1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 4096
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp;
83 u32 tail ____cacheline_aligned_in_smp;
87 * This data is shared with the application through the mmap at offset
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head and the application controls tail.
102 * Bitmask to apply to head and tail offsets (constant, equals
106 /* Ring size (constant, power of 2) */
109 * Number of invalid entries dropped by the kernel due to
110 * invalid index stored in array
112 * Written by the kernel, shouldn't be modified by the
113 * application (i.e. get number of "new events" by comparing to
116 * After a new SQ head value was read by the application this
117 * counter includes all submissions that were dropped reaching
118 * the new SQ head (and possibly more).
124 * Written by the kernel, shouldn't be modified by the
127 * The application needs a full memory barrier before checking
128 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
132 * Ring buffer of indices into array of io_uring_sqe, which is
133 * mmapped by the application using the IORING_OFF_SQES offset.
135 * This indirection could e.g. be used to assign fixed
136 * io_uring_sqe entries to operations and only submit them to
137 * the queue when needed.
139 * The kernel modifies neither the indices array nor the entries
146 * This data is shared with the application through the mmap at offset
147 * IORING_OFF_CQ_RING.
149 * The offsets to the member fields are published through struct
150 * io_cqring_offsets when calling io_uring_setup.
154 * Head and tail offsets into the ring; the offsets need to be
155 * masked to get valid indices.
157 * The application controls head and the kernel tail.
161 * Bitmask to apply to head and tail offsets (constant, equals
165 /* Ring size (constant, power of 2) */
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending thatn there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[];
191 struct io_mapped_ubuf {
194 struct bio_vec *bvec;
195 unsigned int nr_bvecs;
201 struct list_head list;
210 struct percpu_ref refs;
211 } ____cacheline_aligned_in_smp;
219 struct io_sq_ring *sq_ring;
220 unsigned cached_sq_head;
223 unsigned sq_thread_idle;
224 struct io_uring_sqe *sq_sqes;
226 struct list_head defer_list;
227 } ____cacheline_aligned_in_smp;
230 struct workqueue_struct *sqo_wq;
231 struct task_struct *sqo_thread; /* if using sq thread polling */
232 struct mm_struct *sqo_mm;
233 wait_queue_head_t sqo_wait;
237 struct io_cq_ring *cq_ring;
238 unsigned cached_cq_tail;
241 struct wait_queue_head cq_wait;
242 struct fasync_struct *cq_fasync;
243 struct eventfd_ctx *cq_ev_fd;
244 } ____cacheline_aligned_in_smp;
247 * If used, fixed file set. Writers must ensure that ->refs is dead,
248 * readers must ensure that ->refs is alive as long as the file* is
249 * used. Only updated through io_uring_register(2).
251 struct file **user_files;
252 unsigned nr_user_files;
254 /* if used, fixed mapped user buffers */
255 unsigned nr_user_bufs;
256 struct io_mapped_ubuf *user_bufs;
258 struct user_struct *user;
260 struct completion ctx_done;
263 struct mutex uring_lock;
264 wait_queue_head_t wait;
265 } ____cacheline_aligned_in_smp;
268 spinlock_t completion_lock;
269 bool poll_multi_file;
271 * ->poll_list is protected by the ctx->uring_lock for
272 * io_uring instances that don't use IORING_SETUP_SQPOLL.
273 * For SQPOLL, only the single threaded io_sq_thread() will
274 * manipulate the list, hence no extra locking is needed there.
276 struct list_head poll_list;
277 struct list_head cancel_list;
278 } ____cacheline_aligned_in_smp;
280 struct async_list pending_async[2];
282 #if defined(CONFIG_UNIX)
283 struct socket *ring_sock;
288 const struct io_uring_sqe *sqe;
289 unsigned short index;
292 bool needs_fixed_file;
296 * First field must be the file pointer in all the
297 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
299 struct io_poll_iocb {
301 struct wait_queue_head *head;
305 struct wait_queue_entry wait;
309 * NOTE! Each of the iocb union members has the file pointer
310 * as the first entry in their struct definition. So you can
311 * access the file pointer through any of the sub-structs,
312 * or directly as just 'ki_filp' in this struct.
