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 struct io_uring_sqe *sq_sqes;
202 struct list_head defer_list;
203 struct list_head timeout_list;
204 } ____cacheline_aligned_in_smp;
207 struct workqueue_struct *sqo_wq[2];
208 struct task_struct *sqo_thread; /* if using sq thread polling */
209 struct mm_struct *sqo_mm;
210 wait_queue_head_t sqo_wait;
211 struct completion sqo_thread_started;
214 unsigned cached_cq_tail;
217 struct wait_queue_head cq_wait;
218 struct fasync_struct *cq_fasync;
219 struct eventfd_ctx *cq_ev_fd;
220 atomic_t cq_timeouts;
221 } ____cacheline_aligned_in_smp;
223 struct io_rings *rings;
226 * If used, fixed file set. Writers must ensure that ->refs is dead,
227 * readers must ensure that ->refs is alive as long as the file* is
228 * used. Only updated through io_uring_register(2).
230 struct file **user_files;
231 unsigned nr_user_files;
233 /* if used, fixed mapped user buffers */
234 unsigned nr_user_bufs;
235 struct io_mapped_ubuf *user_bufs;
237 struct user_struct *user;
239 struct completion ctx_done;
242 struct mutex uring_lock;
243 wait_queue_head_t wait;
244 } ____cacheline_aligned_in_smp;
247 spinlock_t completion_lock;
248 bool poll_multi_file;
250 * ->poll_list is protected by the ctx->uring_lock for
251 * io_uring instances that don't use IORING_SETUP_SQPOLL.
252 * For SQPOLL, only the single threaded io_sq_thread() will
253 * manipulate the list, hence no extra locking is needed there.
255 struct list_head poll_list;
256 struct list_head cancel_list;
257 } ____cacheline_aligned_in_smp;
259 struct async_list pending_async[2];
261 #if defined(CONFIG_UNIX)
262 struct socket *ring_sock;
267 const struct io_uring_sqe *sqe;
268 unsigned short index;
272 bool needs_fixed_file;
276 * First field must be the file pointer in all the
277 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
279 struct io_poll_iocb {
281 struct wait_queue_head *head;
285 struct wait_queue_entry wait;
290 struct hrtimer timer;
294 * NOTE! Each of the iocb union members has the file pointer
295 * as the first entry in their struct definition. So you can
296 * access the file pointer through any of the sub-structs,
297 * or directly as just 'ki_filp' in this struct.
303 struct io_poll_iocb poll;
304 struct io_timeout timeout;
307 struct sqe_submit submit;
309 struct io_ring_ctx *ctx;
310 struct list_head list;
311 struct list_head link_list;
314 #define REQ_F_NOWAIT 1 /* must not punt to workers */
315 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
316 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
317 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
318 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
319 #define REQ_F_IO_DRAINED 32 /* drain done */
320 #define REQ_F_LINK 64 /* linked sqes */
321 #define REQ_F_LINK_DONE 128 /* linked sqes done */
322 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
323 #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */
324 #define REQ_F_TIMEOUT 1024 /* timeout request */
329 struct work_struct work;
332 #define IO_PLUG_THRESHOLD 2
333 #define IO_IOPOLL_BATCH 8
335 struct io_submit_state {
336 struct blk_plug plug;
339 * io_kiocb alloc cache
341 void *reqs[IO_IOPOLL_BATCH];
342 unsigned int free_reqs;
343 unsigned int cur_req;
346 * File reference cache
350 unsigned int has_refs;
351 unsigned int used_refs;
352 unsigned int ios_left;
355 static void io_sq_wq_submit_work(struct work_struct *work);
356 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
358 static void __io_free_req(struct io_kiocb *req);
360 static struct kmem_cache *req_cachep;
362 static const struct file_operations io_uring_fops;
364 struct sock *io_uring_get_socket(struct file *file)
366 #if defined(CONFIG_UNIX)
367 if (file->f_op == &io_uring_fops) {
368 struct io_ring_ctx *ctx = file->private_data;
370 return ctx->ring_sock->sk;
375 EXPORT_SYMBOL(io_uring_get_socket);
377 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
379 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
381 complete(&ctx->ctx_done);
384 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
386 struct io_ring_ctx *ctx;
389 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
393 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
394 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
399 ctx->flags = p->flags;
400 init_waitqueue_head(&ctx->cq_wait);
401 init_completion(&ctx->ctx_done);
402 init_completion(&ctx->sqo_thread_started);
403 mutex_init(&ctx->uring_lock);
404 init_waitqueue_head(&ctx->wait);
405 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
406 spin_lock_init(&ctx->pending_async[i].lock);
407 INIT_LIST_HEAD(&ctx->pending_async[i].list);
408 atomic_set(&ctx->pending_async[i].cnt, 0);
410 spin_lock_init(&ctx->completion_lock);
411 INIT_LIST_HEAD(&ctx->poll_list);
412 INIT_LIST_HEAD(&ctx->cancel_list);
413 INIT_LIST_HEAD(&ctx->defer_list);
414 INIT_LIST_HEAD(&ctx->timeout_list);
418 static inline bool __io_sequence_defer(struct io_ring_ctx *ctx,
419 struct io_kiocb *req)
421 return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped;
424 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
425 struct io_kiocb *req)
427 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
430 return __io_sequence_defer(ctx, req);
433 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
435 struct io_kiocb *req;
437 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
438 if (req && !io_sequence_defer(ctx, req)) {
439 list_del_init(&req->list);
446 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
448 struct io_kiocb *req;
450 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
451 if (req && !__io_sequence_defer(ctx, req)) {
452 list_del_init(&req->list);
459 static void __io_commit_cqring(struct io_ring_ctx *ctx)
461 struct io_rings *rings = ctx->rings;
463 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
464 /* order cqe stores with ring update */
465 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
467 if (wq_has_sleeper(&ctx->cq_wait)) {
468 wake_up_interruptible(&ctx->cq_wait);
469 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
474 static inline void io_queue_async_work(struct io_ring_ctx *ctx,
475 struct io_kiocb *req)
479 if (req->submit.sqe) {
480 switch (req->submit.sqe->opcode) {
481 case IORING_OP_WRITEV:
482 case IORING_OP_WRITE_FIXED:
483 rw = !(req->rw.ki_flags & IOCB_DIRECT);
488 queue_work(ctx->sqo_wq[rw], &req->work);
491 static void io_kill_timeout(struct io_kiocb *req)
495 ret = hrtimer_try_to_cancel(&req->timeout.timer);
497 atomic_inc(&req->ctx->cq_timeouts);
498 list_del(&req->list);
499 io_cqring_fill_event(req->ctx, req->user_data, 0);
504 static void io_kill_timeouts(struct io_ring_ctx *ctx)
506 struct io_kiocb *req, *tmp;
508 spin_lock_irq(&ctx->completion_lock);
509 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
510 io_kill_timeout(req);
511 spin_unlock_irq(&ctx->completion_lock);
514 static void io_commit_cqring(struct io_ring_ctx *ctx)
516 struct io_kiocb *req;
518 while ((req = io_get_timeout_req(ctx)) != NULL)
519 io_kill_timeout(req);
521 __io_commit_cqring(ctx);
523 while ((req = io_get_deferred_req(ctx)) != NULL) {
524 if (req->flags & REQ_F_SHADOW_DRAIN) {
525 /* Just for drain, free it. */
529 req->flags |= REQ_F_IO_DRAINED;
530 io_queue_async_work(ctx, req);
534 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
536 struct io_rings *rings = ctx->rings;
539 tail = ctx->cached_cq_tail;
541 * writes to the cq entry need to come after reading head; the
542 * control dependency is enough as we're using WRITE_ONCE to
545 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
548 ctx->cached_cq_tail++;
549 return &rings->cqes[tail & ctx->cq_mask];
552 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
555 struct io_uring_cqe *cqe;
558 * If we can't get a cq entry, userspace overflowed the
559 * submission (by quite a lot). Increment the overflow count in
562 cqe = io_get_cqring(ctx);
564 WRITE_ONCE(cqe->user_data, ki_user_data);
565 WRITE_ONCE(cqe->res, res);
566 WRITE_ONCE(cqe->flags, 0);
568 unsigned overflow = READ_ONCE(ctx->rings->cq_overflow);
570 WRITE_ONCE(ctx->rings->cq_overflow, overflow + 1);
574 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
576 if (waitqueue_active(&ctx->wait))
578 if (waitqueue_active(&ctx->sqo_wait))
579 wake_up(&ctx->sqo_wait);
581 eventfd_signal(ctx->cq_ev_fd, 1);
584 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
589 spin_lock_irqsave(&ctx->completion_lock, flags);
590 io_cqring_fill_event(ctx, user_data, res);
591 io_commit_cqring(ctx);
592 spin_unlock_irqrestore(&ctx->completion_lock, flags);
594 io_cqring_ev_posted(ctx);
597 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
598 struct io_submit_state *state)
600 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
601 struct io_kiocb *req;
603 if (!percpu_ref_tryget(&ctx->refs))
607 req = kmem_cache_alloc(req_cachep, gfp);
610 } else if (!state->free_reqs) {
614 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
615 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
618 * Bulk alloc is all-or-nothing. If we fail to get a batch,
619 * retry single alloc to be on the safe side.
