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,
403 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
408 ctx->flags = p->flags;
409 init_waitqueue_head(&ctx->cq_wait);
410 init_completion(&ctx->ctx_done);
411 init_completion(&ctx->sqo_thread_started);
412 mutex_init(&ctx->uring_lock);
413 init_waitqueue_head(&ctx->wait);
414 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
415 spin_lock_init(&ctx->pending_async[i].lock);
416 INIT_LIST_HEAD(&ctx->pending_async[i].list);
417 atomic_set(&ctx->pending_async[i].cnt, 0);
419 spin_lock_init(&ctx->completion_lock);
420 INIT_LIST_HEAD(&ctx->poll_list);
421 INIT_LIST_HEAD(&ctx->cancel_list);
422 INIT_LIST_HEAD(&ctx->defer_list);
426 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
427 struct io_kiocb *req)
429 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
432 return req->sequence > ctx->cached_cq_tail + ctx->sq_ring->dropped;
435 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
437 struct io_kiocb *req;
439 if (list_empty(&ctx->defer_list))
442 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
443 if (!io_sequence_defer(ctx, req)) {
444 list_del_init(&req->list);
451 static void __io_commit_cqring(struct io_ring_ctx *ctx)
453 struct io_cq_ring *ring = ctx->cq_ring;
455 if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) {
456 /* order cqe stores with ring update */
457 smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
459 if (wq_has_sleeper(&ctx->cq_wait)) {
460 wake_up_interruptible(&ctx->cq_wait);
461 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
466 static void io_commit_cqring(struct io_ring_ctx *ctx)
468 struct io_kiocb *req;
470 __io_commit_cqring(ctx);
472 while ((req = io_get_deferred_req(ctx)) != NULL) {
473 req->flags |= REQ_F_IO_DRAINED;
474 queue_work(ctx->sqo_wq, &req->work);
478 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
480 struct io_cq_ring *ring = ctx->cq_ring;
483 tail = ctx->cached_cq_tail;
485 * writes to the cq entry need to come after reading head; the
486 * control dependency is enough as we're using WRITE_ONCE to
489 if (tail - READ_ONCE(ring->r.head) == ring->ring_entries)
492 ctx->cached_cq_tail++;
493 return &ring->cqes[tail & ctx->cq_mask];
496 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
499 struct io_uring_cqe *cqe;
502 * If we can't get a cq entry, userspace overflowed the
503 * submission (by quite a lot). Increment the overflow count in
506 cqe = io_get_cqring(ctx);
508 WRITE_ONCE(cqe->user_data, ki_user_data);
509 WRITE_ONCE(cqe->res, res);
510 WRITE_ONCE(cqe->flags, 0);
512 unsigned overflow = READ_ONCE(ctx->cq_ring->overflow);
514 WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1);
518 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
520 if (waitqueue_active(&ctx->wait))
522 if (waitqueue_active(&ctx->sqo_wait))
523 wake_up(&ctx->sqo_wait);
525 eventfd_signal(ctx->cq_ev_fd, 1);
528 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
533 spin_lock_irqsave(&ctx->completion_lock, flags);
534 io_cqring_fill_event(ctx, user_data, res);
535 io_commit_cqring(ctx);
536 spin_unlock_irqrestore(&ctx->completion_lock, flags);
538 io_cqring_ev_posted(ctx);
541 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
543 percpu_ref_put_many(&ctx->refs, refs);
545 if (waitqueue_active(&ctx->wait))
549 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
550 struct io_submit_state *state)
552 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
553 struct io_kiocb *req;
555 if (!percpu_ref_tryget(&ctx->refs))
559 req = kmem_cache_alloc(req_cachep, gfp);
562 } else if (!state->free_reqs) {
566 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
567 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
570 * Bulk alloc is all-or-nothing. If we fail to get a batch,
571 * retry single alloc to be on the safe side.
573 if (unlikely(ret <= 0)) {
574 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
579 state->free_reqs = ret - 1;
581 req = state->reqs[0];
583 req = state->reqs[state->cur_req];
591 /* one is dropped after submission, the other at completion */
592 refcount_set(&req->refs, 2);
596 io_ring_drop_ctx_refs(ctx, 1);
600 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
603 kmem_cache_free_bulk(req_cachep, *nr, reqs);
604 io_ring_drop_ctx_refs(ctx, *nr);
609 static void __io_free_req(struct io_kiocb *req)
611 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
613 io_ring_drop_ctx_refs(req->ctx, 1);
614 kmem_cache_free(req_cachep, req);
617 static void io_req_link_next(struct io_kiocb *req)
619 struct io_kiocb *nxt;
622 * The list should never be empty when we are called here. But could
623 * potentially happen if the chain is messed up, check to be on the
626 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
628 list_del(&nxt->list);
629 if (!list_empty(&req->link_list)) {
630 INIT_LIST_HEAD(&nxt->link_list);
631 list_splice(&req->link_list, &nxt->link_list);
632 nxt->flags |= REQ_F_LINK;
635 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
636 queue_work(req->ctx->sqo_wq, &nxt->work);
641 * Called if REQ_F_LINK is set, and we fail the head request
643 static void io_fail_links(struct io_kiocb *req)
645 struct io_kiocb *link;
647 while (!list_empty(&req->link_list)) {
648 link = list_first_entry(&req->link_list, struct io_kiocb, list);
649 list_del(&link->list);
651 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
656 static void io_free_req(struct io_kiocb *req)
659 * If LINK is set, we have dependent requests in this chain. If we
660 * didn't fail this request, queue the first one up, moving any other
661 * dependencies to the next request. In case of failure, fail the rest
664 if (req->flags & REQ_F_LINK) {
665 if (req->flags & REQ_F_FAIL_LINK)
668 io_req_link_next(req);
674 static void io_put_req(struct io_kiocb *req)
676 if (refcount_dec_and_test(&req->refs))
681 * Find and free completed poll iocbs
683 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
684 struct list_head *done)
686 void *reqs[IO_IOPOLL_BATCH];
687 struct io_kiocb *req;
691 while (!list_empty(done)) {
692 req = list_first_entry(done, struct io_kiocb, list);
693 list_del(&req->list);
695 io_cqring_fill_event(ctx, req->user_data, req->result);
698 if (refcount_dec_and_test(&req->refs)) {
699 /* If we're not using fixed files, we have to pair the
700 * completion part with the file put. Use regular
701 * completions for those, only batch free for fixed
702 * file and non-linked commands.
704 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
706 reqs[to_free++] = req;
707 if (to_free == ARRAY_SIZE(reqs))
708 io_free_req_many(ctx, reqs, &to_free);
715 io_commit_cqring(ctx);
716 io_free_req_many(ctx, reqs, &to_free);
719 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
722 struct io_kiocb *req, *tmp;
728 * Only spin for completions if we don't have multiple devices hanging
729 * off our complete list, and we're under the requested amount.
731 spin = !ctx->poll_multi_file && *nr_events < min;
734 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
735 struct kiocb *kiocb = &req->rw;
738 * Move completed entries to our local list. If we find a
739 * request that requires polling, break out and complete
740 * the done list first, if we have entries there.
742 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
743 list_move_tail(&req->list, &done);
746 if (!list_empty(&done))
749 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
758 if (!list_empty(&done))
759 io_iopoll_complete(ctx, nr_events, &done);
765 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
766 * non-spinning poll check - we'll still enter the driver poll loop, but only
767 * as a non-spinning completion check.
769 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
772 while (!list_empty(&ctx->poll_list)) {
775 ret = io_do_iopoll(ctx, nr_events, min);
778 if (!min || *nr_events >= min)
786 * We can't just wait for polled events to come to us, we have to actively
787 * find and complete them.
