2 * An async IO implementation for Linux
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/percpu.h>
31 #include <linux/slab.h>
32 #include <linux/timer.h>
33 #include <linux/aio.h>
34 #include <linux/highmem.h>
35 #include <linux/workqueue.h>
36 #include <linux/security.h>
37 #include <linux/eventfd.h>
38 #include <linux/blkdev.h>
39 #include <linux/compat.h>
40 #include <linux/migrate.h>
41 #include <linux/ramfs.h>
42 #include <linux/percpu-refcount.h>
43 #include <linux/mount.h>
44 #include <linux/pseudo_fs.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[];
71 }; /* 128 bytes + ring size */
74 * Plugging is meant to work with larger batches of IOs. If we don't
75 * have more than the below, then don't bother setting up a plug.
77 #define AIO_PLUG_THRESHOLD 2
79 #define AIO_RING_PAGES 8
84 struct kioctx __rcu *table[];
88 unsigned reqs_available;
92 struct completion comp;
97 struct percpu_ref users;
100 struct percpu_ref reqs;
102 unsigned long user_id;
104 struct __percpu kioctx_cpu *cpu;
107 * For percpu reqs_available, number of slots we move to/from global
112 * This is what userspace passed to io_setup(), it's not used for
113 * anything but counting against the global max_reqs quota.
115 * The real limit is nr_events - 1, which will be larger (see
120 /* Size of ringbuffer, in units of struct io_event */
123 unsigned long mmap_base;
124 unsigned long mmap_size;
126 struct page **ring_pages;
129 struct rcu_work free_rwork; /* see free_ioctx() */
132 * signals when all in-flight requests are done
134 struct ctx_rq_wait *rq_wait;
138 * This counts the number of available slots in the ringbuffer,
139 * so we avoid overflowing it: it's decremented (if positive)
140 * when allocating a kiocb and incremented when the resulting
141 * io_event is pulled off the ringbuffer.
143 * We batch accesses to it with a percpu version.
145 atomic_t reqs_available;
146 } ____cacheline_aligned_in_smp;
150 struct list_head active_reqs; /* used for cancellation */
151 } ____cacheline_aligned_in_smp;
154 struct mutex ring_lock;
155 wait_queue_head_t wait;
156 } ____cacheline_aligned_in_smp;
160 unsigned completed_events;
161 spinlock_t completion_lock;
162 } ____cacheline_aligned_in_smp;
164 struct page *internal_pages[AIO_RING_PAGES];
165 struct file *aio_ring_file;
171 * First field must be the file pointer in all the
172 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
176 struct work_struct work;
183 struct wait_queue_head *head;
187 struct wait_queue_entry wait;
188 struct work_struct work;
192 * NOTE! Each of the iocb union members has the file pointer
193 * as the first entry in their struct definition. So you can
194 * access the file pointer through any of the sub-structs,
195 * or directly as just 'ki_filp' in this struct.
199 struct file *ki_filp;
201 struct fsync_iocb fsync;
202 struct poll_iocb poll;
205 struct kioctx *ki_ctx;
206 kiocb_cancel_fn *ki_cancel;
208 struct io_event ki_res;
210 struct list_head ki_list; /* the aio core uses this
211 * for cancellation */
212 refcount_t ki_refcnt;
215 * If the aio_resfd field of the userspace iocb is not zero,
216 * this is the underlying eventfd context to deliver events to.
218 struct eventfd_ctx *ki_eventfd;
221 /*------ sysctl variables----*/
222 static DEFINE_SPINLOCK(aio_nr_lock);
223 unsigned long aio_nr; /* current system wide number of aio requests */
224 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct kmem_cache *kiocb_cachep;
228 static struct kmem_cache *kioctx_cachep;
230 static struct vfsmount *aio_mnt;
232 static const struct file_operations aio_ring_fops;
233 static const struct address_space_operations aio_ctx_aops;
235 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
238 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
240 return ERR_CAST(inode);
242 inode->i_mapping->a_ops = &aio_ctx_aops;
243 inode->i_mapping->private_data = ctx;
244 inode->i_size = PAGE_SIZE * nr_pages;
246 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
247 O_RDWR, &aio_ring_fops);
253 static int aio_init_fs_context(struct fs_context *fc)
255 if (!init_pseudo(fc, AIO_RING_MAGIC))
257 fc->s_iflags |= SB_I_NOEXEC;
262 * Creates the slab caches used by the aio routines, panic on
263 * failure as this is done early during the boot sequence.
265 static int __init aio_setup(void)
267 static struct file_system_type aio_fs = {
269 .init_fs_context = aio_init_fs_context,
270 .kill_sb = kill_anon_super,
272 aio_mnt = kern_mount(&aio_fs);
274 panic("Failed to create aio fs mount.");
276 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
277 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
280 __initcall(aio_setup);
282 static void put_aio_ring_file(struct kioctx *ctx)
284 struct file *aio_ring_file = ctx->aio_ring_file;
285 struct address_space *i_mapping;
288 truncate_setsize(file_inode(aio_ring_file), 0);
290 /* Prevent further access to the kioctx from migratepages */
291 i_mapping = aio_ring_file->f_mapping;
292 spin_lock(&i_mapping->private_lock);
293 i_mapping->private_data = NULL;
294 ctx->aio_ring_file = NULL;
295 spin_unlock(&i_mapping->private_lock);
301 static void aio_free_ring(struct kioctx *ctx)
305 /* Disconnect the kiotx from the ring file. This prevents future
306 * accesses to the kioctx from page migration.