621 if (unlikely(ret <= 0)) {
622 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
627 state->free_reqs = ret - 1;
629 req = state->reqs[0];
631 req = state->reqs[state->cur_req];
639 /* one is dropped after submission, the other at completion */
640 refcount_set(&req->refs, 2);
644 percpu_ref_put(&ctx->refs);
648 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
651 kmem_cache_free_bulk(req_cachep, *nr, reqs);
652 percpu_ref_put_many(&ctx->refs, *nr);
657 static void __io_free_req(struct io_kiocb *req)
659 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
661 percpu_ref_put(&req->ctx->refs);
662 kmem_cache_free(req_cachep, req);
665 static void io_req_link_next(struct io_kiocb *req)
667 struct io_kiocb *nxt;
670 * The list should never be empty when we are called here. But could
671 * potentially happen if the chain is messed up, check to be on the
674 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
676 list_del(&nxt->list);
677 if (!list_empty(&req->link_list)) {
678 INIT_LIST_HEAD(&nxt->link_list);
679 list_splice(&req->link_list, &nxt->link_list);
680 nxt->flags |= REQ_F_LINK;
683 nxt->flags |= REQ_F_LINK_DONE;
684 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
685 io_queue_async_work(req->ctx, nxt);
690 * Called if REQ_F_LINK is set, and we fail the head request
692 static void io_fail_links(struct io_kiocb *req)
694 struct io_kiocb *link;
696 while (!list_empty(&req->link_list)) {
697 link = list_first_entry(&req->link_list, struct io_kiocb, list);
698 list_del(&link->list);
700 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
705 static void io_free_req(struct io_kiocb *req)
708 * If LINK is set, we have dependent requests in this chain. If we
709 * didn't fail this request, queue the first one up, moving any other
710 * dependencies to the next request. In case of failure, fail the rest
713 if (req->flags & REQ_F_LINK) {
714 if (req->flags & REQ_F_FAIL_LINK)
717 io_req_link_next(req);
723 static void io_put_req(struct io_kiocb *req)
725 if (refcount_dec_and_test(&req->refs))
729 static unsigned io_cqring_events(struct io_rings *rings)
731 /* See comment at the top of this file */
733 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
737 * Find and free completed poll iocbs
739 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
740 struct list_head *done)
742 void *reqs[IO_IOPOLL_BATCH];
743 struct io_kiocb *req;
747 while (!list_empty(done)) {
748 req = list_first_entry(done, struct io_kiocb, list);
749 list_del(&req->list);
751 io_cqring_fill_event(ctx, req->user_data, req->result);
754 if (refcount_dec_and_test(&req->refs)) {
755 /* If we're not using fixed files, we have to pair the
756 * completion part with the file put. Use regular
757 * completions for those, only batch free for fixed
758 * file and non-linked commands.
760 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
762 reqs[to_free++] = req;
763 if (to_free == ARRAY_SIZE(reqs))
764 io_free_req_many(ctx, reqs, &to_free);
771 io_commit_cqring(ctx);
772 io_free_req_many(ctx, reqs, &to_free);
775 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
778 struct io_kiocb *req, *tmp;
784 * Only spin for completions if we don't have multiple devices hanging
785 * off our complete list, and we're under the requested amount.
787 spin = !ctx->poll_multi_file && *nr_events < min;
790 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
791 struct kiocb *kiocb = &req->rw;
794 * Move completed entries to our local list. If we find a
795 * request that requires polling, break out and complete
796 * the done list first, if we have entries there.
798 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
799 list_move_tail(&req->list, &done);
802 if (!list_empty(&done))
805 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
814 if (!list_empty(&done))
815 io_iopoll_complete(ctx, nr_events, &done);
821 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
822 * non-spinning poll check - we'll still enter the driver poll loop, but only
823 * as a non-spinning completion check.
825 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
828 while (!list_empty(&ctx->poll_list) && !need_resched()) {
831 ret = io_do_iopoll(ctx, nr_events, min);
834 if (!min || *nr_events >= min)
842 * We can't just wait for polled events to come to us, we have to actively
843 * find and complete them.
845 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
847 if (!(ctx->flags & IORING_SETUP_IOPOLL))
850 mutex_lock(&ctx->uring_lock);
851 while (!list_empty(&ctx->poll_list)) {
852 unsigned int nr_events = 0;
854 io_iopoll_getevents(ctx, &nr_events, 1);
857 * Ensure we allow local-to-the-cpu processing to take place,
858 * in this case we need to ensure that we reap all events.
862 mutex_unlock(&ctx->uring_lock);
865 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
871 * We disallow the app entering submit/complete with polling, but we
872 * still need to lock the ring to prevent racing with polled issue
873 * that got punted to a workqueue.
875 mutex_lock(&ctx->uring_lock);
882 * Don't enter poll loop if we already have events pending.
883 * If we do, we can potentially be spinning for commands that
884 * already triggered a CQE (eg in error).
886 if (io_cqring_events(ctx->rings))
890 * If a submit got punted to a workqueue, we can have the
891 * application entering polling for a command before it gets
892 * issued. That app will hold the uring_lock for the duration
893 * of the poll right here, so we need to take a breather every
894 * now and then to ensure that the issue has a chance to add
895 * the poll to the issued list. Otherwise we can spin here
896 * forever, while the workqueue is stuck trying to acquire the
899 if (!(++iters & 7)) {
900 mutex_unlock(&ctx->uring_lock);
901 mutex_lock(&ctx->uring_lock);
904 if (*nr_events < min)
905 tmin = min - *nr_events;
907 ret = io_iopoll_getevents(ctx, nr_events, tmin);
911 } while (min && !*nr_events && !need_resched());
913 mutex_unlock(&ctx->uring_lock);
917 static void kiocb_end_write(struct kiocb *kiocb)
919 if (kiocb->ki_flags & IOCB_WRITE) {
920 struct inode *inode = file_inode(kiocb->ki_filp);
923 * Tell lockdep we inherited freeze protection from submission
926 if (S_ISREG(inode->i_mode))
927 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
928 file_end_write(kiocb->ki_filp);
932 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
934 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
936 kiocb_end_write(kiocb);
938 if ((req->flags & REQ_F_LINK) && res != req->result)
939 req->flags |= REQ_F_FAIL_LINK;
940 io_cqring_add_event(req->ctx, req->user_data, res);
944 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
946 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
948 kiocb_end_write(kiocb);
950 if ((req->flags & REQ_F_LINK) && res != req->result)
951 req->flags |= REQ_F_FAIL_LINK;
954 req->flags |= REQ_F_IOPOLL_COMPLETED;
958 * After the iocb has been issued, it's safe to be found on the poll list.
959 * Adding the kiocb to the list AFTER submission ensures that we don't
960 * find it from a io_iopoll_getevents() thread before the issuer is done
961 * accessing the kiocb cookie.
963 static void io_iopoll_req_issued(struct io_kiocb *req)
965 struct io_ring_ctx *ctx = req->ctx;
968 * Track whether we have multiple files in our lists. This will impact
969 * how we do polling eventually, not spinning if we're on potentially
972 if (list_empty(&ctx->poll_list)) {
973 ctx->poll_multi_file = false;
974 } else if (!ctx->poll_multi_file) {
975 struct io_kiocb *list_req;
977 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
979 if (list_req->rw.ki_filp != req->rw.ki_filp)
980 ctx->poll_multi_file = true;
984 * For fast devices, IO may have already completed. If it has, add
985 * it to the front so we find it first.
987 if (req->flags & REQ_F_IOPOLL_COMPLETED)
988 list_add(&req->list, &ctx->poll_list);
990 list_add_tail(&req->list, &ctx->poll_list);
993 static void io_file_put(struct io_submit_state *state)
996 int diff = state->has_refs - state->used_refs;
999 fput_many(state->file, diff);
1005 * Get as many references to a file as we have IOs left in this submission,
1006 * assuming most submissions are for one file, or at least that each file
1007 * has more than one submission.
1009 static struct file *io_file_get(struct io_submit_state *state, int fd)
1015 if (state->fd == fd) {
1022 state->file = fget_many(fd, state->ios_left);
1027 state->has_refs = state->ios_left;
1028 state->used_refs = 1;
1034 * If we tracked the file through the SCM inflight mechanism, we could support
1035 * any file. For now, just ensure that anything potentially problematic is done
1038 static bool io_file_supports_async(struct file *file)
1040 umode_t mode = file_inode(file)->i_mode;
1042 if (S_ISBLK(mode) || S_ISCHR(mode))
1044 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1050 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
1051 bool force_nonblock)
1053 const struct io_uring_sqe *sqe = s->sqe;
1054 struct io_ring_ctx *ctx = req->ctx;
1055 struct kiocb *kiocb = &req->rw;
1062 if (force_nonblock && !io_file_supports_async(req->file))
1063 force_nonblock = false;
1065 kiocb->ki_pos = READ_ONCE(sqe->off);
1066 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1067 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1069 ioprio = READ_ONCE(sqe->ioprio);
1071 ret = ioprio_check_cap(ioprio);
1075 kiocb->ki_ioprio = ioprio;
1077 kiocb->ki_ioprio = get_current_ioprio();
1079 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1083 /* don't allow async punt if RWF_NOWAIT was requested */
1084 if (kiocb->ki_flags & IOCB_NOWAIT)
1085 req->flags |= REQ_F_NOWAIT;
1088 kiocb->ki_flags |= IOCB_NOWAIT;
1090 if (ctx->flags & IORING_SETUP_IOPOLL) {
1091 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1092 !kiocb->ki_filp->f_op->iopoll)
1095 kiocb->ki_flags |= IOCB_HIPRI;
1096 kiocb->ki_complete = io_complete_rw_iopoll;
1098 if (kiocb->ki_flags & IOCB_HIPRI)
1100 kiocb->ki_complete = io_complete_rw;
1105 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1111 case -ERESTARTNOINTR:
1112 case -ERESTARTNOHAND:
1113 case -ERESTART_RESTARTBLOCK:
1115 * We can't just restart the syscall, since previously
1116 * submitted sqes may already be in progress. Just fail this
1122 kiocb->ki_complete(kiocb, ret, 0);
1126 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1127 const struct io_uring_sqe *sqe,
1128 struct iov_iter *iter)
1130 size_t len = READ_ONCE(sqe->len);
1131 struct io_mapped_ubuf *imu;
1132 unsigned index, buf_index;
1136 /* attempt to use fixed buffers without having provided iovecs */
1137 if (unlikely(!ctx->user_bufs))
1140 buf_index = READ_ONCE(sqe->buf_index);
1141 if (unlikely(buf_index >= ctx->nr_user_bufs))
1144 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1145 imu = &ctx->user_bufs[index];
1146 buf_addr = READ_ONCE(sqe->addr);
1149 if (buf_addr + len < buf_addr)
1151 /* not inside the mapped region */
1152 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1156 * May not be a start of buffer, set size appropriately
1157 * and advance us to the beginning.