789 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
791 if (!(ctx->flags & IORING_SETUP_IOPOLL))
794 mutex_lock(&ctx->uring_lock);
795 while (!list_empty(&ctx->poll_list)) {
796 unsigned int nr_events = 0;
798 io_iopoll_getevents(ctx, &nr_events, 1);
800 mutex_unlock(&ctx->uring_lock);
803 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
811 if (*nr_events < min)
812 tmin = min - *nr_events;
814 ret = io_iopoll_getevents(ctx, nr_events, tmin);
818 } while (min && !*nr_events && !need_resched());
823 static void kiocb_end_write(struct kiocb *kiocb)
825 if (kiocb->ki_flags & IOCB_WRITE) {
826 struct inode *inode = file_inode(kiocb->ki_filp);
829 * Tell lockdep we inherited freeze protection from submission
832 if (S_ISREG(inode->i_mode))
833 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
834 file_end_write(kiocb->ki_filp);
838 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
840 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
842 kiocb_end_write(kiocb);
844 if ((req->flags & REQ_F_LINK) && res != req->result)
845 req->flags |= REQ_F_FAIL_LINK;
846 io_cqring_add_event(req->ctx, req->user_data, res);
850 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
852 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
854 kiocb_end_write(kiocb);
856 if ((req->flags & REQ_F_LINK) && res != req->result)
857 req->flags |= REQ_F_FAIL_LINK;
860 req->flags |= REQ_F_IOPOLL_COMPLETED;
864 * After the iocb has been issued, it's safe to be found on the poll list.
865 * Adding the kiocb to the list AFTER submission ensures that we don't
866 * find it from a io_iopoll_getevents() thread before the issuer is done
867 * accessing the kiocb cookie.
869 static void io_iopoll_req_issued(struct io_kiocb *req)
871 struct io_ring_ctx *ctx = req->ctx;
874 * Track whether we have multiple files in our lists. This will impact
875 * how we do polling eventually, not spinning if we're on potentially
878 if (list_empty(&ctx->poll_list)) {
879 ctx->poll_multi_file = false;
880 } else if (!ctx->poll_multi_file) {
881 struct io_kiocb *list_req;
883 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
885 if (list_req->rw.ki_filp != req->rw.ki_filp)
886 ctx->poll_multi_file = true;
890 * For fast devices, IO may have already completed. If it has, add
891 * it to the front so we find it first.
893 if (req->flags & REQ_F_IOPOLL_COMPLETED)
894 list_add(&req->list, &ctx->poll_list);
896 list_add_tail(&req->list, &ctx->poll_list);
899 static void io_file_put(struct io_submit_state *state)
902 int diff = state->has_refs - state->used_refs;
905 fput_many(state->file, diff);
911 * Get as many references to a file as we have IOs left in this submission,
912 * assuming most submissions are for one file, or at least that each file
913 * has more than one submission.
915 static struct file *io_file_get(struct io_submit_state *state, int fd)
921 if (state->fd == fd) {
928 state->file = fget_many(fd, state->ios_left);
933 state->has_refs = state->ios_left;
934 state->used_refs = 1;
940 * If we tracked the file through the SCM inflight mechanism, we could support
941 * any file. For now, just ensure that anything potentially problematic is done
944 static bool io_file_supports_async(struct file *file)
946 umode_t mode = file_inode(file)->i_mode;
948 if (S_ISBLK(mode) || S_ISCHR(mode))
950 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
956 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
959 const struct io_uring_sqe *sqe = s->sqe;
960 struct io_ring_ctx *ctx = req->ctx;
961 struct kiocb *kiocb = &req->rw;
968 if (force_nonblock && !io_file_supports_async(req->file))
969 force_nonblock = false;
971 kiocb->ki_pos = READ_ONCE(sqe->off);
972 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
973 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
975 ioprio = READ_ONCE(sqe->ioprio);
977 ret = ioprio_check_cap(ioprio);
981 kiocb->ki_ioprio = ioprio;
983 kiocb->ki_ioprio = get_current_ioprio();
985 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
989 /* don't allow async punt if RWF_NOWAIT was requested */
990 if (kiocb->ki_flags & IOCB_NOWAIT)
991 req->flags |= REQ_F_NOWAIT;
994 kiocb->ki_flags |= IOCB_NOWAIT;
996 if (ctx->flags & IORING_SETUP_IOPOLL) {
997 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
998 !kiocb->ki_filp->f_op->iopoll)
1001 kiocb->ki_flags |= IOCB_HIPRI;
1002 kiocb->ki_complete = io_complete_rw_iopoll;
1004 if (kiocb->ki_flags & IOCB_HIPRI)
1006 kiocb->ki_complete = io_complete_rw;
1011 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1017 case -ERESTARTNOINTR:
1018 case -ERESTARTNOHAND:
1019 case -ERESTART_RESTARTBLOCK:
1021 * We can't just restart the syscall, since previously
1022 * submitted sqes may already be in progress. Just fail this
1028 kiocb->ki_complete(kiocb, ret, 0);
1032 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1033 const struct io_uring_sqe *sqe,
1034 struct iov_iter *iter)
1036 size_t len = READ_ONCE(sqe->len);
1037 struct io_mapped_ubuf *imu;
1038 unsigned index, buf_index;
1042 /* attempt to use fixed buffers without having provided iovecs */
1043 if (unlikely(!ctx->user_bufs))
1046 buf_index = READ_ONCE(sqe->buf_index);
1047 if (unlikely(buf_index >= ctx->nr_user_bufs))
1050 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1051 imu = &ctx->user_bufs[index];
1052 buf_addr = READ_ONCE(sqe->addr);
1055 if (buf_addr + len < buf_addr)
1057 /* not inside the mapped region */
1058 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1062 * May not be a start of buffer, set size appropriately
1063 * and advance us to the beginning.
1065 offset = buf_addr - imu->ubuf;
1066 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1068 iov_iter_advance(iter, offset);
1072 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1073 const struct sqe_submit *s, struct iovec **iovec,
1074 struct iov_iter *iter)
1076 const struct io_uring_sqe *sqe = s->sqe;
1077 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1078 size_t sqe_len = READ_ONCE(sqe->len);
1082 * We're reading ->opcode for the second time, but the first read
1083 * doesn't care whether it's _FIXED or not, so it doesn't matter
1084 * whether ->opcode changes concurrently. The first read does care
1085 * about whether it is a READ or a WRITE, so we don't trust this read
1086 * for that purpose and instead let the caller pass in the read/write
1089 opcode = READ_ONCE(sqe->opcode);
1090 if (opcode == IORING_OP_READ_FIXED ||
1091 opcode == IORING_OP_WRITE_FIXED) {
1092 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1100 #ifdef CONFIG_COMPAT
1102 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1106 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1110 * Make a note of the last file/offset/direction we punted to async
1111 * context. We'll use this information to see if we can piggy back a
1112 * sequential request onto the previous one, if it's still hasn't been
1113 * completed by the async worker.
