1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 4096
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp;
83 u32 tail ____cacheline_aligned_in_smp;
87 * This data is shared with the application through the mmap at offset
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head and the application controls tail.
102 * Bitmask to apply to head and tail offsets (constant, equals
106 /* Ring size (constant, power of 2) */
109 * Number of invalid entries dropped by the kernel due to
110 * invalid index stored in array
112 * Written by the kernel, shouldn't be modified by the
113 * application (i.e. get number of "new events" by comparing to
116 * After a new SQ head value was read by the application this
117 * counter includes all submissions that were dropped reaching
118 * the new SQ head (and possibly more).
124 * Written by the kernel, shouldn't be modified by the
127 * The application needs a full memory barrier before checking
128 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
132 * Ring buffer of indices into array of io_uring_sqe, which is
133 * mmapped by the application using the IORING_OFF_SQES offset.
135 * This indirection could e.g. be used to assign fixed
136 * io_uring_sqe entries to operations and only submit them to
137 * the queue when needed.
139 * The kernel modifies neither the indices array nor the entries
146 * This data is shared with the application through the mmap at offset
147 * IORING_OFF_CQ_RING.
149 * The offsets to the member fields are published through struct
150 * io_cqring_offsets when calling io_uring_setup.
154 * Head and tail offsets into the ring; the offsets need to be
155 * masked to get valid indices.
157 * The application controls head and the kernel tail.
161 * Bitmask to apply to head and tail offsets (constant, equals
165 /* Ring size (constant, power of 2) */
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending thatn there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[];
191 struct io_mapped_ubuf {
194 struct bio_vec *bvec;
195 unsigned int nr_bvecs;
201 struct list_head list;
210 struct percpu_ref refs;
211 } ____cacheline_aligned_in_smp;
219 struct io_sq_ring *sq_ring;
220 unsigned cached_sq_head;
223 unsigned sq_thread_idle;
224 struct io_uring_sqe *sq_sqes;
226 struct list_head defer_list;
227 } ____cacheline_aligned_in_smp;
230 struct workqueue_struct *sqo_wq;
231 struct task_struct *sqo_thread; /* if using sq thread polling */
232 struct mm_struct *sqo_mm;
233 wait_queue_head_t sqo_wait;
234 struct completion sqo_thread_started;
238 struct io_cq_ring *cq_ring;
239 unsigned cached_cq_tail;
242 struct wait_queue_head cq_wait;
243 struct fasync_struct *cq_fasync;
244 struct eventfd_ctx *cq_ev_fd;
245 } ____cacheline_aligned_in_smp;
248 * If used, fixed file set. Writers must ensure that ->refs is dead,
249 * readers must ensure that ->refs is alive as long as the file* is
250 * used. Only updated through io_uring_register(2).
252 struct file **user_files;
253 unsigned nr_user_files;
255 /* if used, fixed mapped user buffers */
256 unsigned nr_user_bufs;
257 struct io_mapped_ubuf *user_bufs;
259 struct user_struct *user;
261 struct completion ctx_done;
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
269 spinlock_t completion_lock;
270 bool poll_multi_file;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list;
278 struct list_head cancel_list;
279 } ____cacheline_aligned_in_smp;
281 struct async_list pending_async[2];
283 #if defined(CONFIG_UNIX)
284 struct socket *ring_sock;
289 const struct io_uring_sqe *sqe;
290 unsigned short index;
293 bool needs_fixed_file;
297 * First field must be the file pointer in all the
298 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
300 struct io_poll_iocb {
302 struct wait_queue_head *head;
306 struct wait_queue_entry wait;
310 * NOTE! Each of the iocb union members has the file pointer
311 * as the first entry in their struct definition. So you can
312 * access the file pointer through any of the sub-structs,
313 * or directly as just 'ki_filp' in this struct.
319 struct io_poll_iocb poll;
322 struct sqe_submit submit;
324 struct io_ring_ctx *ctx;
325 struct list_head list;
326 struct list_head link_list;
329 #define REQ_F_NOWAIT 1 /* must not punt to workers */
330 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
331 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
332 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
333 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
334 #define REQ_F_IO_DRAINED 32 /* drain done */
335 #define REQ_F_LINK 64 /* linked sqes */
336 #define REQ_F_FAIL_LINK 128 /* fail rest of links */
341 struct work_struct work;
344 #define IO_PLUG_THRESHOLD 2
345 #define IO_IOPOLL_BATCH 8
347 struct io_submit_state {
348 struct blk_plug plug;
351 * io_kiocb alloc cache
353 void *reqs[IO_IOPOLL_BATCH];
354 unsigned int free_reqs;
355 unsigned int cur_req;
358 * File reference cache
362 unsigned int has_refs;
363 unsigned int used_refs;
364 unsigned int ios_left;
367 static void io_sq_wq_submit_work(struct work_struct *work);
369 static struct kmem_cache *req_cachep;
371 static const struct file_operations io_uring_fops;
373 struct sock *io_uring_get_socket(struct file *file)
375 #if defined(CONFIG_UNIX)
376 if (file->f_op == &io_uring_fops) {
377 struct io_ring_ctx *ctx = file->private_data;
379 return ctx->ring_sock->sk;
384 EXPORT_SYMBOL(io_uring_get_socket);
386 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
388 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
390 complete(&ctx->ctx_done);
393 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
395 struct io_ring_ctx *ctx;
398 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
402 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) {
407 ctx->flags = p->flags;
408 init_waitqueue_head(&ctx->cq_wait);
409 init_completion(&ctx->ctx_done);
410 init_completion(&ctx->sqo_thread_started);
411 mutex_init(&ctx->uring_lock);
412 init_waitqueue_head(&ctx->wait);
413 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
414 spin_lock_init(&ctx->pending_async[i].lock);
415 INIT_LIST_HEAD(&ctx->pending_async[i].list);
416 atomic_set(&ctx->pending_async[i].cnt, 0);
418 spin_lock_init(&ctx->completion_lock);
419 INIT_LIST_HEAD(&ctx->poll_list);
420 INIT_LIST_HEAD(&ctx->cancel_list);
421 INIT_LIST_HEAD(&ctx->defer_list);
425 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
426 struct io_kiocb *req)
428 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
431 return req->sequence > ctx->cached_cq_tail + ctx->sq_ring->dropped;
434 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
436 struct io_kiocb *req;
438 if (list_empty(&ctx->defer_list))
441 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
442 if (!io_sequence_defer(ctx, req)) {
443 list_del_init(&req->list);
450 static void __io_commit_cqring(struct io_ring_ctx *ctx)
452 struct io_cq_ring *ring = ctx->cq_ring;
454 if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) {
455 /* order cqe stores with ring update */
456 smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
458 if (wq_has_sleeper(&ctx->cq_wait)) {
459 wake_up_interruptible(&ctx->cq_wait);
460 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
465 static void io_commit_cqring(struct io_ring_ctx *ctx)
467 struct io_kiocb *req;
469 __io_commit_cqring(ctx);
471 while ((req = io_get_deferred_req(ctx)) != NULL) {
472 req->flags |= REQ_F_IO_DRAINED;
473 queue_work(ctx->sqo_wq, &req->work);
477 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
479 struct io_cq_ring *ring = ctx->cq_ring;
482 tail = ctx->cached_cq_tail;
484 * writes to the cq entry need to come after reading head; the
485 * control dependency is enough as we're using WRITE_ONCE to
488 if (tail - READ_ONCE(ring->r.head) == ring->ring_entries)
491 ctx->cached_cq_tail++;
492 return &ring->cqes[tail & ctx->cq_mask];
495 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
498 struct io_uring_cqe *cqe;
501 * If we can't get a cq entry, userspace overflowed the
502 * submission (by quite a lot). Increment the overflow count in
505 cqe = io_get_cqring(ctx);
507 WRITE_ONCE(cqe->user_data, ki_user_data);
508 WRITE_ONCE(cqe->res, res);
509 WRITE_ONCE(cqe->flags, 0);
511 unsigned overflow = READ_ONCE(ctx->cq_ring->overflow);
513 WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1);
517 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
519 if (waitqueue_active(&ctx->wait))
521 if (waitqueue_active(&ctx->sqo_wait))
522 wake_up(&ctx->sqo_wait);
524 eventfd_signal(ctx->cq_ev_fd, 1);
527 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
532 spin_lock_irqsave(&ctx->completion_lock, flags);
533 io_cqring_fill_event(ctx, user_data, res);
534 io_commit_cqring(ctx);
535 spin_unlock_irqrestore(&ctx->completion_lock, flags);
537 io_cqring_ev_posted(ctx);
540 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
542 percpu_ref_put_many(&ctx->refs, refs);
544 if (waitqueue_active(&ctx->wait))
548 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
549 struct io_submit_state *state)
551 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
552 struct io_kiocb *req;
554 if (!percpu_ref_tryget(&ctx->refs))
558 req = kmem_cache_alloc(req_cachep, gfp);
561 } else if (!state->free_reqs) {
565 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
566 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
569 * Bulk alloc is all-or-nothing. If we fail to get a batch,
570 * retry single alloc to be on the safe side.