318 struct io_poll_iocb poll;
321 struct sqe_submit submit;
323 struct io_ring_ctx *ctx;
324 struct list_head list;
327 #define REQ_F_NOWAIT 1 /* must not punt to workers */
328 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
329 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
330 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
331 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
332 #define REQ_F_IO_DRAINED 32 /* drain done */
334 u32 error; /* iopoll result from callback */
337 struct work_struct work;
340 #define IO_PLUG_THRESHOLD 2
341 #define IO_IOPOLL_BATCH 8
343 struct io_submit_state {
344 struct blk_plug plug;
347 * io_kiocb alloc cache
349 void *reqs[IO_IOPOLL_BATCH];
350 unsigned int free_reqs;
351 unsigned int cur_req;
354 * File reference cache
358 unsigned int has_refs;
359 unsigned int used_refs;
360 unsigned int ios_left;
363 static void io_sq_wq_submit_work(struct work_struct *work);
365 static struct kmem_cache *req_cachep;
367 static const struct file_operations io_uring_fops;
369 struct sock *io_uring_get_socket(struct file *file)
371 #if defined(CONFIG_UNIX)
372 if (file->f_op == &io_uring_fops) {
373 struct io_ring_ctx *ctx = file->private_data;
375 return ctx->ring_sock->sk;
380 EXPORT_SYMBOL(io_uring_get_socket);
382 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
384 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
386 complete(&ctx->ctx_done);
389 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
391 struct io_ring_ctx *ctx;
394 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
398 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) {
403 ctx->flags = p->flags;
404 init_waitqueue_head(&ctx->cq_wait);
405 init_completion(&ctx->ctx_done);
406 mutex_init(&ctx->uring_lock);
407 init_waitqueue_head(&ctx->wait);
408 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
409 spin_lock_init(&ctx->pending_async[i].lock);
410 INIT_LIST_HEAD(&ctx->pending_async[i].list);
411 atomic_set(&ctx->pending_async[i].cnt, 0);
413 spin_lock_init(&ctx->completion_lock);
414 INIT_LIST_HEAD(&ctx->poll_list);
415 INIT_LIST_HEAD(&ctx->cancel_list);
416 INIT_LIST_HEAD(&ctx->defer_list);
420 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
421 struct io_kiocb *req)
423 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
426 return req->sequence > ctx->cached_cq_tail + ctx->sq_ring->dropped;
429 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
431 struct io_kiocb *req;
433 if (list_empty(&ctx->defer_list))
436 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
437 if (!io_sequence_defer(ctx, req)) {
438 list_del_init(&req->list);
445 static void __io_commit_cqring(struct io_ring_ctx *ctx)
447 struct io_cq_ring *ring = ctx->cq_ring;
449 if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) {
450 /* order cqe stores with ring update */
451 smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
453 if (wq_has_sleeper(&ctx->cq_wait)) {
454 wake_up_interruptible(&ctx->cq_wait);
455 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
460 static void io_commit_cqring(struct io_ring_ctx *ctx)
462 struct io_kiocb *req;
464 __io_commit_cqring(ctx);
466 while ((req = io_get_deferred_req(ctx)) != NULL) {
467 req->flags |= REQ_F_IO_DRAINED;
468 queue_work(ctx->sqo_wq, &req->work);
472 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
474 struct io_cq_ring *ring = ctx->cq_ring;
477 tail = ctx->cached_cq_tail;
479 * writes to the cq entry need to come after reading head; the
480 * control dependency is enough as we're using WRITE_ONCE to
483 if (tail - READ_ONCE(ring->r.head) == ring->ring_entries)
486 ctx->cached_cq_tail++;
487 return &ring->cqes[tail & ctx->cq_mask];
490 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
493 struct io_uring_cqe *cqe;
496 * If we can't get a cq entry, userspace overflowed the
497 * submission (by quite a lot). Increment the overflow count in
500 cqe = io_get_cqring(ctx);
502 WRITE_ONCE(cqe->user_data, ki_user_data);
503 WRITE_ONCE(cqe->res, res);
504 WRITE_ONCE(cqe->flags, 0);
506 unsigned overflow = READ_ONCE(ctx->cq_ring->overflow);
508 WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1);
512 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
514 if (waitqueue_active(&ctx->wait))
516 if (waitqueue_active(&ctx->sqo_wait))
517 wake_up(&ctx->sqo_wait);
519 eventfd_signal(ctx->cq_ev_fd, 1);
522 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
527 spin_lock_irqsave(&ctx->completion_lock, flags);
528 io_cqring_fill_event(ctx, user_data, res);
529 io_commit_cqring(ctx);
530 spin_unlock_irqrestore(&ctx->completion_lock, flags);
532 io_cqring_ev_posted(ctx);
535 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
537 percpu_ref_put_many(&ctx->refs, refs);
539 if (waitqueue_active(&ctx->wait))
543 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
544 struct io_submit_state *state)
546 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
547 struct io_kiocb *req;
549 if (!percpu_ref_tryget(&ctx->refs))
553 req = kmem_cache_alloc(req_cachep, gfp);
556 } else if (!state->free_reqs) {
560 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
561 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
564 * Bulk alloc is all-or-nothing. If we fail to get a batch,
565 * retry single alloc to be on the safe side.
