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 32768
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 offsets
88 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
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 of the sq ring and the tail of the cq ring,
99 * and the application controls tail of the sq ring and the head of the
102 struct io_uring sq, cq;
104 * Bitmasks to apply to head and tail offsets (constant, equals
107 u32 sq_ring_mask, cq_ring_mask;
108 /* Ring sizes (constant, power of 2) */
109 u32 sq_ring_entries, cq_ring_entries;
111 * Number of invalid entries dropped by the kernel due to
112 * invalid index stored in array
114 * Written by the kernel, shouldn't be modified by the
115 * application (i.e. get number of "new events" by comparing to
118 * After a new SQ head value was read by the application this
119 * counter includes all submissions that were dropped reaching
120 * the new SQ head (and possibly more).
126 * Written by the kernel, shouldn't be modified by the
129 * The application needs a full memory barrier before checking
130 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
134 * Number of completion events lost because the queue was full;
135 * this should be avoided by the application by making sure
136 * there are not more requests pending thatn there is space in
137 * the completion queue.
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * As completion events come in out of order this counter is not
144 * ordered with any other data.
148 * Ring buffer of completion events.
150 * The kernel writes completion events fresh every time they are
151 * produced, so the application is allowed to modify pending
154 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
157 struct io_mapped_ubuf {
160 struct bio_vec *bvec;
161 unsigned int nr_bvecs;
167 struct list_head list;
176 struct percpu_ref refs;
177 } ____cacheline_aligned_in_smp;
185 * Ring buffer of indices into array of io_uring_sqe, which is
186 * mmapped by the application using the IORING_OFF_SQES offset.
188 * This indirection could e.g. be used to assign fixed
189 * io_uring_sqe entries to operations and only submit them to
190 * the queue when needed.
192 * The kernel modifies neither the indices array nor the entries
196 unsigned cached_sq_head;
199 unsigned sq_thread_idle;
200 unsigned cached_sq_dropped;
201 struct io_uring_sqe *sq_sqes;
203 struct list_head defer_list;
204 struct list_head timeout_list;
205 } ____cacheline_aligned_in_smp;
208 struct workqueue_struct *sqo_wq[2];
209 struct task_struct *sqo_thread; /* if using sq thread polling */
210 struct mm_struct *sqo_mm;
211 wait_queue_head_t sqo_wait;
212 struct completion sqo_thread_started;
215 unsigned cached_cq_tail;
216 atomic_t cached_cq_overflow;
219 struct wait_queue_head cq_wait;
220 struct fasync_struct *cq_fasync;
221 struct eventfd_ctx *cq_ev_fd;
222 atomic_t cq_timeouts;
223 } ____cacheline_aligned_in_smp;
225 struct io_rings *rings;
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
232 struct file **user_files;
233 unsigned nr_user_files;
235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
239 struct user_struct *user;
241 struct completion ctx_done;
244 struct mutex uring_lock;
245 wait_queue_head_t wait;
246 } ____cacheline_aligned_in_smp;
249 spinlock_t completion_lock;
250 bool poll_multi_file;
252 * ->poll_list is protected by the ctx->uring_lock for
253 * io_uring instances that don't use IORING_SETUP_SQPOLL.
254 * For SQPOLL, only the single threaded io_sq_thread() will
255 * manipulate the list, hence no extra locking is needed there.
257 struct list_head poll_list;
258 struct list_head cancel_list;
259 } ____cacheline_aligned_in_smp;
261 struct async_list pending_async[2];
263 #if defined(CONFIG_UNIX)
264 struct socket *ring_sock;
269 const struct io_uring_sqe *sqe;
270 unsigned short index;
274 bool needs_fixed_file;
278 * First field must be the file pointer in all the
279 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
281 struct io_poll_iocb {
283 struct wait_queue_head *head;
287 struct wait_queue_entry wait;
292 struct hrtimer timer;
296 * NOTE! Each of the iocb union members has the file pointer
297 * as the first entry in their struct definition. So you can
298 * access the file pointer through any of the sub-structs,
299 * or directly as just 'ki_filp' in this struct.
305 struct io_poll_iocb poll;
306 struct io_timeout timeout;
309 struct sqe_submit submit;
311 struct io_ring_ctx *ctx;
312 struct list_head list;
313 struct list_head link_list;
316 #define REQ_F_NOWAIT 1 /* must not punt to workers */
317 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
318 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
319 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
320 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
321 #define REQ_F_IO_DRAINED 32 /* drain done */
322 #define REQ_F_LINK 64 /* linked sqes */
323 #define REQ_F_LINK_DONE 128 /* linked sqes done */
324 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
325 #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */
326 #define REQ_F_TIMEOUT 1024 /* timeout request */
327 #define REQ_F_ISREG 2048 /* regular file */
328 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
333 struct work_struct work;
336 #define IO_PLUG_THRESHOLD 2
337 #define IO_IOPOLL_BATCH 8
339 struct io_submit_state {
340 struct blk_plug plug;
343 * io_kiocb alloc cache
345 void *reqs[IO_IOPOLL_BATCH];
346 unsigned int free_reqs;
347 unsigned int cur_req;
350 * File reference cache
354 unsigned int has_refs;
355 unsigned int used_refs;
356 unsigned int ios_left;
359 static void io_sq_wq_submit_work(struct work_struct *work);
360 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
362 static void __io_free_req(struct io_kiocb *req);
364 static struct kmem_cache *req_cachep;
366 static const struct file_operations io_uring_fops;
368 struct sock *io_uring_get_socket(struct file *file)
370 #if defined(CONFIG_UNIX)
371 if (file->f_op == &io_uring_fops) {
372 struct io_ring_ctx *ctx = file->private_data;
374 return ctx->ring_sock->sk;
379 EXPORT_SYMBOL(io_uring_get_socket);
381 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
383 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
385 complete(&ctx->ctx_done);
388 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
390 struct io_ring_ctx *ctx;
393 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
397 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
398 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
403 ctx->flags = p->flags;
404 init_waitqueue_head(&ctx->cq_wait);
405 init_completion(&ctx->ctx_done);
406 init_completion(&ctx->sqo_thread_started);
407 mutex_init(&ctx->uring_lock);
408 init_waitqueue_head(&ctx->wait);
409 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
410 spin_lock_init(&ctx->pending_async[i].lock);
411 INIT_LIST_HEAD(&ctx->pending_async[i].list);
412 atomic_set(&ctx->pending_async[i].cnt, 0);
414 spin_lock_init(&ctx->completion_lock);
415 INIT_LIST_HEAD(&ctx->poll_list);
416 INIT_LIST_HEAD(&ctx->cancel_list);
417 INIT_LIST_HEAD(&ctx->defer_list);
418 INIT_LIST_HEAD(&ctx->timeout_list);
422 static inline bool __io_sequence_defer(struct io_ring_ctx *ctx,
423 struct io_kiocb *req)
425 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
426 + atomic_read(&ctx->cached_cq_overflow);
429 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
430 struct io_kiocb *req)
432 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
435 return __io_sequence_defer(ctx, req);
438 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
440 struct io_kiocb *req;
442 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
443 if (req && !io_sequence_defer(ctx, req)) {
444 list_del_init(&req->list);
451 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
453 struct io_kiocb *req;
455 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
456 if (req && !__io_sequence_defer(ctx, req)) {
457 list_del_init(&req->list);
464 static void __io_commit_cqring(struct io_ring_ctx *ctx)
466 struct io_rings *rings = ctx->rings;
468 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
469 /* order cqe stores with ring update */
470 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
472 if (wq_has_sleeper(&ctx->cq_wait)) {
473 wake_up_interruptible(&ctx->cq_wait);
474 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
479 static inline void io_queue_async_work(struct io_ring_ctx *ctx,
480 struct io_kiocb *req)
484 if (req->submit.sqe) {
485 switch (req->submit.sqe->opcode) {
486 case IORING_OP_WRITEV:
487 case IORING_OP_WRITE_FIXED:
488 rw = !(req->rw.ki_flags & IOCB_DIRECT);
493 queue_work(ctx->sqo_wq[rw], &req->work);
496 static void io_kill_timeout(struct io_kiocb *req)
500 ret = hrtimer_try_to_cancel(&req->timeout.timer);
502 atomic_inc(&req->ctx->cq_timeouts);
503 list_del(&req->list);
504 io_cqring_fill_event(req->ctx, req->user_data, 0);
509 static void io_kill_timeouts(struct io_ring_ctx *ctx)
511 struct io_kiocb *req, *tmp;
513 spin_lock_irq(&ctx->completion_lock);
514 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
515 io_kill_timeout(req);
516 spin_unlock_irq(&ctx->completion_lock);
519 static void io_commit_cqring(struct io_ring_ctx *ctx)
521 struct io_kiocb *req;
523 while ((req = io_get_timeout_req(ctx)) != NULL)
524 io_kill_timeout(req);
526 __io_commit_cqring(ctx);
528 while ((req = io_get_deferred_req(ctx)) != NULL) {
529 if (req->flags & REQ_F_SHADOW_DRAIN) {
530 /* Just for drain, free it. */
534 req->flags |= REQ_F_IO_DRAINED;
535 io_queue_async_work(ctx, req);
539 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
541 struct io_rings *rings = ctx->rings;
544 tail = ctx->cached_cq_tail;
546 * writes to the cq entry need to come after reading head; the
547 * control dependency is enough as we're using WRITE_ONCE to
550 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
553 ctx->cached_cq_tail++;
554 return &rings->cqes[tail & ctx->cq_mask];
557 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
560 struct io_uring_cqe *cqe;
563 * If we can't get a cq entry, userspace overflowed the
564 * submission (by quite a lot). Increment the overflow count in
567 cqe = io_get_cqring(ctx);
569 WRITE_ONCE(cqe->user_data, ki_user_data);
570 WRITE_ONCE(cqe->res, res);
571 WRITE_ONCE(cqe->flags, 0);
573 WRITE_ONCE(ctx->rings->cq_overflow,
574 atomic_inc_return(&ctx->cached_cq_overflow));
578 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
580 if (waitqueue_active(&ctx->wait))
582 if (waitqueue_active(&ctx->sqo_wait))
583 wake_up(&ctx->sqo_wait);
585 eventfd_signal(ctx->cq_ev_fd, 1);
588 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
593 spin_lock_irqsave(&ctx->completion_lock, flags);
594 io_cqring_fill_event(ctx, user_data, res);
595 io_commit_cqring(ctx);
596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
598 io_cqring_ev_posted(ctx);
601 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
602 struct io_submit_state *state)
604 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
605 struct io_kiocb *req;
607 if (!percpu_ref_tryget(&ctx->refs))
611 req = kmem_cache_alloc(req_cachep, gfp);
614 } else if (!state->free_reqs) {
618 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
619 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
622 * Bulk alloc is all-or-nothing. If we fail to get a batch,
623 * retry single alloc to be on the safe side.