572 if (unlikely(ret <= 0)) {
573 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
578 state->free_reqs = ret - 1;
580 req = state->reqs[0];
582 req = state->reqs[state->cur_req];
590 /* one is dropped after submission, the other at completion */
591 refcount_set(&req->refs, 2);
595 io_ring_drop_ctx_refs(ctx, 1);
599 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
602 kmem_cache_free_bulk(req_cachep, *nr, reqs);
603 io_ring_drop_ctx_refs(ctx, *nr);
608 static void __io_free_req(struct io_kiocb *req)
610 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
612 io_ring_drop_ctx_refs(req->ctx, 1);
613 kmem_cache_free(req_cachep, req);
616 static void io_req_link_next(struct io_kiocb *req)
618 struct io_kiocb *nxt;
621 * The list should never be empty when we are called here. But could
622 * potentially happen if the chain is messed up, check to be on the
625 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
627 list_del(&nxt->list);
628 if (!list_empty(&req->link_list)) {
629 INIT_LIST_HEAD(&nxt->link_list);
630 list_splice(&req->link_list, &nxt->link_list);
631 nxt->flags |= REQ_F_LINK;
634 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
635 queue_work(req->ctx->sqo_wq, &nxt->work);
640 * Called if REQ_F_LINK is set, and we fail the head request
642 static void io_fail_links(struct io_kiocb *req)
644 struct io_kiocb *link;
646 while (!list_empty(&req->link_list)) {
647 link = list_first_entry(&req->link_list, struct io_kiocb, list);
648 list_del(&link->list);
650 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
655 static void io_free_req(struct io_kiocb *req)
658 * If LINK is set, we have dependent requests in this chain. If we
659 * didn't fail this request, queue the first one up, moving any other
660 * dependencies to the next request. In case of failure, fail the rest
663 if (req->flags & REQ_F_LINK) {
664 if (req->flags & REQ_F_FAIL_LINK)
667 io_req_link_next(req);
673 static void io_put_req(struct io_kiocb *req)
675 if (refcount_dec_and_test(&req->refs))
680 * Find and free completed poll iocbs
682 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
683 struct list_head *done)
685 void *reqs[IO_IOPOLL_BATCH];
686 struct io_kiocb *req;
690 while (!list_empty(done)) {
691 req = list_first_entry(done, struct io_kiocb, list);
692 list_del(&req->list);
694 io_cqring_fill_event(ctx, req->user_data, req->result);
697 if (refcount_dec_and_test(&req->refs)) {
698 /* If we're not using fixed files, we have to pair the
699 * completion part with the file put. Use regular
700 * completions for those, only batch free for fixed
701 * file and non-linked commands.
703 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
705 reqs[to_free++] = req;
706 if (to_free == ARRAY_SIZE(reqs))
707 io_free_req_many(ctx, reqs, &to_free);
714 io_commit_cqring(ctx);
715 io_free_req_many(ctx, reqs, &to_free);
718 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
721 struct io_kiocb *req, *tmp;
727 * Only spin for completions if we don't have multiple devices hanging
728 * off our complete list, and we're under the requested amount.
730 spin = !ctx->poll_multi_file && *nr_events < min;
733 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
734 struct kiocb *kiocb = &req->rw;
737 * Move completed entries to our local list. If we find a
738 * request that requires polling, break out and complete
739 * the done list first, if we have entries there.
741 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
742 list_move_tail(&req->list, &done);
745 if (!list_empty(&done))
748 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
757 if (!list_empty(&done))
758 io_iopoll_complete(ctx, nr_events, &done);
764 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
765 * non-spinning poll check - we'll still enter the driver poll loop, but only
766 * as a non-spinning completion check.
768 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
771 while (!list_empty(&ctx->poll_list)) {
774 ret = io_do_iopoll(ctx, nr_events, min);
777 if (!min || *nr_events >= min)
785 * We can't just wait for polled events to come to us, we have to actively
786 * find and complete them.
788 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
790 if (!(ctx->flags & IORING_SETUP_IOPOLL))
793 mutex_lock(&ctx->uring_lock);
794 while (!list_empty(&ctx->poll_list)) {
795 unsigned int nr_events = 0;
797 io_iopoll_getevents(ctx, &nr_events, 1);
799 mutex_unlock(&ctx->uring_lock);
802 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
810 if (*nr_events < min)
811 tmin = min - *nr_events;
813 ret = io_iopoll_getevents(ctx, nr_events, tmin);
817 } while (min && !*nr_events && !need_resched());
822 static void kiocb_end_write(struct kiocb *kiocb)
824 if (kiocb->ki_flags & IOCB_WRITE) {
825 struct inode *inode = file_inode(kiocb->ki_filp);
828 * Tell lockdep we inherited freeze protection from submission
831 if (S_ISREG(inode->i_mode))
832 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
833 file_end_write(kiocb->ki_filp);
837 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
839 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
841 kiocb_end_write(kiocb);
843 if ((req->flags & REQ_F_LINK) && res != req->result)
844 req->flags |= REQ_F_FAIL_LINK;
845 io_cqring_add_event(req->ctx, req->user_data, res);
849 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
851 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
853 kiocb_end_write(kiocb);
855 if ((req->flags & REQ_F_LINK) && res != req->result)
856 req->flags |= REQ_F_FAIL_LINK;
859 req->flags |= REQ_F_IOPOLL_COMPLETED;
863 * After the iocb has been issued, it's safe to be found on the poll list.
864 * Adding the kiocb to the list AFTER submission ensures that we don't
865 * find it from a io_iopoll_getevents() thread before the issuer is done
866 * accessing the kiocb cookie.
868 static void io_iopoll_req_issued(struct io_kiocb *req)
870 struct io_ring_ctx *ctx = req->ctx;
873 * Track whether we have multiple files in our lists. This will impact
874 * how we do polling eventually, not spinning if we're on potentially
877 if (list_empty(&ctx->poll_list)) {
878 ctx->poll_multi_file = false;
879 } else if (!ctx->poll_multi_file) {
880 struct io_kiocb *list_req;
882 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
884 if (list_req->rw.ki_filp != req->rw.ki_filp)
885 ctx->poll_multi_file = true;
889 * For fast devices, IO may have already completed. If it has, add
890 * it to the front so we find it first.
892 if (req->flags & REQ_F_IOPOLL_COMPLETED)
893 list_add(&req->list, &ctx->poll_list);
895 list_add_tail(&req->list, &ctx->poll_list);
898 static void io_file_put(struct io_submit_state *state)
901 int diff = state->has_refs - state->used_refs;
904 fput_many(state->file, diff);
910 * Get as many references to a file as we have IOs left in this submission,
911 * assuming most submissions are for one file, or at least that each file
912 * has more than one submission.