567 if (unlikely(ret <= 0)) {
568 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
573 state->free_reqs = ret - 1;
575 req = state->reqs[0];
577 req = state->reqs[state->cur_req];
584 /* one is dropped after submission, the other at completion */
585 refcount_set(&req->refs, 2);
588 io_ring_drop_ctx_refs(ctx, 1);
592 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
595 kmem_cache_free_bulk(req_cachep, *nr, reqs);
596 io_ring_drop_ctx_refs(ctx, *nr);
601 static void io_free_req(struct io_kiocb *req)
603 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
605 io_ring_drop_ctx_refs(req->ctx, 1);
606 kmem_cache_free(req_cachep, req);
609 static void io_put_req(struct io_kiocb *req)
611 if (refcount_dec_and_test(&req->refs))
616 * Find and free completed poll iocbs
618 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
619 struct list_head *done)
621 void *reqs[IO_IOPOLL_BATCH];
622 struct io_kiocb *req;
626 while (!list_empty(done)) {
627 req = list_first_entry(done, struct io_kiocb, list);
628 list_del(&req->list);
630 io_cqring_fill_event(ctx, req->user_data, req->error);
633 if (refcount_dec_and_test(&req->refs)) {
634 /* If we're not using fixed files, we have to pair the
635 * completion part with the file put. Use regular
636 * completions for those, only batch free for fixed
639 if (req->flags & REQ_F_FIXED_FILE) {
640 reqs[to_free++] = req;
641 if (to_free == ARRAY_SIZE(reqs))
642 io_free_req_many(ctx, reqs, &to_free);
649 io_commit_cqring(ctx);
650 io_free_req_many(ctx, reqs, &to_free);
653 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
656 struct io_kiocb *req, *tmp;
662 * Only spin for completions if we don't have multiple devices hanging
663 * off our complete list, and we're under the requested amount.
665 spin = !ctx->poll_multi_file && *nr_events < min;
668 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
669 struct kiocb *kiocb = &req->rw;
672 * Move completed entries to our local list. If we find a
673 * request that requires polling, break out and complete
674 * the done list first, if we have entries there.
676 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
677 list_move_tail(&req->list, &done);
680 if (!list_empty(&done))
683 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
692 if (!list_empty(&done))
693 io_iopoll_complete(ctx, nr_events, &done);
699 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
700 * non-spinning poll check - we'll still enter the driver poll loop, but only
701 * as a non-spinning completion check.
703 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
706 while (!list_empty(&ctx->poll_list)) {
709 ret = io_do_iopoll(ctx, nr_events, min);
712 if (!min || *nr_events >= min)
720 * We can't just wait for polled events to come to us, we have to actively
721 * find and complete them.
723 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
725 if (!(ctx->flags & IORING_SETUP_IOPOLL))
728 mutex_lock(&ctx->uring_lock);
729 while (!list_empty(&ctx->poll_list)) {
730 unsigned int nr_events = 0;
732 io_iopoll_getevents(ctx, &nr_events, 1);
734 mutex_unlock(&ctx->uring_lock);
737 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
745 if (*nr_events < min)
746 tmin = min - *nr_events;
748 ret = io_iopoll_getevents(ctx, nr_events, tmin);
752 } while (min && !*nr_events && !need_resched());
757 static void kiocb_end_write(struct kiocb *kiocb)
759 if (kiocb->ki_flags & IOCB_WRITE) {
760 struct inode *inode = file_inode(kiocb->ki_filp);
763 * Tell lockdep we inherited freeze protection from submission
766 if (S_ISREG(inode->i_mode))
767 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
768 file_end_write(kiocb->ki_filp);
772 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
774 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
776 kiocb_end_write(kiocb);
778 io_cqring_add_event(req->ctx, req->user_data, res);
782 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
784 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
786 kiocb_end_write(kiocb);
790 req->flags |= REQ_F_IOPOLL_COMPLETED;
794 * After the iocb has been issued, it's safe to be found on the poll list.
