inverts the normal
1763 * locking order. Use irqsave/irqrestore because not all
1764 * filesystems (e.g. fuse) call this function with IRQs disabled,
1765 * yet IRQs have to be disabled before ctx_lock is obtained.
1767 if (mask && !req->work_scheduled &&
1768 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1769 struct kioctx *ctx = iocb->ki_ctx;
1771 list_del_init(&req->wait.entry);
1772 list_del(&iocb->ki_list);
1773 iocb->ki_res.res = mangle_poll(mask);
1774 if (iocb->ki_eventfd && eventfd_signal_count()) {
1776 INIT_WORK(&req->work, aio_poll_put_work);
1777 schedule_work(&req->work);
1779 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1784 * Schedule the completion work if needed. If it was already
1785 * scheduled, record that another wakeup came in.
1787 * Don't remove the request from the waitqueue here, as it might
1788 * not actually be complete yet (we won't know until vfs_poll()
1789 * is called), and we must not miss any wakeups. POLLFREE is an
1790 * exception to this; see below.
1792 if (req->work_scheduled) {
1793 req->work_need_resched = true;
1795 schedule_work(&req->work);
1796 req->work_scheduled = true;
1800 * If the waitqueue is being freed early but we can't complete
1801 * the request inline, we have to tear down the request as best
1802 * we can. That means immediately removing the request from its
1803 * waitqueue and preventing all further accesses to the
1804 * waitqueue via the request. We also need to schedule the
1805 * completion work (done above). Also mark the request as
1806 * cancelled, to potentially skip an unneeded call to ->poll().
1808 if (mask & POLLFREE) {
1809 WRITE_ONCE(req->cancelled, true);
1810 list_del_init(&req->wait.entry);
1813 * Careful: this *must* be the last step, since as soon
1814 * as req->head is NULL'ed out, the request can be
1815 * completed and freed, since aio_poll_complete_work()
1816 * will no longer need to take the waitqueue lock.
1818 smp_store_release(&req->head, NULL);
1824 struct aio_poll_table {
1825 struct poll_table_struct pt;
1826 struct aio_kiocb *iocb;
1832 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1833 struct poll_table_struct *p)
1835 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1837 /* multiple wait queues per file are not supported */
1838 if (unlikely(pt->queued)) {
1839 pt->error = -EINVAL;
1845 pt->iocb->poll.head = head;
1846 add_wait_queue(head, &pt->iocb->poll.wait);
1849 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1851 struct kioctx *ctx = aiocb->ki_ctx;
1852 struct poll_iocb *req = &aiocb->poll;
1853 struct aio_poll_table apt;
1854 bool cancel = false;
1857 /* reject any unknown events outside the normal event mask. */
1858 if ((u16)iocb->aio_buf != iocb->aio_buf)
1860 /* reject fields that are not defined for poll */
1861 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1864 INIT_WORK(&req->work, aio_poll_complete_work);
1865 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1868 req->cancelled = false;
1869 req->work_scheduled = false;
1870 req->work_need_resched = false;
1872 apt.pt._qproc = aio_poll_queue_proc;
1873 apt.pt._key = req->events;
1876 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1878 /* initialized the list so that we can do list_empty checks */
1879 INIT_LIST_HEAD(&req->wait.entry);
1880 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1882 mask = vfs_poll(req->file, &apt.pt) & req->events;
1883 spin_lock_irq(&ctx->ctx_lock);
1884 if (likely(apt.queued)) {
1885 bool on_queue = poll_iocb_lock_wq(req);
1887 if (!on_queue || req->work_scheduled) {
1889 * aio_poll_wake() already either scheduled the async
1890 * completion work, or completed the request inline.
1892 if (apt.error) /* unsupported case: multiple queues */
1897 if (mask || apt.error) {
1898 /* Steal to complete synchronously. */
1899 list_del_init(&req->wait.entry);
1900 } else if (cancel) {
1901 /* Cancel if possible (may be too late though). */
1902 WRITE_ONCE(req->cancelled, true);
1903 } else if (on_queue) {
1905 * Actually waiting for an event, so add the request to
1906 * active_reqs so that it can be cancelled if needed.
1908 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1909 aiocb->ki_cancel = aio_poll_cancel;
1912 poll_iocb_unlock_wq(req);
1914 if (mask) { /* no async, we'd stolen it */
1915 aiocb->ki_res.res = mangle_poll(mask);
1918 spin_unlock_irq(&ctx->ctx_lock);
1924 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1925 struct iocb __user *user_iocb, struct aio_kiocb *req,
1928 req->ki_filp = fget(iocb->aio_fildes);
1929 if (unlikely(!req->ki_filp))
1932 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1933 struct eventfd_ctx *eventfd;
1935 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1936 * instance of the file* now. The file descriptor must be
1937 * an eventfd() fd, and will be signaled for each completed
1938 * event using the eventfd_signal() function.