914 static struct file *io_file_get(struct io_submit_state *state, int fd)
920 if (state->fd == fd) {
927 state->file = fget_many(fd, state->ios_left);
932 state->has_refs = state->ios_left;
933 state->used_refs = 1;
939 * If we tracked the file through the SCM inflight mechanism, we could support
940 * any file. For now, just ensure that anything potentially problematic is done
943 static bool io_file_supports_async(struct file *file)
945 umode_t mode = file_inode(file)->i_mode;
947 if (S_ISBLK(mode) || S_ISCHR(mode))
949 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
955 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
958 const struct io_uring_sqe *sqe = s->sqe;
959 struct io_ring_ctx *ctx = req->ctx;
960 struct kiocb *kiocb = &req->rw;
967 if (force_nonblock && !io_file_supports_async(req->file))
968 force_nonblock = false;
970 kiocb->ki_pos = READ_ONCE(sqe->off);
971 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
972 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
974 ioprio = READ_ONCE(sqe->ioprio);
976 ret = ioprio_check_cap(ioprio);
980 kiocb->ki_ioprio = ioprio;
982 kiocb->ki_ioprio = get_current_ioprio();
984 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
988 /* don't allow async punt if RWF_NOWAIT was requested */
989 if (kiocb->ki_flags & IOCB_NOWAIT)
990 req->flags |= REQ_F_NOWAIT;
993 kiocb->ki_flags |= IOCB_NOWAIT;
995 if (ctx->flags & IORING_SETUP_IOPOLL) {
996 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
997 !kiocb->ki_filp->f_op->iopoll)
1000 kiocb->ki_flags |= IOCB_HIPRI;
1001 kiocb->ki_complete = io_complete_rw_iopoll;
1003 if (kiocb->ki_flags & IOCB_HIPRI)
1005 kiocb->ki_complete = io_complete_rw;
1010 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1016 case -ERESTARTNOINTR:
1017 case -ERESTARTNOHAND:
1018 case -ERESTART_RESTARTBLOCK:
1020 * We can't just restart the syscall, since previously
1021 * submitted sqes may already be in progress. Just fail this
1027 kiocb->ki_complete(kiocb, ret, 0);
1031 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1032 const struct io_uring_sqe *sqe,
1033 struct iov_iter *iter)
1035 size_t len = READ_ONCE(sqe->len);
1036 struct io_mapped_ubuf *imu;
1037 unsigned index, buf_index;
1041 /* attempt to use fixed buffers without having provided iovecs */
1042 if (unlikely(!ctx->user_bufs))
1045 buf_index = READ_ONCE(sqe->buf_index);
1046 if (unlikely(buf_index >= ctx->nr_user_bufs))
1049 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1050 imu = &ctx->user_bufs[index];
1051 buf_addr = READ_ONCE(sqe->addr);
1054 if (buf_addr + len < buf_addr)
1056 /* not inside the mapped region */
1057 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1061 * May not be a start of buffer, set size appropriately
1062 * and advance us to the beginning.
1064 offset = buf_addr - imu->ubuf;
1065 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1067 iov_iter_advance(iter, offset);
1071 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1072 const struct sqe_submit *s, struct iovec **iovec,
1073 struct iov_iter *iter)
1075 const struct io_uring_sqe *sqe = s->sqe;
1076 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1077 size_t sqe_len = READ_ONCE(sqe->len);
1081 * We're reading ->opcode for the second time, but the first read
1082 * doesn't care whether it's _FIXED or not, so it doesn't matter
1083 * whether ->opcode changes concurrently. The first read does care
1084 * about whether it is a READ or a WRITE, so we don't trust this read
1085 * for that purpose and instead let the caller pass in the read/write
1088 opcode = READ_ONCE(sqe->opcode);
1089 if (opcode == IORING_OP_READ_FIXED ||
1090 opcode == IORING_OP_WRITE_FIXED) {
1091 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1099 #ifdef CONFIG_COMPAT
1101 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1105 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1109 * Make a note of the last file/offset/direction we punted to async
1110 * context. We'll use this information to see if we can piggy back a
1111 * sequential request onto the previous one, if it's still hasn't been
1112 * completed by the async worker.
1114 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1116 struct async_list *async_list = &req->ctx->pending_async[rw];
1117 struct kiocb *kiocb = &req->rw;
1118 struct file *filp = kiocb->ki_filp;
1119 off_t io_end = kiocb->ki_pos + len;
1121 if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) {
1122 unsigned long max_pages;
1124 /* Use 8x RA size as a decent limiter for both reads/writes */
1125 max_pages = filp->f_ra.ra_pages;
1127 max_pages = VM_READAHEAD_PAGES;
1130 /* If max pages are exceeded, reset the state */
1132 if (async_list->io_pages + len <= max_pages) {
1133 req->flags |= REQ_F_SEQ_PREV;
1134 async_list->io_pages += len;
1137 async_list->io_pages = 0;
1141 /* New file? Reset state. */
1142 if (async_list->file != filp) {
1143 async_list->io_pages = 0;
1144 async_list->file = filp;
1146 async_list->io_end = io_end;
1149 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1150 bool force_nonblock)
1152 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1153 struct kiocb *kiocb = &req->rw;
1154 struct iov_iter iter;
1157 ssize_t read_size, ret;
1159 ret = io_prep_rw(req, s, force_nonblock);
1162 file = kiocb->ki_filp;
1164 if (unlikely(!(file->f_mode & FMODE_READ)))
1166 if (unlikely(!file->f_op->read_iter))
1169 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1174 if (req->flags & REQ_F_LINK)
1175 req->result = read_size;
1177 iov_count = iov_iter_count(&iter);
1178 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1182 ret2 = call_read_iter(file, kiocb, &iter);
1184 * In case of a short read, punt to async. This can happen
1185 * if we have data partially cached. Alternatively we can
1186 * return the short read, in which case the application will
1187 * need to issue another SQE and wait for it. That SQE will
1188 * need async punt anyway, so it's more efficient to do it
1191 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1193 /* Catch -EAGAIN return for forced non-blocking submission */
1194 if (!force_nonblock || ret2 != -EAGAIN) {
1195 io_rw_done(kiocb, ret2);
1198 * If ->needs_lock is true, we're already in async
1202 io_async_list_note(READ, req, iov_count);
1210 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1211 bool force_nonblock)
1213 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1214 struct kiocb *kiocb = &req->rw;
1215 struct iov_iter iter;
1220 ret = io_prep_rw(req, s, force_nonblock);
1224 file = kiocb->ki_filp;
1225 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1227 if (unlikely(!file->f_op->write_iter))
1230 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1234 if (req->flags & REQ_F_LINK)
1237 iov_count = iov_iter_count(&iter);
1240 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1241 /* If ->needs_lock is true, we're already in async context. */
1243 io_async_list_note(WRITE, req, iov_count);
1247 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1252 * Open-code file_start_write here to grab freeze protection,
1253 * which will be released by another thread in
1254 * io_complete_rw(). Fool lockdep by telling it the lock got
1255 * released so that it doesn't complain about the held lock when
1256 * we return to userspace.
1258 if (S_ISREG(file_inode(file)->i_mode)) {
1259 __sb_start_write(file_inode(file)->i_sb,
1260 SB_FREEZE_WRITE, true);
1261 __sb_writers_release(file_inode(file)->i_sb,
1264 kiocb->ki_flags |= IOCB_WRITE;
1266 ret2 = call_write_iter(file, kiocb, &iter);
1267 if (!force_nonblock || ret2 != -EAGAIN) {
1268 io_rw_done(kiocb, ret2);
1271 * If ->needs_lock is true, we're already in async
1275 io_async_list_note(WRITE, req, iov_count);
1285 * IORING_OP_NOP just posts a completion event, nothing else.