1115 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1117 struct async_list *async_list = &req->ctx->pending_async[rw];
1118 struct kiocb *kiocb = &req->rw;
1119 struct file *filp = kiocb->ki_filp;
1120 off_t io_end = kiocb->ki_pos + len;
1122 if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) {
1123 unsigned long max_pages;
1125 /* Use 8x RA size as a decent limiter for both reads/writes */
1126 max_pages = filp->f_ra.ra_pages;
1128 max_pages = VM_READAHEAD_PAGES;
1131 /* If max pages are exceeded, reset the state */
1133 if (async_list->io_pages + len <= max_pages) {
1134 req->flags |= REQ_F_SEQ_PREV;
1135 async_list->io_pages += len;
1138 async_list->io_pages = 0;
1142 /* New file? Reset state. */
1143 if (async_list->file != filp) {
1144 async_list->io_pages = 0;
1145 async_list->file = filp;
1147 async_list->io_end = io_end;
1150 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1151 bool force_nonblock)
1153 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1154 struct kiocb *kiocb = &req->rw;
1155 struct iov_iter iter;
1158 ssize_t read_size, ret;
1160 ret = io_prep_rw(req, s, force_nonblock);
1163 file = kiocb->ki_filp;
1165 if (unlikely(!(file->f_mode & FMODE_READ)))
1167 if (unlikely(!file->f_op->read_iter))
1170 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1175 if (req->flags & REQ_F_LINK)
1176 req->result = read_size;
1178 iov_count = iov_iter_count(&iter);
1179 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1183 ret2 = call_read_iter(file, kiocb, &iter);
1185 * In case of a short read, punt to async. This can happen
1186 * if we have data partially cached. Alternatively we can
1187 * return the short read, in which case the application will
1188 * need to issue another SQE and wait for it. That SQE will
1189 * need async punt anyway, so it's more efficient to do it
1192 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1194 /* Catch -EAGAIN return for forced non-blocking submission */
1195 if (!force_nonblock || ret2 != -EAGAIN) {
1196 io_rw_done(kiocb, ret2);
1199 * If ->needs_lock is true, we're already in async
1203 io_async_list_note(READ, req, iov_count);
1211 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1212 bool force_nonblock)
1214 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1215 struct kiocb *kiocb = &req->rw;
1216 struct iov_iter iter;
1221 ret = io_prep_rw(req, s, force_nonblock);
1225 file = kiocb->ki_filp;
1226 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1228 if (unlikely(!file->f_op->write_iter))
1231 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1235 if (req->flags & REQ_F_LINK)
1238 iov_count = iov_iter_count(&iter);
1241 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1242 /* If ->needs_lock is true, we're already in async context. */
1244 io_async_list_note(WRITE, req, iov_count);
1248 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1253 * Open-code file_start_write here to grab freeze protection,
1254 * which will be released by another thread in
1255 * io_complete_rw(). Fool lockdep by telling it the lock got
1256 * released so that it doesn't complain about the held lock when
1257 * we return to userspace.
1259 if (S_ISREG(file_inode(file)->i_mode)) {
1260 __sb_start_write(file_inode(file)->i_sb,
1261 SB_FREEZE_WRITE, true);
1262 __sb_writers_release(file_inode(file)->i_sb,
1265 kiocb->ki_flags |= IOCB_WRITE;
1267 ret2 = call_write_iter(file, kiocb, &iter);
1268 if (!force_nonblock || ret2 != -EAGAIN) {
1269 io_rw_done(kiocb, ret2);
1272 * If ->needs_lock is true, we're already in async
1276 io_async_list_note(WRITE, req, iov_count);
1286 * IORING_OP_NOP just posts a completion event, nothing else.
1288 static int io_nop(struct io_kiocb *req, u64 user_data)
1290 struct io_ring_ctx *ctx = req->ctx;
1293 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1296 io_cqring_add_event(ctx, user_data, err);
1301 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1303 struct io_ring_ctx *ctx = req->ctx;
1308 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1310 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1316 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1317 bool force_nonblock)
1319 loff_t sqe_off = READ_ONCE(sqe->off);
1320 loff_t sqe_len = READ_ONCE(sqe->len);
1321 loff_t end = sqe_off + sqe_len;
1322 unsigned fsync_flags;
1325 fsync_flags = READ_ONCE(sqe->fsync_flags);
1326 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1329 ret = io_prep_fsync(req, sqe);
1333 /* fsync always requires a blocking context */
1337 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1338 end > 0 ? end : LLONG_MAX,
1339 fsync_flags & IORING_FSYNC_DATASYNC);
1341 if (ret < 0 && (req->flags & REQ_F_LINK))
1342 req->flags |= REQ_F_FAIL_LINK;
1343 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1348 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1350 struct io_ring_ctx *ctx = req->ctx;
1356 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1358 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1364 static int io_sync_file_range(struct io_kiocb *req,
1365 const struct io_uring_sqe *sqe,
1366 bool force_nonblock)
1373 ret = io_prep_sfr(req, sqe);
1377 /* sync_file_range always requires a blocking context */
1381 sqe_off = READ_ONCE(sqe->off);
1382 sqe_len = READ_ONCE(sqe->len);
1383 flags = READ_ONCE(sqe->sync_range_flags);
1385 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1387 if (ret < 0 && (req->flags & REQ_F_LINK))
1388 req->flags |= REQ_F_FAIL_LINK;
1389 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1394 #if defined(CONFIG_NET)
1395 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1396 bool force_nonblock,
1397 long (*fn)(struct socket *, struct user_msghdr __user *,
1400 struct socket *sock;
1403 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1406 sock = sock_from_file(req->file, &ret);
1408 struct user_msghdr __user *msg;
1411 flags = READ_ONCE(sqe->msg_flags);
1412 if (flags & MSG_DONTWAIT)
1413 req->flags |= REQ_F_NOWAIT;
1414 else if (force_nonblock)
1415 flags |= MSG_DONTWAIT;
1417 msg = (struct user_msghdr __user *) (unsigned long)
1418 READ_ONCE(sqe->addr);
1420 ret = fn(sock, msg, flags);
1421 if (force_nonblock && ret == -EAGAIN)
1425 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1431 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1432 bool force_nonblock)
1434 #if defined(CONFIG_NET)
1435 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1441 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1442 bool force_nonblock)
1444 #if defined(CONFIG_NET)
1445 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1451 static void io_poll_remove_one(struct io_kiocb *req)
1453 struct io_poll_iocb *poll = &req->poll;
1455 spin_lock(&poll->head->lock);
1456 WRITE_ONCE(poll->canceled, true);
1457 if (!list_empty(&poll->wait.entry)) {
1458 list_del_init(&poll->wait.entry);
1459 queue_work(req->ctx->sqo_wq, &req->work);
1461 spin_unlock(&poll->head->lock);
1463 list_del_init(&req->list);
1466 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1468 struct io_kiocb *req;
1470 spin_lock_irq(&ctx->completion_lock);
1471 while (!list_empty(&ctx->cancel_list)) {
1472 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1473 io_poll_remove_one(req);
1475 spin_unlock_irq(&ctx->completion_lock);
1479 * Find a running poll command that matches one specified in sqe->addr,
1480 * and remove it if found.
1482 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1484 struct io_ring_ctx *ctx = req->ctx;
1485 struct io_kiocb *poll_req, *next;
1488 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1490 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1494 spin_lock_irq(&ctx->completion_lock);
1495 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1496 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1497 io_poll_remove_one(poll_req);
1502 spin_unlock_irq(&ctx->completion_lock);
1504 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1509 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1512 req->poll.done = true;
1513 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1514 io_commit_cqring(ctx);
1517 static void io_poll_complete_work(struct work_struct *work)
1519 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1520 struct io_poll_iocb *poll = &req->poll;
1521 struct poll_table_struct pt = { ._key = poll->events };
1522 struct io_ring_ctx *ctx = req->ctx;
1525 if (!READ_ONCE(poll->canceled))
1526 mask = vfs_poll(poll->file, &pt) & poll->events;
1529 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1530 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1531 * synchronize with them. In the cancellation case the list_del_init
1532 * itself is not actually needed, but harmless so we keep it in to
1533 * avoid further branches in the fast path.
