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/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 <linux/uaccess.h>
47 #include <linux/nospec.h>
53 #define AIO_RING_MAGIC 0xa10a10a1
54 #define AIO_RING_COMPAT_FEATURES 1
55 #define AIO_RING_INCOMPAT_FEATURES 0
57 unsigned id; /* kernel internal index number */
58 unsigned nr; /* number of io_events */
59 unsigned head; /* Written to by userland or under ring_lock
60 * mutex by aio_read_events_ring(). */
64 unsigned compat_features;
65 unsigned incompat_features;
66 unsigned header_length; /* size of aio_ring */
69 struct io_event io_events[];
70 }; /* 128 bytes + ring size */
73 * Plugging is meant to work with larger batches of IOs. If we don't
74 * have more than the below, then don't bother setting up a plug.
76 #define AIO_PLUG_THRESHOLD 2
78 #define AIO_RING_PAGES 8
83 struct kioctx __rcu *table[] __counted_by(nr);
87 unsigned reqs_available;
91 struct completion comp;
96 struct percpu_ref users;
99 struct percpu_ref reqs;
101 unsigned long user_id;
103 struct __percpu kioctx_cpu *cpu;
106 * For percpu reqs_available, number of slots we move to/from global
111 * This is what userspace passed to io_setup(), it's not used for
112 * anything but counting against the global max_reqs quota.
114 * The real limit is nr_events - 1, which will be larger (see
119 /* Size of ringbuffer, in units of struct io_event */
122 unsigned long mmap_base;
123 unsigned long mmap_size;
125 struct page **ring_pages;
128 struct rcu_work free_rwork; /* see free_ioctx() */
131 * signals when all in-flight requests are done
133 struct ctx_rq_wait *rq_wait;
137 * This counts the number of available slots in the ringbuffer,
138 * so we avoid overflowing it: it's decremented (if positive)
139 * when allocating a kiocb and incremented when the resulting
140 * io_event is pulled off the ringbuffer.
142 * We batch accesses to it with a percpu version.
144 atomic_t reqs_available;
145 } ____cacheline_aligned_in_smp;
149 struct list_head active_reqs; /* used for cancellation */
150 } ____cacheline_aligned_in_smp;
153 struct mutex ring_lock;
154 wait_queue_head_t wait;
155 } ____cacheline_aligned_in_smp;
159 unsigned completed_events;
160 spinlock_t completion_lock;
161 } ____cacheline_aligned_in_smp;
163 struct page *internal_pages[AIO_RING_PAGES];
164 struct file *aio_ring_file;
170 * First field must be the file pointer in all the
171 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
175 struct work_struct work;
182 struct wait_queue_head *head;
186 bool work_need_resched;
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 static unsigned long aio_nr; /* current system wide number of aio requests */
224 static unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
225 /*----end sysctl variables---*/
227 static struct ctl_table aio_sysctls[] = {
229 .procname = "aio-nr",
231 .maxlen = sizeof(aio_nr),
233 .proc_handler = proc_doulongvec_minmax,
236 .procname = "aio-max-nr",
238 .maxlen = sizeof(aio_max_nr),
240 .proc_handler = proc_doulongvec_minmax,
244 static void __init aio_sysctl_init(void)
246 register_sysctl_init("fs", aio_sysctls);
249 #define aio_sysctl_init() do { } while (0)
252 static struct kmem_cache *kiocb_cachep;
253 static struct kmem_cache *kioctx_cachep;
255 static struct vfsmount *aio_mnt;
257 static const struct file_operations aio_ring_fops;
258 static const struct address_space_operations aio_ctx_aops;
260 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
263 struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
265 return ERR_CAST(inode);
267 inode->i_mapping->a_ops = &aio_ctx_aops;
268 inode->i_mapping->i_private_data = ctx;
269 inode->i_size = PAGE_SIZE * nr_pages;
271 file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
272 O_RDWR, &aio_ring_fops);
278 static int aio_init_fs_context(struct fs_context *fc)
280 if (!init_pseudo(fc, AIO_RING_MAGIC))
282 fc->s_iflags |= SB_I_NOEXEC;
287 * Creates the slab caches used by the aio routines, panic on
288 * failure as this is done early during the boot sequence.
290 static int __init aio_setup(void)
292 static struct file_system_type aio_fs = {
294 .init_fs_context = aio_init_fs_context,
295 .kill_sb = kill_anon_super,
297 aio_mnt = kern_mount(&aio_fs);
299 panic("Failed to create aio fs mount.");
301 kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
302 kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
306 __initcall(aio_setup);
308 static void put_aio_ring_file(struct kioctx *ctx)
310 struct file *aio_ring_file = ctx->aio_ring_file;
311 struct address_space *i_mapping;
314 truncate_setsize(file_inode(aio_ring_file), 0);
316 /* Prevent further access to the kioctx from migratepages */
317 i_mapping = aio_ring_file->f_mapping;
318 spin_lock(&i_mapping->i_private_lock);
319 i_mapping->i_private_data = NULL;
320 ctx->aio_ring_file = NULL;
321 spin_unlock(&i_mapping->i_private_lock);
327 static void aio_free_ring(struct kioctx *ctx)
331 /* Disconnect the kiotx from the ring file. This prevents future
332 * accesses to the kioctx from page migration.
334 put_aio_ring_file(ctx);
336 for (i = 0; i < ctx->nr_pages; i++) {
338 pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
339 page_count(ctx->ring_pages[i]));
340 page = ctx->ring_pages[i];
343 ctx->ring_pages[i] = NULL;
347 if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
348 kfree(ctx->ring_pages);
349 ctx->ring_pages = NULL;
353 static int aio_ring_mremap(struct vm_area_struct *vma)
355 struct file *file = vma->vm_file;
356 struct mm_struct *mm = vma->vm_mm;
357 struct kioctx_table *table;
358 int i, res = -EINVAL;
360 spin_lock(&mm->ioctx_lock);
362 table = rcu_dereference(mm->ioctx_table);
366 for (i = 0; i < table->nr; i++) {
369 ctx = rcu_dereference(table->table[i]);
370 if (ctx && ctx->aio_ring_file == file) {
371 if (!atomic_read(&ctx->dead)) {
372 ctx->user_id = ctx->mmap_base = vma->vm_start;
381 spin_unlock(&mm->ioctx_lock);
385 static const struct vm_operations_struct aio_ring_vm_ops = {
386 .mremap = aio_ring_mremap,
387 #if IS_ENABLED(CONFIG_MMU)
388 .fault = filemap_fault,
389 .map_pages = filemap_map_pages,
390 .page_mkwrite = filemap_page_mkwrite,
394 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
396 vm_flags_set(vma, VM_DONTEXPAND);
397 vma->vm_ops = &aio_ring_vm_ops;
401 static const struct file_operations aio_ring_fops = {
402 .mmap = aio_ring_mmap,
405 #if IS_ENABLED(CONFIG_MIGRATION)
406 static int aio_migrate_folio(struct address_space *mapping, struct folio *dst,
407 struct folio *src, enum migrate_mode mode)
415 * We cannot support the _NO_COPY case here, because copy needs to
416 * happen under the ctx->completion_lock. That does not work with the
417 * migration workflow of MIGRATE_SYNC_NO_COPY.