625 if (unlikely(ret <= 0)) {
626 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
631 state->free_reqs = ret - 1;
633 req = state->reqs[0];
635 req = state->reqs[state->cur_req];
643 /* one is dropped after submission, the other at completion */
644 refcount_set(&req->refs, 2);
648 percpu_ref_put(&ctx->refs);
652 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
655 kmem_cache_free_bulk(req_cachep, *nr, reqs);
656 percpu_ref_put_many(&ctx->refs, *nr);
661 static void __io_free_req(struct io_kiocb *req)
663 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
665 percpu_ref_put(&req->ctx->refs);
666 kmem_cache_free(req_cachep, req);
669 static void io_req_link_next(struct io_kiocb *req)
671 struct io_kiocb *nxt;
674 * The list should never be empty when we are called here. But could
675 * potentially happen if the chain is messed up, check to be on the
678 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
680 list_del(&nxt->list);
681 if (!list_empty(&req->link_list)) {
682 INIT_LIST_HEAD(&nxt->link_list);
683 list_splice(&req->link_list, &nxt->link_list);
684 nxt->flags |= REQ_F_LINK;
687 nxt->flags |= REQ_F_LINK_DONE;
688 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
689 io_queue_async_work(req->ctx, nxt);
694 * Called if REQ_F_LINK is set, and we fail the head request
696 static void io_fail_links(struct io_kiocb *req)
698 struct io_kiocb *link;
700 while (!list_empty(&req->link_list)) {
701 link = list_first_entry(&req->link_list, struct io_kiocb, list);
702 list_del(&link->list);
704 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
709 static void io_free_req(struct io_kiocb *req)
712 * If LINK is set, we have dependent requests in this chain. If we
713 * didn't fail this request, queue the first one up, moving any other
714 * dependencies to the next request. In case of failure, fail the rest
717 if (req->flags & REQ_F_LINK) {
718 if (req->flags & REQ_F_FAIL_LINK)
721 io_req_link_next(req);
727 static void io_put_req(struct io_kiocb *req)
729 if (refcount_dec_and_test(&req->refs))
733 static unsigned io_cqring_events(struct io_rings *rings)
735 /* See comment at the top of this file */
737 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
740 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
742 struct io_rings *rings = ctx->rings;
744 /* make sure SQ entry isn't read before tail */
745 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
749 * Find and free completed poll iocbs
751 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
752 struct list_head *done)
754 void *reqs[IO_IOPOLL_BATCH];
755 struct io_kiocb *req;
759 while (!list_empty(done)) {
760 req = list_first_entry(done, struct io_kiocb, list);
761 list_del(&req->list);
763 io_cqring_fill_event(ctx, req->user_data, req->result);
766 if (refcount_dec_and_test(&req->refs)) {
767 /* If we're not using fixed files, we have to pair the
768 * completion part with the file put. Use regular
769 * completions for those, only batch free for fixed
770 * file and non-linked commands.
772 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
774 reqs[to_free++] = req;
775 if (to_free == ARRAY_SIZE(reqs))
776 io_free_req_many(ctx, reqs, &to_free);
783 io_commit_cqring(ctx);
784 io_free_req_many(ctx, reqs, &to_free);
787 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
790 struct io_kiocb *req, *tmp;
796 * Only spin for completions if we don't have multiple devices hanging
797 * off our complete list, and we're under the requested amount.
799 spin = !ctx->poll_multi_file && *nr_events < min;
802 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
803 struct kiocb *kiocb = &req->rw;
806 * Move completed entries to our local list. If we find a
807 * request that requires polling, break out and complete
808 * the done list first, if we have entries there.
810 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
811 list_move_tail(&req->list, &done);
814 if (!list_empty(&done))
817 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
826 if (!list_empty(&done))
827 io_iopoll_complete(ctx, nr_events, &done);
833 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
834 * non-spinning poll check - we'll still enter the driver poll loop, but only
835 * as a non-spinning completion check.
837 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
840 while (!list_empty(&ctx->poll_list) && !need_resched()) {
843 ret = io_do_iopoll(ctx, nr_events, min);
846 if (!min || *nr_events >= min)
854 * We can't just wait for polled events to come to us, we have to actively
855 * find and complete them.
857 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
859 if (!(ctx->flags & IORING_SETUP_IOPOLL))
862 mutex_lock(&ctx->uring_lock);
863 while (!list_empty(&ctx->poll_list)) {
864 unsigned int nr_events = 0;
866 io_iopoll_getevents(ctx, &nr_events, 1);
869 * Ensure we allow local-to-the-cpu processing to take place,
870 * in this case we need to ensure that we reap all events.
874 mutex_unlock(&ctx->uring_lock);
877 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
880 int iters = 0, ret = 0;
886 * Don't enter poll loop if we already have events pending.
887 * If we do, we can potentially be spinning for commands that
888 * already triggered a CQE (eg in error).
890 if (io_cqring_events(ctx->rings))
894 * If a submit got punted to a workqueue, we can have the
895 * application entering polling for a command before it gets
896 * issued. That app will hold the uring_lock for the duration
897 * of the poll right here, so we need to take a breather every
898 * now and then to ensure that the issue has a chance to add
899 * the poll to the issued list. Otherwise we can spin here
900 * forever, while the workqueue is stuck trying to acquire the
903 if (!(++iters & 7)) {
904 mutex_unlock(&ctx->uring_lock);
905 mutex_lock(&ctx->uring_lock);
908 if (*nr_events < min)
909 tmin = min - *nr_events;
911 ret = io_iopoll_getevents(ctx, nr_events, tmin);
915 } while (min && !*nr_events && !need_resched());
920 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
926 * We disallow the app entering submit/complete with polling, but we
927 * still need to lock the ring to prevent racing with polled issue
928 * that got punted to a workqueue.
930 mutex_lock(&ctx->uring_lock);
931 ret = __io_iopoll_check(ctx, nr_events, min);
932 mutex_unlock(&ctx->uring_lock);
936 static void kiocb_end_write(struct io_kiocb *req)
939 * Tell lockdep we inherited freeze protection from submission
942 if (req->flags & REQ_F_ISREG) {
943 struct inode *inode = file_inode(req->file);
945 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
947 file_end_write(req->file);
950 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
952 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
954 if (kiocb->ki_flags & IOCB_WRITE)
955 kiocb_end_write(req);
957 if ((req->flags & REQ_F_LINK) && res != req->result)
958 req->flags |= REQ_F_FAIL_LINK;
959 io_cqring_add_event(req->ctx, req->user_data, res);
963 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
965 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
967 if (kiocb->ki_flags & IOCB_WRITE)
968 kiocb_end_write(req);
970 if ((req->flags & REQ_F_LINK) && res != req->result)
971 req->flags |= REQ_F_FAIL_LINK;
974 req->flags |= REQ_F_IOPOLL_COMPLETED;
978 * After the iocb has been issued, it's safe to be found on the poll list.
979 * Adding the kiocb to the list AFTER submission ensures that we don't
980 * find it from a io_iopoll_getevents() thread before the issuer is done
981 * accessing the kiocb cookie.
983 static void io_iopoll_req_issued(struct io_kiocb *req)
985 struct io_ring_ctx *ctx = req->ctx;
988 * Track whether we have multiple files in our lists. This will impact
989 * how we do polling eventually, not spinning if we're on potentially
992 if (list_empty(&ctx->poll_list)) {
993 ctx->poll_multi_file = false;
994 } else if (!ctx->poll_multi_file) {
995 struct io_kiocb *list_req;
997 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
999 if (list_req->rw.ki_filp != req->rw.ki_filp)
1000 ctx->poll_multi_file = true;
1004 * For fast devices, IO may have already completed. If it has, add
1005 * it to the front so we find it first.
1007 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1008 list_add(&req->list, &ctx->poll_list);
1010 list_add_tail(&req->list, &ctx->poll_list);
1013 static void io_file_put(struct io_submit_state *state)
1016 int diff = state->has_refs - state->used_refs;
1019 fput_many(state->file, diff);
1025 * Get as many references to a file as we have IOs left in this submission,
1026 * assuming most submissions are for one file, or at least that each file
1027 * has more than one submission.