308 put_aio_ring_file(ctx);
310 for (i = 0; i < ctx->nr_pages; i++) {
312 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
313 page_count(ctx->ring_pages[i]));
314 page = ctx->ring_pages[i];
317 ctx->ring_pages[i] = NULL;
321 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
322 kfree(ctx->ring_pages);
323 ctx->ring_pages = NULL;
327 static int aio_ring_mremap(struct vm_area_struct *vma, unsigned long flags)
329 struct file *file = vma->vm_file;
330 struct mm_struct *mm = vma->vm_mm;
331 struct kioctx_table *table;
332 int i, res = -EINVAL;
334 if (flags & MREMAP_DONTUNMAP)
337 spin_lock(&mm->ioctx_lock);
339 table = rcu_dereference(mm->ioctx_table);
340 for (i = 0; i < table->nr; i++) {
343 ctx = rcu_dereference(table->table[i]);
344 if (ctx && ctx->aio_ring_file == file) {
345 if (!atomic_read(&ctx->dead)) {
346 ctx->user_id = ctx->mmap_base = vma->vm_start;
354 spin_unlock(&mm->ioctx_lock);
358 static const struct vm_operations_struct aio_ring_vm_ops = {
359 .mremap = aio_ring_mremap,
360 #if IS_ENABLED(CONFIG_MMU)
361 .fault = filemap_fault,
362 .map_pages = filemap_map_pages,
363 .page_mkwrite = filemap_page_mkwrite,
367 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
369 vma->vm_flags |= VM_DONTEXPAND;
370 vma->vm_ops = &aio_ring_vm_ops;
374 static const struct file_operations aio_ring_fops = {
375 .mmap = aio_ring_mmap,
378 #if IS_ENABLED(CONFIG_MIGRATION)
379 static int aio_migratepage(struct address_space *mapping, struct page *new,
380 struct page *old, enum migrate_mode mode)
388 * We cannot support the _NO_COPY case here, because copy needs to
389 * happen under the ctx->completion_lock. That does not work with the
390 * migration workflow of MIGRATE_SYNC_NO_COPY.
392 if (mode == MIGRATE_SYNC_NO_COPY)
397 /* mapping->private_lock here protects against the kioctx teardown. */
398 spin_lock(&mapping->private_lock);
399 ctx = mapping->private_data;
405 /* The ring_lock mutex. The prevents aio_read_events() from writing
406 * to the ring's head, and prevents page migration from mucking in
407 * a partially initialized kiotx.
409 if (!mutex_trylock(&ctx->ring_lock)) {
415 if (idx < (pgoff_t)ctx->nr_pages) {
416 /* Make sure the old page hasn't already been changed */
417 if (ctx->ring_pages[idx] != old)
425 /* Writeback must be complete */
426 BUG_ON(PageWriteback(old));
429 rc = migrate_page_move_mapping(mapping, new, old, 1);
430 if (rc != MIGRATEPAGE_SUCCESS) {
435 /* Take completion_lock to prevent other writes to the ring buffer
436 * while the old page is copied to the new. This prevents new
437 * events from being lost.
439 spin_lock_irqsave(&ctx->completion_lock, flags);
440 migrate_page_copy(new, old);
441 BUG_ON(ctx->ring_pages[idx] != old);
442 ctx->ring_pages[idx] = new;
443 spin_unlock_irqrestore(&ctx->completion_lock, flags);
445 /* The old page is no longer accessible. */
449 mutex_unlock(&ctx->ring_lock);
451 spin_unlock(&mapping->private_lock);
456 static const struct address_space_operations aio_ctx_aops = {
457 .set_page_dirty = __set_page_dirty_no_writeback,
458 #if IS_ENABLED(CONFIG_MIGRATION)
459 .migratepage = aio_migratepage,
463 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
465 struct aio_ring *ring;
466 struct mm_struct *mm = current->mm;
467 unsigned long size, unused;
472 /* Compensate for the ring buffer's head/tail overlap entry */
473 nr_events += 2; /* 1 is required, 2 for good luck */
475 size = sizeof(struct aio_ring);
476 size += sizeof(struct io_event) * nr_events;
478 nr_pages = PFN_UP(size);
482 file = aio_private_file(ctx, nr_pages);
484 ctx->aio_ring_file = NULL;
488 ctx->aio_ring_file = file;
489 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
490 / sizeof(struct io_event);
492 ctx->ring_pages = ctx->internal_pages;
493 if (nr_pages > AIO_RING_PAGES) {
494 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
496 if (!ctx->ring_pages) {
497 put_aio_ring_file(ctx);
502 for (i = 0; i < nr_pages; i++) {
504 page = find_or_create_page(file->f_mapping,
505 i, GFP_HIGHUSER | __GFP_ZERO);
508 pr_debug("pid(%d) page[%d]->count=%d\n",
509 current->pid, i, page_count(page));
510 SetPageUptodate(page);
513 ctx->ring_pages[i] = page;
517 if (unlikely(i != nr_pages)) {
522 ctx->mmap_size = nr_pages * PAGE_SIZE;
523 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
525 if (mmap_write_lock_killable(mm)) {
531 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
532 PROT_READ | PROT_WRITE,
533 MAP_SHARED, 0, &unused, NULL);
534 mmap_write_unlock(mm);
535 if (IS_ERR((void *)ctx->mmap_base)) {
541 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
543 ctx->user_id = ctx->mmap_base;
544 ctx->nr_events = nr_events; /* trusted copy */
546 ring = kmap_atomic(ctx->ring_pages[0]);
547 ring->nr = nr_events; /* user copy */
549 ring->head = ring->tail = 0;
550 ring->magic = AIO_RING_MAGIC;
551 ring->compat_features = AIO_RING_COMPAT_FEATURES;
552 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
553 ring->header_length = sizeof(struct aio_ring);
555 flush_dcache_page(ctx->ring_pages[0]);
560 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
561 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
562 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
564 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
566 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
567 struct kioctx *ctx = req->ki_ctx;
570 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
573 spin_lock_irqsave(&ctx->ctx_lock, flags);
574 list_add_tail(&req->ki_list, &ctx->active_reqs);
575 req->ki_cancel = cancel;
576 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
578 EXPORT_SYMBOL(kiocb_set_cancel_fn);
581 * free_ioctx() should be RCU delayed to synchronize against the RCU
582 * protected lookup_ioctx() and also needs process context to call
583 * aio_free_ring(). Use rcu_work.
585 static void free_ioctx(struct work_struct *work)
587 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
589 pr_debug("freeing %p\n", ctx);
592 free_percpu(ctx->cpu);
593 percpu_ref_exit(&ctx->reqs);
594 percpu_ref_exit(&ctx->users);
595 kmem_cache_free(kioctx_cachep, ctx);
598 static void free_ioctx_reqs(struct percpu_ref *ref)
600 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
602 /* At this point we know that there are no any in-flight requests */
603 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
604 complete(&ctx->rq_wait->comp);
606 /* Synchronize against RCU protected table->table[] dereferences */
607 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
608 queue_rcu_work(system_wq, &ctx->free_rwork);
612 * When this function runs, the kioctx has been removed from the "hash table"
613 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
614 * now it's safe to cancel any that need to be.