419 if (mode == MIGRATE_SYNC_NO_COPY)
424 /* mapping->i_private_lock here protects against the kioctx teardown. */
425 spin_lock(&mapping->i_private_lock);
426 ctx = mapping->i_private_data;
432 /* The ring_lock mutex. The prevents aio_read_events() from writing
433 * to the ring's head, and prevents page migration from mucking in
434 * a partially initialized kiotx.
436 if (!mutex_trylock(&ctx->ring_lock)) {
442 if (idx < (pgoff_t)ctx->nr_pages) {
443 /* Make sure the old folio hasn't already been changed */
444 if (ctx->ring_pages[idx] != &src->page)
452 /* Writeback must be complete */
453 BUG_ON(folio_test_writeback(src));
456 rc = folio_migrate_mapping(mapping, dst, src, 1);
457 if (rc != MIGRATEPAGE_SUCCESS) {
462 /* Take completion_lock to prevent other writes to the ring buffer
463 * while the old folio is copied to the new. This prevents new
464 * events from being lost.
466 spin_lock_irqsave(&ctx->completion_lock, flags);
467 folio_migrate_copy(dst, src);
468 BUG_ON(ctx->ring_pages[idx] != &src->page);
469 ctx->ring_pages[idx] = &dst->page;
470 spin_unlock_irqrestore(&ctx->completion_lock, flags);
472 /* The old folio is no longer accessible. */
476 mutex_unlock(&ctx->ring_lock);
478 spin_unlock(&mapping->i_private_lock);
482 #define aio_migrate_folio NULL
485 static const struct address_space_operations aio_ctx_aops = {
486 .dirty_folio = noop_dirty_folio,
487 .migrate_folio = aio_migrate_folio,
490 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
492 struct aio_ring *ring;
493 struct mm_struct *mm = current->mm;
494 unsigned long size, unused;
499 /* Compensate for the ring buffer's head/tail overlap entry */
500 nr_events += 2; /* 1 is required, 2 for good luck */
502 size = sizeof(struct aio_ring);
503 size += sizeof(struct io_event) * nr_events;
505 nr_pages = PFN_UP(size);
509 file = aio_private_file(ctx, nr_pages);
511 ctx->aio_ring_file = NULL;
515 ctx->aio_ring_file = file;
516 nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
517 / sizeof(struct io_event);
519 ctx->ring_pages = ctx->internal_pages;
520 if (nr_pages > AIO_RING_PAGES) {
521 ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
523 if (!ctx->ring_pages) {
524 put_aio_ring_file(ctx);
529 for (i = 0; i < nr_pages; i++) {
531 page = find_or_create_page(file->f_mapping,
532 i, GFP_USER | __GFP_ZERO);
535 pr_debug("pid(%d) page[%d]->count=%d\n",
536 current->pid, i, page_count(page));
537 SetPageUptodate(page);
540 ctx->ring_pages[i] = page;
544 if (unlikely(i != nr_pages)) {
549 ctx->mmap_size = nr_pages * PAGE_SIZE;
550 pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
552 if (mmap_write_lock_killable(mm)) {
558 ctx->mmap_base = do_mmap(ctx->aio_ring_file, 0, ctx->mmap_size,
559 PROT_READ | PROT_WRITE,
560 MAP_SHARED, 0, 0, &unused, NULL);
561 mmap_write_unlock(mm);
562 if (IS_ERR((void *)ctx->mmap_base)) {
568 pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
570 ctx->user_id = ctx->mmap_base;
571 ctx->nr_events = nr_events; /* trusted copy */
573 ring = page_address(ctx->ring_pages[0]);
574 ring->nr = nr_events; /* user copy */
576 ring->head = ring->tail = 0;
577 ring->magic = AIO_RING_MAGIC;
578 ring->compat_features = AIO_RING_COMPAT_FEATURES;
579 ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
580 ring->header_length = sizeof(struct aio_ring);
581 flush_dcache_page(ctx->ring_pages[0]);
586 #define AIO_EVENTS_PER_PAGE (PAGE_SIZE / sizeof(struct io_event))
587 #define AIO_EVENTS_FIRST_PAGE ((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
588 #define AIO_EVENTS_OFFSET (AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
590 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
592 struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
593 struct kioctx *ctx = req->ki_ctx;
597 * kiocb didn't come from aio or is neither a read nor a write, hence
600 if (!(iocb->ki_flags & IOCB_AIO_RW))
603 if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
606 spin_lock_irqsave(&ctx->ctx_lock, flags);
607 list_add_tail(&req->ki_list, &ctx->active_reqs);
608 req->ki_cancel = cancel;
609 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
611 EXPORT_SYMBOL(kiocb_set_cancel_fn);
614 * free_ioctx() should be RCU delayed to synchronize against the RCU
615 * protected lookup_ioctx() and also needs process context to call
616 * aio_free_ring(). Use rcu_work.
618 static void free_ioctx(struct work_struct *work)
620 struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
622 pr_debug("freeing %p\n", ctx);
625 free_percpu(ctx->cpu);
626 percpu_ref_exit(&ctx->reqs);
627 percpu_ref_exit(&ctx->users);
628 kmem_cache_free(kioctx_cachep, ctx);
631 static void free_ioctx_reqs(struct percpu_ref *ref)
633 struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
635 /* At this point we know that there are no any in-flight requests */
636 if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
637 complete(&ctx->rq_wait->comp);
639 /* Synchronize against RCU protected table->table[] dereferences */
640 INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
641 queue_rcu_work(system_wq, &ctx->free_rwork);
645 * When this function runs, the kioctx has been removed from the "hash table"
646 * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
647 * now it's safe to cancel any that need to be.
649 static void free_ioctx_users(struct percpu_ref *ref)
651 struct kioctx *ctx = container_of(ref, struct kioctx, users);
652 struct aio_kiocb *req;
654 spin_lock_irq(&ctx->ctx_lock);
656 while (!list_empty(&ctx->active_reqs)) {
657 req = list_first_entry(&ctx->active_reqs,
658 struct aio_kiocb, ki_list);
659 req->ki_cancel(&req->rw);
660 list_del_init(&req->ki_list);
663 spin_unlock_irq(&ctx->ctx_lock);
665 percpu_ref_kill(&ctx->reqs);
666 percpu_ref_put(&ctx->reqs);
669 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
672 struct kioctx_table *table, *old;
673 struct aio_ring *ring;
675 spin_lock(&mm->ioctx_lock);
676 table = rcu_dereference_raw(mm->ioctx_table);
680 for (i = 0; i < table->nr; i++)
681 if (!rcu_access_pointer(table->table[i])) {
683 rcu_assign_pointer(table->table[i], ctx);
684 spin_unlock(&mm->ioctx_lock);
686 /* While kioctx setup is in progress,
687 * we are protected from page migration
688 * changes ring_pages by ->ring_lock.
690 ring = page_address(ctx->ring_pages[0]);
695 new_nr = (table ? table->nr : 1) * 4;
696 spin_unlock(&mm->ioctx_lock);
698 table = kzalloc(struct_size(table, table, new_nr), GFP_KERNEL);
704 spin_lock(&mm->ioctx_lock);
705 old = rcu_dereference_raw(mm->ioctx_table);
708 rcu_assign_pointer(mm->ioctx_table, table);
709 } else if (table->nr > old->nr) {
710 memcpy(table->table, old->table,
711 old->nr * sizeof(struct kioctx *));
713 rcu_assign_pointer(mm->ioctx_table, table);
722 static void aio_nr_sub(unsigned nr)
724 spin_lock(&aio_nr_lock);
725 if (WARN_ON(aio_nr - nr > aio_nr))
729 spin_unlock(&aio_nr_lock);
733 * Allocates and initializes an ioctx. Returns an ERR_PTR if it failed.