1159 offset = buf_addr - imu->ubuf;
1160 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1164 * Don't use iov_iter_advance() here, as it's really slow for
1165 * using the latter parts of a big fixed buffer - it iterates
1166 * over each segment manually. We can cheat a bit here, because
1169 * 1) it's a BVEC iter, we set it up
1170 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1171 * first and last bvec
1173 * So just find our index, and adjust the iterator afterwards.
1174 * If the offset is within the first bvec (or the whole first
1175 * bvec, just use iov_iter_advance(). This makes it easier
1176 * since we can just skip the first segment, which may not
1177 * be PAGE_SIZE aligned.
1179 const struct bio_vec *bvec = imu->bvec;
1181 if (offset <= bvec->bv_len) {
1182 iov_iter_advance(iter, offset);
1184 unsigned long seg_skip;
1186 /* skip first vec */
1187 offset -= bvec->bv_len;
1188 seg_skip = 1 + (offset >> PAGE_SHIFT);
1190 iter->bvec = bvec + seg_skip;
1191 iter->nr_segs -= seg_skip;
1192 iter->count -= bvec->bv_len + offset;
1193 iter->iov_offset = offset & ~PAGE_MASK;
1200 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1201 const struct sqe_submit *s, struct iovec **iovec,
1202 struct iov_iter *iter)
1204 const struct io_uring_sqe *sqe = s->sqe;
1205 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1206 size_t sqe_len = READ_ONCE(sqe->len);
1210 * We're reading ->opcode for the second time, but the first read
1211 * doesn't care whether it's _FIXED or not, so it doesn't matter
1212 * whether ->opcode changes concurrently. The first read does care
1213 * about whether it is a READ or a WRITE, so we don't trust this read
1214 * for that purpose and instead let the caller pass in the read/write
1217 opcode = READ_ONCE(sqe->opcode);
1218 if (opcode == IORING_OP_READ_FIXED ||
1219 opcode == IORING_OP_WRITE_FIXED) {
1220 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1228 #ifdef CONFIG_COMPAT
1230 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1234 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1237 static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb)
1239 if (al->file == kiocb->ki_filp) {
1243 * Allow merging if we're anywhere in the range of the same
1244 * page. Generally this happens for sub-page reads or writes,
1245 * and it's beneficial to allow the first worker to bring the
1246 * page in and the piggy backed work can then work on the
1249 start = al->io_start & PAGE_MASK;
1250 end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK;
1251 if (kiocb->ki_pos >= start && kiocb->ki_pos <= end)
1260 * Make a note of the last file/offset/direction we punted to async
1261 * context. We'll use this information to see if we can piggy back a
1262 * sequential request onto the previous one, if it's still hasn't been
1263 * completed by the async worker.
1265 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1267 struct async_list *async_list = &req->ctx->pending_async[rw];
1268 struct kiocb *kiocb = &req->rw;
1269 struct file *filp = kiocb->ki_filp;
1271 if (io_should_merge(async_list, kiocb)) {
1272 unsigned long max_bytes;
1274 /* Use 8x RA size as a decent limiter for both reads/writes */
1275 max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3);
1277 max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3);
1279 /* If max len are exceeded, reset the state */
1280 if (async_list->io_len + len <= max_bytes) {
1281 req->flags |= REQ_F_SEQ_PREV;
1282 async_list->io_len += len;
1284 async_list->file = NULL;
1288 /* New file? Reset state. */
1289 if (async_list->file != filp) {
1290 async_list->io_start = kiocb->ki_pos;
1291 async_list->io_len = len;
1292 async_list->file = filp;
1297 * For files that don't have ->read_iter() and ->write_iter(), handle them
1298 * by looping over ->read() or ->write() manually.
1300 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1301 struct iov_iter *iter)
1306 * Don't support polled IO through this interface, and we can't
1307 * support non-blocking either. For the latter, this just causes
1308 * the kiocb to be handled from an async context.
1310 if (kiocb->ki_flags & IOCB_HIPRI)
1312 if (kiocb->ki_flags & IOCB_NOWAIT)
1315 while (iov_iter_count(iter)) {
1316 struct iovec iovec = iov_iter_iovec(iter);
1320 nr = file->f_op->read(file, iovec.iov_base,
1321 iovec.iov_len, &kiocb->ki_pos);
1323 nr = file->f_op->write(file, iovec.iov_base,
1324 iovec.iov_len, &kiocb->ki_pos);
1333 if (nr != iovec.iov_len)
1335 iov_iter_advance(iter, nr);
1341 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1342 bool force_nonblock)
1344 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1345 struct kiocb *kiocb = &req->rw;
1346 struct iov_iter iter;
1349 ssize_t read_size, ret;
1351 ret = io_prep_rw(req, s, force_nonblock);
1354 file = kiocb->ki_filp;
1356 if (unlikely(!(file->f_mode & FMODE_READ)))
1359 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1364 if (req->flags & REQ_F_LINK)
1365 req->result = read_size;
1367 iov_count = iov_iter_count(&iter);
1368 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1372 if (file->f_op->read_iter)
1373 ret2 = call_read_iter(file, kiocb, &iter);
1375 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1378 * In case of a short read, punt to async. This can happen
1379 * if we have data partially cached. Alternatively we can
1380 * return the short read, in which case the application will
1381 * need to issue another SQE and wait for it. That SQE will
1382 * need async punt anyway, so it's more efficient to do it
1385 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1387 /* Catch -EAGAIN return for forced non-blocking submission */
1388 if (!force_nonblock || ret2 != -EAGAIN) {
1389 io_rw_done(kiocb, ret2);
1392 * If ->needs_lock is true, we're already in async
1396 io_async_list_note(READ, req, iov_count);
1404 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1405 bool force_nonblock)
1407 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1408 struct kiocb *kiocb = &req->rw;
1409 struct iov_iter iter;
1414 ret = io_prep_rw(req, s, force_nonblock);
1418 file = kiocb->ki_filp;
1419 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1422 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1426 if (req->flags & REQ_F_LINK)
1429 iov_count = iov_iter_count(&iter);
1432 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1433 /* If ->needs_lock is true, we're already in async context. */
1435 io_async_list_note(WRITE, req, iov_count);
1439 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1444 * Open-code file_start_write here to grab freeze protection,
1445 * which will be released by another thread in
1446 * io_complete_rw(). Fool lockdep by telling it the lock got
1447 * released so that it doesn't complain about the held lock when
1448 * we return to userspace.
1450 if (S_ISREG(file_inode(file)->i_mode)) {
1451 __sb_start_write(file_inode(file)->i_sb,
1452 SB_FREEZE_WRITE, true);
1453 __sb_writers_release(file_inode(file)->i_sb,
1456 kiocb->ki_flags |= IOCB_WRITE;
1458 if (file->f_op->write_iter)
1459 ret2 = call_write_iter(file, kiocb, &iter);
1461 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1462 if (!force_nonblock || ret2 != -EAGAIN) {
1463 io_rw_done(kiocb, ret2);
1466 * If ->needs_lock is true, we're already in async
1470 io_async_list_note(WRITE, req, iov_count);
1480 * IORING_OP_NOP just posts a completion event, nothing else.
1482 static int io_nop(struct io_kiocb *req, u64 user_data)
1484 struct io_ring_ctx *ctx = req->ctx;
1487 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1490 io_cqring_add_event(ctx, user_data, err);
1495 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1497 struct io_ring_ctx *ctx = req->ctx;
1502 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1504 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1510 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1511 bool force_nonblock)
1513 loff_t sqe_off = READ_ONCE(sqe->off);
1514 loff_t sqe_len = READ_ONCE(sqe->len);
1515 loff_t end = sqe_off + sqe_len;
1516 unsigned fsync_flags;
1519 fsync_flags = READ_ONCE(sqe->fsync_flags);
1520 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1523 ret = io_prep_fsync(req, sqe);
1527 /* fsync always requires a blocking context */
1531 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1532 end > 0 ? end : LLONG_MAX,
1533 fsync_flags & IORING_FSYNC_DATASYNC);
1535 if (ret < 0 && (req->flags & REQ_F_LINK))
1536 req->flags |= REQ_F_FAIL_LINK;
1537 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1542 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1544 struct io_ring_ctx *ctx = req->ctx;
1550 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1552 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1558 static int io_sync_file_range(struct io_kiocb *req,
1559 const struct io_uring_sqe *sqe,
1560 bool force_nonblock)
1567 ret = io_prep_sfr(req, sqe);
1571 /* sync_file_range always requires a blocking context */
1575 sqe_off = READ_ONCE(sqe->off);
1576 sqe_len = READ_ONCE(sqe->len);
1577 flags = READ_ONCE(sqe->sync_range_flags);
1579 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1581 if (ret < 0 && (req->flags & REQ_F_LINK))
1582 req->flags |= REQ_F_FAIL_LINK;
1583 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1588 #if defined(CONFIG_NET)
1589 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1590 bool force_nonblock,
1591 long (*fn)(struct socket *, struct user_msghdr __user *,
1594 struct socket *sock;
1597 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1600 sock = sock_from_file(req->file, &ret);
1602 struct user_msghdr __user *msg;
1605 flags = READ_ONCE(sqe->msg_flags);
1606 if (flags & MSG_DONTWAIT)
1607 req->flags |= REQ_F_NOWAIT;
1608 else if (force_nonblock)
1609 flags |= MSG_DONTWAIT;
1611 msg = (struct user_msghdr __user *) (unsigned long)
1612 READ_ONCE(sqe->addr);
1614 ret = fn(sock, msg, flags);
1615 if (force_nonblock && ret == -EAGAIN)
1619 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1625 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1626 bool force_nonblock)
1628 #if defined(CONFIG_NET)
1629 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1635 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1636 bool force_nonblock)
1638 #if defined(CONFIG_NET)
1639 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1645 static void io_poll_remove_one(struct io_kiocb *req)
1647 struct io_poll_iocb *poll = &req->poll;
1649 spin_lock(&poll->head->lock);
1650 WRITE_ONCE(poll->canceled, true);
1651 if (!list_empty(&poll->wait.entry)) {
1652 list_del_init(&poll->wait.entry);
1653 io_queue_async_work(req->ctx, req);
1655 spin_unlock(&poll->head->lock);
1657 list_del_init(&req->list);
1660 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1662 struct io_kiocb *req;
1664 spin_lock_irq(&ctx->completion_lock);
1665 while (!list_empty(&ctx->cancel_list)) {
1666 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1667 io_poll_remove_one(req);
1669 spin_unlock_irq(&ctx->completion_lock);
1673 * Find a running poll command that matches one specified in sqe->addr,
1674 * and remove it if found.