795 * Adding the kiocb to the list AFTER submission ensures that we don't
796 * find it from a io_iopoll_getevents() thread before the issuer is done
797 * accessing the kiocb cookie.
799 static void io_iopoll_req_issued(struct io_kiocb *req)
801 struct io_ring_ctx *ctx = req->ctx;
804 * Track whether we have multiple files in our lists. This will impact
805 * how we do polling eventually, not spinning if we're on potentially
808 if (list_empty(&ctx->poll_list)) {
809 ctx->poll_multi_file = false;
810 } else if (!ctx->poll_multi_file) {
811 struct io_kiocb *list_req;
813 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
815 if (list_req->rw.ki_filp != req->rw.ki_filp)
816 ctx->poll_multi_file = true;
820 * For fast devices, IO may have already completed. If it has, add
821 * it to the front so we find it first.
823 if (req->flags & REQ_F_IOPOLL_COMPLETED)
824 list_add(&req->list, &ctx->poll_list);
826 list_add_tail(&req->list, &ctx->poll_list);
829 static void io_file_put(struct io_submit_state *state)
832 int diff = state->has_refs - state->used_refs;
835 fput_many(state->file, diff);
841 * Get as many references to a file as we have IOs left in this submission,
842 * assuming most submissions are for one file, or at least that each file
843 * has more than one submission.
845 static struct file *io_file_get(struct io_submit_state *state, int fd)
851 if (state->fd == fd) {
858 state->file = fget_many(fd, state->ios_left);
863 state->has_refs = state->ios_left;
864 state->used_refs = 1;
870 * If we tracked the file through the SCM inflight mechanism, we could support
871 * any file. For now, just ensure that anything potentially problematic is done
874 static bool io_file_supports_async(struct file *file)
876 umode_t mode = file_inode(file)->i_mode;
878 if (S_ISBLK(mode) || S_ISCHR(mode))
880 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
886 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
889 const struct io_uring_sqe *sqe = s->sqe;
890 struct io_ring_ctx *ctx = req->ctx;
891 struct kiocb *kiocb = &req->rw;
898 if (force_nonblock && !io_file_supports_async(req->file))
899 force_nonblock = false;
901 kiocb->ki_pos = READ_ONCE(sqe->off);
902 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
903 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
905 ioprio = READ_ONCE(sqe->ioprio);
907 ret = ioprio_check_cap(ioprio);
911 kiocb->ki_ioprio = ioprio;
913 kiocb->ki_ioprio = get_current_ioprio();
915 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
919 /* don't allow async punt if RWF_NOWAIT was requested */
920 if (kiocb->ki_flags & IOCB_NOWAIT)
921 req->flags |= REQ_F_NOWAIT;
924 kiocb->ki_flags |= IOCB_NOWAIT;
926 if (ctx->flags & IORING_SETUP_IOPOLL) {
927 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
928 !kiocb->ki_filp->f_op->iopoll)
932 kiocb->ki_flags |= IOCB_HIPRI;
933 kiocb->ki_complete = io_complete_rw_iopoll;
935 if (kiocb->ki_flags & IOCB_HIPRI)
937 kiocb->ki_complete = io_complete_rw;
942 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
948 case -ERESTARTNOINTR:
949 case -ERESTARTNOHAND:
950 case -ERESTART_RESTARTBLOCK:
952 * We can't just restart the syscall, since previously
953 * submitted sqes may already be in progress. Just fail this
959 kiocb->ki_complete(kiocb, ret, 0);
963 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
964 const struct io_uring_sqe *sqe,
965 struct iov_iter *iter)
967 size_t len = READ_ONCE(sqe->len);
968 struct io_mapped_ubuf *imu;
969 unsigned index, buf_index;
973 /* attempt to use fixed buffers without having provided iovecs */
974 if (unlikely(!ctx->user_bufs))
977 buf_index = READ_ONCE(sqe->buf_index);
978 if (unlikely(buf_index >= ctx->nr_user_bufs))
981 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
982 imu = &ctx->user_bufs[index];
983 buf_addr = READ_ONCE(sqe->addr);
986 if (buf_addr + len < buf_addr)
988 /* not inside the mapped region */
989 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
993 * May not be a start of buffer, set size appropriately
994 * and advance us to the beginning.