735 static struct kioctx *ioctx_alloc(unsigned nr_events)
737 struct mm_struct *mm = current->mm;
742 * Store the original nr_events -- what userspace passed to io_setup(),
743 * for counting against the global limit -- before it changes.
745 unsigned int max_reqs = nr_events;
748 * We keep track of the number of available ringbuffer slots, to prevent
749 * overflow (reqs_available), and we also use percpu counters for this.
751 * So since up to half the slots might be on other cpu's percpu counters
752 * and unavailable, double nr_events so userspace sees what they
753 * expected: additionally, we move req_batch slots to/from percpu
754 * counters at a time, so make sure that isn't 0:
756 nr_events = max(nr_events, num_possible_cpus() * 4);
759 /* Prevent overflows */
760 if (nr_events > (0x10000000U / sizeof(struct io_event))) {
761 pr_debug("ENOMEM: nr_events too high\n");
762 return ERR_PTR(-EINVAL);
765 if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
766 return ERR_PTR(-EAGAIN);
768 ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
770 return ERR_PTR(-ENOMEM);
772 ctx->max_reqs = max_reqs;
774 spin_lock_init(&ctx->ctx_lock);
775 spin_lock_init(&ctx->completion_lock);
776 mutex_init(&ctx->ring_lock);
777 /* Protect against page migration throughout kiotx setup by keeping
778 * the ring_lock mutex held until setup is complete. */
779 mutex_lock(&ctx->ring_lock);
780 init_waitqueue_head(&ctx->wait);
782 INIT_LIST_HEAD(&ctx->active_reqs);
784 if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
787 if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
790 ctx->cpu = alloc_percpu(struct kioctx_cpu);
794 err = aio_setup_ring(ctx, nr_events);
798 atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
799 ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
800 if (ctx->req_batch < 1)
803 /* limit the number of system wide aios */
804 spin_lock(&aio_nr_lock);
805 if (aio_nr + ctx->max_reqs > aio_max_nr ||
806 aio_nr + ctx->max_reqs < aio_nr) {
807 spin_unlock(&aio_nr_lock);
811 aio_nr += ctx->max_reqs;
812 spin_unlock(&aio_nr_lock);
814 percpu_ref_get(&ctx->users); /* io_setup() will drop this ref */
815 percpu_ref_get(&ctx->reqs); /* free_ioctx_users() will drop this */
817 err = ioctx_add_table(ctx, mm);
821 /* Release the ring_lock mutex now that all setup is complete. */
822 mutex_unlock(&ctx->ring_lock);
824 pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
825 ctx, ctx->user_id, mm, ctx->nr_events);
829 aio_nr_sub(ctx->max_reqs);
831 atomic_set(&ctx->dead, 1);
833 vm_munmap(ctx->mmap_base, ctx->mmap_size);
836 mutex_unlock(&ctx->ring_lock);
837 free_percpu(ctx->cpu);
838 percpu_ref_exit(&ctx->reqs);
839 percpu_ref_exit(&ctx->users);
840 kmem_cache_free(kioctx_cachep, ctx);
841 pr_debug("error allocating ioctx %d\n", err);
846 * Cancels all outstanding aio requests on an aio context. Used
847 * when the processes owning a context have all exited to encourage
848 * the rapid destruction of the kioctx.
850 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
851 struct ctx_rq_wait *wait)
853 struct kioctx_table *table;
855 spin_lock(&mm->ioctx_lock);
856 if (atomic_xchg(&ctx->dead, 1)) {
857 spin_unlock(&mm->ioctx_lock);
861 table = rcu_dereference_raw(mm->ioctx_table);
862 WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
863 RCU_INIT_POINTER(table->table[ctx->id], NULL);
864 spin_unlock(&mm->ioctx_lock);
866 /* free_ioctx_reqs() will do the necessary RCU synchronization */
867 wake_up_all(&ctx->wait);
870 * It'd be more correct to do this in free_ioctx(), after all
871 * the outstanding kiocbs have finished - but by then io_destroy
872 * has already returned, so io_setup() could potentially return
873 * -EAGAIN with no ioctxs actually in use (as far as userspace
876 aio_nr_sub(ctx->max_reqs);
879 vm_munmap(ctx->mmap_base, ctx->mmap_size);
882 percpu_ref_kill(&ctx->users);
887 * exit_aio: called when the last user of mm goes away. At this point, there is
888 * no way for any new requests to be submited or any of the io_* syscalls to be
889 * called on the context.
891 * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
894 void exit_aio(struct mm_struct *mm)
896 struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
897 struct ctx_rq_wait wait;
903 atomic_set(&wait.count, table->nr);
904 init_completion(&wait.comp);
907 for (i = 0; i < table->nr; ++i) {
909 rcu_dereference_protected(table->table[i], true);
917 * We don't need to bother with munmap() here - exit_mmap(mm)
918 * is coming and it'll unmap everything. And we simply can't,
919 * this is not necessarily our ->mm.
920 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
921 * that it needs to unmap the area, just set it to 0.
924 kill_ioctx(mm, ctx, &wait);
927 if (!atomic_sub_and_test(skipped, &wait.count)) {
928 /* Wait until all IO for the context are done. */
929 wait_for_completion(&wait.comp);
932 RCU_INIT_POINTER(mm->ioctx_table, NULL);
936 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
938 struct kioctx_cpu *kcpu;
941 local_irq_save(flags);
942 kcpu = this_cpu_ptr(ctx->cpu);
943 kcpu->reqs_available += nr;
945 while (kcpu->reqs_available >= ctx->req_batch * 2) {
946 kcpu->reqs_available -= ctx->req_batch;
947 atomic_add(ctx->req_batch, &ctx->reqs_available);
950 local_irq_restore(flags);
953 static bool __get_reqs_available(struct kioctx *ctx)
955 struct kioctx_cpu *kcpu;
959 local_irq_save(flags);
960 kcpu = this_cpu_ptr(ctx->cpu);
961 if (!kcpu->reqs_available) {
962 int avail = atomic_read(&ctx->reqs_available);
965 if (avail < ctx->req_batch)
967 } while (!atomic_try_cmpxchg(&ctx->reqs_available,
968 &avail, avail - ctx->req_batch));
970 kcpu->reqs_available += ctx->req_batch;
974 kcpu->reqs_available--;
976 local_irq_restore(flags);
980 /* refill_reqs_available
981 * Updates the reqs_available reference counts used for tracking the
982 * number of free slots in the completion ring. This can be called
983 * from aio_complete() (to optimistically update reqs_available) or
984 * from aio_get_req() (the we're out of events case). It must be
985 * called holding ctx->completion_lock.
987 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
990 unsigned events_in_ring, completed;
992 /* Clamp head since userland can write to it. */
993 head %= ctx->nr_events;
995 events_in_ring = tail - head;
997 events_in_ring = ctx->nr_events - (head - tail);
999 completed = ctx->completed_events;
1000 if (events_in_ring < completed)
1001 completed -= events_in_ring;
1008 ctx->completed_events -= completed;
1009 put_reqs_available(ctx, completed);
1012 /* user_refill_reqs_available
1013 * Called to refill reqs_available when aio_get_req() encounters an
1014 * out of space in the completion ring.