1029 static struct file *io_file_get(struct io_submit_state *state, int fd)
1035 if (state->fd == fd) {
1042 state->file = fget_many(fd, state->ios_left);
1047 state->has_refs = state->ios_left;
1048 state->used_refs = 1;
1054 * If we tracked the file through the SCM inflight mechanism, we could support
1055 * any file. For now, just ensure that anything potentially problematic is done
1058 static bool io_file_supports_async(struct file *file)
1060 umode_t mode = file_inode(file)->i_mode;
1062 if (S_ISBLK(mode) || S_ISCHR(mode))
1064 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1070 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
1071 bool force_nonblock)
1073 const struct io_uring_sqe *sqe = s->sqe;
1074 struct io_ring_ctx *ctx = req->ctx;
1075 struct kiocb *kiocb = &req->rw;
1082 if (S_ISREG(file_inode(req->file)->i_mode))
1083 req->flags |= REQ_F_ISREG;
1086 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1087 * we know to async punt it even if it was opened O_NONBLOCK
1089 if (force_nonblock && !io_file_supports_async(req->file)) {
1090 req->flags |= REQ_F_MUST_PUNT;
1094 kiocb->ki_pos = READ_ONCE(sqe->off);
1095 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1096 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1098 ioprio = READ_ONCE(sqe->ioprio);
1100 ret = ioprio_check_cap(ioprio);
1104 kiocb->ki_ioprio = ioprio;
1106 kiocb->ki_ioprio = get_current_ioprio();
1108 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1112 /* don't allow async punt if RWF_NOWAIT was requested */
1113 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1114 (req->file->f_flags & O_NONBLOCK))
1115 req->flags |= REQ_F_NOWAIT;
1118 kiocb->ki_flags |= IOCB_NOWAIT;
1120 if (ctx->flags & IORING_SETUP_IOPOLL) {
1121 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1122 !kiocb->ki_filp->f_op->iopoll)
1125 kiocb->ki_flags |= IOCB_HIPRI;
1126 kiocb->ki_complete = io_complete_rw_iopoll;
1128 if (kiocb->ki_flags & IOCB_HIPRI)
1130 kiocb->ki_complete = io_complete_rw;
1135 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1141 case -ERESTARTNOINTR:
1142 case -ERESTARTNOHAND:
1143 case -ERESTART_RESTARTBLOCK:
1145 * We can't just restart the syscall, since previously
1146 * submitted sqes may already be in progress. Just fail this
1152 kiocb->ki_complete(kiocb, ret, 0);
1156 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1157 const struct io_uring_sqe *sqe,
1158 struct iov_iter *iter)
1160 size_t len = READ_ONCE(sqe->len);
1161 struct io_mapped_ubuf *imu;
1162 unsigned index, buf_index;
1166 /* attempt to use fixed buffers without having provided iovecs */
1167 if (unlikely(!ctx->user_bufs))
1170 buf_index = READ_ONCE(sqe->buf_index);
1171 if (unlikely(buf_index >= ctx->nr_user_bufs))
1174 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1175 imu = &ctx->user_bufs[index];
1176 buf_addr = READ_ONCE(sqe->addr);
1179 if (buf_addr + len < buf_addr)
1181 /* not inside the mapped region */
1182 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1186 * May not be a start of buffer, set size appropriately
1187 * and advance us to the beginning.
1189 offset = buf_addr - imu->ubuf;
1190 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1194 * Don't use iov_iter_advance() here, as it's really slow for
1195 * using the latter parts of a big fixed buffer - it iterates
1196 * over each segment manually. We can cheat a bit here, because
1199 * 1) it's a BVEC iter, we set it up
1200 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1201 * first and last bvec
1203 * So just find our index, and adjust the iterator afterwards.
1204 * If the offset is within the first bvec (or the whole first
1205 * bvec, just use iov_iter_advance(). This makes it easier
1206 * since we can just skip the first segment, which may not
1207 * be PAGE_SIZE aligned.
1209 const struct bio_vec *bvec = imu->bvec;
1211 if (offset <= bvec->bv_len) {
1212 iov_iter_advance(iter, offset);
1214 unsigned long seg_skip;
1216 /* skip first vec */
1217 offset -= bvec->bv_len;
1218 seg_skip = 1 + (offset >> PAGE_SHIFT);
1220 iter->bvec = bvec + seg_skip;
1221 iter->nr_segs -= seg_skip;
1222 iter->count -= bvec->bv_len + offset;
1223 iter->iov_offset = offset & ~PAGE_MASK;
1230 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1231 const struct sqe_submit *s, struct iovec **iovec,
1232 struct iov_iter *iter)
1234 const struct io_uring_sqe *sqe = s->sqe;
1235 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1236 size_t sqe_len = READ_ONCE(sqe->len);
1240 * We're reading ->opcode for the second time, but the first read
1241 * doesn't care whether it's _FIXED or not, so it doesn't matter
1242 * whether ->opcode changes concurrently. The first read does care
1243 * about whether it is a READ or a WRITE, so we don't trust this read
1244 * for that purpose and instead let the caller pass in the read/write
1247 opcode = READ_ONCE(sqe->opcode);
1248 if (opcode == IORING_OP_READ_FIXED ||
1249 opcode == IORING_OP_WRITE_FIXED) {
1250 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1258 #ifdef CONFIG_COMPAT
1260 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1264 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1267 static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb)
1269 if (al->file == kiocb->ki_filp) {
1273 * Allow merging if we're anywhere in the range of the same
1274 * page. Generally this happens for sub-page reads or writes,
1275 * and it's beneficial to allow the first worker to bring the
1276 * page in and the piggy backed work can then work on the
1279 start = al->io_start & PAGE_MASK;
1280 end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK;
1281 if (kiocb->ki_pos >= start && kiocb->ki_pos <= end)
1290 * Make a note of the last file/offset/direction we punted to async
1291 * context. We'll use this information to see if we can piggy back a
1292 * sequential request onto the previous one, if it's still hasn't been
1293 * completed by the async worker.
1295 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1297 struct async_list *async_list = &req->ctx->pending_async[rw];
1298 struct kiocb *kiocb = &req->rw;
1299 struct file *filp = kiocb->ki_filp;
1301 if (io_should_merge(async_list, kiocb)) {
1302 unsigned long max_bytes;
1304 /* Use 8x RA size as a decent limiter for both reads/writes */
1305 max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3);
1307 max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3);
1309 /* If max len are exceeded, reset the state */
1310 if (async_list->io_len + len <= max_bytes) {
1311 req->flags |= REQ_F_SEQ_PREV;
1312 async_list->io_len += len;
1314 async_list->file = NULL;
1318 /* New file? Reset state. */
1319 if (async_list->file != filp) {
1320 async_list->io_start = kiocb->ki_pos;
1321 async_list->io_len = len;
1322 async_list->file = filp;
1327 * For files that don't have ->read_iter() and ->write_iter(), handle them
1328 * by looping over ->read() or ->write() manually.
1330 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1331 struct iov_iter *iter)
1336 * Don't support polled IO through this interface, and we can't
1337 * support non-blocking either. For the latter, this just causes
1338 * the kiocb to be handled from an async context.
1340 if (kiocb->ki_flags & IOCB_HIPRI)
1342 if (kiocb->ki_flags & IOCB_NOWAIT)
1345 while (iov_iter_count(iter)) {
1346 struct iovec iovec = iov_iter_iovec(iter);
1350 nr = file->f_op->read(file, iovec.iov_base,
1351 iovec.iov_len, &kiocb->ki_pos);
1353 nr = file->f_op->write(file, iovec.iov_base,
1354 iovec.iov_len, &kiocb->ki_pos);
1363 if (nr != iovec.iov_len)
1365 iov_iter_advance(iter, nr);
1371 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1372 bool force_nonblock)
1374 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1375 struct kiocb *kiocb = &req->rw;
1376 struct iov_iter iter;
1379 ssize_t read_size, ret;
1381 ret = io_prep_rw(req, s, force_nonblock);
1384 file = kiocb->ki_filp;
1386 if (unlikely(!(file->f_mode & FMODE_READ)))
1389 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1394 if (req->flags & REQ_F_LINK)
1395 req->result = read_size;
1397 iov_count = iov_iter_count(&iter);
1398 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1402 if (file->f_op->read_iter)
1403 ret2 = call_read_iter(file, kiocb, &iter);
1405 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1408 * In case of a short read, punt to async. This can happen
1409 * if we have data partially cached. Alternatively we can
1410 * return the short read, in which case the application will
1411 * need to issue another SQE and wait for it. That SQE will
1412 * need async punt anyway, so it's more efficient to do it
1415 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1416 (req->flags & REQ_F_ISREG) &&
1417 ret2 > 0 && ret2 < read_size)
1419 /* Catch -EAGAIN return for forced non-blocking submission */
1420 if (!force_nonblock || ret2 != -EAGAIN) {
1421 io_rw_done(kiocb, ret2);
1424 * If ->needs_lock is true, we're already in async
1428 io_async_list_note(READ, req, iov_count);
1436 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1437 bool force_nonblock)
1439 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1440 struct kiocb *kiocb = &req->rw;
1441 struct iov_iter iter;
1446 ret = io_prep_rw(req, s, force_nonblock);
1450 file = kiocb->ki_filp;
1451 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1454 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1458 if (req->flags & REQ_F_LINK)
1461 iov_count = iov_iter_count(&iter);
1464 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1465 /* If ->needs_lock is true, we're already in async context. */
1467 io_async_list_note(WRITE, req, iov_count);
1471 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1476 * Open-code file_start_write here to grab freeze protection,
1477 * which will be released by another thread in
1478 * io_complete_rw(). Fool lockdep by telling it the lock got
1479 * released so that it doesn't complain about the held lock when
1480 * we return to userspace.
1482 if (req->flags & REQ_F_ISREG) {
1483 __sb_start_write(file_inode(file)->i_sb,
1484 SB_FREEZE_WRITE, true);
1485 __sb_writers_release(file_inode(file)->i_sb,
1488 kiocb->ki_flags |= IOCB_WRITE;
1490 if (file->f_op->write_iter)
1491 ret2 = call_write_iter(file, kiocb, &iter);
1493 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1494 if (!force_nonblock || ret2 != -EAGAIN) {
1495 io_rw_done(kiocb, ret2);
1498 * If ->needs_lock is true, we're already in async
1502 io_async_list_note(WRITE, req, iov_count);
1512 * IORING_OP_NOP just posts a completion event, nothing else.