1535 spin_lock_irq(&ctx->completion_lock);
1536 if (!mask && !READ_ONCE(poll->canceled)) {
1537 add_wait_queue(poll->head, &poll->wait);
1538 spin_unlock_irq(&ctx->completion_lock);
1541 list_del_init(&req->list);
1542 io_poll_complete(ctx, req, mask);
1543 spin_unlock_irq(&ctx->completion_lock);
1545 io_cqring_ev_posted(ctx);
1549 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1552 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1554 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1555 struct io_ring_ctx *ctx = req->ctx;
1556 __poll_t mask = key_to_poll(key);
1557 unsigned long flags;
1559 /* for instances that support it check for an event match first: */
1560 if (mask && !(mask & poll->events))
1563 list_del_init(&poll->wait.entry);
1565 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1566 list_del(&req->list);
1567 io_poll_complete(ctx, req, mask);
1568 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1570 io_cqring_ev_posted(ctx);
1573 queue_work(ctx->sqo_wq, &req->work);
1579 struct io_poll_table {
1580 struct poll_table_struct pt;
1581 struct io_kiocb *req;
1585 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1586 struct poll_table_struct *p)
1588 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1590 if (unlikely(pt->req->poll.head)) {
1591 pt->error = -EINVAL;
1596 pt->req->poll.head = head;
1597 add_wait_queue(head, &pt->req->poll.wait);
1600 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1602 struct io_poll_iocb *poll = &req->poll;
1603 struct io_ring_ctx *ctx = req->ctx;
1604 struct io_poll_table ipt;
1605 bool cancel = false;
1609 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1611 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1616 INIT_WORK(&req->work, io_poll_complete_work);
1617 events = READ_ONCE(sqe->poll_events);
1618 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1622 poll->canceled = false;
1624 ipt.pt._qproc = io_poll_queue_proc;
1625 ipt.pt._key = poll->events;
1627 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1629 /* initialized the list so that we can do list_empty checks */
1630 INIT_LIST_HEAD(&poll->wait.entry);
1631 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1633 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1635 spin_lock_irq(&ctx->completion_lock);
1636 if (likely(poll->head)) {
1637 spin_lock(&poll->head->lock);
1638 if (unlikely(list_empty(&poll->wait.entry))) {
1644 if (mask || ipt.error)
1645 list_del_init(&poll->wait.entry);
1647 WRITE_ONCE(poll->canceled, true);
1648 else if (!poll->done) /* actually waiting for an event */
1649 list_add_tail(&req->list, &ctx->cancel_list);
1650 spin_unlock(&poll->head->lock);
1652 if (mask) { /* no async, we'd stolen it */
1654 io_poll_complete(ctx, req, mask);
1656 spin_unlock_irq(&ctx->completion_lock);
1659 io_cqring_ev_posted(ctx);
1665 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1666 const struct io_uring_sqe *sqe)
1668 struct io_uring_sqe *sqe_copy;
1670 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1673 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1677 spin_lock_irq(&ctx->completion_lock);
1678 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1679 spin_unlock_irq(&ctx->completion_lock);
1684 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1685 req->submit.sqe = sqe_copy;
1687 INIT_WORK(&req->work, io_sq_wq_submit_work);
1688 list_add_tail(&req->list, &ctx->defer_list);
1689 spin_unlock_irq(&ctx->completion_lock);
1690 return -EIOCBQUEUED;
1693 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1694 const struct sqe_submit *s, bool force_nonblock)
1698 req->user_data = READ_ONCE(s->sqe->user_data);
1700 if (unlikely(s->index >= ctx->sq_entries))
1703 opcode = READ_ONCE(s->sqe->opcode);
1706 ret = io_nop(req, req->user_data);
1708 case IORING_OP_READV:
1709 if (unlikely(s->sqe->buf_index))
1711 ret = io_read(req, s, force_nonblock);
1713 case IORING_OP_WRITEV:
1714 if (unlikely(s->sqe->buf_index))
1716 ret = io_write(req, s, force_nonblock);
1718 case IORING_OP_READ_FIXED:
1719 ret = io_read(req, s, force_nonblock);
1721 case IORING_OP_WRITE_FIXED:
1722 ret = io_write(req, s, force_nonblock);
1724 case IORING_OP_FSYNC:
1725 ret = io_fsync(req, s->sqe, force_nonblock);
1727 case IORING_OP_POLL_ADD:
1728 ret = io_poll_add(req, s->sqe);
1730 case IORING_OP_POLL_REMOVE:
1731 ret = io_poll_remove(req, s->sqe);
1733 case IORING_OP_SYNC_FILE_RANGE:
1734 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1736 case IORING_OP_SENDMSG:
1737 ret = io_sendmsg(req, s->sqe, force_nonblock);
1739 case IORING_OP_RECVMSG:
1740 ret = io_recvmsg(req, s->sqe, force_nonblock);
1750 if (ctx->flags & IORING_SETUP_IOPOLL) {
1751 if (req->result == -EAGAIN)
1754 /* workqueue context doesn't hold uring_lock, grab it now */
1756 mutex_lock(&ctx->uring_lock);
1757 io_iopoll_req_issued(req);
1759 mutex_unlock(&ctx->uring_lock);
1765 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1766 const struct io_uring_sqe *sqe)
1768 switch (sqe->opcode) {
1769 case IORING_OP_READV:
1770 case IORING_OP_READ_FIXED:
1771 return &ctx->pending_async[READ];
1772 case IORING_OP_WRITEV:
1773 case IORING_OP_WRITE_FIXED:
1774 return &ctx->pending_async[WRITE];
1780 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1782 u8 opcode = READ_ONCE(sqe->opcode);
1784 return !(opcode == IORING_OP_READ_FIXED ||
1785 opcode == IORING_OP_WRITE_FIXED);
1788 static void io_sq_wq_submit_work(struct work_struct *work)
1790 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1791 struct io_ring_ctx *ctx = req->ctx;
1792 struct mm_struct *cur_mm = NULL;
1793 struct async_list *async_list;
1794 LIST_HEAD(req_list);
1795 mm_segment_t old_fs;
1798 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1801 struct sqe_submit *s = &req->submit;
1802 const struct io_uring_sqe *sqe = s->sqe;
1804 /* Ensure we clear previously set non-block flag */
1805 req->rw.ki_flags &= ~IOCB_NOWAIT;
1808 if (io_sqe_needs_user(sqe) && !cur_mm) {
1809 if (!mmget_not_zero(ctx->sqo_mm)) {
1812 cur_mm = ctx->sqo_mm;
1820 s->has_user = cur_mm != NULL;
1821 s->needs_lock = true;
1823 ret = __io_submit_sqe(ctx, req, s, false);
1825 * We can get EAGAIN for polled IO even though
1826 * we're forcing a sync submission from here,
1827 * since we can't wait for request slots on the
1836 /* drop submission reference */
1840 io_cqring_add_event(ctx, sqe->user_data, ret);
1844 /* async context always use a copy of the sqe */
1849 if (!list_empty(&req_list)) {
1850 req = list_first_entry(&req_list, struct io_kiocb,
1852 list_del(&req->list);
1855 if (list_empty(&async_list->list))
1859 spin_lock(&async_list->lock);
1860 if (list_empty(&async_list->list)) {
1861 spin_unlock(&async_list->lock);
1864 list_splice_init(&async_list->list, &req_list);
1865 spin_unlock(&async_list->lock);
1867 req = list_first_entry(&req_list, struct io_kiocb, list);
1868 list_del(&req->list);
1872 * Rare case of racing with a submitter. If we find the count has
1873 * dropped to zero AND we have pending work items, then restart
1874 * the processing. This is a tiny race window.
1877 ret = atomic_dec_return(&async_list->cnt);
1878 while (!ret && !list_empty(&async_list->list)) {
1879 spin_lock(&async_list->lock);
1880 atomic_inc(&async_list->cnt);
1881 list_splice_init(&async_list->list, &req_list);
1882 spin_unlock(&async_list->lock);
1884 if (!list_empty(&req_list)) {
1885 req = list_first_entry(&req_list,
1886 struct io_kiocb, list);
1887 list_del(&req->list);
1890 ret = atomic_dec_return(&async_list->cnt);
1902 * See if we can piggy back onto previously submitted work, that is still
1903 * running. We currently only allow this if the new request is sequential
1904 * to the previous one we punted.