1940 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1941 if (IS_ERR(eventfd))
1942 return PTR_ERR(eventfd);
1944 req->ki_eventfd = eventfd;
1947 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1948 pr_debug("EFAULT: aio_key\n");
1952 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1953 req->ki_res.data = iocb->aio_data;
1954 req->ki_res.res = 0;
1955 req->ki_res.res2 = 0;
1957 switch (iocb->aio_lio_opcode) {
1958 case IOCB_CMD_PREAD:
1959 return aio_read(&req->rw, iocb, false, compat);
1960 case IOCB_CMD_PWRITE:
1961 return aio_write(&req->rw, iocb, false, compat);
1962 case IOCB_CMD_PREADV:
1963 return aio_read(&req->rw, iocb, true, compat);
1964 case IOCB_CMD_PWRITEV:
1965 return aio_write(&req->rw, iocb, true, compat);
1966 case IOCB_CMD_FSYNC:
1967 return aio_fsync(&req->fsync, iocb, false);
1968 case IOCB_CMD_FDSYNC:
1969 return aio_fsync(&req->fsync, iocb, true);
1971 return aio_poll(req, iocb);
1973 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1978 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1981 struct aio_kiocb *req;
1985 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1988 /* enforce forwards compatibility on users */
1989 if (unlikely(iocb.aio_reserved2)) {
1990 pr_debug("EINVAL: reserve field set\n");
1994 /* prevent overflows */
1996 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1997 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1998 ((ssize_t)iocb.aio_nbytes < 0)
2000 pr_debug("EINVAL: overflow check\n");
2004 req = aio_get_req(ctx);
2008 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2010 /* Done with the synchronous reference */
2014 * If err is 0, we'd either done aio_complete() ourselves or have
2015 * arranged for that to be done asynchronously. Anything non-zero
2016 * means that we need to destroy req ourselves.
2018 if (unlikely(err)) {
2020 put_reqs_available(ctx, 1);
2026 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2027 * the number of iocbs queued. May return -EINVAL if the aio_context
2028 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2029 * *iocbpp[0] is not properly initialized, if the operation specified
2030 * is invalid for the file descriptor in the iocb. May fail with
2031 * -EFAULT if any of the data structures point to invalid data. May
2032 * fail with -EBADF if the file descriptor specified in the first
2033 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2034 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2035 * fail with -ENOSYS if not implemented.
2037 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2038 struct iocb __user * __user *, iocbpp)
2043 struct blk_plug plug;
2045 if (unlikely(nr < 0))
2048 ctx = lookup_ioctx(ctx_id);
2049 if (unlikely(!ctx)) {
2050 pr_debug("EINVAL: invalid context id\n");
2054 if (nr > ctx->nr_events)
2055 nr = ctx->nr_events;
2057 if (nr > AIO_PLUG_THRESHOLD)
2058 blk_start_plug(&plug);
2059 for (i = 0; i < nr; i++) {
2060 struct iocb __user *user_iocb;
2062 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2067 ret = io_submit_one(ctx, user_iocb, false);
2071 if (nr > AIO_PLUG_THRESHOLD)
2072 blk_finish_plug(&plug);
2074 percpu_ref_put(&ctx->users);
2078 #ifdef CONFIG_COMPAT
2079 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2080 int, nr, compat_uptr_t __user *, iocbpp)
2085 struct blk_plug plug;
2087 if (unlikely(nr < 0))
2090 ctx = lookup_ioctx(ctx_id);
2091 if (unlikely(!ctx)) {
2092 pr_debug("EINVAL: invalid context id\n");
2096 if (nr > ctx->nr_events)
2097 nr = ctx->nr_events;
2099 if (nr > AIO_PLUG_THRESHOLD)
2100 blk_start_plug(&plug);
2101 for (i = 0; i < nr; i++) {
2102 compat_uptr_t user_iocb;
2104 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2109 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2113 if (nr > AIO_PLUG_THRESHOLD)
2114 blk_finish_plug(&plug);
2116 percpu_ref_put(&ctx->users);
2122 * Attempts to cancel an iocb previously passed to io_submit. If
2123 * the operation is successfully cancelled, the resulting event is
2124 * copied into the memory pointed to by result without being placed
2125 * into the completion queue and 0 is returned. May fail with
2126 * -EFAULT if any of the data structures pointed to are invalid.
2127 * May fail with -EINVAL if aio_context specified by ctx_id is
2128 * invalid. May fail with -EAGAIN if the iocb specified was not
2129 * cancelled. Will fail with -ENOSYS if not implemented.