1676 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1678 struct io_ring_ctx *ctx = req->ctx;
1679 struct io_kiocb *poll_req, *next;
1682 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1684 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1688 spin_lock_irq(&ctx->completion_lock);
1689 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1690 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1691 io_poll_remove_one(poll_req);
1696 spin_unlock_irq(&ctx->completion_lock);
1698 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1703 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1706 req->poll.done = true;
1707 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1708 io_commit_cqring(ctx);
1711 static void io_poll_complete_work(struct work_struct *work)
1713 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1714 struct io_poll_iocb *poll = &req->poll;
1715 struct poll_table_struct pt = { ._key = poll->events };
1716 struct io_ring_ctx *ctx = req->ctx;
1719 if (!READ_ONCE(poll->canceled))
1720 mask = vfs_poll(poll->file, &pt) & poll->events;
1723 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1724 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1725 * synchronize with them. In the cancellation case the list_del_init
1726 * itself is not actually needed, but harmless so we keep it in to
1727 * avoid further branches in the fast path.
1729 spin_lock_irq(&ctx->completion_lock);
1730 if (!mask && !READ_ONCE(poll->canceled)) {
1731 add_wait_queue(poll->head, &poll->wait);
1732 spin_unlock_irq(&ctx->completion_lock);
1735 list_del_init(&req->list);
1736 io_poll_complete(ctx, req, mask);
1737 spin_unlock_irq(&ctx->completion_lock);
1739 io_cqring_ev_posted(ctx);
1743 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1746 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1748 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1749 struct io_ring_ctx *ctx = req->ctx;
1750 __poll_t mask = key_to_poll(key);
1751 unsigned long flags;
1753 /* for instances that support it check for an event match first: */
1754 if (mask && !(mask & poll->events))
1757 list_del_init(&poll->wait.entry);
1759 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1760 list_del(&req->list);
1761 io_poll_complete(ctx, req, mask);
1762 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1764 io_cqring_ev_posted(ctx);
1767 io_queue_async_work(ctx, req);
1773 struct io_poll_table {
1774 struct poll_table_struct pt;
1775 struct io_kiocb *req;
1779 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1780 struct poll_table_struct *p)
1782 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1784 if (unlikely(pt->req->poll.head)) {
1785 pt->error = -EINVAL;
1790 pt->req->poll.head = head;
1791 add_wait_queue(head, &pt->req->poll.wait);
1794 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1796 struct io_poll_iocb *poll = &req->poll;
1797 struct io_ring_ctx *ctx = req->ctx;
1798 struct io_poll_table ipt;
1799 bool cancel = false;
1803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1805 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1810 req->submit.sqe = NULL;
1811 INIT_WORK(&req->work, io_poll_complete_work);
1812 events = READ_ONCE(sqe->poll_events);
1813 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1817 poll->canceled = false;
1819 ipt.pt._qproc = io_poll_queue_proc;
1820 ipt.pt._key = poll->events;
1822 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1824 /* initialized the list so that we can do list_empty checks */
1825 INIT_LIST_HEAD(&poll->wait.entry);
1826 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1828 INIT_LIST_HEAD(&req->list);
1830 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1832 spin_lock_irq(&ctx->completion_lock);
1833 if (likely(poll->head)) {
1834 spin_lock(&poll->head->lock);
1835 if (unlikely(list_empty(&poll->wait.entry))) {
1841 if (mask || ipt.error)
1842 list_del_init(&poll->wait.entry);
1844 WRITE_ONCE(poll->canceled, true);
1845 else if (!poll->done) /* actually waiting for an event */
1846 list_add_tail(&req->list, &ctx->cancel_list);
1847 spin_unlock(&poll->head->lock);
1849 if (mask) { /* no async, we'd stolen it */
1851 io_poll_complete(ctx, req, mask);
1853 spin_unlock_irq(&ctx->completion_lock);
1856 io_cqring_ev_posted(ctx);
1862 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
1864 struct io_ring_ctx *ctx;
1865 struct io_kiocb *req;
1866 unsigned long flags;
1868 req = container_of(timer, struct io_kiocb, timeout.timer);
1870 atomic_inc(&ctx->cq_timeouts);
1872 spin_lock_irqsave(&ctx->completion_lock, flags);
1873 list_del(&req->list);
1875 io_cqring_fill_event(ctx, req->user_data, -ETIME);
1876 io_commit_cqring(ctx);
1877 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1879 io_cqring_ev_posted(ctx);
1882 return HRTIMER_NORESTART;
1885 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1887 unsigned count, req_dist, tail_index;
1888 struct io_ring_ctx *ctx = req->ctx;
1889 struct list_head *entry;
1890 struct timespec64 ts;
1892 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1894 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->timeout_flags ||
1898 if (get_timespec64(&ts, u64_to_user_ptr(sqe->addr)))
1902 * sqe->off holds how many events that need to occur for this
1903 * timeout event to be satisfied.
1905 count = READ_ONCE(sqe->off);
1909 req->sequence = ctx->cached_sq_head + count - 1;
1910 req->flags |= REQ_F_TIMEOUT;
1913 * Insertion sort, ensuring the first entry in the list is always
1914 * the one we need first.
1916 tail_index = ctx->cached_cq_tail - ctx->rings->sq_dropped;
1917 req_dist = req->sequence - tail_index;
1918 spin_lock_irq(&ctx->completion_lock);
1919 list_for_each_prev(entry, &ctx->timeout_list) {
1920 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
1923 dist = nxt->sequence - tail_index;
1924 if (req_dist >= dist)
1927 list_add(&req->list, entry);
1928 spin_unlock_irq(&ctx->completion_lock);
1930 hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1931 req->timeout.timer.function = io_timeout_fn;
1932 hrtimer_start(&req->timeout.timer, timespec64_to_ktime(ts),
1937 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1938 const struct io_uring_sqe *sqe)
1940 struct io_uring_sqe *sqe_copy;
1942 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1945 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1949 spin_lock_irq(&ctx->completion_lock);
1950 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1951 spin_unlock_irq(&ctx->completion_lock);
1956 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1957 req->submit.sqe = sqe_copy;
1959 INIT_WORK(&req->work, io_sq_wq_submit_work);
1960 list_add_tail(&req->list, &ctx->defer_list);
1961 spin_unlock_irq(&ctx->completion_lock);
1962 return -EIOCBQUEUED;
1965 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1966 const struct sqe_submit *s, bool force_nonblock)
1970 req->user_data = READ_ONCE(s->sqe->user_data);
1972 if (unlikely(s->index >= ctx->sq_entries))
1975 opcode = READ_ONCE(s->sqe->opcode);
1978 ret = io_nop(req, req->user_data);
1980 case IORING_OP_READV:
1981 if (unlikely(s->sqe->buf_index))
1983 ret = io_read(req, s, force_nonblock);
1985 case IORING_OP_WRITEV:
1986 if (unlikely(s->sqe->buf_index))
1988 ret = io_write(req, s, force_nonblock);
1990 case IORING_OP_READ_FIXED:
1991 ret = io_read(req, s, force_nonblock);
1993 case IORING_OP_WRITE_FIXED:
1994 ret = io_write(req, s, force_nonblock);
1996 case IORING_OP_FSYNC:
1997 ret = io_fsync(req, s->sqe, force_nonblock);
1999 case IORING_OP_POLL_ADD:
2000 ret = io_poll_add(req, s->sqe);
2002 case IORING_OP_POLL_REMOVE:
2003 ret = io_poll_remove(req, s->sqe);
2005 case IORING_OP_SYNC_FILE_RANGE:
2006 ret = io_sync_file_range(req, s->sqe, force_nonblock);
2008 case IORING_OP_SENDMSG:
2009 ret = io_sendmsg(req, s->sqe, force_nonblock);
2011 case IORING_OP_RECVMSG:
2012 ret = io_recvmsg(req, s->sqe, force_nonblock);
2014 case IORING_OP_TIMEOUT:
2015 ret = io_timeout(req, s->sqe);
2025 if (ctx->flags & IORING_SETUP_IOPOLL) {
2026 if (req->result == -EAGAIN)
2029 /* workqueue context doesn't hold uring_lock, grab it now */
2031 mutex_lock(&ctx->uring_lock);
2032 io_iopoll_req_issued(req);
2034 mutex_unlock(&ctx->uring_lock);
2040 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
2041 const struct io_uring_sqe *sqe)
2043 switch (sqe->opcode) {
2044 case IORING_OP_READV:
2045 case IORING_OP_READ_FIXED:
2046 return &ctx->pending_async[READ];
2047 case IORING_OP_WRITEV:
2048 case IORING_OP_WRITE_FIXED:
2049 return &ctx->pending_async[WRITE];
2055 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
2057 u8 opcode = READ_ONCE(sqe->opcode);
2059 return !(opcode == IORING_OP_READ_FIXED ||
2060 opcode == IORING_OP_WRITE_FIXED);
2063 static void io_sq_wq_submit_work(struct work_struct *work)
2065 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2066 struct io_ring_ctx *ctx = req->ctx;
2067 struct mm_struct *cur_mm = NULL;
2068 struct async_list *async_list;
2069 LIST_HEAD(req_list);
2070 mm_segment_t old_fs;
2073 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
2076 struct sqe_submit *s = &req->submit;
2077 const struct io_uring_sqe *sqe = s->sqe;
2078 unsigned int flags = req->flags;
2080 /* Ensure we clear previously set non-block flag */
2081 req->rw.ki_flags &= ~IOCB_NOWAIT;
2084 if (io_sqe_needs_user(sqe) && !cur_mm) {
2085 if (!mmget_not_zero(ctx->sqo_mm)) {
2088 cur_mm = ctx->sqo_mm;
2096 s->has_user = cur_mm != NULL;
2097 s->needs_lock = true;
2099 ret = __io_submit_sqe(ctx, req, s, false);
2101 * We can get EAGAIN for polled IO even though
2102 * we're forcing a sync submission from here,
2103 * since we can't wait for request slots on the
2112 /* drop submission reference */
2116 io_cqring_add_event(ctx, sqe->user_data, ret);
2120 /* async context always use a copy of the sqe */
2123 /* req from defer and link list needn't decrease async cnt */
2124 if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE))
2129 if (!list_empty(&req_list)) {
2130 req = list_first_entry(&req_list, struct io_kiocb,
2132 list_del(&req->list);
2135 if (list_empty(&async_list->list))
2139 spin_lock(&async_list->lock);
2140 if (list_empty(&async_list->list)) {
2141 spin_unlock(&async_list->lock);
2144 list_splice_init(&async_list->list, &req_list);
2145 spin_unlock(&async_list->lock);
2147 req = list_first_entry(&req_list, struct io_kiocb, list);
2148 list_del(&req->list);
2152 * Rare case of racing with a submitter. If we find the count has
2153 * dropped to zero AND we have pending work items, then restart
2154 * the processing. This is a tiny race window.