1287 static int io_nop(struct io_kiocb *req, u64 user_data)
1289 struct io_ring_ctx *ctx = req->ctx;
1292 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1295 io_cqring_add_event(ctx, user_data, err);
1300 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1302 struct io_ring_ctx *ctx = req->ctx;
1307 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1309 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1315 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1316 bool force_nonblock)
1318 loff_t sqe_off = READ_ONCE(sqe->off);
1319 loff_t sqe_len = READ_ONCE(sqe->len);
1320 loff_t end = sqe_off + sqe_len;
1321 unsigned fsync_flags;
1324 fsync_flags = READ_ONCE(sqe->fsync_flags);
1325 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1328 ret = io_prep_fsync(req, sqe);
1332 /* fsync always requires a blocking context */
1336 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1337 end > 0 ? end : LLONG_MAX,
1338 fsync_flags & IORING_FSYNC_DATASYNC);
1340 if (ret < 0 && (req->flags & REQ_F_LINK))
1341 req->flags |= REQ_F_FAIL_LINK;
1342 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1347 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1349 struct io_ring_ctx *ctx = req->ctx;
1355 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1357 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1363 static int io_sync_file_range(struct io_kiocb *req,
1364 const struct io_uring_sqe *sqe,
1365 bool force_nonblock)
1372 ret = io_prep_sfr(req, sqe);
1376 /* sync_file_range always requires a blocking context */
1380 sqe_off = READ_ONCE(sqe->off);
1381 sqe_len = READ_ONCE(sqe->len);
1382 flags = READ_ONCE(sqe->sync_range_flags);
1384 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1386 if (ret < 0 && (req->flags & REQ_F_LINK))
1387 req->flags |= REQ_F_FAIL_LINK;
1388 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1393 #if defined(CONFIG_NET)
1394 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1395 bool force_nonblock,
1396 long (*fn)(struct socket *, struct user_msghdr __user *,
1399 struct socket *sock;
1402 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1405 sock = sock_from_file(req->file, &ret);
1407 struct user_msghdr __user *msg;
1410 flags = READ_ONCE(sqe->msg_flags);
1411 if (flags & MSG_DONTWAIT)
1412 req->flags |= REQ_F_NOWAIT;
1413 else if (force_nonblock)
1414 flags |= MSG_DONTWAIT;
1416 msg = (struct user_msghdr __user *) (unsigned long)
1417 READ_ONCE(sqe->addr);
1419 ret = fn(sock, msg, flags);
1420 if (force_nonblock && ret == -EAGAIN)
1424 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1430 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1431 bool force_nonblock)
1433 #if defined(CONFIG_NET)
1434 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1440 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1441 bool force_nonblock)
1443 #if defined(CONFIG_NET)
1444 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1450 static void io_poll_remove_one(struct io_kiocb *req)
1452 struct io_poll_iocb *poll = &req->poll;
1454 spin_lock(&poll->head->lock);
1455 WRITE_ONCE(poll->canceled, true);
1456 if (!list_empty(&poll->wait.entry)) {
1457 list_del_init(&poll->wait.entry);
1458 queue_work(req->ctx->sqo_wq, &req->work);
1460 spin_unlock(&poll->head->lock);
1462 list_del_init(&req->list);
1465 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1467 struct io_kiocb *req;
1469 spin_lock_irq(&ctx->completion_lock);
1470 while (!list_empty(&ctx->cancel_list)) {
1471 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1472 io_poll_remove_one(req);
1474 spin_unlock_irq(&ctx->completion_lock);
1478 * Find a running poll command that matches one specified in sqe->addr,
1479 * and remove it if found.
1481 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1483 struct io_ring_ctx *ctx = req->ctx;
1484 struct io_kiocb *poll_req, *next;
1487 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1489 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1493 spin_lock_irq(&ctx->completion_lock);
1494 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1495 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1496 io_poll_remove_one(poll_req);
1501 spin_unlock_irq(&ctx->completion_lock);
1503 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1508 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1511 req->poll.done = true;
1512 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1513 io_commit_cqring(ctx);
1516 static void io_poll_complete_work(struct work_struct *work)
1518 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1519 struct io_poll_iocb *poll = &req->poll;
1520 struct poll_table_struct pt = { ._key = poll->events };
1521 struct io_ring_ctx *ctx = req->ctx;
1524 if (!READ_ONCE(poll->canceled))
1525 mask = vfs_poll(poll->file, &pt) & poll->events;
1528 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1529 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1530 * synchronize with them. In the cancellation case the list_del_init
1531 * itself is not actually needed, but harmless so we keep it in to
1532 * avoid further branches in the fast path.
1534 spin_lock_irq(&ctx->completion_lock);
1535 if (!mask && !READ_ONCE(poll->canceled)) {
1536 add_wait_queue(poll->head, &poll->wait);
1537 spin_unlock_irq(&ctx->completion_lock);
1540 list_del_init(&req->list);
1541 io_poll_complete(ctx, req, mask);
1542 spin_unlock_irq(&ctx->completion_lock);
1544 io_cqring_ev_posted(ctx);
1548 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1551 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1553 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1554 struct io_ring_ctx *ctx = req->ctx;
1555 __poll_t mask = key_to_poll(key);
1556 unsigned long flags;
1558 /* for instances that support it check for an event match first: */
1559 if (mask && !(mask & poll->events))
1562 list_del_init(&poll->wait.entry);
1564 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1565 list_del(&req->list);
1566 io_poll_complete(ctx, req, mask);
1567 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1569 io_cqring_ev_posted(ctx);
1572 queue_work(ctx->sqo_wq, &req->work);
1578 struct io_poll_table {
1579 struct poll_table_struct pt;
1580 struct io_kiocb *req;
1584 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1585 struct poll_table_struct *p)
1587 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1589 if (unlikely(pt->req->poll.head)) {
1590 pt->error = -EINVAL;
1595 pt->req->poll.head = head;
1596 add_wait_queue(head, &pt->req->poll.wait);
1599 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1601 struct io_poll_iocb *poll = &req->poll;
1602 struct io_ring_ctx *ctx = req->ctx;
1603 struct io_poll_table ipt;
1604 bool cancel = false;
1608 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1610 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1615 INIT_WORK(&req->work, io_poll_complete_work);
1616 events = READ_ONCE(sqe->poll_events);
1617 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1621 poll->canceled = false;
1623 ipt.pt._qproc = io_poll_queue_proc;
1624 ipt.pt._key = poll->events;
1626 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1628 /* initialized the list so that we can do list_empty checks */
1629 INIT_LIST_HEAD(&poll->wait.entry);
1630 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1632 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1634 spin_lock_irq(&ctx->completion_lock);
1635 if (likely(poll->head)) {
1636 spin_lock(&poll->head->lock);
1637 if (unlikely(list_empty(&poll->wait.entry))) {
1643 if (mask || ipt.error)
1644 list_del_init(&poll->wait.entry);
1646 WRITE_ONCE(poll->canceled, true);
1647 else if (!poll->done) /* actually waiting for an event */
1648 list_add_tail(&req->list, &ctx->cancel_list);
1649 spin_unlock(&poll->head->lock);
1651 if (mask) { /* no async, we'd stolen it */
1653 io_poll_complete(ctx, req, mask);
1655 spin_unlock_irq(&ctx->completion_lock);
1658 io_cqring_ev_posted(ctx);
1664 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1665 const struct io_uring_sqe *sqe)
1667 struct io_uring_sqe *sqe_copy;
1669 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1672 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1676 spin_lock_irq(&ctx->completion_lock);
1677 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1678 spin_unlock_irq(&ctx->completion_lock);
1683 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1684 req->submit.sqe = sqe_copy;
1686 INIT_WORK(&req->work, io_sq_wq_submit_work);
1687 list_add_tail(&req->list, &ctx->defer_list);
1688 spin_unlock_irq(&ctx->completion_lock);
1689 return -EIOCBQUEUED;
1692 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1693 const struct sqe_submit *s, bool force_nonblock)
1697 req->user_data = READ_ONCE(s->sqe->user_data);
1699 if (unlikely(s->index >= ctx->sq_entries))
1702 opcode = READ_ONCE(s->sqe->opcode);
1705 ret = io_nop(req, req->user_data);
1707 case IORING_OP_READV:
1708 if (unlikely(s->sqe->buf_index))
1710 ret = io_read(req, s, force_nonblock);
1712 case IORING_OP_WRITEV:
1713 if (unlikely(s->sqe->buf_index))
1715 ret = io_write(req, s, force_nonblock);
1717 case IORING_OP_READ_FIXED:
1718 ret = io_read(req, s, force_nonblock);
1720 case IORING_OP_WRITE_FIXED:
1721 ret = io_write(req, s, force_nonblock);
1723 case IORING_OP_FSYNC:
1724 ret = io_fsync(req, s->sqe, force_nonblock);
1726 case IORING_OP_POLL_ADD:
1727 ret = io_poll_add(req, s->sqe);
1729 case IORING_OP_POLL_REMOVE:
1730 ret = io_poll_remove(req, s->sqe);
1732 case IORING_OP_SYNC_FILE_RANGE:
1733 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1735 case IORING_OP_SENDMSG:
1736 ret = io_sendmsg(req, s->sqe, force_nonblock);
1738 case IORING_OP_RECVMSG:
1739 ret = io_recvmsg(req, s->sqe, force_nonblock);
1749 if (ctx->flags & IORING_SETUP_IOPOLL) {
1750 if (req->result == -EAGAIN)
1753 /* workqueue context doesn't hold uring_lock, grab it now */
1755 mutex_lock(&ctx->uring_lock);
1756 io_iopoll_req_issued(req);
1758 mutex_unlock(&ctx->uring_lock);
1764 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1765 const struct io_uring_sqe *sqe)
1767 switch (sqe->opcode) {
1768 case IORING_OP_READV:
1769 case IORING_OP_READ_FIXED:
1770 return &ctx->pending_async[READ];
1771 case IORING_OP_WRITEV:
1772 case IORING_OP_WRITE_FIXED:
1773 return &ctx->pending_async[WRITE];
1779 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1781 u8 opcode = READ_ONCE(sqe->opcode);
1783 return !(opcode == IORING_OP_READ_FIXED ||
1784 opcode == IORING_OP_WRITE_FIXED);
1787 static void io_sq_wq_submit_work(struct work_struct *work)
1789 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1790 struct io_ring_ctx *ctx = req->ctx;
1791 struct mm_struct *cur_mm = NULL;
1792 struct async_list *async_list;
1793 LIST_HEAD(req_list);
1794 mm_segment_t old_fs;
1797 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1800 struct sqe_submit *s = &req->submit;
1801 const struct io_uring_sqe *sqe = s->sqe;
1803 /* Ensure we clear previously set non-block flag */
1804 req->rw.ki_flags &= ~IOCB_NOWAIT;
1807 if (io_sqe_needs_user(sqe) && !cur_mm) {
1808 if (!mmget_not_zero(ctx->sqo_mm)) {
1811 cur_mm = ctx->sqo_mm;
1819 s->has_user = cur_mm != NULL;
1820 s->needs_lock = true;
1822 ret = __io_submit_sqe(ctx, req, s, false);
1824 * We can get EAGAIN for polled IO even though
1825 * we're forcing a sync submission from here,
1826 * since we can't wait for request slots on the
1835 /* drop submission reference */
1839 io_cqring_add_event(ctx, sqe->user_data, ret);
1843 /* async context always use a copy of the sqe */
1848 if (!list_empty(&req_list)) {
1849 req = list_first_entry(&req_list, struct io_kiocb,
1851 list_del(&req->list);
1854 if (list_empty(&async_list->list))
1858 spin_lock(&async_list->lock);
1859 if (list_empty(&async_list->list)) {
1860 spin_unlock(&async_list->lock);
1863 list_splice_init(&async_list->list, &req_list);
1864 spin_unlock(&async_list->lock);
1866 req = list_first_entry(&req_list, struct io_kiocb, list);
1867 list_del(&req->list);
1871 * Rare case of racing with a submitter. If we find the count has
1872 * dropped to zero AND we have pending work items, then restart
1873 * the processing. This is a tiny race window.