1514 static int io_nop(struct io_kiocb *req, u64 user_data)
1516 struct io_ring_ctx *ctx = req->ctx;
1519 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1522 io_cqring_add_event(ctx, user_data, err);
1527 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1529 struct io_ring_ctx *ctx = req->ctx;
1534 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1536 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1542 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1543 bool force_nonblock)
1545 loff_t sqe_off = READ_ONCE(sqe->off);
1546 loff_t sqe_len = READ_ONCE(sqe->len);
1547 loff_t end = sqe_off + sqe_len;
1548 unsigned fsync_flags;
1551 fsync_flags = READ_ONCE(sqe->fsync_flags);
1552 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1555 ret = io_prep_fsync(req, sqe);
1559 /* fsync always requires a blocking context */
1563 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1564 end > 0 ? end : LLONG_MAX,
1565 fsync_flags & IORING_FSYNC_DATASYNC);
1567 if (ret < 0 && (req->flags & REQ_F_LINK))
1568 req->flags |= REQ_F_FAIL_LINK;
1569 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1574 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1576 struct io_ring_ctx *ctx = req->ctx;
1582 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1584 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1590 static int io_sync_file_range(struct io_kiocb *req,
1591 const struct io_uring_sqe *sqe,
1592 bool force_nonblock)
1599 ret = io_prep_sfr(req, sqe);
1603 /* sync_file_range always requires a blocking context */
1607 sqe_off = READ_ONCE(sqe->off);
1608 sqe_len = READ_ONCE(sqe->len);
1609 flags = READ_ONCE(sqe->sync_range_flags);
1611 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1613 if (ret < 0 && (req->flags & REQ_F_LINK))
1614 req->flags |= REQ_F_FAIL_LINK;
1615 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1620 #if defined(CONFIG_NET)
1621 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1622 bool force_nonblock,
1623 long (*fn)(struct socket *, struct user_msghdr __user *,
1626 struct socket *sock;
1629 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1632 sock = sock_from_file(req->file, &ret);
1634 struct user_msghdr __user *msg;
1637 flags = READ_ONCE(sqe->msg_flags);
1638 if (flags & MSG_DONTWAIT)
1639 req->flags |= REQ_F_NOWAIT;
1640 else if (force_nonblock)
1641 flags |= MSG_DONTWAIT;
1643 msg = (struct user_msghdr __user *) (unsigned long)
1644 READ_ONCE(sqe->addr);
1646 ret = fn(sock, msg, flags);
1647 if (force_nonblock && ret == -EAGAIN)
1651 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1657 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1658 bool force_nonblock)
1660 #if defined(CONFIG_NET)
1661 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1667 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1668 bool force_nonblock)
1670 #if defined(CONFIG_NET)
1671 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1677 static void io_poll_remove_one(struct io_kiocb *req)
1679 struct io_poll_iocb *poll = &req->poll;
1681 spin_lock(&poll->head->lock);
1682 WRITE_ONCE(poll->canceled, true);
1683 if (!list_empty(&poll->wait.entry)) {
1684 list_del_init(&poll->wait.entry);
1685 io_queue_async_work(req->ctx, req);
1687 spin_unlock(&poll->head->lock);
1689 list_del_init(&req->list);
1692 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1694 struct io_kiocb *req;
1696 spin_lock_irq(&ctx->completion_lock);
1697 while (!list_empty(&ctx->cancel_list)) {
1698 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1699 io_poll_remove_one(req);
1701 spin_unlock_irq(&ctx->completion_lock);
1705 * Find a running poll command that matches one specified in sqe->addr,
1706 * and remove it if found.
1708 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1710 struct io_ring_ctx *ctx = req->ctx;
1711 struct io_kiocb *poll_req, *next;
1714 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1716 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1720 spin_lock_irq(&ctx->completion_lock);
1721 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1722 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1723 io_poll_remove_one(poll_req);
1728 spin_unlock_irq(&ctx->completion_lock);
1730 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1735 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1738 req->poll.done = true;
1739 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1740 io_commit_cqring(ctx);
1743 static void io_poll_complete_work(struct work_struct *work)
1745 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1746 struct io_poll_iocb *poll = &req->poll;
1747 struct poll_table_struct pt = { ._key = poll->events };
1748 struct io_ring_ctx *ctx = req->ctx;
1751 if (!READ_ONCE(poll->canceled))
1752 mask = vfs_poll(poll->file, &pt) & poll->events;
1755 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1756 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1757 * synchronize with them. In the cancellation case the list_del_init
1758 * itself is not actually needed, but harmless so we keep it in to
1759 * avoid further branches in the fast path.
1761 spin_lock_irq(&ctx->completion_lock);
1762 if (!mask && !READ_ONCE(poll->canceled)) {
1763 add_wait_queue(poll->head, &poll->wait);
1764 spin_unlock_irq(&ctx->completion_lock);
1767 list_del_init(&req->list);
1768 io_poll_complete(ctx, req, mask);
1769 spin_unlock_irq(&ctx->completion_lock);
1771 io_cqring_ev_posted(ctx);
1775 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1778 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1780 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1781 struct io_ring_ctx *ctx = req->ctx;
1782 __poll_t mask = key_to_poll(key);
1783 unsigned long flags;
1785 /* for instances that support it check for an event match first: */
1786 if (mask && !(mask & poll->events))
1789 list_del_init(&poll->wait.entry);
1791 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1792 list_del(&req->list);
1793 io_poll_complete(ctx, req, mask);
1794 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1796 io_cqring_ev_posted(ctx);
1799 io_queue_async_work(ctx, req);
1805 struct io_poll_table {
1806 struct poll_table_struct pt;
1807 struct io_kiocb *req;
1811 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1812 struct poll_table_struct *p)
1814 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1816 if (unlikely(pt->req->poll.head)) {
1817 pt->error = -EINVAL;
1822 pt->req->poll.head = head;
1823 add_wait_queue(head, &pt->req->poll.wait);
1826 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1828 struct io_poll_iocb *poll = &req->poll;
1829 struct io_ring_ctx *ctx = req->ctx;
1830 struct io_poll_table ipt;
1831 bool cancel = false;
1835 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1837 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1842 req->submit.sqe = NULL;
1843 INIT_WORK(&req->work, io_poll_complete_work);
1844 events = READ_ONCE(sqe->poll_events);
1845 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1849 poll->canceled = false;
1851 ipt.pt._qproc = io_poll_queue_proc;
1852 ipt.pt._key = poll->events;
1854 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1856 /* initialized the list so that we can do list_empty checks */
1857 INIT_LIST_HEAD(&poll->wait.entry);
1858 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1860 INIT_LIST_HEAD(&req->list);
1862 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1864 spin_lock_irq(&ctx->completion_lock);
1865 if (likely(poll->head)) {
1866 spin_lock(&poll->head->lock);
1867 if (unlikely(list_empty(&poll->wait.entry))) {
1873 if (mask || ipt.error)
1874 list_del_init(&poll->wait.entry);
1876 WRITE_ONCE(poll->canceled, true);
1877 else if (!poll->done) /* actually waiting for an event */
1878 list_add_tail(&req->list, &ctx->cancel_list);
1879 spin_unlock(&poll->head->lock);
1881 if (mask) { /* no async, we'd stolen it */
1883 io_poll_complete(ctx, req, mask);
1885 spin_unlock_irq(&ctx->completion_lock);
1888 io_cqring_ev_posted(ctx);
1894 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
1896 struct io_ring_ctx *ctx;
1897 struct io_kiocb *req, *prev;
1898 unsigned long flags;
1900 req = container_of(timer, struct io_kiocb, timeout.timer);
1902 atomic_inc(&ctx->cq_timeouts);
1904 spin_lock_irqsave(&ctx->completion_lock, flags);
1906 * Adjust the reqs sequence before the current one because it
1907 * will consume a slot in the cq_ring and the the cq_tail pointer
1908 * will be increased, otherwise other timeout reqs may return in
1909 * advance without waiting for enough wait_nr.
1912 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
1914 list_del(&req->list);
1916 io_cqring_fill_event(ctx, req->user_data, -ETIME);
1917 io_commit_cqring(ctx);
1918 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1920 io_cqring_ev_posted(ctx);
1923 return HRTIMER_NORESTART;
1926 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1929 struct io_ring_ctx *ctx = req->ctx;
1930 struct list_head *entry;
1931 struct timespec64 ts;
1934 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1936 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->timeout_flags ||
1940 if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr)))
1944 * sqe->off holds how many events that need to occur for this
1945 * timeout event to be satisfied.
1947 count = READ_ONCE(sqe->off);
1951 req->sequence = ctx->cached_sq_head + count - 1;
1952 /* reuse it to store the count */
1953 req->submit.sequence = count;
1954 req->flags |= REQ_F_TIMEOUT;
1957 * Insertion sort, ensuring the first entry in the list is always
1958 * the one we need first.
1960 spin_lock_irq(&ctx->completion_lock);
1961 list_for_each_prev(entry, &ctx->timeout_list) {
1962 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
1963 unsigned nxt_sq_head;
1964 long long tmp, tmp_nxt;
1967 * Since cached_sq_head + count - 1 can overflow, use type long
1970 tmp = (long long)ctx->cached_sq_head + count - 1;
1971 nxt_sq_head = nxt->sequence - nxt->submit.sequence + 1;
1972 tmp_nxt = (long long)nxt_sq_head + nxt->submit.sequence - 1;
1975 * cached_sq_head may overflow, and it will never overflow twice
1976 * once there is some timeout req still be valid.
1978 if (ctx->cached_sq_head < nxt_sq_head)
1985 * Sequence of reqs after the insert one and itself should
1986 * be adjusted because each timeout req consumes a slot.