616 static void free_ioctx_users(struct percpu_ref *ref)
618 struct kioctx *ctx = container_of(ref, struct kioctx, users);
619 struct aio_kiocb *req;
621 spin_lock_irq(&ctx->ctx_lock);
623 while (!list_empty(&ctx->active_reqs)) {
624 req = list_first_entry(&ctx->active_reqs,
625 struct aio_kiocb, ki_list);
626 req->ki_cancel(&req->rw);
627 list_del_init(&req->ki_list);
630 spin_unlock_irq(&ctx->ctx_lock);
632 percpu_ref_kill(&ctx->reqs);
633 percpu_ref_put(&ctx->reqs);
636 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
639 struct kioctx_table *table, *old;
640 struct aio_ring *ring;
642 spin_lock(&mm->ioctx_lock);
643 table = rcu_dereference_raw(mm->ioctx_table);
647 for (i = 0; i < table->nr; i++)
648 if (!rcu_access_pointer(table->table[i])) {
650 rcu_assign_pointer(table->table[i], ctx);
651 spin_unlock(&mm->ioctx_lock);
653 /* While kioctx setup is in progress,
654 * we are protected from page migration
655 * changes ring_pages by ->ring_lock.
657 ring = kmap_atomic(ctx->ring_pages[0]);
663 new_nr = (table ? table->nr : 1) * 4;
664 spin_unlock(&mm->ioctx_lock);
666 table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
673 spin_lock(&mm->ioctx_lock);
674 old = rcu_dereference_raw(mm->ioctx_table);
677 rcu_assign_pointer(mm->ioctx_table, table);
678 } else if (table->nr > old->nr) {
679 memcpy(table->table, old->table,
680 old->nr * sizeof(struct kioctx *));
682 rcu_assign_pointer(mm->ioctx_table, table);
691 static void aio_nr_sub(unsigned nr)
693 spin_lock(&aio_nr_lock);
694 if (WARN_ON(aio_nr - nr > aio_nr))
698 spin_unlock(&aio_nr_lock);
702 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
704 static struct kioctx *ioctx_alloc(unsigned nr_events)
706 struct mm_struct *mm = current->mm;
711 * Store the original nr_events -- what userspace passed to io_setup(),
712 * for counting against the global limit -- before it changes.
714 unsigned int max_reqs = nr_events;
717 * We keep track of the number of available ringbuffer slots, to prevent
718 * overflow (reqs_available), and we also use percpu counters for this.
720 * So since up to half the slots might be on other cpu's percpu counters
721 * and unavailable, double nr_events so userspace sees what they
722 * expected: additionally, we move req_batch slots to/from percpu
723 * counters at a time, so make sure that isn't 0:
725 nr_events = max(nr_events, num_possible_cpus() * 4);
728 /* Prevent overflows */
729 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
730 pr_debug("ENOMEM: nr_events too high\n");
731 return ERR_PTR(-EINVAL);
734 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
735 return ERR_PTR(-EAGAIN);
737 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
739 return ERR_PTR(-ENOMEM);
741 ctx->max_reqs = max_reqs;
743 spin_lock_init(&ctx->ctx_lock);
744 spin_lock_init(&ctx->completion_lock);
745 mutex_init(&ctx->ring_lock);
746 /* Protect against page migration throughout kiotx setup by keeping
747 * the ring_lock mutex held until setup is complete. */
748 mutex_lock(&ctx->ring_lock);
749 init_waitqueue_head(&ctx->wait);
751 INIT_LIST_HEAD(&ctx->active_reqs);
753 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
756 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
759 ctx->cpu = alloc_percpu(struct kioctx_cpu);
763 err = aio_setup_ring(ctx, nr_events);
767 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
768 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
769 if (ctx->req_batch < 1)
772 /* limit the number of system wide aios */
773 spin_lock(&aio_nr_lock);
774 if (aio_nr + ctx->max_reqs > aio_max_nr ||
775 aio_nr + ctx->max_reqs < aio_nr) {
776 spin_unlock(&aio_nr_lock);
780 aio_nr += ctx->max_reqs;
781 spin_unlock(&aio_nr_lock);
783 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
784 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
786 err = ioctx_add_table(ctx, mm);
790 /* Release the ring_lock mutex now that all setup is complete. */
791 mutex_unlock(&ctx->ring_lock);
793 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
794 ctx, ctx->user_id, mm, ctx->nr_events);
798 aio_nr_sub(ctx->max_reqs);
800 atomic_set(&ctx->dead, 1);
802 vm_munmap(ctx->mmap_base, ctx->mmap_size);
805 mutex_unlock(&ctx->ring_lock);
806 free_percpu(ctx->cpu);
807 percpu_ref_exit(&ctx->reqs);
808 percpu_ref_exit(&ctx->users);
809 kmem_cache_free(kioctx_cachep, ctx);
810 pr_debug("error allocating ioctx %d\n", err);
815 * Cancels all outstanding aio requests on an aio context. Used
816 * when the processes owning a context have all exited to encourage
817 * the rapid destruction of the kioctx.
819 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
820 struct ctx_rq_wait *wait)
822 struct kioctx_table *table;
824 spin_lock(&mm->ioctx_lock);
825 if (atomic_xchg(&ctx->dead, 1)) {
826 spin_unlock(&mm->ioctx_lock);
830 table = rcu_dereference_raw(mm->ioctx_table);
831 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
832 RCU_INIT_POINTER(table->table[ctx->id], NULL);
833 spin_unlock(&mm->ioctx_lock);
835 /* free_ioctx_reqs() will do the necessary RCU synchronization */
836 wake_up_all(&ctx->wait);
839 * It'd be more correct to do this in free_ioctx(), after all
840 * the outstanding kiocbs have finished - but by then io_destroy
841 * has already returned, so io_setup() could potentially return
842 * -EAGAIN with no ioctxs actually in use (as far as userspace
845 aio_nr_sub(ctx->max_reqs);
848 vm_munmap(ctx->mmap_base, ctx->mmap_size);
851 percpu_ref_kill(&ctx->users);
856 * exit_aio: called when the last user of mm goes away. At this point, there is
857 * no way for any new requests to be submited or any of the io_* syscalls to be
858 * called on the context.