1876 ret = atomic_dec_return(&async_list->cnt);
1877 while (!ret && !list_empty(&async_list->list)) {
1878 spin_lock(&async_list->lock);
1879 atomic_inc(&async_list->cnt);
1880 list_splice_init(&async_list->list, &req_list);
1881 spin_unlock(&async_list->lock);
1883 if (!list_empty(&req_list)) {
1884 req = list_first_entry(&req_list,
1885 struct io_kiocb, list);
1886 list_del(&req->list);
1889 ret = atomic_dec_return(&async_list->cnt);
1901 * See if we can piggy back onto previously submitted work, that is still
1902 * running. We currently only allow this if the new request is sequential
1903 * to the previous one we punted.
1905 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
1911 if (!(req->flags & REQ_F_SEQ_PREV))
1913 if (!atomic_read(&list->cnt))
1917 spin_lock(&list->lock);
1918 list_add_tail(&req->list, &list->list);
1919 if (!atomic_read(&list->cnt)) {
1920 list_del_init(&req->list);
1923 spin_unlock(&list->lock);
1927 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
1929 int op = READ_ONCE(sqe->opcode);
1933 case IORING_OP_POLL_REMOVE:
1940 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
1941 struct io_submit_state *state, struct io_kiocb *req)
1946 flags = READ_ONCE(s->sqe->flags);
1947 fd = READ_ONCE(s->sqe->fd);
1949 if (flags & IOSQE_IO_DRAIN) {
1950 req->flags |= REQ_F_IO_DRAIN;
1951 req->sequence = ctx->cached_sq_head - 1;
1954 if (!io_op_needs_file(s->sqe))
1957 if (flags & IOSQE_FIXED_FILE) {
1958 if (unlikely(!ctx->user_files ||
1959 (unsigned) fd >= ctx->nr_user_files))
1961 req->file = ctx->user_files[fd];
1962 req->flags |= REQ_F_FIXED_FILE;
1964 if (s->needs_fixed_file)
1966 req->file = io_file_get(state, fd);
1967 if (unlikely(!req->file))
1974 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1975 struct sqe_submit *s)
1979 ret = __io_submit_sqe(ctx, req, s, true);
1980 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
1981 struct io_uring_sqe *sqe_copy;
1983 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1985 struct async_list *list;
1987 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
1990 memcpy(&req->submit, s, sizeof(*s));
1991 list = io_async_list_from_sqe(ctx, s->sqe);
1992 if (!io_add_to_prev_work(list, req)) {
1994 atomic_inc(&list->cnt);
1995 INIT_WORK(&req->work, io_sq_wq_submit_work);
1996 queue_work(ctx->sqo_wq, &req->work);
2000 * Queued up for async execution, worker will release
2001 * submit reference when the iocb is actually submitted.
2007 /* drop submission reference */
2010 /* and drop final reference, if we failed */
2012 io_cqring_add_event(ctx, req->user_data, ret);
2013 if (req->flags & REQ_F_LINK)
2014 req->flags |= REQ_F_FAIL_LINK;
2021 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2023 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2024 struct io_submit_state *state, struct io_kiocb **link)
2026 struct io_uring_sqe *sqe_copy;
2027 struct io_kiocb *req;
2030 /* enforce forwards compatibility on users */
2031 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2036 req = io_get_req(ctx, state);
2037 if (unlikely(!req)) {
2042 ret = io_req_set_file(ctx, s, state, req);
2043 if (unlikely(ret)) {
2047 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2051 ret = io_req_defer(ctx, req, s->sqe);
2053 if (ret != -EIOCBQUEUED)
2059 * If we already have a head request, queue this one for async
2060 * submittal once the head completes. If we don't have a head but
2061 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2062 * submitted sync once the chain is complete. If none of those
2063 * conditions are true (normal request), then just queue it.
2066 struct io_kiocb *prev = *link;
2068 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2075 memcpy(&req->submit, s, sizeof(*s));
2076 list_add_tail(&req->list, &prev->link_list);
2077 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2078 req->flags |= REQ_F_LINK;
2080 memcpy(&req->submit, s, sizeof(*s));
2081 INIT_LIST_HEAD(&req->link_list);
2084 io_queue_sqe(ctx, req, s);
2089 * Batched submission is done, ensure local IO is flushed out.
2091 static void io_submit_state_end(struct io_submit_state *state)
2093 blk_finish_plug(&state->plug);
2095 if (state->free_reqs)
2096 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2097 &state->reqs[state->cur_req]);
2101 * Start submission side cache.
2103 static void io_submit_state_start(struct io_submit_state *state,
2104 struct io_ring_ctx *ctx, unsigned max_ios)
2106 blk_start_plug(&state->plug);
2107 state->free_reqs = 0;
2109 state->ios_left = max_ios;
2112 static void io_commit_sqring(struct io_ring_ctx *ctx)
2114 struct io_sq_ring *ring = ctx->sq_ring;
2116 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
2118 * Ensure any loads from the SQEs are done at this point,
2119 * since once we write the new head, the application could
2120 * write new data to them.
2122 smp_store_release(&ring->r.head, ctx->cached_sq_head);
2127 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2128 * that is mapped by userspace. This means that care needs to be taken to
2129 * ensure that reads are stable, as we cannot rely on userspace always
2130 * being a good citizen. If members of the sqe are validated and then later
2131 * used, it's important that those reads are done through READ_ONCE() to
2132 * prevent a re-load down the line.
2134 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2136 struct io_sq_ring *ring = ctx->sq_ring;
2140 * The cached sq head (or cq tail) serves two purposes:
2142 * 1) allows us to batch the cost of updating the user visible
2144 * 2) allows the kernel side to track the head on its own, even
2145 * though the application is the one updating it.
2147 head = ctx->cached_sq_head;
2148 /* make sure SQ entry isn't read before tail */
2149 if (head == smp_load_acquire(&ring->r.tail))
2152 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
2153 if (head < ctx->sq_entries) {
2155 s->sqe = &ctx->sq_sqes[head];
2156 ctx->cached_sq_head++;
2160 /* drop invalid entries */
2161 ctx->cached_sq_head++;
2166 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2167 unsigned int nr, bool has_user, bool mm_fault)
2169 struct io_submit_state state, *statep = NULL;
2170 struct io_kiocb *link = NULL;
2171 bool prev_was_link = false;
2172 int i, submitted = 0;
2174 if (nr > IO_PLUG_THRESHOLD) {
2175 io_submit_state_start(&state, ctx, nr);
2179 for (i = 0; i < nr; i++) {
2181 * If previous wasn't linked and we have a linked command,
2182 * that's the end of the chain. Submit the previous link.