996 offset = buf_addr - imu->ubuf;
997 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
999 iov_iter_advance(iter, offset);
1001 /* don't drop a reference to these pages */
1002 iter->type |= ITER_BVEC_FLAG_NO_REF;
1006 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1007 const struct sqe_submit *s, struct iovec **iovec,
1008 struct iov_iter *iter)
1010 const struct io_uring_sqe *sqe = s->sqe;
1011 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1012 size_t sqe_len = READ_ONCE(sqe->len);
1016 * We're reading ->opcode for the second time, but the first read
1017 * doesn't care whether it's _FIXED or not, so it doesn't matter
1018 * whether ->opcode changes concurrently. The first read does care
1019 * about whether it is a READ or a WRITE, so we don't trust this read
1020 * for that purpose and instead let the caller pass in the read/write
1023 opcode = READ_ONCE(sqe->opcode);
1024 if (opcode == IORING_OP_READ_FIXED ||
1025 opcode == IORING_OP_WRITE_FIXED) {
1026 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1034 #ifdef CONFIG_COMPAT
1036 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1040 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1044 * Make a note of the last file/offset/direction we punted to async
1045 * context. We'll use this information to see if we can piggy back a
1046 * sequential request onto the previous one, if it's still hasn't been
1047 * completed by the async worker.
1049 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1051 struct async_list *async_list = &req->ctx->pending_async[rw];
1052 struct kiocb *kiocb = &req->rw;
1053 struct file *filp = kiocb->ki_filp;
1054 off_t io_end = kiocb->ki_pos + len;
1056 if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) {
1057 unsigned long max_pages;
1059 /* Use 8x RA size as a decent limiter for both reads/writes */
1060 max_pages = filp->f_ra.ra_pages;
1062 max_pages = VM_READAHEAD_PAGES;
1065 /* If max pages are exceeded, reset the state */
1067 if (async_list->io_pages + len <= max_pages) {
1068 req->flags |= REQ_F_SEQ_PREV;
1069 async_list->io_pages += len;
1072 async_list->io_pages = 0;
1076 /* New file? Reset state. */
1077 if (async_list->file != filp) {
1078 async_list->io_pages = 0;
1079 async_list->file = filp;
1081 async_list->io_end = io_end;
1084 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1085 bool force_nonblock)
1087 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1088 struct kiocb *kiocb = &req->rw;
1089 struct iov_iter iter;
1092 ssize_t read_size, ret;
1094 ret = io_prep_rw(req, s, force_nonblock);
1097 file = kiocb->ki_filp;
1099 if (unlikely(!(file->f_mode & FMODE_READ)))
1101 if (unlikely(!file->f_op->read_iter))
1104 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1109 iov_count = iov_iter_count(&iter);
1110 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1114 ret2 = call_read_iter(file, kiocb, &iter);
1116 * In case of a short read, punt to async. This can happen
1117 * if we have data partially cached. Alternatively we can
1118 * return the short read, in which case the application will
1119 * need to issue another SQE and wait for it. That SQE will
1120 * need async punt anyway, so it's more efficient to do it
1123 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1125 /* Catch -EAGAIN return for forced non-blocking submission */
1126 if (!force_nonblock || ret2 != -EAGAIN) {
1127 io_rw_done(kiocb, ret2);
1130 * If ->needs_lock is true, we're already in async
1134 io_async_list_note(READ, req, iov_count);
1142 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1143 bool force_nonblock)
1145 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1146 struct kiocb *kiocb = &req->rw;
1147 struct iov_iter iter;
1152 ret = io_prep_rw(req, s, force_nonblock);
1156 file = kiocb->ki_filp;
1157 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1159 if (unlikely(!file->f_op->write_iter))
1162 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1166 iov_count = iov_iter_count(&iter);
1169 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1170 /* If ->needs_lock is true, we're already in async context. */
1172 io_async_list_note(WRITE, req, iov_count);
1176 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1181 * Open-code file_start_write here to grab freeze protection,
1182 * which will be released by another thread in
1183 * io_complete_rw(). Fool lockdep by telling it the lock got
1184 * released so that it doesn't complain about the held lock when
1185 * we return to userspace.