860 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
863 void exit_aio(struct mm_struct *mm)
865 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
866 struct ctx_rq_wait wait;
872 atomic_set(&wait.count, table->nr);
873 init_completion(&wait.comp);
876 for (i = 0; i < table->nr; ++i) {
878 rcu_dereference_protected(table->table[i], true);
886 * We don't need to bother with munmap() here - exit_mmap(mm)
887 * is coming and it'll unmap everything. And we simply can't,
888 * this is not necessarily our ->mm.
889 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
890 * that it needs to unmap the area, just set it to 0.
893 kill_ioctx(mm, ctx, &wait);
896 if (!atomic_sub_and_test(skipped, &wait.count)) {
897 /* Wait until all IO for the context are done. */
898 wait_for_completion(&wait.comp);
901 RCU_INIT_POINTER(mm->ioctx_table, NULL);
905 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
907 struct kioctx_cpu *kcpu;
910 local_irq_save(flags);
911 kcpu = this_cpu_ptr(ctx->cpu);
912 kcpu->reqs_available += nr;
914 while (kcpu->reqs_available >= ctx->req_batch * 2) {
915 kcpu->reqs_available -= ctx->req_batch;
916 atomic_add(ctx->req_batch, &ctx->reqs_available);
919 local_irq_restore(flags);
922 static bool __get_reqs_available(struct kioctx *ctx)
924 struct kioctx_cpu *kcpu;
928 local_irq_save(flags);
929 kcpu = this_cpu_ptr(ctx->cpu);
930 if (!kcpu->reqs_available) {
931 int old, avail = atomic_read(&ctx->reqs_available);
934 if (avail < ctx->req_batch)
938 avail = atomic_cmpxchg(&ctx->reqs_available,
939 avail, avail - ctx->req_batch);
940 } while (avail != old);
942 kcpu->reqs_available += ctx->req_batch;
946 kcpu->reqs_available--;
948 local_irq_restore(flags);
952 /* refill_reqs_available
953 * Updates the reqs_available reference counts used for tracking the
954 * number of free slots in the completion ring. This can be called
955 * from aio_complete() (to optimistically update reqs_available) or
956 * from aio_get_req() (the we're out of events case). It must be
957 * called holding ctx->completion_lock.
959 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
962 unsigned events_in_ring, completed;
964 /* Clamp head since userland can write to it. */
965 head %= ctx->nr_events;
967 events_in_ring = tail - head;
969 events_in_ring = ctx->nr_events - (head - tail);
971 completed = ctx->completed_events;
972 if (events_in_ring < completed)
973 completed -= events_in_ring;
980 ctx->completed_events -= completed;
981 put_reqs_available(ctx, completed);
984 /* user_refill_reqs_available
985 * Called to refill reqs_available when aio_get_req() encounters an
986 * out of space in the completion ring.
988 static void user_refill_reqs_available(struct kioctx *ctx)
990 spin_lock_irq(&ctx->completion_lock);
991 if (ctx->completed_events) {
992 struct aio_ring *ring;
995 /* Access of ring->head may race with aio_read_events_ring()
996 * here, but that's okay since whether we read the old version
997 * or the new version, and either will be valid. The important
998 * part is that head cannot pass tail since we prevent
999 * aio_complete() from updating tail by holding
1000 * ctx->completion_lock. Even if head is invalid, the check
1001 * against ctx->completed_events below will make sure we do the
1004 ring = kmap_atomic(ctx->ring_pages[0]);
1006 kunmap_atomic(ring);
1008 refill_reqs_available(ctx, head, ctx->tail);
1011 spin_unlock_irq(&ctx->completion_lock);
1014 static bool get_reqs_available(struct kioctx *ctx)
1016 if (__get_reqs_available(ctx))
1018 user_refill_reqs_available(ctx);
1019 return __get_reqs_available(ctx);
1023 * Allocate a slot for an aio request.
1024 * Returns NULL if no requests are free.
1026 * The refcount is initialized to 2 - one for the async op completion,
1027 * one for the synchronous code that does this.
1029 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1031 struct aio_kiocb *req;
1033 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1037 if (unlikely(!get_reqs_available(ctx))) {
1038 kmem_cache_free(kiocb_cachep, req);
1042 percpu_ref_get(&ctx->reqs);
1044 INIT_LIST_HEAD(&req->ki_list);
1045 refcount_set(&req->ki_refcnt, 2);
1046 req->ki_eventfd = NULL;
1050 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1052 struct aio_ring __user *ring = (void __user *)ctx_id;
1053 struct mm_struct *mm = current->mm;
1054 struct kioctx *ctx, *ret = NULL;
1055 struct kioctx_table *table;
1058 if (get_user(id, &ring->id))
1062 table = rcu_dereference(mm->ioctx_table);
1064 if (!table || id >= table->nr)
1067 id = array_index_nospec(id, table->nr);
1068 ctx = rcu_dereference(table->table[id]);
1069 if (ctx && ctx->user_id == ctx_id) {
1070 if (percpu_ref_tryget_live(&ctx->users))
1078 static inline void iocb_destroy(struct aio_kiocb *iocb)
1080 if (iocb->ki_eventfd)
1081 eventfd_ctx_put(iocb->ki_eventfd);
1083 fput(iocb->ki_filp);
1084 percpu_ref_put(&iocb->ki_ctx->reqs);
1085 kmem_cache_free(kiocb_cachep, iocb);
1089 * Called when the io request on the given iocb is complete.
1091 static void aio_complete(struct aio_kiocb *iocb)
1093 struct kioctx *ctx = iocb->ki_ctx;
1094 struct aio_ring *ring;
1095 struct io_event *ev_page, *event;
1096 unsigned tail, pos, head;
1097 unsigned long flags;
1100 * Add a completion event to the ring buffer. Must be done holding
1101 * ctx->completion_lock to prevent other code from messing with the tail
1102 * pointer since we might be called from irq context.