1991 req->sequence -= span;
1992 list_add(&req->list, entry);
1993 spin_unlock_irq(&ctx->completion_lock);
1995 hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1996 req->timeout.timer.function = io_timeout_fn;
1997 hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts),
2002 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
2003 const struct io_uring_sqe *sqe)
2005 struct io_uring_sqe *sqe_copy;
2007 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
2010 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
2014 spin_lock_irq(&ctx->completion_lock);
2015 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
2016 spin_unlock_irq(&ctx->completion_lock);
2021 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
2022 req->submit.sqe = sqe_copy;
2024 INIT_WORK(&req->work, io_sq_wq_submit_work);
2025 list_add_tail(&req->list, &ctx->defer_list);
2026 spin_unlock_irq(&ctx->completion_lock);
2027 return -EIOCBQUEUED;
2030 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2031 const struct sqe_submit *s, bool force_nonblock)
2035 req->user_data = READ_ONCE(s->sqe->user_data);
2037 if (unlikely(s->index >= ctx->sq_entries))
2040 opcode = READ_ONCE(s->sqe->opcode);
2043 ret = io_nop(req, req->user_data);
2045 case IORING_OP_READV:
2046 if (unlikely(s->sqe->buf_index))
2048 ret = io_read(req, s, force_nonblock);
2050 case IORING_OP_WRITEV:
2051 if (unlikely(s->sqe->buf_index))
2053 ret = io_write(req, s, force_nonblock);
2055 case IORING_OP_READ_FIXED:
2056 ret = io_read(req, s, force_nonblock);
2058 case IORING_OP_WRITE_FIXED:
2059 ret = io_write(req, s, force_nonblock);
2061 case IORING_OP_FSYNC:
2062 ret = io_fsync(req, s->sqe, force_nonblock);
2064 case IORING_OP_POLL_ADD:
2065 ret = io_poll_add(req, s->sqe);
2067 case IORING_OP_POLL_REMOVE:
2068 ret = io_poll_remove(req, s->sqe);
2070 case IORING_OP_SYNC_FILE_RANGE:
2071 ret = io_sync_file_range(req, s->sqe, force_nonblock);
2073 case IORING_OP_SENDMSG:
2074 ret = io_sendmsg(req, s->sqe, force_nonblock);
2076 case IORING_OP_RECVMSG:
2077 ret = io_recvmsg(req, s->sqe, force_nonblock);
2079 case IORING_OP_TIMEOUT:
2080 ret = io_timeout(req, s->sqe);
2090 if (ctx->flags & IORING_SETUP_IOPOLL) {
2091 if (req->result == -EAGAIN)
2094 /* workqueue context doesn't hold uring_lock, grab it now */
2096 mutex_lock(&ctx->uring_lock);
2097 io_iopoll_req_issued(req);
2099 mutex_unlock(&ctx->uring_lock);
2105 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
2106 const struct io_uring_sqe *sqe)
2108 switch (sqe->opcode) {
2109 case IORING_OP_READV:
2110 case IORING_OP_READ_FIXED:
2111 return &ctx->pending_async[READ];
2112 case IORING_OP_WRITEV:
2113 case IORING_OP_WRITE_FIXED:
2114 return &ctx->pending_async[WRITE];
2120 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
2122 u8 opcode = READ_ONCE(sqe->opcode);
2124 return !(opcode == IORING_OP_READ_FIXED ||
2125 opcode == IORING_OP_WRITE_FIXED);
2128 static void io_sq_wq_submit_work(struct work_struct *work)
2130 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2131 struct io_ring_ctx *ctx = req->ctx;
2132 struct mm_struct *cur_mm = NULL;
2133 struct async_list *async_list;
2134 LIST_HEAD(req_list);
2135 mm_segment_t old_fs;
2138 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
2141 struct sqe_submit *s = &req->submit;
2142 const struct io_uring_sqe *sqe = s->sqe;
2143 unsigned int flags = req->flags;
2145 /* Ensure we clear previously set non-block flag */
2146 req->rw.ki_flags &= ~IOCB_NOWAIT;
2149 if (io_sqe_needs_user(sqe) && !cur_mm) {
2150 if (!mmget_not_zero(ctx->sqo_mm)) {
2153 cur_mm = ctx->sqo_mm;
2161 s->has_user = cur_mm != NULL;
2162 s->needs_lock = true;
2164 ret = __io_submit_sqe(ctx, req, s, false);
2166 * We can get EAGAIN for polled IO even though
2167 * we're forcing a sync submission from here,
2168 * since we can't wait for request slots on the
2177 /* drop submission reference */
2181 io_cqring_add_event(ctx, sqe->user_data, ret);
2185 /* async context always use a copy of the sqe */
2188 /* req from defer and link list needn't decrease async cnt */
2189 if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE))
2194 if (!list_empty(&req_list)) {
2195 req = list_first_entry(&req_list, struct io_kiocb,
2197 list_del(&req->list);
2200 if (list_empty(&async_list->list))
2204 spin_lock(&async_list->lock);
2205 if (list_empty(&async_list->list)) {
2206 spin_unlock(&async_list->lock);
2209 list_splice_init(&async_list->list, &req_list);
2210 spin_unlock(&async_list->lock);
2212 req = list_first_entry(&req_list, struct io_kiocb, list);
2213 list_del(&req->list);
2217 * Rare case of racing with a submitter. If we find the count has
2218 * dropped to zero AND we have pending work items, then restart
2219 * the processing. This is a tiny race window.
2222 ret = atomic_dec_return(&async_list->cnt);
2223 while (!ret && !list_empty(&async_list->list)) {
2224 spin_lock(&async_list->lock);
2225 atomic_inc(&async_list->cnt);
2226 list_splice_init(&async_list->list, &req_list);
2227 spin_unlock(&async_list->lock);
2229 if (!list_empty(&req_list)) {
2230 req = list_first_entry(&req_list,
2231 struct io_kiocb, list);
2232 list_del(&req->list);
2235 ret = atomic_dec_return(&async_list->cnt);
2248 * See if we can piggy back onto previously submitted work, that is still
2249 * running. We currently only allow this if the new request is sequential
2250 * to the previous one we punted.
2252 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
2258 if (!(req->flags & REQ_F_SEQ_PREV))
2260 if (!atomic_read(&list->cnt))
2264 spin_lock(&list->lock);
2265 list_add_tail(&req->list, &list->list);
2267 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
2270 if (!atomic_read(&list->cnt)) {
2271 list_del_init(&req->list);
2274 spin_unlock(&list->lock);
2278 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2280 int op = READ_ONCE(sqe->opcode);
2284 case IORING_OP_POLL_REMOVE:
2291 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2292 struct io_submit_state *state, struct io_kiocb *req)
2297 flags = READ_ONCE(s->sqe->flags);
2298 fd = READ_ONCE(s->sqe->fd);
2300 if (flags & IOSQE_IO_DRAIN)
2301 req->flags |= REQ_F_IO_DRAIN;
2303 * All io need record the previous position, if LINK vs DARIN,
2304 * it can be used to mark the position of the first IO in the
2307 req->sequence = s->sequence;
2309 if (!io_op_needs_file(s->sqe))
2312 if (flags & IOSQE_FIXED_FILE) {
2313 if (unlikely(!ctx->user_files ||
2314 (unsigned) fd >= ctx->nr_user_files))
2316 req->file = ctx->user_files[fd];
2317 req->flags |= REQ_F_FIXED_FILE;
2319 if (s->needs_fixed_file)
2321 req->file = io_file_get(state, fd);
2322 if (unlikely(!req->file))
2329 static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2330 struct sqe_submit *s)
2334 ret = __io_submit_sqe(ctx, req, s, true);
2337 * We async punt it if the file wasn't marked NOWAIT, or if the file
2338 * doesn't support non-blocking read/write attempts
2340 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
2341 (req->flags & REQ_F_MUST_PUNT))) {
2342 struct io_uring_sqe *sqe_copy;
2344 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2346 struct async_list *list;
2349 memcpy(&req->submit, s, sizeof(*s));
2350 list = io_async_list_from_sqe(ctx, s->sqe);
2351 if (!io_add_to_prev_work(list, req)) {
2353 atomic_inc(&list->cnt);
2354 INIT_WORK(&req->work, io_sq_wq_submit_work);
2355 io_queue_async_work(ctx, req);
2359 * Queued up for async execution, worker will release
2360 * submit reference when the iocb is actually submitted.
2366 /* drop submission reference */
2369 /* and drop final reference, if we failed */
2371 io_cqring_add_event(ctx, req->user_data, ret);
2372 if (req->flags & REQ_F_LINK)
2373 req->flags |= REQ_F_FAIL_LINK;
2380 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2381 struct sqe_submit *s)
2385 ret = io_req_defer(ctx, req, s->sqe);
2387 if (ret != -EIOCBQUEUED) {
2389 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2394 return __io_queue_sqe(ctx, req, s);
2397 static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
2398 struct sqe_submit *s, struct io_kiocb *shadow)
2401 int need_submit = false;
2404 return io_queue_sqe(ctx, req, s);
2407 * Mark the first IO in link list as DRAIN, let all the following
2408 * IOs enter the defer list. all IO needs to be completed before link
2411 req->flags |= REQ_F_IO_DRAIN;
2412 ret = io_req_defer(ctx, req, s->sqe);
2414 if (ret != -EIOCBQUEUED) {
2416 __io_free_req(shadow);
2417 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2422 * If ret == 0 means that all IOs in front of link io are
2423 * running done. let's queue link head.
2428 /* Insert shadow req to defer_list, blocking next IOs */
2429 spin_lock_irq(&ctx->completion_lock);
2430 list_add_tail(&shadow->list, &ctx->defer_list);
2431 spin_unlock_irq(&ctx->completion_lock);
2434 return __io_queue_sqe(ctx, req, s);
2439 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2441 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2442 struct io_submit_state *state, struct io_kiocb **link)
2444 struct io_uring_sqe *sqe_copy;
2445 struct io_kiocb *req;
2448 /* enforce forwards compatibility on users */
2449 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2454 req = io_get_req(ctx, state);
2455 if (unlikely(!req)) {
2460 ret = io_req_set_file(ctx, s, state, req);
2461 if (unlikely(ret)) {
2465 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2469 req->user_data = s->sqe->user_data;
2472 * If we already have a head request, queue this one for async
2473 * submittal once the head completes. If we don't have a head but
2474 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2475 * submitted sync once the chain is complete. If none of those
2476 * conditions are true (normal request), then just queue it.
2479 struct io_kiocb *prev = *link;
2481 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2488 memcpy(&req->submit, s, sizeof(*s));
2489 list_add_tail(&req->list, &prev->link_list);
2490 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2491 req->flags |= REQ_F_LINK;
2493 memcpy(&req->submit, s, sizeof(*s));
2494 INIT_LIST_HEAD(&req->link_list);
2497 io_queue_sqe(ctx, req, s);
2502 * Batched submission is done, ensure local IO is flushed out.
2504 static void io_submit_state_end(struct io_submit_state *state)
2506 blk_finish_plug(&state->plug);
2508 if (state->free_reqs)
2509 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2510 &state->reqs[state->cur_req]);
2514 * Start submission side cache.