1016 static void user_refill_reqs_available(struct kioctx *ctx)
1018 spin_lock_irq(&ctx->completion_lock);
1019 if (ctx->completed_events) {
1020 struct aio_ring *ring;
1023 /* Access of ring->head may race with aio_read_events_ring()
1024 * here, but that's okay since whether we read the old version
1025 * or the new version, and either will be valid. The important
1026 * part is that head cannot pass tail since we prevent
1027 * aio_complete() from updating tail by holding
1028 * ctx->completion_lock. Even if head is invalid, the check
1029 * against ctx->completed_events below will make sure we do the
1032 ring = page_address(ctx->ring_pages[0]);
1035 refill_reqs_available(ctx, head, ctx->tail);
1038 spin_unlock_irq(&ctx->completion_lock);
1041 static bool get_reqs_available(struct kioctx *ctx)
1043 if (__get_reqs_available(ctx))
1045 user_refill_reqs_available(ctx);
1046 return __get_reqs_available(ctx);
1050 * Allocate a slot for an aio request.
1051 * Returns NULL if no requests are free.
1053 * The refcount is initialized to 2 - one for the async op completion,
1054 * one for the synchronous code that does this.
1056 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1058 struct aio_kiocb *req;
1060 req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
1064 if (unlikely(!get_reqs_available(ctx))) {
1065 kmem_cache_free(kiocb_cachep, req);
1069 percpu_ref_get(&ctx->reqs);
1071 INIT_LIST_HEAD(&req->ki_list);
1072 refcount_set(&req->ki_refcnt, 2);
1073 req->ki_eventfd = NULL;
1077 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1079 struct aio_ring __user *ring = (void __user *)ctx_id;
1080 struct mm_struct *mm = current->mm;
1081 struct kioctx *ctx, *ret = NULL;
1082 struct kioctx_table *table;
1085 if (get_user(id, &ring->id))
1089 table = rcu_dereference(mm->ioctx_table);
1091 if (!table || id >= table->nr)
1094 id = array_index_nospec(id, table->nr);
1095 ctx = rcu_dereference(table->table[id]);
1096 if (ctx && ctx->user_id == ctx_id) {
1097 if (percpu_ref_tryget_live(&ctx->users))
1105 static inline void iocb_destroy(struct aio_kiocb *iocb)
1107 if (iocb->ki_eventfd)
1108 eventfd_ctx_put(iocb->ki_eventfd);
1110 fput(iocb->ki_filp);
1111 percpu_ref_put(&iocb->ki_ctx->reqs);
1112 kmem_cache_free(kiocb_cachep, iocb);
1116 struct wait_queue_entry w;
1121 * Called when the io request on the given iocb is complete.
1123 static void aio_complete(struct aio_kiocb *iocb)
1125 struct kioctx *ctx = iocb->ki_ctx;
1126 struct aio_ring *ring;
1127 struct io_event *ev_page, *event;
1128 unsigned tail, pos, head, avail;
1129 unsigned long flags;
1132 * Add a completion event to the ring buffer. Must be done holding
1133 * ctx->completion_lock to prevent other code from messing with the tail
1134 * pointer since we might be called from irq context.
1136 spin_lock_irqsave(&ctx->completion_lock, flags);
1139 pos = tail + AIO_EVENTS_OFFSET;
1141 if (++tail >= ctx->nr_events)
1144 ev_page = page_address(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1145 event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1147 *event = iocb->ki_res;
1149 flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1151 pr_debug("%p[%u]: %p: %p %Lx %Lx %Lx\n", ctx, tail, iocb,
1152 (void __user *)(unsigned long)iocb->ki_res.obj,
1153 iocb->ki_res.data, iocb->ki_res.res, iocb->ki_res.res2);
1155 /* after flagging the request as done, we
1156 * must never even look at it again
1158 smp_wmb(); /* make event visible before updating tail */
1162 ring = page_address(ctx->ring_pages[0]);
1165 flush_dcache_page(ctx->ring_pages[0]);
1167 ctx->completed_events++;
1168 if (ctx->completed_events > 1)
1169 refill_reqs_available(ctx, head, tail);
1173 : tail + ctx->nr_events - head;
1174 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1176 pr_debug("added to ring %p at [%u]\n", iocb, tail);
1179 * Check if the user asked us to deliver the result through an
1180 * eventfd. The eventfd_signal() function is safe to be called
1183 if (iocb->ki_eventfd)
1184 eventfd_signal(iocb->ki_eventfd);
1187 * We have to order our ring_info tail store above and test
1188 * of the wait list below outside the wait lock. This is
1189 * like in wake_up_bit() where clearing a bit has to be
1190 * ordered with the unlocked test.
1194 if (waitqueue_active(&ctx->wait)) {
1195 struct aio_waiter *curr, *next;
1196 unsigned long flags;
1198 spin_lock_irqsave(&ctx->wait.lock, flags);
1199 list_for_each_entry_safe(curr, next, &ctx->wait.head, w.entry)
1200 if (avail >= curr->min_nr) {
1201 list_del_init_careful(&curr->w.entry);
1202 wake_up_process(curr->w.private);
1204 spin_unlock_irqrestore(&ctx->wait.lock, flags);
1208 static inline void iocb_put(struct aio_kiocb *iocb)
1210 if (refcount_dec_and_test(&iocb->ki_refcnt)) {
1216 /* aio_read_events_ring
1217 * Pull an event off of the ioctx's event ring. Returns the number of
1220 static long aio_read_events_ring(struct kioctx *ctx,
1221 struct io_event __user *event, long nr)
1223 struct aio_ring *ring;
1224 unsigned head, tail, pos;
1229 * The mutex can block and wake us up and that will cause
1230 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1231 * and repeat. This should be rare enough that it doesn't cause
1232 * peformance issues. See the comment in read_events() for more detail.
1234 sched_annotate_sleep();
1235 mutex_lock(&ctx->ring_lock);
1237 /* Access to ->ring_pages here is protected by ctx->ring_lock. */
1238 ring = page_address(ctx->ring_pages[0]);
1243 * Ensure that once we've read the current tail pointer, that
1244 * we also see the events that were stored up to the tail.
1248 pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1253 head %= ctx->nr_events;
1254 tail %= ctx->nr_events;
1258 struct io_event *ev;
1261 avail = (head <= tail ? tail : ctx->nr_events) - head;
1265 pos = head + AIO_EVENTS_OFFSET;
1266 page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1267 pos %= AIO_EVENTS_PER_PAGE;
1269 avail = min(avail, nr - ret);
1270 avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1272 ev = page_address(page);
1273 copy_ret = copy_to_user(event + ret, ev + pos,
1274 sizeof(*ev) * avail);
1276 if (unlikely(copy_ret)) {
1283 head %= ctx->nr_events;
1286 ring = page_address(ctx->ring_pages[0]);
1288 flush_dcache_page(ctx->ring_pages[0]);
1290 pr_debug("%li h%u t%u\n", ret, head, tail);
1292 mutex_unlock(&ctx->ring_lock);
1297 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1298 struct io_event __user *event, long *i)
1300 long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1305 if (unlikely(atomic_read(&ctx->dead)))
1311 return ret < 0 || *i >= min_nr;
1314 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1315 struct io_event __user *event,
1318 struct hrtimer_sleeper t;
1319 struct aio_waiter w;
1320 long ret = 0, ret2 = 0;
1323 * Note that aio_read_events() is being called as the conditional - i.e.