1906 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
1912 if (!(req->flags & REQ_F_SEQ_PREV))
1914 if (!atomic_read(&list->cnt))
1918 spin_lock(&list->lock);
1919 list_add_tail(&req->list, &list->list);
1920 if (!atomic_read(&list->cnt)) {
1921 list_del_init(&req->list);
1924 spin_unlock(&list->lock);
1928 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
1930 int op = READ_ONCE(sqe->opcode);
1934 case IORING_OP_POLL_REMOVE:
1941 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
1942 struct io_submit_state *state, struct io_kiocb *req)
1947 flags = READ_ONCE(s->sqe->flags);
1948 fd = READ_ONCE(s->sqe->fd);
1950 if (flags & IOSQE_IO_DRAIN) {
1951 req->flags |= REQ_F_IO_DRAIN;
1952 req->sequence = ctx->cached_sq_head - 1;
1955 if (!io_op_needs_file(s->sqe))
1958 if (flags & IOSQE_FIXED_FILE) {
1959 if (unlikely(!ctx->user_files ||
1960 (unsigned) fd >= ctx->nr_user_files))
1962 req->file = ctx->user_files[fd];
1963 req->flags |= REQ_F_FIXED_FILE;
1965 if (s->needs_fixed_file)
1967 req->file = io_file_get(state, fd);
1968 if (unlikely(!req->file))
1975 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1976 struct sqe_submit *s)
1980 ret = __io_submit_sqe(ctx, req, s, true);
1981 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
1982 struct io_uring_sqe *sqe_copy;
1984 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1986 struct async_list *list;
1988 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
1991 memcpy(&req->submit, s, sizeof(*s));
1992 list = io_async_list_from_sqe(ctx, s->sqe);
1993 if (!io_add_to_prev_work(list, req)) {
1995 atomic_inc(&list->cnt);
1996 INIT_WORK(&req->work, io_sq_wq_submit_work);
1997 queue_work(ctx->sqo_wq, &req->work);
2001 * Queued up for async execution, worker will release
2002 * submit reference when the iocb is actually submitted.
2008 /* drop submission reference */
2011 /* and drop final reference, if we failed */
2013 io_cqring_add_event(ctx, req->user_data, ret);
2014 if (req->flags & REQ_F_LINK)
2015 req->flags |= REQ_F_FAIL_LINK;
2022 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2024 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2025 struct io_submit_state *state, struct io_kiocb **link)
2027 struct io_uring_sqe *sqe_copy;
2028 struct io_kiocb *req;
2031 /* enforce forwards compatibility on users */
2032 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2037 req = io_get_req(ctx, state);
2038 if (unlikely(!req)) {
2043 ret = io_req_set_file(ctx, s, state, req);
2044 if (unlikely(ret)) {
2048 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2052 ret = io_req_defer(ctx, req, s->sqe);
2054 if (ret != -EIOCBQUEUED)
2060 * If we already have a head request, queue this one for async
2061 * submittal once the head completes. If we don't have a head but
2062 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2063 * submitted sync once the chain is complete. If none of those
2064 * conditions are true (normal request), then just queue it.
2067 struct io_kiocb *prev = *link;
2069 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2076 memcpy(&req->submit, s, sizeof(*s));
2077 list_add_tail(&req->list, &prev->link_list);
2078 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2079 req->flags |= REQ_F_LINK;
2081 memcpy(&req->submit, s, sizeof(*s));
2082 INIT_LIST_HEAD(&req->link_list);
2085 io_queue_sqe(ctx, req, s);
2090 * Batched submission is done, ensure local IO is flushed out.
2092 static void io_submit_state_end(struct io_submit_state *state)
2094 blk_finish_plug(&state->plug);
2096 if (state->free_reqs)
2097 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2098 &state->reqs[state->cur_req]);
2102 * Start submission side cache.
2104 static void io_submit_state_start(struct io_submit_state *state,
2105 struct io_ring_ctx *ctx, unsigned max_ios)
2107 blk_start_plug(&state->plug);
2108 state->free_reqs = 0;
2110 state->ios_left = max_ios;
2113 static void io_commit_sqring(struct io_ring_ctx *ctx)
2115 struct io_sq_ring *ring = ctx->sq_ring;
2117 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
2119 * Ensure any loads from the SQEs are done at this point,
2120 * since once we write the new head, the application could
2121 * write new data to them.
2123 smp_store_release(&ring->r.head, ctx->cached_sq_head);
2128 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2129 * that is mapped by userspace. This means that care needs to be taken to
2130 * ensure that reads are stable, as we cannot rely on userspace always
2131 * being a good citizen. If members of the sqe are validated and then later
2132 * used, it's important that those reads are done through READ_ONCE() to
2133 * prevent a re-load down the line.
2135 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2137 struct io_sq_ring *ring = ctx->sq_ring;
2141 * The cached sq head (or cq tail) serves two purposes:
2143 * 1) allows us to batch the cost of updating the user visible
2145 * 2) allows the kernel side to track the head on its own, even
2146 * though the application is the one updating it.
2148 head = ctx->cached_sq_head;
2149 /* make sure SQ entry isn't read before tail */
2150 if (head == smp_load_acquire(&ring->r.tail))
2153 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
2154 if (head < ctx->sq_entries) {
2156 s->sqe = &ctx->sq_sqes[head];
2157 ctx->cached_sq_head++;
2161 /* drop invalid entries */
2162 ctx->cached_sq_head++;
2167 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2168 unsigned int nr, bool has_user, bool mm_fault)
2170 struct io_submit_state state, *statep = NULL;
2171 struct io_kiocb *link = NULL;
2172 bool prev_was_link = false;
2173 int i, submitted = 0;
2175 if (nr > IO_PLUG_THRESHOLD) {
2176 io_submit_state_start(&state, ctx, nr);
2180 for (i = 0; i < nr; i++) {
2182 * If previous wasn't linked and we have a linked command,
2183 * that's the end of the chain. Submit the previous link.
2185 if (!prev_was_link && link) {
2186 io_queue_sqe(ctx, link, &link->submit);
2189 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2191 if (unlikely(mm_fault)) {
2192 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2195 sqes[i].has_user = has_user;
2196 sqes[i].needs_lock = true;
2197 sqes[i].needs_fixed_file = true;
2198 io_submit_sqe(ctx, &sqes[i], statep, &link);
2204 io_queue_sqe(ctx, link, &link->submit);
2206 io_submit_state_end(&state);
2211 static int io_sq_thread(void *data)
2213 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2214 struct io_ring_ctx *ctx = data;
2215 struct mm_struct *cur_mm = NULL;
2216 mm_segment_t old_fs;
2219 unsigned long timeout;
2221 complete(&ctx->sqo_thread_started);
2226 timeout = inflight = 0;
2227 while (!kthread_should_park()) {
2228 bool all_fixed, mm_fault = false;
2232 unsigned nr_events = 0;
2234 if (ctx->flags & IORING_SETUP_IOPOLL) {
2236 * We disallow the app entering submit/complete
2237 * with polling, but we still need to lock the
2238 * ring to prevent racing with polled issue
2239 * that got punted to a workqueue.
2241 mutex_lock(&ctx->uring_lock);
2242 io_iopoll_check(ctx, &nr_events, 0);
2243 mutex_unlock(&ctx->uring_lock);
2246 * Normal IO, just pretend everything completed.
2247 * We don't have to poll completions for that.