2157 ret = atomic_dec_return(&async_list->cnt);
2158 while (!ret && !list_empty(&async_list->list)) {
2159 spin_lock(&async_list->lock);
2160 atomic_inc(&async_list->cnt);
2161 list_splice_init(&async_list->list, &req_list);
2162 spin_unlock(&async_list->lock);
2164 if (!list_empty(&req_list)) {
2165 req = list_first_entry(&req_list,
2166 struct io_kiocb, list);
2167 list_del(&req->list);
2170 ret = atomic_dec_return(&async_list->cnt);
2183 * See if we can piggy back onto previously submitted work, that is still
2184 * running. We currently only allow this if the new request is sequential
2185 * to the previous one we punted.
2187 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
2193 if (!(req->flags & REQ_F_SEQ_PREV))
2195 if (!atomic_read(&list->cnt))
2199 spin_lock(&list->lock);
2200 list_add_tail(&req->list, &list->list);
2202 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
2205 if (!atomic_read(&list->cnt)) {
2206 list_del_init(&req->list);
2209 spin_unlock(&list->lock);
2213 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2215 int op = READ_ONCE(sqe->opcode);
2219 case IORING_OP_POLL_REMOVE:
2226 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2227 struct io_submit_state *state, struct io_kiocb *req)
2232 flags = READ_ONCE(s->sqe->flags);
2233 fd = READ_ONCE(s->sqe->fd);
2235 if (flags & IOSQE_IO_DRAIN)
2236 req->flags |= REQ_F_IO_DRAIN;
2238 * All io need record the previous position, if LINK vs DARIN,
2239 * it can be used to mark the position of the first IO in the
2242 req->sequence = s->sequence;
2244 if (!io_op_needs_file(s->sqe))
2247 if (flags & IOSQE_FIXED_FILE) {
2248 if (unlikely(!ctx->user_files ||
2249 (unsigned) fd >= ctx->nr_user_files))
2251 req->file = ctx->user_files[fd];
2252 req->flags |= REQ_F_FIXED_FILE;
2254 if (s->needs_fixed_file)
2256 req->file = io_file_get(state, fd);
2257 if (unlikely(!req->file))
2264 static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2265 struct sqe_submit *s, bool force_nonblock)
2269 ret = __io_submit_sqe(ctx, req, s, force_nonblock);
2270 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
2271 struct io_uring_sqe *sqe_copy;
2273 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2275 struct async_list *list;
2278 memcpy(&req->submit, s, sizeof(*s));
2279 list = io_async_list_from_sqe(ctx, s->sqe);
2280 if (!io_add_to_prev_work(list, req)) {
2282 atomic_inc(&list->cnt);
2283 INIT_WORK(&req->work, io_sq_wq_submit_work);
2284 io_queue_async_work(ctx, req);
2288 * Queued up for async execution, worker will release
2289 * submit reference when the iocb is actually submitted.
2295 /* drop submission reference */
2298 /* and drop final reference, if we failed */
2300 io_cqring_add_event(ctx, req->user_data, ret);
2301 if (req->flags & REQ_F_LINK)
2302 req->flags |= REQ_F_FAIL_LINK;
2309 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2310 struct sqe_submit *s, bool force_nonblock)
2314 ret = io_req_defer(ctx, req, s->sqe);
2316 if (ret != -EIOCBQUEUED) {
2318 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2323 return __io_queue_sqe(ctx, req, s, force_nonblock);
2326 static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
2327 struct sqe_submit *s, struct io_kiocb *shadow,
2328 bool force_nonblock)
2331 int need_submit = false;
2334 return io_queue_sqe(ctx, req, s, force_nonblock);
2337 * Mark the first IO in link list as DRAIN, let all the following
2338 * IOs enter the defer list. all IO needs to be completed before link
2341 req->flags |= REQ_F_IO_DRAIN;
2342 ret = io_req_defer(ctx, req, s->sqe);
2344 if (ret != -EIOCBQUEUED) {
2346 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2351 * If ret == 0 means that all IOs in front of link io are
2352 * running done. let's queue link head.
2357 /* Insert shadow req to defer_list, blocking next IOs */
2358 spin_lock_irq(&ctx->completion_lock);
2359 list_add_tail(&shadow->list, &ctx->defer_list);
2360 spin_unlock_irq(&ctx->completion_lock);
2363 return __io_queue_sqe(ctx, req, s, force_nonblock);
2368 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2370 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2371 struct io_submit_state *state, struct io_kiocb **link,
2372 bool force_nonblock)
2374 struct io_uring_sqe *sqe_copy;
2375 struct io_kiocb *req;
2378 /* enforce forwards compatibility on users */
2379 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2384 req = io_get_req(ctx, state);
2385 if (unlikely(!req)) {
2390 ret = io_req_set_file(ctx, s, state, req);
2391 if (unlikely(ret)) {
2395 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2400 * If we already have a head request, queue this one for async
2401 * submittal once the head completes. If we don't have a head but
2402 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2403 * submitted sync once the chain is complete. If none of those
2404 * conditions are true (normal request), then just queue it.
2407 struct io_kiocb *prev = *link;
2409 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2416 memcpy(&req->submit, s, sizeof(*s));
2417 list_add_tail(&req->list, &prev->link_list);
2418 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2419 req->flags |= REQ_F_LINK;
2421 memcpy(&req->submit, s, sizeof(*s));
2422 INIT_LIST_HEAD(&req->link_list);
2425 io_queue_sqe(ctx, req, s, force_nonblock);
2430 * Batched submission is done, ensure local IO is flushed out.
2432 static void io_submit_state_end(struct io_submit_state *state)
2434 blk_finish_plug(&state->plug);
2436 if (state->free_reqs)
2437 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2438 &state->reqs[state->cur_req]);
2442 * Start submission side cache.
2444 static void io_submit_state_start(struct io_submit_state *state,
2445 struct io_ring_ctx *ctx, unsigned max_ios)
2447 blk_start_plug(&state->plug);
2448 state->free_reqs = 0;
2450 state->ios_left = max_ios;
2453 static void io_commit_sqring(struct io_ring_ctx *ctx)
2455 struct io_rings *rings = ctx->rings;
2457 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
2459 * Ensure any loads from the SQEs are done at this point,
2460 * since once we write the new head, the application could
2461 * write new data to them.
2463 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2468 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2469 * that is mapped by userspace. This means that care needs to be taken to
2470 * ensure that reads are stable, as we cannot rely on userspace always
2471 * being a good citizen. If members of the sqe are validated and then later
2472 * used, it's important that those reads are done through READ_ONCE() to
2473 * prevent a re-load down the line.
2475 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2477 struct io_rings *rings = ctx->rings;
2478 u32 *sq_array = ctx->sq_array;
2482 * The cached sq head (or cq tail) serves two purposes:
2484 * 1) allows us to batch the cost of updating the user visible
2486 * 2) allows the kernel side to track the head on its own, even
2487 * though the application is the one updating it.