1324 * we're calling it after prepare_to_wait() has set task state to
1325 * TASK_INTERRUPTIBLE.
1327 * But aio_read_events() can block, and if it blocks it's going to flip
1328 * the task state back to TASK_RUNNING.
1330 * This should be ok, provided it doesn't flip the state back to
1331 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1332 * will only happen if the mutex_lock() call blocks, and we then find
1333 * the ringbuffer empty. So in practice we should be ok, but it's
1334 * something to be aware of when touching this code.
1336 aio_read_events(ctx, min_nr, nr, event, &ret);
1337 if (until == 0 || ret < 0 || ret >= min_nr)
1340 hrtimer_init_sleeper_on_stack(&t, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1341 if (until != KTIME_MAX) {
1342 hrtimer_set_expires_range_ns(&t.timer, until, current->timer_slack_ns);
1343 hrtimer_sleeper_start_expires(&t, HRTIMER_MODE_REL);
1349 unsigned long nr_got = ret;
1351 w.min_nr = min_nr - ret;
1353 ret2 = prepare_to_wait_event(&ctx->wait, &w.w, TASK_INTERRUPTIBLE);
1354 if (!ret2 && !t.task)
1357 if (aio_read_events(ctx, min_nr, nr, event, &ret) || ret2)
1364 finish_wait(&ctx->wait, &w.w);
1365 hrtimer_cancel(&t.timer);
1366 destroy_hrtimer_on_stack(&t.timer);
1372 * Create an aio_context capable of receiving at least nr_events.
1373 * ctxp must not point to an aio_context that already exists, and
1374 * must be initialized to 0 prior to the call. On successful
1375 * creation of the aio_context, *ctxp is filled in with the resulting
1376 * handle. May fail with -EINVAL if *ctxp is not initialized,
1377 * if the specified nr_events exceeds internal limits. May fail
1378 * with -EAGAIN if the specified nr_events exceeds the user's limit
1379 * of available events. May fail with -ENOMEM if insufficient kernel
1380 * resources are available. May fail with -EFAULT if an invalid
1381 * pointer is passed for ctxp. Will fail with -ENOSYS if not
1384 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1386 struct kioctx *ioctx = NULL;
1390 ret = get_user(ctx, ctxp);
1395 if (unlikely(ctx || nr_events == 0)) {
1396 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1401 ioctx = ioctx_alloc(nr_events);
1402 ret = PTR_ERR(ioctx);
1403 if (!IS_ERR(ioctx)) {
1404 ret = put_user(ioctx->user_id, ctxp);
1406 kill_ioctx(current->mm, ioctx, NULL);
1407 percpu_ref_put(&ioctx->users);
1414 #ifdef CONFIG_COMPAT
1415 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1417 struct kioctx *ioctx = NULL;
1421 ret = get_user(ctx, ctx32p);
1426 if (unlikely(ctx || nr_events == 0)) {
1427 pr_debug("EINVAL: ctx %lu nr_events %u\n",
1432 ioctx = ioctx_alloc(nr_events);
1433 ret = PTR_ERR(ioctx);
1434 if (!IS_ERR(ioctx)) {
1435 /* truncating is ok because it's a user address */
1436 ret = put_user((u32)ioctx->user_id, ctx32p);
1438 kill_ioctx(current->mm, ioctx, NULL);
1439 percpu_ref_put(&ioctx->users);
1448 * Destroy the aio_context specified. May cancel any outstanding
1449 * AIOs and block on completion. Will fail with -ENOSYS if not
1450 * implemented. May fail with -EINVAL if the context pointed to
1453 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1455 struct kioctx *ioctx = lookup_ioctx(ctx);
1456 if (likely(NULL != ioctx)) {
1457 struct ctx_rq_wait wait;
1460 init_completion(&wait.comp);
1461 atomic_set(&wait.count, 1);
1463 /* Pass requests_done to kill_ioctx() where it can be set
1464 * in a thread-safe way. If we try to set it here then we have
1465 * a race condition if two io_destroy() called simultaneously.
1467 ret = kill_ioctx(current->mm, ioctx, &wait);
1468 percpu_ref_put(&ioctx->users);
1470 /* Wait until all IO for the context are done. Otherwise kernel
1471 * keep using user-space buffers even if user thinks the context
1475 wait_for_completion(&wait.comp);
1479 pr_debug("EINVAL: invalid context id\n");
1483 static void aio_remove_iocb(struct aio_kiocb *iocb)
1485 struct kioctx *ctx = iocb->ki_ctx;
1486 unsigned long flags;
1488 spin_lock_irqsave(&ctx->ctx_lock, flags);
1489 list_del(&iocb->ki_list);
1490 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1493 static void aio_complete_rw(struct kiocb *kiocb, long res)
1495 struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1497 if (!list_empty_careful(&iocb->ki_list))
1498 aio_remove_iocb(iocb);
1500 if (kiocb->ki_flags & IOCB_WRITE) {
1501 struct inode *inode = file_inode(kiocb->ki_filp);
1503 if (S_ISREG(inode->i_mode))
1504 kiocb_end_write(kiocb);
1507 iocb->ki_res.res = res;
1508 iocb->ki_res.res2 = 0;
1512 static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
1516 req->ki_complete = aio_complete_rw;
1517 req->private = NULL;
1518 req->ki_pos = iocb->aio_offset;
1519 req->ki_flags = req->ki_filp->f_iocb_flags | IOCB_AIO_RW;
1520 if (iocb->aio_flags & IOCB_FLAG_RESFD)
1521 req->ki_flags |= IOCB_EVENTFD;
1522 if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1524 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1525 * aio_reqprio is interpreted as an I/O scheduling
1526 * class and priority.
1528 ret = ioprio_check_cap(iocb->aio_reqprio);
1530 pr_debug("aio ioprio check cap error: %d\n", ret);
1534 req->ki_ioprio = iocb->aio_reqprio;
1536 req->ki_ioprio = get_current_ioprio();
1538 ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1542 req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
1546 static ssize_t aio_setup_rw(int rw, const struct iocb *iocb,
1547 struct iovec **iovec, bool vectored, bool compat,
1548 struct iov_iter *iter)
1550 void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1551 size_t len = iocb->aio_nbytes;
1554 ssize_t ret = import_ubuf(rw, buf, len, iter);
1559 return __import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter, compat);
1562 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1568 case -ERESTARTNOINTR:
1569 case -ERESTARTNOHAND:
1570 case -ERESTART_RESTARTBLOCK:
1572 * There's no easy way to restart the syscall since other AIO's
1573 * may be already running. Just fail this IO with EINTR.