2516 static void io_submit_state_start(struct io_submit_state *state,
2517 struct io_ring_ctx *ctx, unsigned max_ios)
2519 blk_start_plug(&state->plug);
2520 state->free_reqs = 0;
2522 state->ios_left = max_ios;
2525 static void io_commit_sqring(struct io_ring_ctx *ctx)
2527 struct io_rings *rings = ctx->rings;
2529 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
2531 * Ensure any loads from the SQEs are done at this point,
2532 * since once we write the new head, the application could
2533 * write new data to them.
2535 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2540 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2541 * that is mapped by userspace. This means that care needs to be taken to
2542 * ensure that reads are stable, as we cannot rely on userspace always
2543 * being a good citizen. If members of the sqe are validated and then later
2544 * used, it's important that those reads are done through READ_ONCE() to
2545 * prevent a re-load down the line.
2547 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2549 struct io_rings *rings = ctx->rings;
2550 u32 *sq_array = ctx->sq_array;
2554 * The cached sq head (or cq tail) serves two purposes:
2556 * 1) allows us to batch the cost of updating the user visible
2558 * 2) allows the kernel side to track the head on its own, even
2559 * though the application is the one updating it.
2561 head = ctx->cached_sq_head;
2562 /* make sure SQ entry isn't read before tail */
2563 if (head == smp_load_acquire(&rings->sq.tail))
2566 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
2567 if (head < ctx->sq_entries) {
2569 s->sqe = &ctx->sq_sqes[head];
2570 s->sequence = ctx->cached_sq_head;
2571 ctx->cached_sq_head++;
2575 /* drop invalid entries */
2576 ctx->cached_sq_head++;
2577 ctx->cached_sq_dropped++;
2578 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
2582 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
2583 bool has_user, bool mm_fault)
2585 struct io_submit_state state, *statep = NULL;
2586 struct io_kiocb *link = NULL;
2587 struct io_kiocb *shadow_req = NULL;
2588 bool prev_was_link = false;
2589 int i, submitted = 0;
2591 if (nr > IO_PLUG_THRESHOLD) {
2592 io_submit_state_start(&state, ctx, nr);
2596 for (i = 0; i < nr; i++) {
2597 struct sqe_submit s;
2599 if (!io_get_sqring(ctx, &s))
2603 * If previous wasn't linked and we have a linked command,
2604 * that's the end of the chain. Submit the previous link.
2606 if (!prev_was_link && link) {
2607 io_queue_link_head(ctx, link, &link->submit, shadow_req);
2611 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2613 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
2615 shadow_req = io_get_req(ctx, NULL);
2616 if (unlikely(!shadow_req))
2618 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2619 refcount_dec(&shadow_req->refs);
2621 shadow_req->sequence = s.sequence;
2625 if (unlikely(mm_fault)) {
2626 io_cqring_add_event(ctx, s.sqe->user_data,
2629 s.has_user = has_user;
2630 s.needs_lock = true;
2631 s.needs_fixed_file = true;
2632 io_submit_sqe(ctx, &s, statep, &link);
2638 io_queue_link_head(ctx, link, &link->submit, shadow_req);
2640 io_submit_state_end(&state);
2645 static int io_sq_thread(void *data)
2647 struct io_ring_ctx *ctx = data;
2648 struct mm_struct *cur_mm = NULL;
2649 mm_segment_t old_fs;
2652 unsigned long timeout;
2654 complete(&ctx->sqo_thread_started);
2659 timeout = inflight = 0;
2660 while (!kthread_should_park()) {
2661 bool mm_fault = false;
2662 unsigned int to_submit;
2665 unsigned nr_events = 0;
2667 if (ctx->flags & IORING_SETUP_IOPOLL) {
2669 * inflight is the count of the maximum possible
2670 * entries we submitted, but it can be smaller
2671 * if we dropped some of them. If we don't have
2672 * poll entries available, then we know that we
2673 * have nothing left to poll for. Reset the
2674 * inflight count to zero in that case.
2676 mutex_lock(&ctx->uring_lock);
2677 if (!list_empty(&ctx->poll_list))
2678 __io_iopoll_check(ctx, &nr_events, 0);
2681 mutex_unlock(&ctx->uring_lock);
2684 * Normal IO, just pretend everything completed.
2685 * We don't have to poll completions for that.
2687 nr_events = inflight;
2690 inflight -= nr_events;
2692 timeout = jiffies + ctx->sq_thread_idle;
2695 to_submit = io_sqring_entries(ctx);
2698 * We're polling. If we're within the defined idle
2699 * period, then let us spin without work before going
2702 if (inflight || !time_after(jiffies, timeout)) {
2708 * Drop cur_mm before scheduling, we can't hold it for
2709 * long periods (or over schedule()). Do this before
2710 * adding ourselves to the waitqueue, as the unuse/drop
2719 prepare_to_wait(&ctx->sqo_wait, &wait,
2720 TASK_INTERRUPTIBLE);
2722 /* Tell userspace we may need a wakeup call */
2723 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
2724 /* make sure to read SQ tail after writing flags */
2727 to_submit = io_sqring_entries(ctx);
2729 if (kthread_should_park()) {
2730 finish_wait(&ctx->sqo_wait, &wait);
2733 if (signal_pending(current))
2734 flush_signals(current);
2736 finish_wait(&ctx->sqo_wait, &wait);
2738 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2741 finish_wait(&ctx->sqo_wait, &wait);
2743 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2746 /* Unless all new commands are FIXED regions, grab mm */
2748 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2750 use_mm(ctx->sqo_mm);
2751 cur_mm = ctx->sqo_mm;
2755 to_submit = min(to_submit, ctx->sq_entries);
2756 inflight += io_submit_sqes(ctx, to_submit, cur_mm != NULL,
2759 /* Commit SQ ring head once we've consumed all SQEs */
2760 io_commit_sqring(ctx);
2774 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2776 struct io_submit_state state, *statep = NULL;
2777 struct io_kiocb *link = NULL;
2778 struct io_kiocb *shadow_req = NULL;
2779 bool prev_was_link = false;
2782 if (to_submit > IO_PLUG_THRESHOLD) {
2783 io_submit_state_start(&state, ctx, to_submit);
2787 for (i = 0; i < to_submit; i++) {
2788 struct sqe_submit s;
2790 if (!io_get_sqring(ctx, &s))
2794 * If previous wasn't linked and we have a linked command,
2795 * that's the end of the chain. Submit the previous link.
2797 if (!prev_was_link && link) {
2798 io_queue_link_head(ctx, link, &link->submit, shadow_req);
2802 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2804 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
2806 shadow_req = io_get_req(ctx, NULL);
2807 if (unlikely(!shadow_req))
2809 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2810 refcount_dec(&shadow_req->refs);
2812 shadow_req->sequence = s.sequence;
2817 s.needs_lock = false;
2818 s.needs_fixed_file = false;
2820 io_submit_sqe(ctx, &s, statep, &link);
2824 io_queue_link_head(ctx, link, &link->submit, shadow_req);
2826 io_submit_state_end(statep);
2828 io_commit_sqring(ctx);
2833 struct io_wait_queue {
2834 struct wait_queue_entry wq;
2835 struct io_ring_ctx *ctx;
2837 unsigned nr_timeouts;
2840 static inline bool io_should_wake(struct io_wait_queue *iowq)
2842 struct io_ring_ctx *ctx = iowq->ctx;
2845 * Wake up if we have enough events, or if a timeout occured since we
2846 * started waiting. For timeouts, we always want to return to userspace,
2847 * regardless of event count.
2849 return io_cqring_events(ctx->rings) >= iowq->to_wait ||
2850 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2853 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2854 int wake_flags, void *key)
2856 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2859 if (!io_should_wake(iowq))
2862 return autoremove_wake_function(curr, mode, wake_flags, key);
2866 * Wait until events become available, if we don't already have some. The
2867 * application must reap them itself, as they reside on the shared cq ring.
2869 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2870 const sigset_t __user *sig, size_t sigsz)
2872 struct io_wait_queue iowq = {
2875 .func = io_wake_function,
2876 .entry = LIST_HEAD_INIT(iowq.wq.entry),
2879 .to_wait = min_events,
2881 struct io_rings *rings = ctx->rings;
2884 if (io_cqring_events(rings) >= min_events)
2888 #ifdef CONFIG_COMPAT
2889 if (in_compat_syscall())
2890 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2894 ret = set_user_sigmask(sig, sigsz);
2901 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2903 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
2904 TASK_INTERRUPTIBLE);
2905 if (io_should_wake(&iowq))
2908 if (signal_pending(current)) {
2913 finish_wait(&ctx->wait, &iowq.wq);
2915 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
2916 if (ret == -ERESTARTSYS)
2919 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2922 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2924 #if defined(CONFIG_UNIX)
2925 if (ctx->ring_sock) {
2926 struct sock *sock = ctx->ring_sock->sk;
2927 struct sk_buff *skb;
2929 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2935 for (i = 0; i < ctx->nr_user_files; i++)
2936 fput(ctx->user_files[i]);
2940 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2942 if (!ctx->user_files)
2945 __io_sqe_files_unregister(ctx);
2946 kfree(ctx->user_files);
2947 ctx->user_files = NULL;
2948 ctx->nr_user_files = 0;
2952 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2954 if (ctx->sqo_thread) {
2955 wait_for_completion(&ctx->sqo_thread_started);
2957 * The park is a bit of a work-around, without it we get
2958 * warning spews on shutdown with SQPOLL set and affinity
2959 * set to a single CPU.
2961 kthread_park(ctx->sqo_thread);
2962 kthread_stop(ctx->sqo_thread);
2963 ctx->sqo_thread = NULL;
2967 static void io_finish_async(struct io_ring_ctx *ctx)
2971 io_sq_thread_stop(ctx);
2973 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) {
2974 if (ctx->sqo_wq[i]) {
2975 destroy_workqueue(ctx->sqo_wq[i]);
2976 ctx->sqo_wq[i] = NULL;
2981 #if defined(CONFIG_UNIX)
2982 static void io_destruct_skb(struct sk_buff *skb)
2984 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2987 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++)
2989 flush_workqueue(ctx->sqo_wq[i]);
2991 unix_destruct_scm(skb);
2995 * Ensure the UNIX gc is aware of our file set, so we are certain that
2996 * the io_uring can be safely unregistered on process exit, even if we have
2997 * loops in the file referencing.