1187 if (S_ISREG(file_inode(file)->i_mode)) {
1188 __sb_start_write(file_inode(file)->i_sb,
1189 SB_FREEZE_WRITE, true);
1190 __sb_writers_release(file_inode(file)->i_sb,
1193 kiocb->ki_flags |= IOCB_WRITE;
1195 ret2 = call_write_iter(file, kiocb, &iter);
1196 if (!force_nonblock || ret2 != -EAGAIN) {
1197 io_rw_done(kiocb, ret2);
1200 * If ->needs_lock is true, we're already in async
1204 io_async_list_note(WRITE, req, iov_count);
1214 * IORING_OP_NOP just posts a completion event, nothing else.
1216 static int io_nop(struct io_kiocb *req, u64 user_data)
1218 struct io_ring_ctx *ctx = req->ctx;
1221 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1224 io_cqring_add_event(ctx, user_data, err);
1229 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1231 struct io_ring_ctx *ctx = req->ctx;
1236 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1238 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1244 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1245 bool force_nonblock)
1247 loff_t sqe_off = READ_ONCE(sqe->off);
1248 loff_t sqe_len = READ_ONCE(sqe->len);
1249 loff_t end = sqe_off + sqe_len;
1250 unsigned fsync_flags;
1253 fsync_flags = READ_ONCE(sqe->fsync_flags);
1254 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1257 ret = io_prep_fsync(req, sqe);
1261 /* fsync always requires a blocking context */
1265 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1266 end > 0 ? end : LLONG_MAX,
1267 fsync_flags & IORING_FSYNC_DATASYNC);
1269 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1274 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1276 struct io_ring_ctx *ctx = req->ctx;
1282 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1284 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1290 static int io_sync_file_range(struct io_kiocb *req,
1291 const struct io_uring_sqe *sqe,
1292 bool force_nonblock)
1299 ret = io_prep_sfr(req, sqe);
1303 /* sync_file_range always requires a blocking context */
1307 sqe_off = READ_ONCE(sqe->off);
1308 sqe_len = READ_ONCE(sqe->len);
1309 flags = READ_ONCE(sqe->sync_range_flags);
1311 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1313 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1318 static void io_poll_remove_one(struct io_kiocb *req)
1320 struct io_poll_iocb *poll = &req->poll;
1322 spin_lock(&poll->head->lock);
1323 WRITE_ONCE(poll->canceled, true);
1324 if (!list_empty(&poll->wait.entry)) {
1325 list_del_init(&poll->wait.entry);
1326 queue_work(req->ctx->sqo_wq, &req->work);
1328 spin_unlock(&poll->head->lock);
1330 list_del_init(&req->list);
1333 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1335 struct io_kiocb *req;
1337 spin_lock_irq(&ctx->completion_lock);
1338 while (!list_empty(&ctx->cancel_list)) {
1339 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1340 io_poll_remove_one(req);
1342 spin_unlock_irq(&ctx->completion_lock);
1346 * Find a running poll command that matches one specified in sqe->addr,
1347 * and remove it if found.
1349 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1351 struct io_ring_ctx *ctx = req->ctx;
1352 struct io_kiocb *poll_req, *next;
1355 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1357 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1361 spin_lock_irq(&ctx->completion_lock);
1362 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1363 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1364 io_poll_remove_one(poll_req);
1369 spin_unlock_irq(&ctx->completion_lock);
1371 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1376 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1379 req->poll.done = true;
1380 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1381 io_commit_cqring(ctx);
1384 static void io_poll_complete_work(struct work_struct *work)
1386 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1387 struct io_poll_iocb *poll = &req->poll;
1388 struct poll_table_struct pt = { ._key = poll->events };
1389 struct io_ring_ctx *ctx = req->ctx;
1392 if (!READ_ONCE(poll->canceled))
1393 mask = vfs_poll(poll->file, &pt) & poll->events;
1396 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1397 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1398 * synchronize with them. In the cancellation case the list_del_init
1399 * itself is not actually needed, but harmless so we keep it in to
1400 * avoid further branches in the fast path.