1104 spin_lock_irqsave(&ctx->completion_lock, flags);
1107 pos = tail + AIO_EVENTS_OFFSET;
1109 if (++tail >= ctx->nr_events)
1112 ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1113 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1115 *event = iocb->ki_res;
1117 kunmap_atomic(ev_page);
1118 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1120 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1121 (void __user *)(unsigned long)iocb->ki_res.obj,
1122 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1124 /* after flagging the request as done, we
1125 * must never even look at it again
1127 smp_wmb(); /* make event visible before updating tail */
1131 ring = kmap_atomic(ctx->ring_pages[0]);
1134 kunmap_atomic(ring);
1135 flush_dcache_page(ctx->ring_pages[0]);
1137 ctx->completed_events++;
1138 if (ctx->completed_events > 1)
1139 refill_reqs_available(ctx, head, tail);
1140 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1142 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1145 * Check if the user asked us to deliver the result through an
1146 * eventfd. The eventfd_signal() function is safe to be called
1149 if (iocb->ki_eventfd)
1150 eventfd_signal(iocb->ki_eventfd, 1);
1153 * We have to order our ring_info tail store above and test
1154 * of the wait list below outside the wait lock. This is
1155 * like in wake_up_bit() where clearing a bit has to be
1156 * ordered with the unlocked test.
1160 if (waitqueue_active(&ctx->wait))
1161 wake_up(&ctx->wait);
1164 static inline void iocb_put(struct aio_kiocb *iocb)
1166 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1172 /* aio_read_events_ring
1173 * Pull an event off of the ioctx's event ring. Returns the number of
1176 static long aio_read_events_ring(struct kioctx *ctx,
1177 struct io_event __user *event, long nr)
1179 struct aio_ring *ring;
1180 unsigned head, tail, pos;
1185 * The mutex can block and wake us up and that will cause
1186 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1187 * and repeat. This should be rare enough that it doesn't cause
1188 * peformance issues. See the comment in read_events() for more detail.
1190 sched_annotate_sleep();
1191 mutex_lock(&ctx->ring_lock);
1193 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1194 ring = kmap_atomic(ctx->ring_pages[0]);
1197 kunmap_atomic(ring);
1200 * Ensure that once we've read the current tail pointer, that
1201 * we also see the events that were stored up to the tail.
1205 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1210 head %= ctx->nr_events;
1211 tail %= ctx->nr_events;
1215 struct io_event *ev;
1218 avail = (head <= tail ? tail : ctx->nr_events) - head;
1222 pos = head + AIO_EVENTS_OFFSET;
1223 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1224 pos %= AIO_EVENTS_PER_PAGE;
1226 avail = min(avail, nr - ret);
1227 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1230 copy_ret = copy_to_user(event + ret, ev + pos,
1231 sizeof(*ev) * avail);
1234 if (unlikely(copy_ret)) {
1241 head %= ctx->nr_events;
1244 ring = kmap_atomic(ctx->ring_pages[0]);
1246 kunmap_atomic(ring);
1247 flush_dcache_page(ctx->ring_pages[0]);
1249 pr_debug("%li h%u t%u\n", ret, head, tail);
1251 mutex_unlock(&ctx->ring_lock);
1256 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1257 struct io_event __user *event, long *i)
1259 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1264 if (unlikely(atomic_read(&ctx->dead)))
1270 return ret < 0 || *i >= min_nr;
1273 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1274 struct io_event __user *event,
1280 * Note that aio_read_events() is being called as the conditional - i.e.
1281 * we're calling it after prepare_to_wait() has set task state to
1282 * TASK_INTERRUPTIBLE.
1284 * But aio_read_events() can block, and if it blocks it's going to flip
1285 * the task state back to TASK_RUNNING.
1287 * This should be ok, provided it doesn't flip the state back to
1288 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1289 * will only happen if the mutex_lock() call blocks, and we then find
1290 * the ringbuffer empty. So in practice we should be ok, but it's
1291 * something to be aware of when touching this code.
1294 aio_read_events(ctx, min_nr, nr, event, &ret);
1296 wait_event_interruptible_hrtimeout(ctx->wait,
1297 aio_read_events(ctx, min_nr, nr, event, &ret),
1303 * Create an aio_context capable of receiving at least nr_events.
1304 * ctxp must not point to an aio_context that already exists, and
1305 * must be initialized to 0 prior to the call. On successful
1306 * creation of the aio_context, *ctxp is filled in with the resulting
1307 * handle. May fail with -EINVAL if *ctxp is not initialized,
1308 * if the specified nr_events exceeds internal limits. May fail
1309 * with -EAGAIN if the specified nr_events exceeds the user's limit
1310 * of available events. May fail with -ENOMEM if insufficient kernel
1311 * resources are available. May fail with -EFAULT if an invalid
1312 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1315 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1317 struct kioctx *ioctx = NULL;
1321 ret = get_user(ctx, ctxp);
1326 if (unlikely(ctx || nr_events == 0)) {
1327 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1332 ioctx = ioctx_alloc(nr_events);
1333 ret = PTR_ERR(ioctx);
1334 if (!IS_ERR(ioctx)) {
1335 ret = put_user(ioctx->user_id, ctxp);
1337 kill_ioctx(current->mm, ioctx, NULL);
1338 percpu_ref_put(&ioctx->users);
1345 #ifdef CONFIG_COMPAT
1346 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1348 struct kioctx *ioctx = NULL;
1352 ret = get_user(ctx, ctx32p);
1357 if (unlikely(ctx || nr_events == 0)) {
1358 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1363 ioctx = ioctx_alloc(nr_events);
1364 ret = PTR_ERR(ioctx);
1365 if (!IS_ERR(ioctx)) {
1366 /* truncating is ok because it's a user address */
1367 ret = put_user((u32)ioctx->user_id, ctx32p);
1369 kill_ioctx(current->mm, ioctx, NULL);
1370 percpu_ref_put(&ioctx->users);
1379 * Destroy the aio_context specified. May cancel any outstanding
1380 * AIOs and block on completion. Will fail with -ENOSYS if not
1381 * implemented. May fail with -EINVAL if the context pointed to
1384 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1386 struct kioctx *ioctx = lookup_ioctx(ctx);
1387 if (likely(NULL != ioctx)) {
1388 struct ctx_rq_wait wait;
1391 init_completion(&wait.comp);
1392 atomic_set(&wait.count, 1);
1394 /* Pass requests_done to kill_ioctx() where it can be set
1395 * in a thread-safe way. If we try to set it here then we have
1396 * a race condition if two io_destroy() called simultaneously.