2249 nr_events = inflight;
2252 inflight -= nr_events;
2254 timeout = jiffies + ctx->sq_thread_idle;
2257 if (!io_get_sqring(ctx, &sqes[0])) {
2259 * We're polling. If we're within the defined idle
2260 * period, then let us spin without work before going
2263 if (inflight || !time_after(jiffies, timeout)) {
2269 * Drop cur_mm before scheduling, we can't hold it for
2270 * long periods (or over schedule()). Do this before
2271 * adding ourselves to the waitqueue, as the unuse/drop
2280 prepare_to_wait(&ctx->sqo_wait, &wait,
2281 TASK_INTERRUPTIBLE);
2283 /* Tell userspace we may need a wakeup call */
2284 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
2285 /* make sure to read SQ tail after writing flags */
2288 if (!io_get_sqring(ctx, &sqes[0])) {
2289 if (kthread_should_park()) {
2290 finish_wait(&ctx->sqo_wait, &wait);
2293 if (signal_pending(current))
2294 flush_signals(current);
2296 finish_wait(&ctx->sqo_wait, &wait);
2298 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2301 finish_wait(&ctx->sqo_wait, &wait);
2303 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2309 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2313 if (i == ARRAY_SIZE(sqes))
2315 } while (io_get_sqring(ctx, &sqes[i]));
2317 /* Unless all new commands are FIXED regions, grab mm */
2318 if (!all_fixed && !cur_mm) {
2319 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2321 use_mm(ctx->sqo_mm);
2322 cur_mm = ctx->sqo_mm;
2326 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2329 /* Commit SQ ring head once we've consumed all SQEs */
2330 io_commit_sqring(ctx);
2344 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2346 struct io_submit_state state, *statep = NULL;
2347 struct io_kiocb *link = NULL;
2348 bool prev_was_link = false;
2351 if (to_submit > IO_PLUG_THRESHOLD) {
2352 io_submit_state_start(&state, ctx, to_submit);
2356 for (i = 0; i < to_submit; i++) {
2357 struct sqe_submit s;
2359 if (!io_get_sqring(ctx, &s))
2363 * If previous wasn't linked and we have a linked command,
2364 * that's the end of the chain. Submit the previous link.
2366 if (!prev_was_link && link) {
2367 io_queue_sqe(ctx, link, &link->submit);
2370 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2373 s.needs_lock = false;
2374 s.needs_fixed_file = false;
2376 io_submit_sqe(ctx, &s, statep, &link);
2378 io_commit_sqring(ctx);
2381 io_queue_sqe(ctx, link, &link->submit);
2383 io_submit_state_end(statep);
2388 static unsigned io_cqring_events(struct io_cq_ring *ring)
2390 /* See comment at the top of this file */
2392 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
2396 * Wait until events become available, if we don't already have some. The
2397 * application must reap them itself, as they reside on the shared cq ring.
2399 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2400 const sigset_t __user *sig, size_t sigsz)
2402 struct io_cq_ring *ring = ctx->cq_ring;
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,
2415 ret = set_user_sigmask(sig, sigsz);
2421 ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events);
2422 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
2423 if (ret == -ERESTARTSYS)
2426 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
2429 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2431 #if defined(CONFIG_UNIX)
2432 if (ctx->ring_sock) {
2433 struct sock *sock = ctx->ring_sock->sk;
2434 struct sk_buff *skb;
2436 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2442 for (i = 0; i < ctx->nr_user_files; i++)
2443 fput(ctx->user_files[i]);
2447 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2449 if (!ctx->user_files)
2452 __io_sqe_files_unregister(ctx);
2453 kfree(ctx->user_files);
2454 ctx->user_files = NULL;
2455 ctx->nr_user_files = 0;
2459 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2461 if (ctx->sqo_thread) {
2462 wait_for_completion(&ctx->sqo_thread_started);
2464 * The park is a bit of a work-around, without it we get
2465 * warning spews on shutdown with SQPOLL set and affinity
2466 * set to a single CPU.
2468 kthread_park(ctx->sqo_thread);
2469 kthread_stop(ctx->sqo_thread);
2470 ctx->sqo_thread = NULL;
2474 static void io_finish_async(struct io_ring_ctx *ctx)
2476 io_sq_thread_stop(ctx);
2479 destroy_workqueue(ctx->sqo_wq);
2484 #if defined(CONFIG_UNIX)
2485 static void io_destruct_skb(struct sk_buff *skb)
2487 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2489 io_finish_async(ctx);
2490 unix_destruct_scm(skb);
2494 * Ensure the UNIX gc is aware of our file set, so we are certain that
2495 * the io_uring can be safely unregistered on process exit, even if we have
2496 * loops in the file referencing.
2498 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2500 struct sock *sk = ctx->ring_sock->sk;
2501 struct scm_fp_list *fpl;
2502 struct sk_buff *skb;
2505 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2506 unsigned long inflight = ctx->user->unix_inflight + nr;
2508 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2512 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2516 skb = alloc_skb(0, GFP_KERNEL);
2523 skb->destructor = io_destruct_skb;
2525 fpl->user = get_uid(ctx->user);
2526 for (i = 0; i < nr; i++) {
2527 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2528 unix_inflight(fpl->user, fpl->fp[i]);
2531 fpl->max = fpl->count = nr;
2532 UNIXCB(skb).fp = fpl;
2533 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2534 skb_queue_head(&sk->sk_receive_queue, skb);
2536 for (i = 0; i < nr; i++)
2543 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2544 * causes regular reference counting to break down. We rely on the UNIX
2545 * garbage collection to take care of this problem for us.
2547 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2549 unsigned left, total;
2553 left = ctx->nr_user_files;
2555 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2557 ret = __io_sqe_files_scm(ctx, this_files, total);
2561 total += this_files;
2567 while (total < ctx->nr_user_files) {
2568 fput(ctx->user_files[total]);
2575 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2581 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2584 __s32 __user *fds = (__s32 __user *) arg;
2588 if (ctx->user_files)
2592 if (nr_args > IORING_MAX_FIXED_FILES)
2595 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2596 if (!ctx->user_files)
2599 for (i = 0; i < nr_args; i++) {
2601 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2604 ctx->user_files[i] = fget(fd);
2607 if (!ctx->user_files[i])
2610 * Don't allow io_uring instances to be registered. If UNIX
2611 * isn't enabled, then this causes a reference cycle and this
2612 * instance can never get freed. If UNIX is enabled we'll
2613 * handle it just fine, but there's still no point in allowing
2614 * a ring fd as it doesn't support regular read/write anyway.