1578 req->ki_complete(req, ret);
1582 static int aio_read(struct kiocb *req, const struct iocb *iocb,
1583 bool vectored, bool compat)
1585 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1586 struct iov_iter iter;
1590 ret = aio_prep_rw(req, iocb);
1593 file = req->ki_filp;
1594 if (unlikely(!(file->f_mode & FMODE_READ)))
1596 if (unlikely(!file->f_op->read_iter))
1599 ret = aio_setup_rw(ITER_DEST, iocb, &iovec, vectored, compat, &iter);
1602 ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1604 aio_rw_done(req, call_read_iter(file, req, &iter));
1609 static int aio_write(struct kiocb *req, const struct iocb *iocb,
1610 bool vectored, bool compat)
1612 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1613 struct iov_iter iter;
1617 ret = aio_prep_rw(req, iocb);
1620 file = req->ki_filp;
1622 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1624 if (unlikely(!file->f_op->write_iter))
1627 ret = aio_setup_rw(ITER_SOURCE, iocb, &iovec, vectored, compat, &iter);
1630 ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1632 if (S_ISREG(file_inode(file)->i_mode))
1633 kiocb_start_write(req);
1634 req->ki_flags |= IOCB_WRITE;
1635 aio_rw_done(req, call_write_iter(file, req, &iter));
1641 static void aio_fsync_work(struct work_struct *work)
1643 struct aio_kiocb *iocb = container_of(work, struct aio_kiocb, fsync.work);
1644 const struct cred *old_cred = override_creds(iocb->fsync.creds);
1646 iocb->ki_res.res = vfs_fsync(iocb->fsync.file, iocb->fsync.datasync);
1647 revert_creds(old_cred);
1648 put_cred(iocb->fsync.creds);
1652 static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
1655 if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1656 iocb->aio_rw_flags))
1659 if (unlikely(!req->file->f_op->fsync))
1662 req->creds = prepare_creds();
1666 req->datasync = datasync;
1667 INIT_WORK(&req->work, aio_fsync_work);
1668 schedule_work(&req->work);
1672 static void aio_poll_put_work(struct work_struct *work)
1674 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1675 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1681 * Safely lock the waitqueue which the request is on, synchronizing with the
1682 * case where the ->poll() provider decides to free its waitqueue early.
1684 * Returns true on success, meaning that req->head->lock was locked, req->wait
1685 * is on req->head, and an RCU read lock was taken. Returns false if the
1686 * request was already removed from its waitqueue (which might no longer exist).
1688 static bool poll_iocb_lock_wq(struct poll_iocb *req)
1690 wait_queue_head_t *head;
1693 * While we hold the waitqueue lock and the waitqueue is nonempty,
1694 * wake_up_pollfree() will wait for us. However, taking the waitqueue
1695 * lock in the first place can race with the waitqueue being freed.
1697 * We solve this as eventpoll does: by taking advantage of the fact that
1698 * all users of wake_up_pollfree() will RCU-delay the actual free. If
1699 * we enter rcu_read_lock() and see that the pointer to the queue is
1700 * non-NULL, we can then lock it without the memory being freed out from
1701 * under us, then check whether the request is still on the queue.
1703 * Keep holding rcu_read_lock() as long as we hold the queue lock, in
1704 * case the caller deletes the entry from the queue, leaving it empty.
1705 * In that case, only RCU prevents the queue memory from being freed.
1708 head = smp_load_acquire(&req->head);
1710 spin_lock(&head->lock);
1711 if (!list_empty(&req->wait.entry))
1713 spin_unlock(&head->lock);
1719 static void poll_iocb_unlock_wq(struct poll_iocb *req)
1721 spin_unlock(&req->head->lock);
1725 static void aio_poll_complete_work(struct work_struct *work)
1727 struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1728 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1729 struct poll_table_struct pt = { ._key = req->events };
1730 struct kioctx *ctx = iocb->ki_ctx;
1733 if (!READ_ONCE(req->cancelled))
1734 mask = vfs_poll(req->file, &pt) & req->events;
1737 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1738 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1739 * synchronize with them. In the cancellation case the list_del_init
1740 * itself is not actually needed, but harmless so we keep it in to
1741 * avoid further branches in the fast path.
1743 spin_lock_irq(&ctx->ctx_lock);
1744 if (poll_iocb_lock_wq(req)) {
1745 if (!mask && !READ_ONCE(req->cancelled)) {
1747 * The request isn't actually ready to be completed yet.
1748 * Reschedule completion if another wakeup came in.
1750 if (req->work_need_resched) {
1751 schedule_work(&req->work);
1752 req->work_need_resched = false;
1754 req->work_scheduled = false;
1756 poll_iocb_unlock_wq(req);
1757 spin_unlock_irq(&ctx->ctx_lock);
1760 list_del_init(&req->wait.entry);
1761 poll_iocb_unlock_wq(req);
1762 } /* else, POLLFREE has freed the waitqueue, so we must complete */
1763 list_del_init(&iocb->ki_list);
1764 iocb->ki_res.res = mangle_poll(mask);
1765 spin_unlock_irq(&ctx->ctx_lock);
1770 /* assumes we are called with irqs disabled */
1771 static int aio_poll_cancel(struct kiocb *iocb)
1773 struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1774 struct poll_iocb *req = &aiocb->poll;
1776 if (poll_iocb_lock_wq(req)) {
1777 WRITE_ONCE(req->cancelled, true);
1778 if (!req->work_scheduled) {
1779 schedule_work(&aiocb->poll.work);
1780 req->work_scheduled = true;
1782 poll_iocb_unlock_wq(req);
1783 } /* else, the request was force-cancelled by POLLFREE already */
1788 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1791 struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1792 struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1793 __poll_t mask = key_to_poll(key);
1794 unsigned long flags;
1796 /* for instances that support it check for an event match first: */
1797 if (mask && !(mask & req->events))
1801 * Complete the request inline if possible. This requires that three
1802 * conditions be met:
1803 * 1. An event mask must have been passed. If a plain wakeup was done
1804 * instead, then mask == 0 and we have to call vfs_poll() to get
1805 * the events, so inline completion isn't possible.
1806 * 2. The completion work must not have already been scheduled.
1807 * 3. ctx_lock must not be busy. We have to use trylock because we
1808 * already hold the waitqueue lock, so this inverts the normal
1809 * locking order. Use irqsave/irqrestore because not all
1810 * filesystems (e.g. fuse) call this function with IRQs disabled,
1811 * yet IRQs have to be disabled before ctx_lock is obtained.
1813 if (mask && !req->work_scheduled &&
1814 spin_trylock_irqsave(&iocb->ki_ctx->ctx_lock, flags)) {
1815 struct kioctx *ctx = iocb->ki_ctx;
1817 list_del_init(&req->wait.entry);
1818 list_del(&iocb->ki_list);
1819 iocb->ki_res.res = mangle_poll(mask);
1820 if (iocb->ki_eventfd && !eventfd_signal_allowed()) {
1822 INIT_WORK(&req->work, aio_poll_put_work);
1823 schedule_work(&req->work);
1825 spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1830 * Schedule the completion work if needed. If it was already
1831 * scheduled, record that another wakeup came in.
1833 * Don't remove the request from the waitqueue here, as it might
1834 * not actually be complete yet (we won't know until vfs_poll()
1835 * is called), and we must not miss any wakeups. POLLFREE is an
1836 * exception to this; see below.