1398 ret = kill_ioctx(current->mm, ioctx, &wait);
1399 percpu_ref_put(&ioctx->users);
1401 /* Wait until all IO for the context are done. Otherwise kernel
1402 * keep using user-space buffers even if user thinks the context
1406 wait_for_completion(&wait.comp);
1410 pr_debug("EINVAL: invalid context id\n");
1414 static void aio_remove_iocb(struct aio_kiocb *iocb)
1416 struct kioctx *ctx = iocb->ki_ctx;
1417 unsigned long flags;
1419 spin_lock_irqsave(&ctx->ctx_lock, flags);
1420 list_del(&iocb->ki_list);
1421 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1424 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1426 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1428 if (!list_empty_careful(&iocb->ki_list))
1429 aio_remove_iocb(iocb);
1431 if (kiocb->ki_flags & IOCB_WRITE) {
1432 struct inode *inode = file_inode(kiocb->ki_filp);
1435 * Tell lockdep we inherited freeze protection from submission
1438 if (S_ISREG(inode->i_mode))
1439 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1440 file_end_write(kiocb->ki_filp);
1443 iocb->ki_res.res = res;
1444 iocb->ki_res.res2 = res2;
1448 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1452 req->ki_complete = aio_complete_rw;
1453 req->private = NULL;
1454 req->ki_pos = iocb->aio_offset;
1455 req->ki_flags = iocb_flags(req->ki_filp);
1456 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1457 req->ki_flags |= IOCB_EVENTFD;
1458 req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1459 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1461 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1462 * aio_reqprio is interpreted as an I/O scheduling
1463 * class and priority.
1465 ret = ioprio_check_cap(iocb->aio_reqprio);
1467 pr_debug("aio ioprio check cap error: %d\n", ret);
1471 req->ki_ioprio = iocb->aio_reqprio;
1473 req->ki_ioprio = get_current_ioprio();
1475 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1479 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1483 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1484 struct iovec **iovec, bool vectored, bool compat,
1485 struct iov_iter *iter)
1487 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1488 size_t len = iocb->aio_nbytes;
1491 ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1496 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1499 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1505 case -ERESTARTNOINTR:
1506 case -ERESTARTNOHAND:
1507 case -ERESTART_RESTARTBLOCK:
1509 * There's no easy way to restart the syscall since other AIO's
1510 * may be already running. Just fail this IO with EINTR.
1515 req->ki_complete(req, ret, 0);
1519 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1520 bool vectored, bool compat)
1522 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1523 struct iov_iter iter;
1527 ret = aio_prep_rw(req, iocb);
1530 file = req->ki_filp;
1531 if (unlikely(!(file->f_mode & FMODE_READ)))
1534 if (unlikely(!file->f_op->read_iter))
1537 ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1540 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1542 aio_rw_done(req, call_read_iter(file, req, &iter));
1547 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1548 bool vectored, bool compat)
1550 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1551 struct iov_iter iter;
1555 ret = aio_prep_rw(req, iocb);
1558 file = req->ki_filp;
1560 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1562 if (unlikely(!file->f_op->write_iter))
1565 ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1568 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1571 * Open-code file_start_write here to grab freeze protection,
1572 * which will be released by another thread in
1573 * aio_complete_rw(). Fool lockdep by telling it the lock got
1574 * released so that it doesn't complain about the held lock when
1575 * we return to userspace.
1577 if (S_ISREG(file_inode(file)->i_mode)) {
1578 sb_start_write(file_inode(file)->i_sb);
1579 __sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1581 req->ki_flags |= IOCB_WRITE;
1582 aio_rw_done(req, call_write_iter(file, req, &iter));
1588 static void aio_fsync_work(struct work_struct *work)
1590 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1591 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1593 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1594 revert_creds(old_cred);
1595 put_cred(iocb->fsync.creds);
1599 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1602 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1603 iocb->aio_rw_flags))
1606 if (unlikely(!req->file->f_op->fsync))
1609 req->creds = prepare_creds();
1613 req->datasync = datasync;
1614 INIT_WORK(&req->work, aio_fsync_work);
1615 schedule_work(&req->work);
1619 static void aio_poll_put_work(struct work_struct *work)
1621 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1622 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1627 static void aio_poll_complete_work(struct work_struct *work)
1629 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1630 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1631 struct poll_table_struct pt = { ._key = req->events };
1632 struct kioctx *ctx = iocb->ki_ctx;
1635 if (!READ_ONCE(req->cancelled))
1636 mask = vfs_poll(req->file, &pt) & req->events;
1639 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1640 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1641 * synchronize with them. In the cancellation case the list_del_init
1642 * itself is not actually needed, but harmless so we keep it in to
1643 * avoid further branches in the fast path.
1645 spin_lock_irq(&ctx->ctx_lock);
1646 if (!mask && !READ_ONCE(req->cancelled)) {
1647 add_wait_queue(req->head, &req->wait);
1648 spin_unlock_irq(&ctx->ctx_lock);
1651 list_del_init(&iocb->ki_list);
1652 iocb->ki_res.res = mangle_poll(mask);
1654 spin_unlock_irq(&ctx->ctx_lock);
1659 /* assumes we are called with irqs disabled */
1660 static int aio_poll_cancel(struct kiocb *iocb)
1662 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1663 struct poll_iocb *req = &aiocb->poll;
1665 spin_lock(&req->head->lock);
1666 WRITE_ONCE(req->cancelled, true);
1667 if (!list_empty(&req->wait.entry)) {
1668 list_del_init(&req->wait.entry);
1669 schedule_work(&aiocb->poll.work);
1671 spin_unlock(&req->head->lock);
1676 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1679 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1680 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1681 __poll_t mask = key_to_poll(key);
1682 unsigned long flags;
1684 /* for instances that support it check for an event match first: */
1685 if (mask && !(mask & req->events))
1688 list_del_init(&req->wait.entry);
1690 if (mask && spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1691 struct kioctx *ctx = iocb->ki_ctx;
1694 * Try to complete the iocb inline if we can. Use
1695 * irqsave/irqrestore because not all filesystems (e.g. fuse)
1696 * call this function with IRQs disabled and because IRQs
1697 * have to be disabled before ctx_lock is obtained.