2489 head = ctx->cached_sq_head;
2490 /* make sure SQ entry isn't read before tail */
2491 if (head == smp_load_acquire(&rings->sq.tail))
2494 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
2495 if (head < ctx->sq_entries) {
2497 s->sqe = &ctx->sq_sqes[head];
2498 s->sequence = ctx->cached_sq_head;
2499 ctx->cached_sq_head++;
2503 /* drop invalid entries */
2504 ctx->cached_sq_head++;
2505 rings->sq_dropped++;
2509 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2510 unsigned int nr, bool has_user, bool mm_fault)
2512 struct io_submit_state state, *statep = NULL;
2513 struct io_kiocb *link = NULL;
2514 struct io_kiocb *shadow_req = NULL;
2515 bool prev_was_link = false;
2516 int i, submitted = 0;
2518 if (nr > IO_PLUG_THRESHOLD) {
2519 io_submit_state_start(&state, ctx, nr);
2523 for (i = 0; i < nr; i++) {
2525 * If previous wasn't linked and we have a linked command,
2526 * that's the end of the chain. Submit the previous link.
2528 if (!prev_was_link && link) {
2529 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2534 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2536 if (link && (sqes[i].sqe->flags & IOSQE_IO_DRAIN)) {
2538 shadow_req = io_get_req(ctx, NULL);
2539 if (unlikely(!shadow_req))
2541 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2542 refcount_dec(&shadow_req->refs);
2544 shadow_req->sequence = sqes[i].sequence;
2548 if (unlikely(mm_fault)) {
2549 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2552 sqes[i].has_user = has_user;
2553 sqes[i].needs_lock = true;
2554 sqes[i].needs_fixed_file = true;
2555 io_submit_sqe(ctx, &sqes[i], statep, &link, true);
2561 io_queue_link_head(ctx, link, &link->submit, shadow_req, true);
2563 io_submit_state_end(&state);
2568 static int io_sq_thread(void *data)
2570 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2571 struct io_ring_ctx *ctx = data;
2572 struct mm_struct *cur_mm = NULL;
2573 mm_segment_t old_fs;
2576 unsigned long timeout;
2578 complete(&ctx->sqo_thread_started);
2583 timeout = inflight = 0;
2584 while (!kthread_should_park()) {
2585 bool all_fixed, mm_fault = false;
2589 unsigned nr_events = 0;
2591 if (ctx->flags & IORING_SETUP_IOPOLL) {
2592 io_iopoll_check(ctx, &nr_events, 0);
2595 * Normal IO, just pretend everything completed.
2596 * We don't have to poll completions for that.
2598 nr_events = inflight;
2601 inflight -= nr_events;
2603 timeout = jiffies + ctx->sq_thread_idle;
2606 if (!io_get_sqring(ctx, &sqes[0])) {
2608 * We're polling. If we're within the defined idle
2609 * period, then let us spin without work before going
2612 if (inflight || !time_after(jiffies, timeout)) {
2618 * Drop cur_mm before scheduling, we can't hold it for
2619 * long periods (or over schedule()). Do this before
2620 * adding ourselves to the waitqueue, as the unuse/drop
2629 prepare_to_wait(&ctx->sqo_wait, &wait,
2630 TASK_INTERRUPTIBLE);
2632 /* Tell userspace we may need a wakeup call */
2633 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
2634 /* make sure to read SQ tail after writing flags */
2637 if (!io_get_sqring(ctx, &sqes[0])) {
2638 if (kthread_should_park()) {
2639 finish_wait(&ctx->sqo_wait, &wait);
2642 if (signal_pending(current))
2643 flush_signals(current);
2645 finish_wait(&ctx->sqo_wait, &wait);
2647 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2650 finish_wait(&ctx->sqo_wait, &wait);
2652 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2658 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2662 if (i == ARRAY_SIZE(sqes))
2664 } while (io_get_sqring(ctx, &sqes[i]));
2666 /* Unless all new commands are FIXED regions, grab mm */
2667 if (!all_fixed && !cur_mm) {
2668 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2670 use_mm(ctx->sqo_mm);
2671 cur_mm = ctx->sqo_mm;
2675 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2678 /* Commit SQ ring head once we've consumed all SQEs */
2679 io_commit_sqring(ctx);
2693 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit,
2694 bool block_for_last)
2696 struct io_submit_state state, *statep = NULL;
2697 struct io_kiocb *link = NULL;
2698 struct io_kiocb *shadow_req = NULL;
2699 bool prev_was_link = false;
2702 if (to_submit > IO_PLUG_THRESHOLD) {
2703 io_submit_state_start(&state, ctx, to_submit);
2707 for (i = 0; i < to_submit; i++) {
2708 bool force_nonblock = true;
2709 struct sqe_submit s;
2711 if (!io_get_sqring(ctx, &s))
2715 * If previous wasn't linked and we have a linked command,
2716 * that's the end of the chain. Submit the previous link.
2718 if (!prev_was_link && link) {
2719 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2724 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2726 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
2728 shadow_req = io_get_req(ctx, NULL);
2729 if (unlikely(!shadow_req))
2731 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2732 refcount_dec(&shadow_req->refs);
2734 shadow_req->sequence = s.sequence;
2739 s.needs_lock = false;
2740 s.needs_fixed_file = false;
2744 * The caller will block for events after submit, submit the
2745 * last IO non-blocking. This is either the only IO it's
2746 * submitting, or it already submitted the previous ones. This
2747 * improves performance by avoiding an async punt that we don't
2750 if (block_for_last && submit == to_submit)
2751 force_nonblock = false;
2753 io_submit_sqe(ctx, &s, statep, &link, force_nonblock);
2755 io_commit_sqring(ctx);
2758 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2761 io_submit_state_end(statep);
2766 struct io_wait_queue {
2767 struct wait_queue_entry wq;
2768 struct io_ring_ctx *ctx;
2770 unsigned nr_timeouts;
2773 static inline bool io_should_wake(struct io_wait_queue *iowq)
2775 struct io_ring_ctx *ctx = iowq->ctx;
2778 * Wake up if we have enough events, or if a timeout occured since we
2779 * started waiting. For timeouts, we always want to return to userspace,
2780 * regardless of event count.
2782 return io_cqring_events(ctx->rings) >= iowq->to_wait ||
2783 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
2786 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
2787 int wake_flags, void *key)
2789 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
2792 if (!io_should_wake(iowq))
2795 return autoremove_wake_function(curr, mode, wake_flags, key);
2799 * Wait until events become available, if we don't already have some. The
2800 * application must reap them itself, as they reside on the shared cq ring.
2802 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2803 const sigset_t __user *sig, size_t sigsz)
2805 struct io_wait_queue iowq = {
2808 .func = io_wake_function,
2809 .entry = LIST_HEAD_INIT(iowq.wq.entry),
2812 .to_wait = min_events,
2814 struct io_rings *rings = ctx->rings;
2817 if (io_cqring_events(rings) >= min_events)
2821 #ifdef CONFIG_COMPAT
2822 if (in_compat_syscall())
2823 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2827 ret = set_user_sigmask(sig, sigsz);
2834 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
2836 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
2837 TASK_INTERRUPTIBLE);
2838 if (io_should_wake(&iowq))
2841 if (signal_pending(current)) {
2846 finish_wait(&ctx->wait, &iowq.wq);
2848 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
2849 if (ret == -ERESTARTSYS)
2852 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2855 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2857 #if defined(CONFIG_UNIX)
2858 if (ctx->ring_sock) {
2859 struct sock *sock = ctx->ring_sock->sk;
2860 struct sk_buff *skb;
2862 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2868 for (i = 0; i < ctx->nr_user_files; i++)
2869 fput(ctx->user_files[i]);
2873 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2875 if (!ctx->user_files)
2878 __io_sqe_files_unregister(ctx);
2879 kfree(ctx->user_files);
2880 ctx->user_files = NULL;
2881 ctx->nr_user_files = 0;
2885 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2887 if (ctx->sqo_thread) {
2888 wait_for_completion(&ctx->sqo_thread_started);
2890 * The park is a bit of a work-around, without it we get
2891 * warning spews on shutdown with SQPOLL set and affinity
2892 * set to a single CPU.
2894 kthread_park(ctx->sqo_thread);
2895 kthread_stop(ctx->sqo_thread);
2896 ctx->sqo_thread = NULL;
2900 static void io_finish_async(struct io_ring_ctx *ctx)
2904 io_sq_thread_stop(ctx);
2906 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) {
2907 if (ctx->sqo_wq[i]) {
2908 destroy_workqueue(ctx->sqo_wq[i]);
2909 ctx->sqo_wq[i] = NULL;
2914 #if defined(CONFIG_UNIX)
2915 static void io_destruct_skb(struct sk_buff *skb)
2917 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2920 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++)
2922 flush_workqueue(ctx->sqo_wq[i]);
2924 unix_destruct_scm(skb);
2928 * Ensure the UNIX gc is aware of our file set, so we are certain that
2929 * the io_uring can be safely unregistered on process exit, even if we have
2930 * loops in the file referencing.
2932 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2934 struct sock *sk = ctx->ring_sock->sk;
2935 struct scm_fp_list *fpl;
2936 struct sk_buff *skb;
2939 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2940 unsigned long inflight = ctx->user->unix_inflight + nr;
2942 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2946 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2950 skb = alloc_skb(0, GFP_KERNEL);
2957 skb->destructor = io_destruct_skb;
2959 fpl->user = get_uid(ctx->user);
2960 for (i = 0; i < nr; i++) {
2961 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2962 unix_inflight(fpl->user, fpl->fp[i]);
2965 fpl->max = fpl->count = nr;
2966 UNIXCB(skb).fp = fpl;
2967 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2968 skb_queue_head(&sk->sk_receive_queue, skb);
2970 for (i = 0; i < nr; i++)
2977 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2978 * causes regular reference counting to break down. We rely on the UNIX
2979 * garbage collection to take care of this problem for us.