2616 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2617 fput(ctx->user_files[i]);
2620 ctx->nr_user_files++;
2625 for (i = 0; i < ctx->nr_user_files; i++)
2626 fput(ctx->user_files[i]);
2628 kfree(ctx->user_files);
2629 ctx->user_files = NULL;
2630 ctx->nr_user_files = 0;
2634 ret = io_sqe_files_scm(ctx);
2636 io_sqe_files_unregister(ctx);
2641 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2642 struct io_uring_params *p)
2646 init_waitqueue_head(&ctx->sqo_wait);
2647 mmgrab(current->mm);
2648 ctx->sqo_mm = current->mm;
2650 if (ctx->flags & IORING_SETUP_SQPOLL) {
2652 if (!capable(CAP_SYS_ADMIN))
2655 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2656 if (!ctx->sq_thread_idle)
2657 ctx->sq_thread_idle = HZ;
2659 if (p->flags & IORING_SETUP_SQ_AFF) {
2660 int cpu = p->sq_thread_cpu;
2663 if (cpu >= nr_cpu_ids)
2665 if (!cpu_online(cpu))
2668 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2672 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2675 if (IS_ERR(ctx->sqo_thread)) {
2676 ret = PTR_ERR(ctx->sqo_thread);
2677 ctx->sqo_thread = NULL;
2680 wake_up_process(ctx->sqo_thread);
2681 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2682 /* Can't have SQ_AFF without SQPOLL */
2687 /* Do QD, or 2 * CPUS, whatever is smallest */
2688 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
2689 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2697 io_sq_thread_stop(ctx);
2698 mmdrop(ctx->sqo_mm);
2703 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2705 atomic_long_sub(nr_pages, &user->locked_vm);
2708 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2710 unsigned long page_limit, cur_pages, new_pages;
2712 /* Don't allow more pages than we can safely lock */
2713 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2716 cur_pages = atomic_long_read(&user->locked_vm);
2717 new_pages = cur_pages + nr_pages;
2718 if (new_pages > page_limit)
2720 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2721 new_pages) != cur_pages);
2726 static void io_mem_free(void *ptr)
2733 page = virt_to_head_page(ptr);
2734 if (put_page_testzero(page))
2735 free_compound_page(page);
2738 static void *io_mem_alloc(size_t size)
2740 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2743 return (void *) __get_free_pages(gfp_flags, get_order(size));
2746 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2748 struct io_sq_ring *sq_ring;
2749 struct io_cq_ring *cq_ring;
2752 bytes = struct_size(sq_ring, array, sq_entries);
2753 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
2754 bytes += struct_size(cq_ring, cqes, cq_entries);
2756 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
2759 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
2763 if (!ctx->user_bufs)
2766 for (i = 0; i < ctx->nr_user_bufs; i++) {
2767 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2769 for (j = 0; j < imu->nr_bvecs; j++)
2770 put_page(imu->bvec[j].bv_page);
2772 if (ctx->account_mem)
2773 io_unaccount_mem(ctx->user, imu->nr_bvecs);
2778 kfree(ctx->user_bufs);
2779 ctx->user_bufs = NULL;
2780 ctx->nr_user_bufs = 0;
2784 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
2785 void __user *arg, unsigned index)
2787 struct iovec __user *src;
2789 #ifdef CONFIG_COMPAT
2791 struct compat_iovec __user *ciovs;
2792 struct compat_iovec ciov;
2794 ciovs = (struct compat_iovec __user *) arg;
2795 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
2798 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
2799 dst->iov_len = ciov.iov_len;
2803 src = (struct iovec __user *) arg;
2804 if (copy_from_user(dst, &src[index], sizeof(*dst)))
2809 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
2812 struct vm_area_struct **vmas = NULL;
2813 struct page **pages = NULL;
2814 int i, j, got_pages = 0;
2819 if (!nr_args || nr_args > UIO_MAXIOV)
2822 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
2824 if (!ctx->user_bufs)
2827 for (i = 0; i < nr_args; i++) {
2828 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2829 unsigned long off, start, end, ubuf;
2834 ret = io_copy_iov(ctx, &iov, arg, i);
2839 * Don't impose further limits on the size and buffer
2840 * constraints here, we'll -EINVAL later when IO is
2841 * submitted if they are wrong.
2844 if (!iov.iov_base || !iov.iov_len)
2847 /* arbitrary limit, but we need something */
2848 if (iov.iov_len > SZ_1G)
2851 ubuf = (unsigned long) iov.iov_base;
2852 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2853 start = ubuf >> PAGE_SHIFT;
2854 nr_pages = end - start;
2856 if (ctx->account_mem) {
2857 ret = io_account_mem(ctx->user, nr_pages);
2863 if (!pages || nr_pages > got_pages) {
2866 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
2868 vmas = kvmalloc_array(nr_pages,
2869 sizeof(struct vm_area_struct *),
2871 if (!pages || !vmas) {
2873 if (ctx->account_mem)
2874 io_unaccount_mem(ctx->user, nr_pages);
2877 got_pages = nr_pages;
2880 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
2884 if (ctx->account_mem)
2885 io_unaccount_mem(ctx->user, nr_pages);
2890 down_read(¤t->mm->mmap_sem);
2891 pret = get_user_pages(ubuf, nr_pages,
2892 FOLL_WRITE | FOLL_LONGTERM,
2894 if (pret == nr_pages) {
2895 /* don't support file backed memory */
2896 for (j = 0; j < nr_pages; j++) {
2897 struct vm_area_struct *vma = vmas[j];
2900 !is_file_hugepages(vma->vm_file)) {
2906 ret = pret < 0 ? pret : -EFAULT;
2908 up_read(¤t->mm->mmap_sem);
2911 * if we did partial map, or found file backed vmas,
2912 * release any pages we did get
2915 for (j = 0; j < pret; j++)
2918 if (ctx->account_mem)
2919 io_unaccount_mem(ctx->user, nr_pages);
2924 off = ubuf & ~PAGE_MASK;
2926 for (j = 0; j < nr_pages; j++) {
2929 vec_len = min_t(size_t, size, PAGE_SIZE - off);
2930 imu->bvec[j].bv_page = pages[j];
2931 imu->bvec[j].bv_len = vec_len;
2932 imu->bvec[j].bv_offset = off;
2936 /* store original address for later verification */
2938 imu->len = iov.iov_len;
2939 imu->nr_bvecs = nr_pages;
2941 ctx->nr_user_bufs++;
2949 io_sqe_buffer_unregister(ctx);
2953 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
2955 __s32 __user *fds = arg;
2961 if (copy_from_user(&fd, fds, sizeof(*fds)))
2964 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
2965 if (IS_ERR(ctx->cq_ev_fd)) {
2966 int ret = PTR_ERR(ctx->cq_ev_fd);
2967 ctx->cq_ev_fd = NULL;
2974 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2976 if (ctx->cq_ev_fd) {
2977 eventfd_ctx_put(ctx->cq_ev_fd);
2978 ctx->cq_ev_fd = NULL;
2985 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
2987 io_finish_async(ctx);
2989 mmdrop(ctx->sqo_mm);
2991 io_iopoll_reap_events(ctx);
2992 io_sqe_buffer_unregister(ctx);
2993 io_sqe_files_unregister(ctx);
2994 io_eventfd_unregister(ctx);
2996 #if defined(CONFIG_UNIX)
2997 if (ctx->ring_sock) {
2998 ctx->ring_sock->file = NULL; /* so that iput() is called */
2999 sock_release(ctx->ring_sock);
3003 io_mem_free(ctx->sq_ring);
3004 io_mem_free(ctx->sq_sqes);
3005 io_mem_free(ctx->cq_ring);
3007 percpu_ref_exit(&ctx->refs);
3008 if (ctx->account_mem)
3009 io_unaccount_mem(ctx->user,
3010 ring_pages(ctx->sq_entries, ctx->cq_entries));
3011 free_uid(ctx->user);
3015 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3017 struct io_ring_ctx *ctx = file->private_data;
3020 poll_wait(file, &ctx->cq_wait, wait);
3022 * synchronizes with barrier from wq_has_sleeper call in
3026 if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
3027 ctx->sq_ring->ring_entries)
3028 mask |= EPOLLOUT | EPOLLWRNORM;
3029 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
3030 mask |= EPOLLIN | EPOLLRDNORM;
3035 static int io_uring_fasync(int fd, struct file *file, int on)
3037 struct io_ring_ctx *ctx = file->private_data;
3039 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3042 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3044 mutex_lock(&ctx->uring_lock);
3045 percpu_ref_kill(&ctx->refs);
3046 mutex_unlock(&ctx->uring_lock);
3048 io_poll_remove_all(ctx);
3049 io_iopoll_reap_events(ctx);
3050 wait_for_completion(&ctx->ctx_done);
3051 io_ring_ctx_free(ctx);
3054 static int io_uring_release(struct inode *inode, struct file *file)
3056 struct io_ring_ctx *ctx = file->private_data;
3058 file->private_data = NULL;
3059 io_ring_ctx_wait_and_kill(ctx);
3063 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3065 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3066 unsigned long sz = vma->vm_end - vma->vm_start;
3067 struct io_ring_ctx *ctx = file->private_data;
3073 case IORING_OFF_SQ_RING:
3076 case IORING_OFF_SQES:
3079 case IORING_OFF_CQ_RING:
3086 page = virt_to_head_page(ptr);
3087 if (sz > (PAGE_SIZE << compound_order(page)))
3090 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3091 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3094 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3095 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3098 struct io_ring_ctx *ctx;
3103 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3111 if (f.file->f_op != &io_uring_fops)
3115 ctx = f.file->private_data;
3116 if (!percpu_ref_tryget(&ctx->refs))
3120 * For SQ polling, the thread will do all submissions and completions.