1699 list_del(&iocb->ki_list);
1700 iocb->ki_res.res = mangle_poll(mask);
1702 if (iocb->ki_eventfd && eventfd_signal_count()) {
1704 INIT_WORK(&req->work, aio_poll_put_work);
1705 schedule_work(&req->work);
1707 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1711 schedule_work(&req->work);
1716 struct aio_poll_table {
1717 struct poll_table_struct pt;
1718 struct aio_kiocb *iocb;
1723 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1724 struct poll_table_struct *p)
1726 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1728 /* multiple wait queues per file are not supported */
1729 if (unlikely(pt->iocb->poll.head)) {
1730 pt->error = -EINVAL;
1735 pt->iocb->poll.head = head;
1736 add_wait_queue(head, &pt->iocb->poll.wait);
1739 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1741 struct kioctx *ctx = aiocb->ki_ctx;
1742 struct poll_iocb *req = &aiocb->poll;
1743 struct aio_poll_table apt;
1744 bool cancel = false;
1747 /* reject any unknown events outside the normal event mask. */
1748 if ((u16)iocb->aio_buf != iocb->aio_buf)
1750 /* reject fields that are not defined for poll */
1751 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1754 INIT_WORK(&req->work, aio_poll_complete_work);
1755 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1759 req->cancelled = false;
1761 apt.pt._qproc = aio_poll_queue_proc;
1762 apt.pt._key = req->events;
1764 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1766 /* initialized the list so that we can do list_empty checks */
1767 INIT_LIST_HEAD(&req->wait.entry);
1768 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1770 mask = vfs_poll(req->file, &apt.pt) & req->events;
1771 spin_lock_irq(&ctx->ctx_lock);
1772 if (likely(req->head)) {
1773 spin_lock(&req->head->lock);
1774 if (unlikely(list_empty(&req->wait.entry))) {
1780 if (mask || apt.error) {
1781 list_del_init(&req->wait.entry);
1782 } else if (cancel) {
1783 WRITE_ONCE(req->cancelled, true);
1784 } else if (!req->done) { /* actually waiting for an event */
1785 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1786 aiocb->ki_cancel = aio_poll_cancel;
1788 spin_unlock(&req->head->lock);
1790 if (mask) { /* no async, we'd stolen it */
1791 aiocb->ki_res.res = mangle_poll(mask);
1794 spin_unlock_irq(&ctx->ctx_lock);
1800 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1801 struct iocb __user *user_iocb, struct aio_kiocb *req,
1804 req->ki_filp = fget(iocb->aio_fildes);
1805 if (unlikely(!req->ki_filp))
1808 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1809 struct eventfd_ctx *eventfd;
1811 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1812 * instance of the file* now. The file descriptor must be
1813 * an eventfd() fd, and will be signaled for each completed
1814 * event using the eventfd_signal() function.
1816 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1817 if (IS_ERR(eventfd))
1818 return PTR_ERR(eventfd);
1820 req->ki_eventfd = eventfd;
1823 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1824 pr_debug("EFAULT: aio_key\n");
1828 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1829 req->ki_res.data = iocb->aio_data;
1830 req->ki_res.res = 0;
1831 req->ki_res.res2 = 0;
1833 switch (iocb->aio_lio_opcode) {
1834 case IOCB_CMD_PREAD:
1835 return aio_read(&req->rw, iocb, false, compat);
1836 case IOCB_CMD_PWRITE:
1837 return aio_write(&req->rw, iocb, false, compat);
1838 case IOCB_CMD_PREADV:
1839 return aio_read(&req->rw, iocb, true, compat);
1840 case IOCB_CMD_PWRITEV:
1841 return aio_write(&req->rw, iocb, true, compat);
1842 case IOCB_CMD_FSYNC:
1843 return aio_fsync(&req->fsync, iocb, false);
1844 case IOCB_CMD_FDSYNC:
1845 return aio_fsync(&req->fsync, iocb, true);
1847 return aio_poll(req, iocb);
1849 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
1854 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1857 struct aio_kiocb *req;
1861 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1864 /* enforce forwards compatibility on users */
1865 if (unlikely(iocb.aio_reserved2)) {
1866 pr_debug("EINVAL: reserve field set\n");
1870 /* prevent overflows */
1872 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1873 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1874 ((ssize_t)iocb.aio_nbytes < 0)
1876 pr_debug("EINVAL: overflow check\n");
1880 req = aio_get_req(ctx);
1884 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
1886 /* Done with the synchronous reference */
1890 * If err is 0, we'd either done aio_complete() ourselves or have
1891 * arranged for that to be done asynchronously. Anything non-zero
1892 * means that we need to destroy req ourselves.
1894 if (unlikely(err)) {
1896 put_reqs_available(ctx, 1);
1902 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
1903 * the number of iocbs queued. May return -EINVAL if the aio_context
1904 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
1905 * *iocbpp[0] is not properly initialized, if the operation specified
1906 * is invalid for the file descriptor in the iocb. May fail with
1907 * -EFAULT if any of the data structures point to invalid data. May
1908 * fail with -EBADF if the file descriptor specified in the first
1909 * iocb is invalid. May fail with -EAGAIN if insufficient resources
1910 * are available to queue any iocbs. Will return 0 if nr is 0. Will
1911 * fail with -ENOSYS if not implemented.
1913 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1914 struct iocb __user * __user *, iocbpp)
1919 struct blk_plug plug;
1921 if (unlikely(nr < 0))
1924 ctx = lookup_ioctx(ctx_id);
1925 if (unlikely(!ctx)) {
1926 pr_debug("EINVAL: invalid context id\n");
1930 if (nr > ctx->nr_events)
1931 nr = ctx->nr_events;
1933 if (nr > AIO_PLUG_THRESHOLD)
1934 blk_start_plug(&plug);
1935 for (i = 0; i < nr; i++) {
1936 struct iocb __user *user_iocb;
1938 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1943 ret = io_submit_one(ctx, user_iocb, false);
1947 if (nr > AIO_PLUG_THRESHOLD)
1948 blk_finish_plug(&plug);
1950 percpu_ref_put(&ctx->users);
1954 #ifdef CONFIG_COMPAT
1955 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1956 int, nr, compat_uptr_t __user *, iocbpp)
1961 struct blk_plug plug;
1963 if (unlikely(nr < 0))
1966 ctx = lookup_ioctx(ctx_id);
1967 if (unlikely(!ctx)) {
1968 pr_debug("EINVAL: invalid context id\n");
1972 if (nr > ctx->nr_events)
1973 nr = ctx->nr_events;
1975 if (nr > AIO_PLUG_THRESHOLD)
1976 blk_start_plug(&plug);
1977 for (i = 0; i < nr; i++) {
1978 compat_uptr_t user_iocb;
1980 if (unlikely(get_user(user_iocb, iocbpp + i))) {
1985 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1989 if (nr > AIO_PLUG_THRESHOLD)
1990 blk_finish_plug(&plug);
1992 percpu_ref_put(&ctx->users);
1998 * Attempts to cancel an iocb previously passed to io_submit. If
1999 * the operation is successfully cancelled, the resulting event is
2000 * copied into the memory pointed to by result without being placed
2001 * into the completion queue and 0 is returned. May fail with
2002 * -EFAULT if any of the data structures pointed to are invalid.