2999 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3001 struct sock *sk = ctx->ring_sock->sk;
3002 struct scm_fp_list *fpl;
3003 struct sk_buff *skb;
3006 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3007 unsigned long inflight = ctx->user->unix_inflight + nr;
3009 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3013 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3017 skb = alloc_skb(0, GFP_KERNEL);
3024 skb->destructor = io_destruct_skb;
3026 fpl->user = get_uid(ctx->user);
3027 for (i = 0; i < nr; i++) {
3028 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
3029 unix_inflight(fpl->user, fpl->fp[i]);
3032 fpl->max = fpl->count = nr;
3033 UNIXCB(skb).fp = fpl;
3034 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3035 skb_queue_head(&sk->sk_receive_queue, skb);
3037 for (i = 0; i < nr; i++)
3044 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3045 * causes regular reference counting to break down. We rely on the UNIX
3046 * garbage collection to take care of this problem for us.
3048 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3050 unsigned left, total;
3054 left = ctx->nr_user_files;
3056 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
3058 ret = __io_sqe_files_scm(ctx, this_files, total);
3062 total += this_files;
3068 while (total < ctx->nr_user_files) {
3069 fput(ctx->user_files[total]);
3076 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3082 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3085 __s32 __user *fds = (__s32 __user *) arg;
3089 if (ctx->user_files)
3093 if (nr_args > IORING_MAX_FIXED_FILES)
3096 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
3097 if (!ctx->user_files)
3100 for (i = 0; i < nr_args; i++) {
3102 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3105 ctx->user_files[i] = fget(fd);
3108 if (!ctx->user_files[i])
3111 * Don't allow io_uring instances to be registered. If UNIX
3112 * isn't enabled, then this causes a reference cycle and this
3113 * instance can never get freed. If UNIX is enabled we'll
3114 * handle it just fine, but there's still no point in allowing
3115 * a ring fd as it doesn't support regular read/write anyway.
3117 if (ctx->user_files[i]->f_op == &io_uring_fops) {
3118 fput(ctx->user_files[i]);
3121 ctx->nr_user_files++;
3126 for (i = 0; i < ctx->nr_user_files; i++)
3127 fput(ctx->user_files[i]);
3129 kfree(ctx->user_files);
3130 ctx->user_files = NULL;
3131 ctx->nr_user_files = 0;
3135 ret = io_sqe_files_scm(ctx);
3137 io_sqe_files_unregister(ctx);
3142 static int io_sq_offload_start(struct io_ring_ctx *ctx,
3143 struct io_uring_params *p)
3147 init_waitqueue_head(&ctx->sqo_wait);
3148 mmgrab(current->mm);
3149 ctx->sqo_mm = current->mm;
3151 if (ctx->flags & IORING_SETUP_SQPOLL) {
3153 if (!capable(CAP_SYS_ADMIN))
3156 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
3157 if (!ctx->sq_thread_idle)
3158 ctx->sq_thread_idle = HZ;
3160 if (p->flags & IORING_SETUP_SQ_AFF) {
3161 int cpu = p->sq_thread_cpu;
3164 if (cpu >= nr_cpu_ids)
3166 if (!cpu_online(cpu))
3169 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
3173 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
3176 if (IS_ERR(ctx->sqo_thread)) {
3177 ret = PTR_ERR(ctx->sqo_thread);
3178 ctx->sqo_thread = NULL;
3181 wake_up_process(ctx->sqo_thread);
3182 } else if (p->flags & IORING_SETUP_SQ_AFF) {
3183 /* Can't have SQ_AFF without SQPOLL */
3188 /* Do QD, or 2 * CPUS, whatever is smallest */
3189 ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq",
3190 WQ_UNBOUND | WQ_FREEZABLE,
3191 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
3192 if (!ctx->sqo_wq[0]) {
3198 * This is for buffered writes, where we want to limit the parallelism
3199 * due to file locking in file systems. As "normal" buffered writes
3200 * should parellelize on writeout quite nicely, limit us to having 2
3201 * pending. This avoids massive contention on the inode when doing
3202 * buffered async writes.
3204 ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq",
3205 WQ_UNBOUND | WQ_FREEZABLE, 2);
3206 if (!ctx->sqo_wq[1]) {
3213 io_finish_async(ctx);
3214 mmdrop(ctx->sqo_mm);
3219 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
3221 atomic_long_sub(nr_pages, &user->locked_vm);
3224 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
3226 unsigned long page_limit, cur_pages, new_pages;
3228 /* Don't allow more pages than we can safely lock */
3229 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
3232 cur_pages = atomic_long_read(&user->locked_vm);
3233 new_pages = cur_pages + nr_pages;
3234 if (new_pages > page_limit)
3236 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
3237 new_pages) != cur_pages);
3242 static void io_mem_free(void *ptr)
3249 page = virt_to_head_page(ptr);
3250 if (put_page_testzero(page))
3251 free_compound_page(page);
3254 static void *io_mem_alloc(size_t size)
3256 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
3259 return (void *) __get_free_pages(gfp_flags, get_order(size));
3262 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
3265 struct io_rings *rings;
3266 size_t off, sq_array_size;
3268 off = struct_size(rings, cqes, cq_entries);
3269 if (off == SIZE_MAX)
3273 off = ALIGN(off, SMP_CACHE_BYTES);
3278 sq_array_size = array_size(sizeof(u32), sq_entries);
3279 if (sq_array_size == SIZE_MAX)
3282 if (check_add_overflow(off, sq_array_size, &off))
3291 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
3295 pages = (size_t)1 << get_order(
3296 rings_size(sq_entries, cq_entries, NULL));
3297 pages += (size_t)1 << get_order(
3298 array_size(sizeof(struct io_uring_sqe), sq_entries));
3303 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
3307 if (!ctx->user_bufs)
3310 for (i = 0; i < ctx->nr_user_bufs; i++) {
3311 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3313 for (j = 0; j < imu->nr_bvecs; j++)
3314 put_user_page(imu->bvec[j].bv_page);
3316 if (ctx->account_mem)
3317 io_unaccount_mem(ctx->user, imu->nr_bvecs);
3322 kfree(ctx->user_bufs);
3323 ctx->user_bufs = NULL;
3324 ctx->nr_user_bufs = 0;
3328 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
3329 void __user *arg, unsigned index)
3331 struct iovec __user *src;
3333 #ifdef CONFIG_COMPAT
3335 struct compat_iovec __user *ciovs;
3336 struct compat_iovec ciov;
3338 ciovs = (struct compat_iovec __user *) arg;
3339 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
3342 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
3343 dst->iov_len = ciov.iov_len;
3347 src = (struct iovec __user *) arg;
3348 if (copy_from_user(dst, &src[index], sizeof(*dst)))
3353 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
3356 struct vm_area_struct **vmas = NULL;
3357 struct page **pages = NULL;
3358 int i, j, got_pages = 0;
3363 if (!nr_args || nr_args > UIO_MAXIOV)
3366 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
3368 if (!ctx->user_bufs)
3371 for (i = 0; i < nr_args; i++) {
3372 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3373 unsigned long off, start, end, ubuf;
3378 ret = io_copy_iov(ctx, &iov, arg, i);
3383 * Don't impose further limits on the size and buffer
3384 * constraints here, we'll -EINVAL later when IO is
3385 * submitted if they are wrong.