2131 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2132 struct io_event __user *, result)
2135 struct aio_kiocb *kiocb;
2138 u64 obj = (u64)(unsigned long)iocb;
2140 if (unlikely(get_user(key, &iocb->aio_key)))
2142 if (unlikely(key != KIOCB_KEY))
2145 ctx = lookup_ioctx(ctx_id);
2149 spin_lock_irq(&ctx->ctx_lock);
2150 /* TODO: use a hash or array, this sucks. */
2151 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2152 if (kiocb->ki_res.obj == obj) {
2153 ret = kiocb->ki_cancel(&kiocb->rw);
2154 list_del_init(&kiocb->ki_list);
2158 spin_unlock_irq(&ctx->ctx_lock);
2162 * The result argument is no longer used - the io_event is
2163 * always delivered via the ring buffer. -EINPROGRESS indicates
2164 * cancellation is progress:
2169 percpu_ref_put(&ctx->users);
2174 static long do_io_getevents(aio_context_t ctx_id,
2177 struct io_event __user *events,
2178 struct timespec64 *ts)
2180 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2181 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2184 if (likely(ioctx)) {
2185 if (likely(min_nr <= nr && min_nr >= 0))
2186 ret = read_events(ioctx, min_nr, nr, events, until);
2187 percpu_ref_put(&ioctx->users);
2194 * Attempts to read at least min_nr events and up to nr events from
2195 * the completion queue for the aio_context specified by ctx_id. If
2196 * it succeeds, the number of read events is returned. May fail with
2197 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2198 * out of range, if timeout is out of range. May fail with -EFAULT
2199 * if any of the memory specified is invalid. May return 0 or
2200 * < min_nr if the timeout specified by timeout has elapsed
2201 * before sufficient events are available, where timeout == NULL
2202 * specifies an infinite timeout. Note that the timeout pointed to by
2203 * timeout is relative. Will fail with -ENOSYS if not implemented.
2205 #if !defined(CONFIG_64BIT_TIME) || defined(CONFIG_64BIT)
2207 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2210 struct io_event __user *, events,
2211 struct __kernel_timespec __user *, timeout)
2213 struct timespec64 ts;
2216 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2219 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2220 if (!ret && signal_pending(current))
2227 struct __aio_sigset {
2228 const sigset_t __user *sigmask;
2232 SYSCALL_DEFINE6(io_pgetevents,
2233 aio_context_t, ctx_id,
2236 struct io_event __user *, events,
2237 struct __kernel_timespec __user *, timeout,
2238 const struct __aio_sigset __user *, usig)
2240 struct __aio_sigset ksig = { NULL, };
2241 struct timespec64 ts;
2245 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2248 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2251 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2255 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2257 interrupted = signal_pending(current);
2258 restore_saved_sigmask_unless(interrupted);
2259 if (interrupted && !ret)
2260 ret = -ERESTARTNOHAND;
2265 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2267 SYSCALL_DEFINE6(io_pgetevents_time32,
2268 aio_context_t, ctx_id,
2271 struct io_event __user *, events,
2272 struct old_timespec32 __user *, timeout,
2273 const struct __aio_sigset __user *, usig)
2275 struct __aio_sigset ksig = { NULL, };
2276 struct timespec64 ts;
2280 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2283 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2287 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2291 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2293 interrupted = signal_pending(current);
2294 restore_saved_sigmask_unless(interrupted);
2295 if (interrupted && !ret)
2296 ret = -ERESTARTNOHAND;
2303 #if defined(CONFIG_COMPAT_32BIT_TIME)
2305 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2308 struct io_event __user *, events,
2309 struct old_timespec32 __user *, timeout)
2311 struct timespec64 t;
2314 if (timeout && get_old_timespec32(&t, timeout))
2317 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2318 if (!ret && signal_pending(current))
2325 #ifdef CONFIG_COMPAT
2327 struct __compat_aio_sigset {
2328 compat_uptr_t sigmask;
2329 compat_size_t sigsetsize;
2332 #if defined(CONFIG_COMPAT_32BIT_TIME)
2334 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2335 compat_aio_context_t, ctx_id,
2336 compat_long_t, min_nr,
2338 struct io_event __user *, events,
2339 struct old_timespec32 __user *, timeout,
2340 const struct __compat_aio_sigset __user *, usig)
2342 struct __compat_aio_sigset ksig = { 0, };
2343 struct timespec64 t;
2347 if (timeout && get_old_timespec32(&t, timeout))
2350 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2353 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2357 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2359 interrupted = signal_pending(current);
2360 restore_saved_sigmask_unless(interrupted);
2361 if (interrupted && !ret)
2362 ret = -ERESTARTNOHAND;
2369 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2370 compat_aio_context_t, ctx_id,
2371 compat_long_t, min_nr,
2373 struct io_event __user *, events,
2374 struct __kernel_timespec __user *, timeout,
2375 const struct __compat_aio_sigset __user *, usig)
2377 struct __compat_aio_sigset ksig = { 0, };
2378 struct timespec64 t;
2382 if (timeout && get_timespec64(&t, timeout))
2385 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2388 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2392 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2394 interrupted = signal_pending(current);
2395 restore_saved_sigmask_unless(interrupted);
2396 if (interrupted && !ret)
2397 ret = -ERESTARTNOHAND;