3121 * Just return the requested submit count, and wake the thread if
3124 if (ctx->flags & IORING_SETUP_SQPOLL) {
3125 if (flags & IORING_ENTER_SQ_WAKEUP)
3126 wake_up(&ctx->sqo_wait);
3127 submitted = to_submit;
3133 to_submit = min(to_submit, ctx->sq_entries);
3135 mutex_lock(&ctx->uring_lock);
3136 submitted = io_ring_submit(ctx, to_submit);
3137 mutex_unlock(&ctx->uring_lock);
3139 if (flags & IORING_ENTER_GETEVENTS) {
3140 unsigned nr_events = 0;
3142 min_complete = min(min_complete, ctx->cq_entries);
3144 if (ctx->flags & IORING_SETUP_IOPOLL) {
3145 mutex_lock(&ctx->uring_lock);
3146 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3147 mutex_unlock(&ctx->uring_lock);
3149 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3154 io_ring_drop_ctx_refs(ctx, 1);
3157 return submitted ? submitted : ret;
3160 static const struct file_operations io_uring_fops = {
3161 .release = io_uring_release,
3162 .mmap = io_uring_mmap,
3163 .poll = io_uring_poll,
3164 .fasync = io_uring_fasync,
3167 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3168 struct io_uring_params *p)
3170 struct io_sq_ring *sq_ring;
3171 struct io_cq_ring *cq_ring;
3174 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
3178 ctx->sq_ring = sq_ring;
3179 sq_ring->ring_mask = p->sq_entries - 1;
3180 sq_ring->ring_entries = p->sq_entries;
3181 ctx->sq_mask = sq_ring->ring_mask;
3182 ctx->sq_entries = sq_ring->ring_entries;
3184 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3185 if (size == SIZE_MAX)
3188 ctx->sq_sqes = io_mem_alloc(size);
3192 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
3196 ctx->cq_ring = cq_ring;
3197 cq_ring->ring_mask = p->cq_entries - 1;
3198 cq_ring->ring_entries = p->cq_entries;
3199 ctx->cq_mask = cq_ring->ring_mask;
3200 ctx->cq_entries = cq_ring->ring_entries;
3205 * Allocate an anonymous fd, this is what constitutes the application
3206 * visible backing of an io_uring instance. The application mmaps this
3207 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3208 * we have to tie this fd to a socket for file garbage collection purposes.
3210 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3215 #if defined(CONFIG_UNIX)
3216 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3222 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3226 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3227 O_RDWR | O_CLOEXEC);
3230 ret = PTR_ERR(file);
3234 #if defined(CONFIG_UNIX)
3235 ctx->ring_sock->file = file;
3236 ctx->ring_sock->sk->sk_user_data = ctx;
3238 fd_install(ret, file);
3241 #if defined(CONFIG_UNIX)
3242 sock_release(ctx->ring_sock);
3243 ctx->ring_sock = NULL;
3248 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3250 struct user_struct *user = NULL;
3251 struct io_ring_ctx *ctx;
3255 if (!entries || entries > IORING_MAX_ENTRIES)
3259 * Use twice as many entries for the CQ ring. It's possible for the
3260 * application to drive a higher depth than the size of the SQ ring,
3261 * since the sqes are only used at submission time. This allows for
3262 * some flexibility in overcommitting a bit.
3264 p->sq_entries = roundup_pow_of_two(entries);
3265 p->cq_entries = 2 * p->sq_entries;
3267 user = get_uid(current_user());
3268 account_mem = !capable(CAP_IPC_LOCK);
3271 ret = io_account_mem(user,
3272 ring_pages(p->sq_entries, p->cq_entries));
3279 ctx = io_ring_ctx_alloc(p);
3282 io_unaccount_mem(user, ring_pages(p->sq_entries,
3287 ctx->compat = in_compat_syscall();
3288 ctx->account_mem = account_mem;
3291 ret = io_allocate_scq_urings(ctx, p);
3295 ret = io_sq_offload_start(ctx, p);
3299 ret = io_uring_get_fd(ctx);
3303 memset(&p->sq_off, 0, sizeof(p->sq_off));
3304 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
3305 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
3306 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
3307 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
3308 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
3309 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
3310 p->sq_off.array = offsetof(struct io_sq_ring, array);
3312 memset(&p->cq_off, 0, sizeof(p->cq_off));
3313 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
3314 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
3315 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
3316 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
3317 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
3318 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
3321 io_ring_ctx_wait_and_kill(ctx);
3326 * Sets up an aio uring context, and returns the fd. Applications asks for a
3327 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3328 * params structure passed in.
3330 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3332 struct io_uring_params p;
3336 if (copy_from_user(&p, params, sizeof(p)))
3338 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3343 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3344 IORING_SETUP_SQ_AFF))
3347 ret = io_uring_create(entries, &p);
3351 if (copy_to_user(params, &p, sizeof(p)))
3357 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3358 struct io_uring_params __user *, params)
3360 return io_uring_setup(entries, params);
3363 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3364 void __user *arg, unsigned nr_args)
3365 __releases(ctx->uring_lock)
3366 __acquires(ctx->uring_lock)
3371 * We're inside the ring mutex, if the ref is already dying, then
3372 * someone else killed the ctx or is already going through
3373 * io_uring_register().
3375 if (percpu_ref_is_dying(&ctx->refs))
3378 percpu_ref_kill(&ctx->refs);
3381 * Drop uring mutex before waiting for references to exit. If another
3382 * thread is currently inside io_uring_enter() it might need to grab
3383 * the uring_lock to make progress. If we hold it here across the drain
3384 * wait, then we can deadlock. It's safe to drop the mutex here, since
3385 * no new references will come in after we've killed the percpu ref.
3387 mutex_unlock(&ctx->uring_lock);
3388 wait_for_completion(&ctx->ctx_done);
3389 mutex_lock(&ctx->uring_lock);
3392 case IORING_REGISTER_BUFFERS:
3393 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3395 case IORING_UNREGISTER_BUFFERS:
3399 ret = io_sqe_buffer_unregister(ctx);
3401 case IORING_REGISTER_FILES:
3402 ret = io_sqe_files_register(ctx, arg, nr_args);
3404 case IORING_UNREGISTER_FILES:
3408 ret = io_sqe_files_unregister(ctx);
3410 case IORING_REGISTER_EVENTFD:
3414 ret = io_eventfd_register(ctx, arg);
3416 case IORING_UNREGISTER_EVENTFD:
3420 ret = io_eventfd_unregister(ctx);
3427 /* bring the ctx back to life */
3428 reinit_completion(&ctx->ctx_done);
3429 percpu_ref_reinit(&ctx->refs);
3433 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3434 void __user *, arg, unsigned int, nr_args)
3436 struct io_ring_ctx *ctx;
3445 if (f.file->f_op != &io_uring_fops)
3448 ctx = f.file->private_data;
3450 mutex_lock(&ctx->uring_lock);
3451 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3452 mutex_unlock(&ctx->uring_lock);
3458 static int __init io_uring_init(void)
3460 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3463 __initcall(io_uring_init);