2003 * May fail with -EINVAL if aio_context specified by ctx_id is
2004 * invalid. May fail with -EAGAIN if the iocb specified was not
2005 * cancelled. Will fail with -ENOSYS if not implemented.
2007 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2008 struct io_event __user *, result)
2011 struct aio_kiocb *kiocb;
2014 u64 obj = (u64)(unsigned long)iocb;
2016 if (unlikely(get_user(key, &iocb->aio_key)))
2018 if (unlikely(key != KIOCB_KEY))
2021 ctx = lookup_ioctx(ctx_id);
2025 spin_lock_irq(&ctx->ctx_lock);
2026 /* TODO: use a hash or array, this sucks. */
2027 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2028 if (kiocb->ki_res.obj == obj) {
2029 ret = kiocb->ki_cancel(&kiocb->rw);
2030 list_del_init(&kiocb->ki_list);
2034 spin_unlock_irq(&ctx->ctx_lock);
2038 * The result argument is no longer used - the io_event is
2039 * always delivered via the ring buffer. -EINPROGRESS indicates
2040 * cancellation is progress:
2045 percpu_ref_put(&ctx->users);
2050 static long do_io_getevents(aio_context_t ctx_id,
2053 struct io_event __user *events,
2054 struct timespec64 *ts)
2056 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2057 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2060 if (likely(ioctx)) {
2061 if (likely(min_nr <= nr && min_nr >= 0))
2062 ret = read_events(ioctx, min_nr, nr, events, until);
2063 percpu_ref_put(&ioctx->users);
2070 * Attempts to read at least min_nr events and up to nr events from
2071 * the completion queue for the aio_context specified by ctx_id. If
2072 * it succeeds, the number of read events is returned. May fail with
2073 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2074 * out of range, if timeout is out of range. May fail with -EFAULT
2075 * if any of the memory specified is invalid. May return 0 or
2076 * < min_nr if the timeout specified by timeout has elapsed
2077 * before sufficient events are available, where timeout == NULL
2078 * specifies an infinite timeout. Note that the timeout pointed to by
2079 * timeout is relative. Will fail with -ENOSYS if not implemented.
2083 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2086 struct io_event __user *, events,
2087 struct __kernel_timespec __user *, timeout)
2089 struct timespec64 ts;
2092 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2095 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2096 if (!ret && signal_pending(current))
2103 struct __aio_sigset {
2104 const sigset_t __user *sigmask;
2108 SYSCALL_DEFINE6(io_pgetevents,
2109 aio_context_t, ctx_id,
2112 struct io_event __user *, events,
2113 struct __kernel_timespec __user *, timeout,
2114 const struct __aio_sigset __user *, usig)
2116 struct __aio_sigset ksig = { NULL, };
2117 struct timespec64 ts;
2121 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2124 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2127 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2131 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2133 interrupted = signal_pending(current);
2134 restore_saved_sigmask_unless(interrupted);
2135 if (interrupted && !ret)
2136 ret = -ERESTARTNOHAND;
2141 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2143 SYSCALL_DEFINE6(io_pgetevents_time32,
2144 aio_context_t, ctx_id,
2147 struct io_event __user *, events,
2148 struct old_timespec32 __user *, timeout,
2149 const struct __aio_sigset __user *, usig)
2151 struct __aio_sigset ksig = { NULL, };
2152 struct timespec64 ts;
2156 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2159 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2163 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2167 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2169 interrupted = signal_pending(current);
2170 restore_saved_sigmask_unless(interrupted);
2171 if (interrupted && !ret)
2172 ret = -ERESTARTNOHAND;
2179 #if defined(CONFIG_COMPAT_32BIT_TIME)
2181 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2184 struct io_event __user *, events,
2185 struct old_timespec32 __user *, timeout)
2187 struct timespec64 t;
2190 if (timeout && get_old_timespec32(&t, timeout))
2193 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2194 if (!ret && signal_pending(current))
2201 #ifdef CONFIG_COMPAT
2203 struct __compat_aio_sigset {
2204 compat_uptr_t sigmask;
2205 compat_size_t sigsetsize;
2208 #if defined(CONFIG_COMPAT_32BIT_TIME)
2210 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2211 compat_aio_context_t, ctx_id,
2212 compat_long_t, min_nr,
2214 struct io_event __user *, events,
2215 struct old_timespec32 __user *, timeout,
2216 const struct __compat_aio_sigset __user *, usig)
2218 struct __compat_aio_sigset ksig = { 0, };
2219 struct timespec64 t;
2223 if (timeout && get_old_timespec32(&t, timeout))
2226 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2229 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2233 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2235 interrupted = signal_pending(current);
2236 restore_saved_sigmask_unless(interrupted);
2237 if (interrupted && !ret)
2238 ret = -ERESTARTNOHAND;
2245 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2246 compat_aio_context_t, ctx_id,
2247 compat_long_t, min_nr,
2249 struct io_event __user *, events,
2250 struct __kernel_timespec __user *, timeout,
2251 const struct __compat_aio_sigset __user *, usig)
2253 struct __compat_aio_sigset ksig = { 0, };
2254 struct timespec64 t;
2258 if (timeout && get_timespec64(&t, timeout))
2261 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2264 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2268 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2270 interrupted = signal_pending(current);
2271 restore_saved_sigmask_unless(interrupted);
2272 if (interrupted && !ret)
2273 ret = -ERESTARTNOHAND;