2184 if (!prev_was_link && link) {
2185 io_queue_sqe(ctx, link, &link->submit);
2188 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2190 if (unlikely(mm_fault)) {
2191 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2194 sqes[i].has_user = has_user;
2195 sqes[i].needs_lock = true;
2196 sqes[i].needs_fixed_file = true;
2197 io_submit_sqe(ctx, &sqes[i], statep, &link);
2203 io_queue_sqe(ctx, link, &link->submit);
2205 io_submit_state_end(&state);
2210 static int io_sq_thread(void *data)
2212 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2213 struct io_ring_ctx *ctx = data;
2214 struct mm_struct *cur_mm = NULL;
2215 mm_segment_t old_fs;
2218 unsigned long timeout;
2220 complete(&ctx->sqo_thread_started);
2225 timeout = inflight = 0;
2226 while (!kthread_should_park()) {
2227 bool all_fixed, mm_fault = false;
2231 unsigned nr_events = 0;
2233 if (ctx->flags & IORING_SETUP_IOPOLL) {
2235 * We disallow the app entering submit/complete
2236 * with polling, but we still need to lock the
2237 * ring to prevent racing with polled issue
2238 * that got punted to a workqueue.
2240 mutex_lock(&ctx->uring_lock);
2241 io_iopoll_check(ctx, &nr_events, 0);
2242 mutex_unlock(&ctx->uring_lock);
2245 * Normal IO, just pretend everything completed.
2246 * We don't have to poll completions for that.
2248 nr_events = inflight;
2251 inflight -= nr_events;
2253 timeout = jiffies + ctx->sq_thread_idle;
2256 if (!io_get_sqring(ctx, &sqes[0])) {
2258 * We're polling. If we're within the defined idle
2259 * period, then let us spin without work before going
2262 if (inflight || !time_after(jiffies, timeout)) {
2268 * Drop cur_mm before scheduling, we can't hold it for
2269 * long periods (or over schedule()). Do this before
2270 * adding ourselves to the waitqueue, as the unuse/drop
2279 prepare_to_wait(&ctx->sqo_wait, &wait,
2280 TASK_INTERRUPTIBLE);
2282 /* Tell userspace we may need a wakeup call */
2283 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
2284 /* make sure to read SQ tail after writing flags */
2287 if (!io_get_sqring(ctx, &sqes[0])) {
2288 if (kthread_should_park()) {
2289 finish_wait(&ctx->sqo_wait, &wait);
2292 if (signal_pending(current))
2293 flush_signals(current);
2295 finish_wait(&ctx->sqo_wait, &wait);
2297 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2300 finish_wait(&ctx->sqo_wait, &wait);
2302 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2308 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2312 if (i == ARRAY_SIZE(sqes))
2314 } while (io_get_sqring(ctx, &sqes[i]));
2316 /* Unless all new commands are FIXED regions, grab mm */
2317 if (!all_fixed && !cur_mm) {
2318 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2320 use_mm(ctx->sqo_mm);
2321 cur_mm = ctx->sqo_mm;
2325 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2328 /* Commit SQ ring head once we've consumed all SQEs */
2329 io_commit_sqring(ctx);
2343 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2345 struct io_submit_state state, *statep = NULL;
2346 struct io_kiocb *link = NULL;
2347 bool prev_was_link = false;
2350 if (to_submit > IO_PLUG_THRESHOLD) {
2351 io_submit_state_start(&state, ctx, to_submit);
2355 for (i = 0; i < to_submit; i++) {
2356 struct sqe_submit s;
2358 if (!io_get_sqring(ctx, &s))
2362 * If previous wasn't linked and we have a linked command,
2363 * that's the end of the chain. Submit the previous link.
2365 if (!prev_was_link && link) {
2366 io_queue_sqe(ctx, link, &link->submit);
2369 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2372 s.needs_lock = false;
2373 s.needs_fixed_file = false;
2375 io_submit_sqe(ctx, &s, statep, &link);
2377 io_commit_sqring(ctx);
2380 io_queue_sqe(ctx, link, &link->submit);
2382 io_submit_state_end(statep);
2387 static unsigned io_cqring_events(struct io_cq_ring *ring)
2389 /* See comment at the top of this file */
2391 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
2395 * Wait until events become available, if we don't already have some. The
2396 * application must reap them itself, as they reside on the shared cq ring.
2398 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2399 const sigset_t __user *sig, size_t sigsz)
2401 struct io_cq_ring *ring = ctx->cq_ring;
2402 sigset_t ksigmask, sigsaved;
2405 if (io_cqring_events(ring) >= min_events)
2409 #ifdef CONFIG_COMPAT
2410 if (in_compat_syscall())
2411 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2412 &ksigmask, &sigsaved, sigsz);
2415 ret = set_user_sigmask(sig, &ksigmask,
2422 ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events);
2425 restore_user_sigmask(sig, &sigsaved, ret == -ERESTARTSYS);
2427 if (ret == -ERESTARTSYS)
2430 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
2433 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2435 #if defined(CONFIG_UNIX)
2436 if (ctx->ring_sock) {
2437 struct sock *sock = ctx->ring_sock->sk;
2438 struct sk_buff *skb;
2440 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2446 for (i = 0; i < ctx->nr_user_files; i++)
2447 fput(ctx->user_files[i]);
2451 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2453 if (!ctx->user_files)
2456 __io_sqe_files_unregister(ctx);
2457 kfree(ctx->user_files);
2458 ctx->user_files = NULL;
2459 ctx->nr_user_files = 0;
2463 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2465 if (ctx->sqo_thread) {
2466 wait_for_completion(&ctx->sqo_thread_started);
2468 * The park is a bit of a work-around, without it we get
2469 * warning spews on shutdown with SQPOLL set and affinity
2470 * set to a single CPU.
2472 kthread_park(ctx->sqo_thread);
2473 kthread_stop(ctx->sqo_thread);
2474 ctx->sqo_thread = NULL;
2478 static void io_finish_async(struct io_ring_ctx *ctx)
2480 io_sq_thread_stop(ctx);
2483 destroy_workqueue(ctx->sqo_wq);
2488 #if defined(CONFIG_UNIX)
2489 static void io_destruct_skb(struct sk_buff *skb)
2491 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2493 io_finish_async(ctx);
2494 unix_destruct_scm(skb);
2498 * Ensure the UNIX gc is aware of our file set, so we are certain that
2499 * the io_uring can be safely unregistered on process exit, even if we have
2500 * loops in the file referencing.
2502 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2504 struct sock *sk = ctx->ring_sock->sk;
2505 struct scm_fp_list *fpl;
2506 struct sk_buff *skb;
2509 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2510 unsigned long inflight = ctx->user->unix_inflight + nr;
2512 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2516 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2520 skb = alloc_skb(0, GFP_KERNEL);
2527 skb->destructor = io_destruct_skb;
2529 fpl->user = get_uid(ctx->user);
2530 for (i = 0; i < nr; i++) {
2531 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2532 unix_inflight(fpl->user, fpl->fp[i]);
2535 fpl->max = fpl->count = nr;
2536 UNIXCB(skb).fp = fpl;
2537 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2538 skb_queue_head(&sk->sk_receive_queue, skb);
2540 for (i = 0; i < nr; i++)
2547 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2548 * causes regular reference counting to break down. We rely on the UNIX
2549 * garbage collection to take care of this problem for us.
2551 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2553 unsigned left, total;
2557 left = ctx->nr_user_files;
2559 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2561 ret = __io_sqe_files_scm(ctx, this_files, total);
2565 total += this_files;
2571 while (total < ctx->nr_user_files) {
2572 fput(ctx->user_files[total]);
2579 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2585 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2588 __s32 __user *fds = (__s32 __user *) arg;
2592 if (ctx->user_files)
2596 if (nr_args > IORING_MAX_FIXED_FILES)
2599 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2600 if (!ctx->user_files)
2603 for (i = 0; i < nr_args; i++) {
2605 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2608 ctx->user_files[i] = fget(fd);
2611 if (!ctx->user_files[i])
2614 * Don't allow io_uring instances to be registered. If UNIX
2615 * isn't enabled, then this causes a reference cycle and this
2616 * instance can never get freed. If UNIX is enabled we'll
2617 * handle it just fine, but there's still no point in allowing
2618 * a ring fd as it doesn't support regular read/write anyway.