2981 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2983 unsigned left, total;
2987 left = ctx->nr_user_files;
2989 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2991 ret = __io_sqe_files_scm(ctx, this_files, total);
2995 total += this_files;
3001 while (total < ctx->nr_user_files) {
3002 fput(ctx->user_files[total]);
3009 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3015 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3018 __s32 __user *fds = (__s32 __user *) arg;
3022 if (ctx->user_files)
3026 if (nr_args > IORING_MAX_FIXED_FILES)
3029 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
3030 if (!ctx->user_files)
3033 for (i = 0; i < nr_args; i++) {
3035 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3038 ctx->user_files[i] = fget(fd);
3041 if (!ctx->user_files[i])
3044 * Don't allow io_uring instances to be registered. If UNIX
3045 * isn't enabled, then this causes a reference cycle and this
3046 * instance can never get freed. If UNIX is enabled we'll
3047 * handle it just fine, but there's still no point in allowing
3048 * a ring fd as it doesn't support regular read/write anyway.
3050 if (ctx->user_files[i]->f_op == &io_uring_fops) {
3051 fput(ctx->user_files[i]);
3054 ctx->nr_user_files++;
3059 for (i = 0; i < ctx->nr_user_files; i++)
3060 fput(ctx->user_files[i]);
3062 kfree(ctx->user_files);
3063 ctx->user_files = NULL;
3064 ctx->nr_user_files = 0;
3068 ret = io_sqe_files_scm(ctx);
3070 io_sqe_files_unregister(ctx);
3075 static int io_sq_offload_start(struct io_ring_ctx *ctx,
3076 struct io_uring_params *p)
3080 init_waitqueue_head(&ctx->sqo_wait);
3081 mmgrab(current->mm);
3082 ctx->sqo_mm = current->mm;
3084 if (ctx->flags & IORING_SETUP_SQPOLL) {
3086 if (!capable(CAP_SYS_ADMIN))
3089 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
3090 if (!ctx->sq_thread_idle)
3091 ctx->sq_thread_idle = HZ;
3093 if (p->flags & IORING_SETUP_SQ_AFF) {
3094 int cpu = p->sq_thread_cpu;
3097 if (cpu >= nr_cpu_ids)
3099 if (!cpu_online(cpu))
3102 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
3106 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
3109 if (IS_ERR(ctx->sqo_thread)) {
3110 ret = PTR_ERR(ctx->sqo_thread);
3111 ctx->sqo_thread = NULL;
3114 wake_up_process(ctx->sqo_thread);
3115 } else if (p->flags & IORING_SETUP_SQ_AFF) {
3116 /* Can't have SQ_AFF without SQPOLL */
3121 /* Do QD, or 2 * CPUS, whatever is smallest */
3122 ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq",
3123 WQ_UNBOUND | WQ_FREEZABLE,
3124 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
3125 if (!ctx->sqo_wq[0]) {
3131 * This is for buffered writes, where we want to limit the parallelism
3132 * due to file locking in file systems. As "normal" buffered writes
3133 * should parellelize on writeout quite nicely, limit us to having 2
3134 * pending. This avoids massive contention on the inode when doing
3135 * buffered async writes.
3137 ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq",
3138 WQ_UNBOUND | WQ_FREEZABLE, 2);
3139 if (!ctx->sqo_wq[1]) {
3146 io_finish_async(ctx);
3147 mmdrop(ctx->sqo_mm);
3152 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
3154 atomic_long_sub(nr_pages, &user->locked_vm);
3157 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
3159 unsigned long page_limit, cur_pages, new_pages;
3161 /* Don't allow more pages than we can safely lock */
3162 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
3165 cur_pages = atomic_long_read(&user->locked_vm);
3166 new_pages = cur_pages + nr_pages;
3167 if (new_pages > page_limit)
3169 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
3170 new_pages) != cur_pages);
3175 static void io_mem_free(void *ptr)
3182 page = virt_to_head_page(ptr);
3183 if (put_page_testzero(page))
3184 free_compound_page(page);
3187 static void *io_mem_alloc(size_t size)
3189 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
3192 return (void *) __get_free_pages(gfp_flags, get_order(size));
3195 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
3198 struct io_rings *rings;
3199 size_t off, sq_array_size;
3201 off = struct_size(rings, cqes, cq_entries);
3202 if (off == SIZE_MAX)
3206 off = ALIGN(off, SMP_CACHE_BYTES);
3211 sq_array_size = array_size(sizeof(u32), sq_entries);
3212 if (sq_array_size == SIZE_MAX)
3215 if (check_add_overflow(off, sq_array_size, &off))
3224 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
3228 pages = (size_t)1 << get_order(
3229 rings_size(sq_entries, cq_entries, NULL));
3230 pages += (size_t)1 << get_order(
3231 array_size(sizeof(struct io_uring_sqe), sq_entries));
3236 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
3240 if (!ctx->user_bufs)
3243 for (i = 0; i < ctx->nr_user_bufs; i++) {
3244 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3246 for (j = 0; j < imu->nr_bvecs; j++)
3247 put_user_page(imu->bvec[j].bv_page);
3249 if (ctx->account_mem)
3250 io_unaccount_mem(ctx->user, imu->nr_bvecs);
3255 kfree(ctx->user_bufs);
3256 ctx->user_bufs = NULL;
3257 ctx->nr_user_bufs = 0;
3261 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
3262 void __user *arg, unsigned index)
3264 struct iovec __user *src;
3266 #ifdef CONFIG_COMPAT
3268 struct compat_iovec __user *ciovs;
3269 struct compat_iovec ciov;
3271 ciovs = (struct compat_iovec __user *) arg;
3272 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
3275 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
3276 dst->iov_len = ciov.iov_len;
3280 src = (struct iovec __user *) arg;
3281 if (copy_from_user(dst, &src[index], sizeof(*dst)))
3286 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
3289 struct vm_area_struct **vmas = NULL;
3290 struct page **pages = NULL;
3291 int i, j, got_pages = 0;
3296 if (!nr_args || nr_args > UIO_MAXIOV)
3299 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
3301 if (!ctx->user_bufs)
3304 for (i = 0; i < nr_args; i++) {
3305 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3306 unsigned long off, start, end, ubuf;
3311 ret = io_copy_iov(ctx, &iov, arg, i);
3316 * Don't impose further limits on the size and buffer
3317 * constraints here, we'll -EINVAL later when IO is
3318 * submitted if they are wrong.
3321 if (!iov.iov_base || !iov.iov_len)
3324 /* arbitrary limit, but we need something */
3325 if (iov.iov_len > SZ_1G)
3328 ubuf = (unsigned long) iov.iov_base;
3329 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3330 start = ubuf >> PAGE_SHIFT;
3331 nr_pages = end - start;
3333 if (ctx->account_mem) {
3334 ret = io_account_mem(ctx->user, nr_pages);
3340 if (!pages || nr_pages > got_pages) {
3343 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
3345 vmas = kvmalloc_array(nr_pages,
3346 sizeof(struct vm_area_struct *),
3348 if (!pages || !vmas) {
3350 if (ctx->account_mem)
3351 io_unaccount_mem(ctx->user, nr_pages);
3354 got_pages = nr_pages;
3357 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
3361 if (ctx->account_mem)
3362 io_unaccount_mem(ctx->user, nr_pages);
3367 down_read(¤t->mm->mmap_sem);
3368 pret = get_user_pages(ubuf, nr_pages,
3369 FOLL_WRITE | FOLL_LONGTERM,
3371 if (pret == nr_pages) {
3372 /* don't support file backed memory */
3373 for (j = 0; j < nr_pages; j++) {
3374 struct vm_area_struct *vma = vmas[j];
3377 !is_file_hugepages(vma->vm_file)) {
3383 ret = pret < 0 ? pret : -EFAULT;
3385 up_read(¤t->mm->mmap_sem);
3388 * if we did partial map, or found file backed vmas,
3389 * release any pages we did get
3392 put_user_pages(pages, pret);
3393 if (ctx->account_mem)
3394 io_unaccount_mem(ctx->user, nr_pages);
3399 off = ubuf & ~PAGE_MASK;
3401 for (j = 0; j < nr_pages; j++) {
3404 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3405 imu->bvec[j].bv_page = pages[j];
3406 imu->bvec[j].bv_len = vec_len;
3407 imu->bvec[j].bv_offset = off;
3411 /* store original address for later verification */
3413 imu->len = iov.iov_len;
3414 imu->nr_bvecs = nr_pages;
3416 ctx->nr_user_bufs++;
3424 io_sqe_buffer_unregister(ctx);
3428 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3430 __s32 __user *fds = arg;
3436 if (copy_from_user(&fd, fds, sizeof(*fds)))
3439 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3440 if (IS_ERR(ctx->cq_ev_fd)) {
3441 int ret = PTR_ERR(ctx->cq_ev_fd);
3442 ctx->cq_ev_fd = NULL;
3449 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3451 if (ctx->cq_ev_fd) {
3452 eventfd_ctx_put(ctx->cq_ev_fd);
3453 ctx->cq_ev_fd = NULL;
3460 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3462 io_finish_async(ctx);
3464 mmdrop(ctx->sqo_mm);
3466 io_iopoll_reap_events(ctx);
3467 io_sqe_buffer_unregister(ctx);
3468 io_sqe_files_unregister(ctx);
3469 io_eventfd_unregister(ctx);
3471 #if defined(CONFIG_UNIX)
3472 if (ctx->ring_sock) {
3473 ctx->ring_sock->file = NULL; /* so that iput() is called */
3474 sock_release(ctx->ring_sock);
3478 io_mem_free(ctx->rings);
3479 io_mem_free(ctx->sq_sqes);
3481 percpu_ref_exit(&ctx->refs);
3482 if (ctx->account_mem)
3483 io_unaccount_mem(ctx->user,
3484 ring_pages(ctx->sq_entries, ctx->cq_entries));
3485 free_uid(ctx->user);
3489 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3491 struct io_ring_ctx *ctx = file->private_data;
3494 poll_wait(file, &ctx->cq_wait, wait);
3496 * synchronizes with barrier from wq_has_sleeper call in
3500 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
3501 ctx->rings->sq_ring_entries)
3502 mask |= EPOLLOUT | EPOLLWRNORM;
3503 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
3504 mask |= EPOLLIN | EPOLLRDNORM;
3509 static int io_uring_fasync(int fd, struct file *file, int on)
3511 struct io_ring_ctx *ctx = file->private_data;
3513 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3516 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3518 mutex_lock(&ctx->uring_lock);
3519 percpu_ref_kill(&ctx->refs);
3520 mutex_unlock(&ctx->uring_lock);
3522 io_kill_timeouts(ctx);
3523 io_poll_remove_all(ctx);
3524 io_iopoll_reap_events(ctx);
3525 wait_for_completion(&ctx->ctx_done);
3526 io_ring_ctx_free(ctx);
3529 static int io_uring_release(struct inode *inode, struct file *file)
3531 struct io_ring_ctx *ctx = file->private_data;
3533 file->private_data = NULL;
3534 io_ring_ctx_wait_and_kill(ctx);
3538 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3540 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3541 unsigned long sz = vma->vm_end - vma->vm_start;
3542 struct io_ring_ctx *ctx = file->private_data;
3548 case IORING_OFF_SQ_RING:
3549 case IORING_OFF_CQ_RING:
3552 case IORING_OFF_SQES:
3559 page = virt_to_head_page(ptr);
3560 if (sz > page_size(page))
3563 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3564 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3567 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3568 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3571 struct io_ring_ctx *ctx;
3576 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3584 if (f.file->f_op != &io_uring_fops)
3588 ctx = f.file->private_data;
3589 if (!percpu_ref_tryget(&ctx->refs))
3593 * For SQ polling, the thread will do all submissions and completions.