3388 if (!iov.iov_base || !iov.iov_len)
3391 /* arbitrary limit, but we need something */
3392 if (iov.iov_len > SZ_1G)
3395 ubuf = (unsigned long) iov.iov_base;
3396 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3397 start = ubuf >> PAGE_SHIFT;
3398 nr_pages = end - start;
3400 if (ctx->account_mem) {
3401 ret = io_account_mem(ctx->user, nr_pages);
3407 if (!pages || nr_pages > got_pages) {
3410 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
3412 vmas = kvmalloc_array(nr_pages,
3413 sizeof(struct vm_area_struct *),
3415 if (!pages || !vmas) {
3417 if (ctx->account_mem)
3418 io_unaccount_mem(ctx->user, nr_pages);
3421 got_pages = nr_pages;
3424 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
3428 if (ctx->account_mem)
3429 io_unaccount_mem(ctx->user, nr_pages);
3434 down_read(¤t->mm->mmap_sem);
3435 pret = get_user_pages(ubuf, nr_pages,
3436 FOLL_WRITE | FOLL_LONGTERM,
3438 if (pret == nr_pages) {
3439 /* don't support file backed memory */
3440 for (j = 0; j < nr_pages; j++) {
3441 struct vm_area_struct *vma = vmas[j];
3444 !is_file_hugepages(vma->vm_file)) {
3450 ret = pret < 0 ? pret : -EFAULT;
3452 up_read(¤t->mm->mmap_sem);
3455 * if we did partial map, or found file backed vmas,
3456 * release any pages we did get
3459 put_user_pages(pages, pret);
3460 if (ctx->account_mem)
3461 io_unaccount_mem(ctx->user, nr_pages);
3466 off = ubuf & ~PAGE_MASK;
3468 for (j = 0; j < nr_pages; j++) {
3471 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3472 imu->bvec[j].bv_page = pages[j];
3473 imu->bvec[j].bv_len = vec_len;
3474 imu->bvec[j].bv_offset = off;
3478 /* store original address for later verification */
3480 imu->len = iov.iov_len;
3481 imu->nr_bvecs = nr_pages;
3483 ctx->nr_user_bufs++;
3491 io_sqe_buffer_unregister(ctx);
3495 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3497 __s32 __user *fds = arg;
3503 if (copy_from_user(&fd, fds, sizeof(*fds)))
3506 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3507 if (IS_ERR(ctx->cq_ev_fd)) {
3508 int ret = PTR_ERR(ctx->cq_ev_fd);
3509 ctx->cq_ev_fd = NULL;
3516 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3518 if (ctx->cq_ev_fd) {
3519 eventfd_ctx_put(ctx->cq_ev_fd);
3520 ctx->cq_ev_fd = NULL;
3527 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3529 io_finish_async(ctx);
3531 mmdrop(ctx->sqo_mm);
3533 io_iopoll_reap_events(ctx);
3534 io_sqe_buffer_unregister(ctx);
3535 io_sqe_files_unregister(ctx);
3536 io_eventfd_unregister(ctx);
3538 #if defined(CONFIG_UNIX)
3539 if (ctx->ring_sock) {
3540 ctx->ring_sock->file = NULL; /* so that iput() is called */
3541 sock_release(ctx->ring_sock);
3545 io_mem_free(ctx->rings);
3546 io_mem_free(ctx->sq_sqes);
3548 percpu_ref_exit(&ctx->refs);
3549 if (ctx->account_mem)
3550 io_unaccount_mem(ctx->user,
3551 ring_pages(ctx->sq_entries, ctx->cq_entries));
3552 free_uid(ctx->user);
3556 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3558 struct io_ring_ctx *ctx = file->private_data;
3561 poll_wait(file, &ctx->cq_wait, wait);
3563 * synchronizes with barrier from wq_has_sleeper call in
3567 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
3568 ctx->rings->sq_ring_entries)
3569 mask |= EPOLLOUT | EPOLLWRNORM;
3570 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
3571 mask |= EPOLLIN | EPOLLRDNORM;
3576 static int io_uring_fasync(int fd, struct file *file, int on)
3578 struct io_ring_ctx *ctx = file->private_data;
3580 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3583 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3585 mutex_lock(&ctx->uring_lock);
3586 percpu_ref_kill(&ctx->refs);
3587 mutex_unlock(&ctx->uring_lock);
3589 io_kill_timeouts(ctx);
3590 io_poll_remove_all(ctx);
3591 io_iopoll_reap_events(ctx);
3592 wait_for_completion(&ctx->ctx_done);
3593 io_ring_ctx_free(ctx);
3596 static int io_uring_release(struct inode *inode, struct file *file)
3598 struct io_ring_ctx *ctx = file->private_data;
3600 file->private_data = NULL;
3601 io_ring_ctx_wait_and_kill(ctx);
3605 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3607 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3608 unsigned long sz = vma->vm_end - vma->vm_start;
3609 struct io_ring_ctx *ctx = file->private_data;
3615 case IORING_OFF_SQ_RING:
3616 case IORING_OFF_CQ_RING:
3619 case IORING_OFF_SQES:
3626 page = virt_to_head_page(ptr);
3627 if (sz > page_size(page))
3630 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3631 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3634 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3635 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3638 struct io_ring_ctx *ctx;
3643 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3651 if (f.file->f_op != &io_uring_fops)
3655 ctx = f.file->private_data;
3656 if (!percpu_ref_tryget(&ctx->refs))
3660 * For SQ polling, the thread will do all submissions and completions.
3661 * Just return the requested submit count, and wake the thread if
3665 if (ctx->flags & IORING_SETUP_SQPOLL) {
3666 if (flags & IORING_ENTER_SQ_WAKEUP)
3667 wake_up(&ctx->sqo_wait);
3668 submitted = to_submit;
3669 } else if (to_submit) {
3670 to_submit = min(to_submit, ctx->sq_entries);
3672 mutex_lock(&ctx->uring_lock);
3673 submitted = io_ring_submit(ctx, to_submit);
3674 mutex_unlock(&ctx->uring_lock);
3676 if (flags & IORING_ENTER_GETEVENTS) {
3677 unsigned nr_events = 0;
3679 min_complete = min(min_complete, ctx->cq_entries);
3681 if (ctx->flags & IORING_SETUP_IOPOLL) {
3682 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3684 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3688 percpu_ref_put(&ctx->refs);
3691 return submitted ? submitted : ret;
3694 static const struct file_operations io_uring_fops = {
3695 .release = io_uring_release,
3696 .mmap = io_uring_mmap,
3697 .poll = io_uring_poll,
3698 .fasync = io_uring_fasync,
3701 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3702 struct io_uring_params *p)
3704 struct io_rings *rings;
3705 size_t size, sq_array_offset;
3707 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
3708 if (size == SIZE_MAX)
3711 rings = io_mem_alloc(size);
3716 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3717 rings->sq_ring_mask = p->sq_entries - 1;
3718 rings->cq_ring_mask = p->cq_entries - 1;
3719 rings->sq_ring_entries = p->sq_entries;
3720 rings->cq_ring_entries = p->cq_entries;
3721 ctx->sq_mask = rings->sq_ring_mask;
3722 ctx->cq_mask = rings->cq_ring_mask;
3723 ctx->sq_entries = rings->sq_ring_entries;
3724 ctx->cq_entries = rings->cq_ring_entries;
3726 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3727 if (size == SIZE_MAX)
3730 ctx->sq_sqes = io_mem_alloc(size);
3738 * Allocate an anonymous fd, this is what constitutes the application
3739 * visible backing of an io_uring instance. The application mmaps this
3740 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3741 * we have to tie this fd to a socket for file garbage collection purposes.
3743 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3748 #if defined(CONFIG_UNIX)
3749 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3755 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3759 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3760 O_RDWR | O_CLOEXEC);
3763 ret = PTR_ERR(file);
3767 #if defined(CONFIG_UNIX)
3768 ctx->ring_sock->file = file;
3769 ctx->ring_sock->sk->sk_user_data = ctx;
3771 fd_install(ret, file);
3774 #if defined(CONFIG_UNIX)
3775 sock_release(ctx->ring_sock);
3776 ctx->ring_sock = NULL;
3781 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3783 struct user_struct *user = NULL;
3784 struct io_ring_ctx *ctx;
3788 if (!entries || entries > IORING_MAX_ENTRIES)
3792 * Use twice as many entries for the CQ ring. It's possible for the
3793 * application to drive a higher depth than the size of the SQ ring,
3794 * since the sqes are only used at submission time. This allows for
3795 * some flexibility in overcommitting a bit.
3797 p->sq_entries = roundup_pow_of_two(entries);
3798 p->cq_entries = 2 * p->sq_entries;
3800 user = get_uid(current_user());
3801 account_mem = !capable(CAP_IPC_LOCK);
3804 ret = io_account_mem(user,
3805 ring_pages(p->sq_entries, p->cq_entries));
3812 ctx = io_ring_ctx_alloc(p);
3815 io_unaccount_mem(user, ring_pages(p->sq_entries,
3820 ctx->compat = in_compat_syscall();
3821 ctx->account_mem = account_mem;
3824 ret = io_allocate_scq_urings(ctx, p);
3828 ret = io_sq_offload_start(ctx, p);
3832 memset(&p->sq_off, 0, sizeof(p->sq_off));
3833 p->sq_off.head = offsetof(struct io_rings, sq.head);
3834 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3835 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3836 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3837 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3838 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3839 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3841 memset(&p->cq_off, 0, sizeof(p->cq_off));
3842 p->cq_off.head = offsetof(struct io_rings, cq.head);
3843 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3844 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3845 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3846 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3847 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3850 * Install ring fd as the very last thing, so we don't risk someone
3851 * having closed it before we finish setup
3853 ret = io_uring_get_fd(ctx);
3857 p->features = IORING_FEAT_SINGLE_MMAP;
3860 io_ring_ctx_wait_and_kill(ctx);
3865 * Sets up an aio uring context, and returns the fd. Applications asks for a
3866 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3867 * params structure passed in.
3869 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3871 struct io_uring_params p;
3875 if (copy_from_user(&p, params, sizeof(p)))
3877 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3882 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3883 IORING_SETUP_SQ_AFF))
3886 ret = io_uring_create(entries, &p);
3890 if (copy_to_user(params, &p, sizeof(p)))
3896 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3897 struct io_uring_params __user *, params)
3899 return io_uring_setup(entries, params);
3902 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3903 void __user *arg, unsigned nr_args)
3904 __releases(ctx->uring_lock)
3905 __acquires(ctx->uring_lock)
3910 * We're inside the ring mutex, if the ref is already dying, then
3911 * someone else killed the ctx or is already going through
3912 * io_uring_register().
3914 if (percpu_ref_is_dying(&ctx->refs))
3917 percpu_ref_kill(&ctx->refs);
3920 * Drop uring mutex before waiting for references to exit. If another
3921 * thread is currently inside io_uring_enter() it might need to grab
3922 * the uring_lock to make progress. If we hold it here across the drain
3923 * wait, then we can deadlock. It's safe to drop the mutex here, since
3924 * no new references will come in after we've killed the percpu ref.
3926 mutex_unlock(&ctx->uring_lock);
3927 wait_for_completion(&ctx->ctx_done);
3928 mutex_lock(&ctx->uring_lock);
3931 case IORING_REGISTER_BUFFERS:
3932 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3934 case IORING_UNREGISTER_BUFFERS:
3938 ret = io_sqe_buffer_unregister(ctx);
3940 case IORING_REGISTER_FILES:
3941 ret = io_sqe_files_register(ctx, arg, nr_args);
3943 case IORING_UNREGISTER_FILES:
3947 ret = io_sqe_files_unregister(ctx);
3949 case IORING_REGISTER_EVENTFD:
3953 ret = io_eventfd_register(ctx, arg);
3955 case IORING_UNREGISTER_EVENTFD:
3959 ret = io_eventfd_unregister(ctx);
3966 /* bring the ctx back to life */
3967 reinit_completion(&ctx->ctx_done);
3968 percpu_ref_reinit(&ctx->refs);
3972 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3973 void __user *, arg, unsigned int, nr_args)
3975 struct io_ring_ctx *ctx;
3984 if (f.file->f_op != &io_uring_fops)
3987 ctx = f.file->private_data;
3989 mutex_lock(&ctx->uring_lock);
3990 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3991 mutex_unlock(&ctx->uring_lock);
3997 static int __init io_uring_init(void)
3999 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
4002 __initcall(io_uring_init);