1838 if (req->work_scheduled) {
1839 req->work_need_resched = true;
1841 schedule_work(&req->work);
1842 req->work_scheduled = true;
1846 * If the waitqueue is being freed early but we can't complete
1847 * the request inline, we have to tear down the request as best
1848 * we can. That means immediately removing the request from its
1849 * waitqueue and preventing all further accesses to the
1850 * waitqueue via the request. We also need to schedule the
1851 * completion work (done above). Also mark the request as
1852 * cancelled, to potentially skip an unneeded call to ->poll().
1854 if (mask & POLLFREE) {
1855 WRITE_ONCE(req->cancelled, true);
1856 list_del_init(&req->wait.entry);
1859 * Careful: this *must* be the last step, since as soon
1860 * as req->head is NULL'ed out, the request can be
1861 * completed and freed, since aio_poll_complete_work()
1862 * will no longer need to take the waitqueue lock.
1864 smp_store_release(&req->head, NULL);
1870 struct aio_poll_table {
1871 struct poll_table_struct pt;
1872 struct aio_kiocb *iocb;
1878 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1879 struct poll_table_struct *p)
1881 struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1883 /* multiple wait queues per file are not supported */
1884 if (unlikely(pt->queued)) {
1885 pt->error = -EINVAL;
1891 pt->iocb->poll.head = head;
1892 add_wait_queue(head, &pt->iocb->poll.wait);
1895 static int aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
1897 struct kioctx *ctx = aiocb->ki_ctx;
1898 struct poll_iocb *req = &aiocb->poll;
1899 struct aio_poll_table apt;
1900 bool cancel = false;
1903 /* reject any unknown events outside the normal event mask. */
1904 if ((u16)iocb->aio_buf != iocb->aio_buf)
1906 /* reject fields that are not defined for poll */
1907 if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1910 INIT_WORK(&req->work, aio_poll_complete_work);
1911 req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1914 req->cancelled = false;
1915 req->work_scheduled = false;
1916 req->work_need_resched = false;
1918 apt.pt._qproc = aio_poll_queue_proc;
1919 apt.pt._key = req->events;
1922 apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1924 /* initialized the list so that we can do list_empty checks */
1925 INIT_LIST_HEAD(&req->wait.entry);
1926 init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1928 mask = vfs_poll(req->file, &apt.pt) & req->events;
1929 spin_lock_irq(&ctx->ctx_lock);
1930 if (likely(apt.queued)) {
1931 bool on_queue = poll_iocb_lock_wq(req);
1933 if (!on_queue || req->work_scheduled) {
1935 * aio_poll_wake() already either scheduled the async
1936 * completion work, or completed the request inline.
1938 if (apt.error) /* unsupported case: multiple queues */
1943 if (mask || apt.error) {
1944 /* Steal to complete synchronously. */
1945 list_del_init(&req->wait.entry);
1946 } else if (cancel) {
1947 /* Cancel if possible (may be too late though). */
1948 WRITE_ONCE(req->cancelled, true);
1949 } else if (on_queue) {
1951 * Actually waiting for an event, so add the request to
1952 * active_reqs so that it can be cancelled if needed.
1954 list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1955 aiocb->ki_cancel = aio_poll_cancel;
1958 poll_iocb_unlock_wq(req);
1960 if (mask) { /* no async, we'd stolen it */
1961 aiocb->ki_res.res = mangle_poll(mask);
1964 spin_unlock_irq(&ctx->ctx_lock);
1970 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
1971 struct iocb __user *user_iocb, struct aio_kiocb *req,
1974 req->ki_filp = fget(iocb->aio_fildes);
1975 if (unlikely(!req->ki_filp))
1978 if (iocb->aio_flags & IOCB_FLAG_RESFD) {
1979 struct eventfd_ctx *eventfd;
1981 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1982 * instance of the file* now. The file descriptor must be
1983 * an eventfd() fd, and will be signaled for each completed
1984 * event using the eventfd_signal() function.
1986 eventfd = eventfd_ctx_fdget(iocb->aio_resfd);
1987 if (IS_ERR(eventfd))
1988 return PTR_ERR(eventfd);
1990 req->ki_eventfd = eventfd;
1993 if (unlikely(put_user(KIOCB_KEY, &user_iocb->aio_key))) {
1994 pr_debug("EFAULT: aio_key\n");
1998 req->ki_res.obj = (u64)(unsigned long)user_iocb;
1999 req->ki_res.data = iocb->aio_data;
2000 req->ki_res.res = 0;
2001 req->ki_res.res2 = 0;
2003 switch (iocb->aio_lio_opcode) {
2004 case IOCB_CMD_PREAD:
2005 return aio_read(&req->rw, iocb, false, compat);
2006 case IOCB_CMD_PWRITE:
2007 return aio_write(&req->rw, iocb, false, compat);
2008 case IOCB_CMD_PREADV:
2009 return aio_read(&req->rw, iocb, true, compat);
2010 case IOCB_CMD_PWRITEV:
2011 return aio_write(&req->rw, iocb, true, compat);
2012 case IOCB_CMD_FSYNC:
2013 return aio_fsync(&req->fsync, iocb, false);
2014 case IOCB_CMD_FDSYNC:
2015 return aio_fsync(&req->fsync, iocb, true);
2017 return aio_poll(req, iocb);
2019 pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
2024 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
2027 struct aio_kiocb *req;
2031 if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
2034 /* enforce forwards compatibility on users */
2035 if (unlikely(iocb.aio_reserved2)) {
2036 pr_debug("EINVAL: reserve field set\n");
2040 /* prevent overflows */
2042 (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
2043 (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
2044 ((ssize_t)iocb.aio_nbytes < 0)
2046 pr_debug("EINVAL: overflow check\n");
2050 req = aio_get_req(ctx);
2054 err = __io_submit_one(ctx, &iocb, user_iocb, req, compat);
2056 /* Done with the synchronous reference */
2060 * If err is 0, we'd either done aio_complete() ourselves or have
2061 * arranged for that to be done asynchronously. Anything non-zero
2062 * means that we need to destroy req ourselves.
2064 if (unlikely(err)) {
2066 put_reqs_available(ctx, 1);
2072 * Queue the nr iocbs pointed to by iocbpp for processing. Returns
2073 * the number of iocbs queued. May return -EINVAL if the aio_context
2074 * specified by ctx_id is invalid, if nr is < 0, if the iocb at
2075 * *iocbpp[0] is not properly initialized, if the operation specified
2076 * is invalid for the file descriptor in the iocb. May fail with
2077 * -EFAULT if any of the data structures point to invalid data. May
2078 * fail with -EBADF if the file descriptor specified in the first
2079 * iocb is invalid. May fail with -EAGAIN if insufficient resources
2080 * are available to queue any iocbs. Will return 0 if nr is 0. Will
2081 * fail with -ENOSYS if not implemented.