3594 * Just return the requested submit count, and wake the thread if
3598 if (ctx->flags & IORING_SETUP_SQPOLL) {
3599 if (flags & IORING_ENTER_SQ_WAKEUP)
3600 wake_up(&ctx->sqo_wait);
3601 submitted = to_submit;
3602 } else if (to_submit) {
3603 bool block_for_last = false;
3605 to_submit = min(to_submit, ctx->sq_entries);
3608 * Allow last submission to block in a series, IFF the caller
3609 * asked to wait for events and we don't currently have
3610 * enough. This potentially avoids an async punt.
3612 if (to_submit == min_complete &&
3613 io_cqring_events(ctx->rings) < min_complete)
3614 block_for_last = true;
3616 mutex_lock(&ctx->uring_lock);
3617 submitted = io_ring_submit(ctx, to_submit, block_for_last);
3618 mutex_unlock(&ctx->uring_lock);
3620 if (flags & IORING_ENTER_GETEVENTS) {
3621 unsigned nr_events = 0;
3623 min_complete = min(min_complete, ctx->cq_entries);
3625 if (ctx->flags & IORING_SETUP_IOPOLL) {
3626 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3628 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3632 percpu_ref_put(&ctx->refs);
3635 return submitted ? submitted : ret;
3638 static const struct file_operations io_uring_fops = {
3639 .release = io_uring_release,
3640 .mmap = io_uring_mmap,
3641 .poll = io_uring_poll,
3642 .fasync = io_uring_fasync,
3645 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3646 struct io_uring_params *p)
3648 struct io_rings *rings;
3649 size_t size, sq_array_offset;
3651 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
3652 if (size == SIZE_MAX)
3655 rings = io_mem_alloc(size);
3660 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3661 rings->sq_ring_mask = p->sq_entries - 1;
3662 rings->cq_ring_mask = p->cq_entries - 1;
3663 rings->sq_ring_entries = p->sq_entries;
3664 rings->cq_ring_entries = p->cq_entries;
3665 ctx->sq_mask = rings->sq_ring_mask;
3666 ctx->cq_mask = rings->cq_ring_mask;
3667 ctx->sq_entries = rings->sq_ring_entries;
3668 ctx->cq_entries = rings->cq_ring_entries;
3670 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3671 if (size == SIZE_MAX)
3674 ctx->sq_sqes = io_mem_alloc(size);
3682 * Allocate an anonymous fd, this is what constitutes the application
3683 * visible backing of an io_uring instance. The application mmaps this
3684 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3685 * we have to tie this fd to a socket for file garbage collection purposes.
3687 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3692 #if defined(CONFIG_UNIX)
3693 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3699 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3703 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3704 O_RDWR | O_CLOEXEC);
3707 ret = PTR_ERR(file);
3711 #if defined(CONFIG_UNIX)
3712 ctx->ring_sock->file = file;
3713 ctx->ring_sock->sk->sk_user_data = ctx;
3715 fd_install(ret, file);
3718 #if defined(CONFIG_UNIX)
3719 sock_release(ctx->ring_sock);
3720 ctx->ring_sock = NULL;
3725 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3727 struct user_struct *user = NULL;
3728 struct io_ring_ctx *ctx;
3732 if (!entries || entries > IORING_MAX_ENTRIES)
3736 * Use twice as many entries for the CQ ring. It's possible for the
3737 * application to drive a higher depth than the size of the SQ ring,
3738 * since the sqes are only used at submission time. This allows for
3739 * some flexibility in overcommitting a bit.
3741 p->sq_entries = roundup_pow_of_two(entries);
3742 p->cq_entries = 2 * p->sq_entries;
3744 user = get_uid(current_user());
3745 account_mem = !capable(CAP_IPC_LOCK);
3748 ret = io_account_mem(user,
3749 ring_pages(p->sq_entries, p->cq_entries));
3756 ctx = io_ring_ctx_alloc(p);
3759 io_unaccount_mem(user, ring_pages(p->sq_entries,
3764 ctx->compat = in_compat_syscall();
3765 ctx->account_mem = account_mem;
3768 ret = io_allocate_scq_urings(ctx, p);
3772 ret = io_sq_offload_start(ctx, p);
3776 ret = io_uring_get_fd(ctx);
3780 memset(&p->sq_off, 0, sizeof(p->sq_off));
3781 p->sq_off.head = offsetof(struct io_rings, sq.head);
3782 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3783 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3784 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3785 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3786 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3787 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3789 memset(&p->cq_off, 0, sizeof(p->cq_off));
3790 p->cq_off.head = offsetof(struct io_rings, cq.head);
3791 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3792 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3793 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3794 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3795 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3797 p->features = IORING_FEAT_SINGLE_MMAP;
3800 io_ring_ctx_wait_and_kill(ctx);
3805 * Sets up an aio uring context, and returns the fd. Applications asks for a
3806 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3807 * params structure passed in.
3809 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3811 struct io_uring_params p;
3815 if (copy_from_user(&p, params, sizeof(p)))
3817 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3822 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3823 IORING_SETUP_SQ_AFF))
3826 ret = io_uring_create(entries, &p);
3830 if (copy_to_user(params, &p, sizeof(p)))
3836 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3837 struct io_uring_params __user *, params)
3839 return io_uring_setup(entries, params);
3842 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3843 void __user *arg, unsigned nr_args)
3844 __releases(ctx->uring_lock)
3845 __acquires(ctx->uring_lock)
3850 * We're inside the ring mutex, if the ref is already dying, then
3851 * someone else killed the ctx or is already going through
3852 * io_uring_register().
3854 if (percpu_ref_is_dying(&ctx->refs))
3857 percpu_ref_kill(&ctx->refs);
3860 * Drop uring mutex before waiting for references to exit. If another
3861 * thread is currently inside io_uring_enter() it might need to grab
3862 * the uring_lock to make progress. If we hold it here across the drain
3863 * wait, then we can deadlock. It's safe to drop the mutex here, since
3864 * no new references will come in after we've killed the percpu ref.
3866 mutex_unlock(&ctx->uring_lock);
3867 wait_for_completion(&ctx->ctx_done);
3868 mutex_lock(&ctx->uring_lock);
3871 case IORING_REGISTER_BUFFERS:
3872 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3874 case IORING_UNREGISTER_BUFFERS:
3878 ret = io_sqe_buffer_unregister(ctx);
3880 case IORING_REGISTER_FILES:
3881 ret = io_sqe_files_register(ctx, arg, nr_args);
3883 case IORING_UNREGISTER_FILES:
3887 ret = io_sqe_files_unregister(ctx);
3889 case IORING_REGISTER_EVENTFD:
3893 ret = io_eventfd_register(ctx, arg);
3895 case IORING_UNREGISTER_EVENTFD:
3899 ret = io_eventfd_unregister(ctx);
3906 /* bring the ctx back to life */
3907 reinit_completion(&ctx->ctx_done);
3908 percpu_ref_reinit(&ctx->refs);
3912 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3913 void __user *, arg, unsigned int, nr_args)
3915 struct io_ring_ctx *ctx;
3924 if (f.file->f_op != &io_uring_fops)
3927 ctx = f.file->private_data;
3929 mutex_lock(&ctx->uring_lock);
3930 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3931 mutex_unlock(&ctx->uring_lock);
3937 static int __init io_uring_init(void)
3939 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3942 __initcall(io_uring_init);