2620 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2621 fput(ctx->user_files[i]);
2624 ctx->nr_user_files++;
2629 for (i = 0; i < ctx->nr_user_files; i++)
2630 fput(ctx->user_files[i]);
2632 kfree(ctx->user_files);
2633 ctx->user_files = NULL;
2634 ctx->nr_user_files = 0;
2638 ret = io_sqe_files_scm(ctx);
2640 io_sqe_files_unregister(ctx);
2645 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2646 struct io_uring_params *p)
2650 init_waitqueue_head(&ctx->sqo_wait);
2651 mmgrab(current->mm);
2652 ctx->sqo_mm = current->mm;
2654 if (ctx->flags & IORING_SETUP_SQPOLL) {
2656 if (!capable(CAP_SYS_ADMIN))
2659 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2660 if (!ctx->sq_thread_idle)
2661 ctx->sq_thread_idle = HZ;
2663 if (p->flags & IORING_SETUP_SQ_AFF) {
2664 int cpu = p->sq_thread_cpu;
2667 if (cpu >= nr_cpu_ids)
2669 if (!cpu_online(cpu))
2672 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2676 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2679 if (IS_ERR(ctx->sqo_thread)) {
2680 ret = PTR_ERR(ctx->sqo_thread);
2681 ctx->sqo_thread = NULL;
2684 wake_up_process(ctx->sqo_thread);
2685 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2686 /* Can't have SQ_AFF without SQPOLL */
2691 /* Do QD, or 2 * CPUS, whatever is smallest */
2692 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
2693 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2701 io_sq_thread_stop(ctx);
2702 mmdrop(ctx->sqo_mm);
2707 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2709 atomic_long_sub(nr_pages, &user->locked_vm);
2712 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2714 unsigned long page_limit, cur_pages, new_pages;
2716 /* Don't allow more pages than we can safely lock */
2717 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2720 cur_pages = atomic_long_read(&user->locked_vm);
2721 new_pages = cur_pages + nr_pages;
2722 if (new_pages > page_limit)
2724 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2725 new_pages) != cur_pages);
2730 static void io_mem_free(void *ptr)
2737 page = virt_to_head_page(ptr);
2738 if (put_page_testzero(page))
2739 free_compound_page(page);
2742 static void *io_mem_alloc(size_t size)
2744 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2747 return (void *) __get_free_pages(gfp_flags, get_order(size));
2750 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2752 struct io_sq_ring *sq_ring;
2753 struct io_cq_ring *cq_ring;
2756 bytes = struct_size(sq_ring, array, sq_entries);
2757 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
2758 bytes += struct_size(cq_ring, cqes, cq_entries);
2760 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
2763 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
2767 if (!ctx->user_bufs)
2770 for (i = 0; i < ctx->nr_user_bufs; i++) {
2771 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2773 for (j = 0; j < imu->nr_bvecs; j++)
2774 put_page(imu->bvec[j].bv_page);
2776 if (ctx->account_mem)
2777 io_unaccount_mem(ctx->user, imu->nr_bvecs);
2782 kfree(ctx->user_bufs);
2783 ctx->user_bufs = NULL;
2784 ctx->nr_user_bufs = 0;
2788 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
2789 void __user *arg, unsigned index)
2791 struct iovec __user *src;
2793 #ifdef CONFIG_COMPAT
2795 struct compat_iovec __user *ciovs;
2796 struct compat_iovec ciov;
2798 ciovs = (struct compat_iovec __user *) arg;
2799 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
2802 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
2803 dst->iov_len = ciov.iov_len;
2807 src = (struct iovec __user *) arg;
2808 if (copy_from_user(dst, &src[index], sizeof(*dst)))
2813 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
2816 struct vm_area_struct **vmas = NULL;
2817 struct page **pages = NULL;
2818 int i, j, got_pages = 0;
2823 if (!nr_args || nr_args > UIO_MAXIOV)
2826 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
2828 if (!ctx->user_bufs)
2831 for (i = 0; i < nr_args; i++) {
2832 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2833 unsigned long off, start, end, ubuf;
2838 ret = io_copy_iov(ctx, &iov, arg, i);
2843 * Don't impose further limits on the size and buffer
2844 * constraints here, we'll -EINVAL later when IO is
2845 * submitted if they are wrong.
2848 if (!iov.iov_base || !iov.iov_len)
2851 /* arbitrary limit, but we need something */
2852 if (iov.iov_len > SZ_1G)
2855 ubuf = (unsigned long) iov.iov_base;
2856 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2857 start = ubuf >> PAGE_SHIFT;
2858 nr_pages = end - start;
2860 if (ctx->account_mem) {
2861 ret = io_account_mem(ctx->user, nr_pages);
2867 if (!pages || nr_pages > got_pages) {
2870 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
2872 vmas = kvmalloc_array(nr_pages,
2873 sizeof(struct vm_area_struct *),
2875 if (!pages || !vmas) {
2877 if (ctx->account_mem)
2878 io_unaccount_mem(ctx->user, nr_pages);
2881 got_pages = nr_pages;
2884 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
2888 if (ctx->account_mem)
2889 io_unaccount_mem(ctx->user, nr_pages);
2894 down_read(¤t->mm->mmap_sem);
2895 pret = get_user_pages(ubuf, nr_pages,
2896 FOLL_WRITE | FOLL_LONGTERM,
2898 if (pret == nr_pages) {
2899 /* don't support file backed memory */
2900 for (j = 0; j < nr_pages; j++) {
2901 struct vm_area_struct *vma = vmas[j];
2904 !is_file_hugepages(vma->vm_file)) {
2910 ret = pret < 0 ? pret : -EFAULT;
2912 up_read(¤t->mm->mmap_sem);
2915 * if we did partial map, or found file backed vmas,
2916 * release any pages we did get
2919 for (j = 0; j < pret; j++)
2922 if (ctx->account_mem)
2923 io_unaccount_mem(ctx->user, nr_pages);
2928 off = ubuf & ~PAGE_MASK;
2930 for (j = 0; j < nr_pages; j++) {
2933 vec_len = min_t(size_t, size, PAGE_SIZE - off);
2934 imu->bvec[j].bv_page = pages[j];
2935 imu->bvec[j].bv_len = vec_len;
2936 imu->bvec[j].bv_offset = off;
2940 /* store original address for later verification */
2942 imu->len = iov.iov_len;
2943 imu->nr_bvecs = nr_pages;
2945 ctx->nr_user_bufs++;
2953 io_sqe_buffer_unregister(ctx);
2957 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
2959 __s32 __user *fds = arg;
2965 if (copy_from_user(&fd, fds, sizeof(*fds)))
2968 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
2969 if (IS_ERR(ctx->cq_ev_fd)) {
2970 int ret = PTR_ERR(ctx->cq_ev_fd);
2971 ctx->cq_ev_fd = NULL;
2978 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2980 if (ctx->cq_ev_fd) {
2981 eventfd_ctx_put(ctx->cq_ev_fd);
2982 ctx->cq_ev_fd = NULL;
2989 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
2991 io_finish_async(ctx);
2993 mmdrop(ctx->sqo_mm);
2995 io_iopoll_reap_events(ctx);
2996 io_sqe_buffer_unregister(ctx);
2997 io_sqe_files_unregister(ctx);
2998 io_eventfd_unregister(ctx);
3000 #if defined(CONFIG_UNIX)
3001 if (ctx->ring_sock) {
3002 ctx->ring_sock->file = NULL; /* so that iput() is called */
3003 sock_release(ctx->ring_sock);
3007 io_mem_free(ctx->sq_ring);
3008 io_mem_free(ctx->sq_sqes);
3009 io_mem_free(ctx->cq_ring);
3011 percpu_ref_exit(&ctx->refs);
3012 if (ctx->account_mem)
3013 io_unaccount_mem(ctx->user,
3014 ring_pages(ctx->sq_entries, ctx->cq_entries));
3015 free_uid(ctx->user);
3019 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3021 struct io_ring_ctx *ctx = file->private_data;
3024 poll_wait(file, &ctx->cq_wait, wait);
3026 * synchronizes with barrier from wq_has_sleeper call in
3030 if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
3031 ctx->sq_ring->ring_entries)
3032 mask |= EPOLLOUT | EPOLLWRNORM;
3033 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
3034 mask |= EPOLLIN | EPOLLRDNORM;
3039 static int io_uring_fasync(int fd, struct file *file, int on)
3041 struct io_ring_ctx *ctx = file->private_data;
3043 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3046 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3048 mutex_lock(&ctx->uring_lock);
3049 percpu_ref_kill(&ctx->refs);
3050 mutex_unlock(&ctx->uring_lock);
3052 io_poll_remove_all(ctx);
3053 io_iopoll_reap_events(ctx);
3054 wait_for_completion(&ctx->ctx_done);
3055 io_ring_ctx_free(ctx);
3058 static int io_uring_release(struct inode *inode, struct file *file)
3060 struct io_ring_ctx *ctx = file->private_data;
3062 file->private_data = NULL;
3063 io_ring_ctx_wait_and_kill(ctx);
3067 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3069 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3070 unsigned long sz = vma->vm_end - vma->vm_start;
3071 struct io_ring_ctx *ctx = file->private_data;
3077 case IORING_OFF_SQ_RING:
3080 case IORING_OFF_SQES:
3083 case IORING_OFF_CQ_RING:
3090 page = virt_to_head_page(ptr);
3091 if (sz > (PAGE_SIZE << compound_order(page)))
3094 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3095 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3098 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3099 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3102 struct io_ring_ctx *ctx;
3107 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3115 if (f.file->f_op != &io_uring_fops)
3119 ctx = f.file->private_data;
3120 if (!percpu_ref_tryget(&ctx->refs))
3124 * For SQ polling, the thread will do all submissions and completions.