2083 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
2084 struct iocb __user * __user *, iocbpp)
2089 struct blk_plug plug;
2091 if (unlikely(nr < 0))
2094 ctx = lookup_ioctx(ctx_id);
2095 if (unlikely(!ctx)) {
2096 pr_debug("EINVAL: invalid context id\n");
2100 if (nr > ctx->nr_events)
2101 nr = ctx->nr_events;
2103 if (nr > AIO_PLUG_THRESHOLD)
2104 blk_start_plug(&plug);
2105 for (i = 0; i < nr; i++) {
2106 struct iocb __user *user_iocb;
2108 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2113 ret = io_submit_one(ctx, user_iocb, false);
2117 if (nr > AIO_PLUG_THRESHOLD)
2118 blk_finish_plug(&plug);
2120 percpu_ref_put(&ctx->users);
2124 #ifdef CONFIG_COMPAT
2125 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
2126 int, nr, compat_uptr_t __user *, iocbpp)
2131 struct blk_plug plug;
2133 if (unlikely(nr < 0))
2136 ctx = lookup_ioctx(ctx_id);
2137 if (unlikely(!ctx)) {
2138 pr_debug("EINVAL: invalid context id\n");
2142 if (nr > ctx->nr_events)
2143 nr = ctx->nr_events;
2145 if (nr > AIO_PLUG_THRESHOLD)
2146 blk_start_plug(&plug);
2147 for (i = 0; i < nr; i++) {
2148 compat_uptr_t user_iocb;
2150 if (unlikely(get_user(user_iocb, iocbpp + i))) {
2155 ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
2159 if (nr > AIO_PLUG_THRESHOLD)
2160 blk_finish_plug(&plug);
2162 percpu_ref_put(&ctx->users);
2168 * Attempts to cancel an iocb previously passed to io_submit(). If the
2169 * operation is successfully cancelled 0 is returned. May fail with
2170 * -EFAULT if any of the data structures pointed to are invalid. May
2171 * fail with -EINVAL if aio_context specified by ctx_id is invalid. Will
2172 * fail with -ENOSYS if not implemented.
2174 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
2175 struct io_event __user *, result)
2178 struct aio_kiocb *kiocb;
2181 u64 obj = (u64)(unsigned long)iocb;
2183 if (unlikely(get_user(key, &iocb->aio_key)))
2185 if (unlikely(key != KIOCB_KEY))
2188 ctx = lookup_ioctx(ctx_id);
2192 spin_lock_irq(&ctx->ctx_lock);
2193 /* TODO: use a hash or array, this sucks. */
2194 list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
2195 if (kiocb->ki_res.obj == obj) {
2196 ret = kiocb->ki_cancel(&kiocb->rw);
2197 list_del_init(&kiocb->ki_list);
2201 spin_unlock_irq(&ctx->ctx_lock);
2204 * The result argument is no longer used - the io_event is always
2205 * delivered via the ring buffer.
2207 if (ret == 0 && kiocb->rw.ki_flags & IOCB_AIO_RW)
2208 aio_complete_rw(&kiocb->rw, -EINTR);
2210 percpu_ref_put(&ctx->users);
2215 static long do_io_getevents(aio_context_t ctx_id,
2218 struct io_event __user *events,
2219 struct timespec64 *ts)
2221 ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2222 struct kioctx *ioctx = lookup_ioctx(ctx_id);
2225 if (likely(ioctx)) {
2226 if (likely(min_nr <= nr && min_nr >= 0))
2227 ret = read_events(ioctx, min_nr, nr, events, until);
2228 percpu_ref_put(&ioctx->users);
2235 * Attempts to read at least min_nr events and up to nr events from
2236 * the completion queue for the aio_context specified by ctx_id. If
2237 * it succeeds, the number of read events is returned. May fail with
2238 * -EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2239 * out of range, if timeout is out of range. May fail with -EFAULT
2240 * if any of the memory specified is invalid. May return 0 or
2241 * < min_nr if the timeout specified by timeout has elapsed
2242 * before sufficient events are available, where timeout == NULL
2243 * specifies an infinite timeout. Note that the timeout pointed to by
2244 * timeout is relative. Will fail with -ENOSYS if not implemented.
2248 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2251 struct io_event __user *, events,
2252 struct __kernel_timespec __user *, timeout)
2254 struct timespec64 ts;
2257 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2260 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2261 if (!ret && signal_pending(current))
2268 struct __aio_sigset {
2269 const sigset_t __user *sigmask;
2273 SYSCALL_DEFINE6(io_pgetevents,
2274 aio_context_t, ctx_id,
2277 struct io_event __user *, events,
2278 struct __kernel_timespec __user *, timeout,
2279 const struct __aio_sigset __user *, usig)
2281 struct __aio_sigset ksig = { NULL, };
2282 struct timespec64 ts;
2286 if (timeout && unlikely(get_timespec64(&ts, timeout)))
2289 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2292 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2296 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2298 interrupted = signal_pending(current);
2299 restore_saved_sigmask_unless(interrupted);
2300 if (interrupted && !ret)
2301 ret = -ERESTARTNOHAND;
2306 #if defined(CONFIG_COMPAT_32BIT_TIME) && !defined(CONFIG_64BIT)
2308 SYSCALL_DEFINE6(io_pgetevents_time32,
2309 aio_context_t, ctx_id,
2312 struct io_event __user *, events,
2313 struct old_timespec32 __user *, timeout,
2314 const struct __aio_sigset __user *, usig)
2316 struct __aio_sigset ksig = { NULL, };
2317 struct timespec64 ts;
2321 if (timeout && unlikely(get_old_timespec32(&ts, timeout)))
2324 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2328 ret = set_user_sigmask(ksig.sigmask, ksig.sigsetsize);
2332 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2334 interrupted = signal_pending(current);
2335 restore_saved_sigmask_unless(interrupted);
2336 if (interrupted && !ret)
2337 ret = -ERESTARTNOHAND;
2344 #if defined(CONFIG_COMPAT_32BIT_TIME)
2346 SYSCALL_DEFINE5(io_getevents_time32, __u32, ctx_id,
2349 struct io_event __user *, events,
2350 struct old_timespec32 __user *, timeout)
2352 struct timespec64 t;
2355 if (timeout && get_old_timespec32(&t, timeout))
2358 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2359 if (!ret && signal_pending(current))
2366 #ifdef CONFIG_COMPAT
2368 struct __compat_aio_sigset {
2369 compat_uptr_t sigmask;
2370 compat_size_t sigsetsize;
2373 #if defined(CONFIG_COMPAT_32BIT_TIME)
2375 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2376 compat_aio_context_t, ctx_id,
2377 compat_long_t, min_nr,
2379 struct io_event __user *, events,
2380 struct old_timespec32 __user *, timeout,
2381 const struct __compat_aio_sigset __user *, usig)
2383 struct __compat_aio_sigset ksig = { 0, };
2384 struct timespec64 t;
2388 if (timeout && get_old_timespec32(&t, timeout))
2391 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2394 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2398 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2400 interrupted = signal_pending(current);
2401 restore_saved_sigmask_unless(interrupted);
2402 if (interrupted && !ret)
2403 ret = -ERESTARTNOHAND;
2410 COMPAT_SYSCALL_DEFINE6(io_pgetevents_time64,
2411 compat_aio_context_t, ctx_id,
2412 compat_long_t, min_nr,
2414 struct io_event __user *, events,
2415 struct __kernel_timespec __user *, timeout,
2416 const struct __compat_aio_sigset __user *, usig)
2418 struct __compat_aio_sigset ksig = { 0, };
2419 struct timespec64 t;
2423 if (timeout && get_timespec64(&t, timeout))
2426 if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2429 ret = set_compat_user_sigmask(compat_ptr(ksig.sigmask), ksig.sigsetsize);
2433 ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2435 interrupted = signal_pending(current);
2436 restore_saved_sigmask_unless(interrupted);
2437 if (interrupted && !ret)
2438 ret = -ERESTARTNOHAND;