3125 * Just return the requested submit count, and wake the thread if
3128 if (ctx->flags & IORING_SETUP_SQPOLL) {
3129 if (flags & IORING_ENTER_SQ_WAKEUP)
3130 wake_up(&ctx->sqo_wait);
3131 submitted = to_submit;
3137 to_submit = min(to_submit, ctx->sq_entries);
3139 mutex_lock(&ctx->uring_lock);
3140 submitted = io_ring_submit(ctx, to_submit);
3141 mutex_unlock(&ctx->uring_lock);
3143 if (flags & IORING_ENTER_GETEVENTS) {
3144 unsigned nr_events = 0;
3146 min_complete = min(min_complete, ctx->cq_entries);
3148 if (ctx->flags & IORING_SETUP_IOPOLL) {
3149 mutex_lock(&ctx->uring_lock);
3150 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3151 mutex_unlock(&ctx->uring_lock);
3153 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3158 io_ring_drop_ctx_refs(ctx, 1);
3161 return submitted ? submitted : ret;
3164 static const struct file_operations io_uring_fops = {
3165 .release = io_uring_release,
3166 .mmap = io_uring_mmap,
3167 .poll = io_uring_poll,
3168 .fasync = io_uring_fasync,
3171 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3172 struct io_uring_params *p)
3174 struct io_sq_ring *sq_ring;
3175 struct io_cq_ring *cq_ring;
3178 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
3182 ctx->sq_ring = sq_ring;
3183 sq_ring->ring_mask = p->sq_entries - 1;
3184 sq_ring->ring_entries = p->sq_entries;
3185 ctx->sq_mask = sq_ring->ring_mask;
3186 ctx->sq_entries = sq_ring->ring_entries;
3188 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3189 if (size == SIZE_MAX)
3192 ctx->sq_sqes = io_mem_alloc(size);
3196 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
3200 ctx->cq_ring = cq_ring;
3201 cq_ring->ring_mask = p->cq_entries - 1;
3202 cq_ring->ring_entries = p->cq_entries;
3203 ctx->cq_mask = cq_ring->ring_mask;
3204 ctx->cq_entries = cq_ring->ring_entries;
3209 * Allocate an anonymous fd, this is what constitutes the application
3210 * visible backing of an io_uring instance. The application mmaps this
3211 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3212 * we have to tie this fd to a socket for file garbage collection purposes.
3214 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3219 #if defined(CONFIG_UNIX)
3220 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3226 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3230 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3231 O_RDWR | O_CLOEXEC);
3234 ret = PTR_ERR(file);
3238 #if defined(CONFIG_UNIX)
3239 ctx->ring_sock->file = file;
3240 ctx->ring_sock->sk->sk_user_data = ctx;
3242 fd_install(ret, file);
3245 #if defined(CONFIG_UNIX)
3246 sock_release(ctx->ring_sock);
3247 ctx->ring_sock = NULL;
3252 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3254 struct user_struct *user = NULL;
3255 struct io_ring_ctx *ctx;
3259 if (!entries || entries > IORING_MAX_ENTRIES)
3263 * Use twice as many entries for the CQ ring. It's possible for the
3264 * application to drive a higher depth than the size of the SQ ring,
3265 * since the sqes are only used at submission time. This allows for
3266 * some flexibility in overcommitting a bit.
3268 p->sq_entries = roundup_pow_of_two(entries);
3269 p->cq_entries = 2 * p->sq_entries;
3271 user = get_uid(current_user());
3272 account_mem = !capable(CAP_IPC_LOCK);
3275 ret = io_account_mem(user,
3276 ring_pages(p->sq_entries, p->cq_entries));
3283 ctx = io_ring_ctx_alloc(p);
3286 io_unaccount_mem(user, ring_pages(p->sq_entries,
3291 ctx->compat = in_compat_syscall();
3292 ctx->account_mem = account_mem;
3295 ret = io_allocate_scq_urings(ctx, p);
3299 ret = io_sq_offload_start(ctx, p);
3303 ret = io_uring_get_fd(ctx);
3307 memset(&p->sq_off, 0, sizeof(p->sq_off));
3308 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
3309 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
3310 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
3311 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
3312 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
3313 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
3314 p->sq_off.array = offsetof(struct io_sq_ring, array);
3316 memset(&p->cq_off, 0, sizeof(p->cq_off));
3317 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
3318 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
3319 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
3320 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
3321 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
3322 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
3325 io_ring_ctx_wait_and_kill(ctx);
3330 * Sets up an aio uring context, and returns the fd. Applications asks for a
3331 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3332 * params structure passed in.
3334 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3336 struct io_uring_params p;
3340 if (copy_from_user(&p, params, sizeof(p)))
3342 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3347 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3348 IORING_SETUP_SQ_AFF))
3351 ret = io_uring_create(entries, &p);
3355 if (copy_to_user(params, &p, sizeof(p)))
3361 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3362 struct io_uring_params __user *, params)
3364 return io_uring_setup(entries, params);
3367 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3368 void __user *arg, unsigned nr_args)
3369 __releases(ctx->uring_lock)
3370 __acquires(ctx->uring_lock)
3375 * We're inside the ring mutex, if the ref is already dying, then
3376 * someone else killed the ctx or is already going through
3377 * io_uring_register().
3379 if (percpu_ref_is_dying(&ctx->refs))
3382 percpu_ref_kill(&ctx->refs);
3385 * Drop uring mutex before waiting for references to exit. If another
3386 * thread is currently inside io_uring_enter() it might need to grab
3387 * the uring_lock to make progress. If we hold it here across the drain
3388 * wait, then we can deadlock. It's safe to drop the mutex here, since
3389 * no new references will come in after we've killed the percpu ref.
3391 mutex_unlock(&ctx->uring_lock);
3392 wait_for_completion(&ctx->ctx_done);
3393 mutex_lock(&ctx->uring_lock);
3396 case IORING_REGISTER_BUFFERS:
3397 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3399 case IORING_UNREGISTER_BUFFERS:
3403 ret = io_sqe_buffer_unregister(ctx);
3405 case IORING_REGISTER_FILES:
3406 ret = io_sqe_files_register(ctx, arg, nr_args);
3408 case IORING_UNREGISTER_FILES:
3412 ret = io_sqe_files_unregister(ctx);
3414 case IORING_REGISTER_EVENTFD:
3418 ret = io_eventfd_register(ctx, arg);
3420 case IORING_UNREGISTER_EVENTFD:
3424 ret = io_eventfd_unregister(ctx);
3431 /* bring the ctx back to life */
3432 reinit_completion(&ctx->ctx_done);
3433 percpu_ref_reinit(&ctx->refs);
3437 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3438 void __user *, arg, unsigned int, nr_args)
3440 struct io_ring_ctx *ctx;
3449 if (f.file->f_op != &io_uring_fops)
3452 ctx = f.file->private_data;
3454 mutex_lock(&ctx->uring_lock);
3455 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3456 mutex_unlock(&ctx->uring_lock);
3462 static int __init io_uring_init(void)
3464 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3467 __initcall(io_uring_init);