1402 spin_lock_irq(&ctx->completion_lock);
1403 if (!mask && !READ_ONCE(poll->canceled)) {
1404 add_wait_queue(poll->head, &poll->wait);
1405 spin_unlock_irq(&ctx->completion_lock);
1408 list_del_init(&req->list);
1409 io_poll_complete(ctx, req, mask);
1410 spin_unlock_irq(&ctx->completion_lock);
1412 io_cqring_ev_posted(ctx);
1416 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1419 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1421 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1422 struct io_ring_ctx *ctx = req->ctx;
1423 __poll_t mask = key_to_poll(key);
1424 unsigned long flags;
1426 /* for instances that support it check for an event match first: */
1427 if (mask && !(mask & poll->events))
1430 list_del_init(&poll->wait.entry);
1432 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1433 list_del(&req->list);
1434 io_poll_complete(ctx, req, mask);
1435 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1437 io_cqring_ev_posted(ctx);
1440 queue_work(ctx->sqo_wq, &req->work);
1446 struct io_poll_table {
1447 struct poll_table_struct pt;
1448 struct io_kiocb *req;
1452 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1453 struct poll_table_struct *p)
1455 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1457 if (unlikely(pt->req->poll.head)) {
1458 pt->error = -EINVAL;
1463 pt->req->poll.head = head;
1464 add_wait_queue(head, &pt->req->poll.wait);
1467 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1469 struct io_poll_iocb *poll = &req->poll;
1470 struct io_ring_ctx *ctx = req->ctx;
1471 struct io_poll_table ipt;
1472 bool cancel = false;
1476 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1478 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1483 INIT_WORK(&req->work, io_poll_complete_work);
1484 events = READ_ONCE(sqe->poll_events);
1485 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1489 poll->canceled = false;
1491 ipt.pt._qproc = io_poll_queue_proc;
1492 ipt.pt._key = poll->events;
1494 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1496 /* initialized the list so that we can do list_empty checks */
1497 INIT_LIST_HEAD(&poll->wait.entry);
1498 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1500 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1502 spin_lock_irq(&ctx->completion_lock);
1503 if (likely(poll->head)) {
1504 spin_lock(&poll->head->lock);
1505 if (unlikely(list_empty(&poll->wait.entry))) {
1511 if (mask || ipt.error)
1512 list_del_init(&poll->wait.entry);
1514 WRITE_ONCE(poll->canceled, true);
1515 else if (!poll->done) /* actually waiting for an event */
1516 list_add_tail(&req->list, &ctx->cancel_list);
1517 spin_unlock(&poll->head->lock);
1519 if (mask) { /* no async, we'd stolen it */
1521 io_poll_complete(ctx, req, mask);
1523 spin_unlock_irq(&ctx->completion_lock);
1526 io_cqring_ev_posted(ctx);
1532 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1533 const struct io_uring_sqe *sqe)
1535 struct io_uring_sqe *sqe_copy;
1537 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1540 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1544 spin_lock_irq(&ctx->completion_lock);
1545 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1546 spin_unlock_irq(&ctx->completion_lock);
1551 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1552 req->submit.sqe = sqe_copy;
1554 INIT_WORK(&req->work, io_sq_wq_submit_work);
1555 list_add_tail(&req->list, &ctx->defer_list);
1556 spin_unlock_irq(&ctx->completion_lock);
1557 return -EIOCBQUEUED;
1560 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1561 const struct sqe_submit *s, bool force_nonblock)
1565 if (unlikely(s->index >= ctx->sq_entries))
1567 req->user_data = READ_ONCE(s->sqe->user_data);
1569 opcode = READ_ONCE(s->sqe->opcode);
1572 ret = io_nop(req, req->user_data);
1574 case IORING_OP_READV:
1575 if (unlikely(s->sqe->buf_index))
1577 ret = io_read(req, s, force_nonblock);
1579 case IORING_OP_WRITEV:
1580 if (unlikely(s->sqe->buf_index))
1582 ret = io_write(req, s, force_nonblock);
1584 case IORING_OP_READ_FIXED:
1585 ret = io_read(req, s, force_nonblock);
1587 case IORING_OP_WRITE_FIXED:
1588 ret = io_write(req, s, force_nonblock);
1590 case IORING_OP_FSYNC:
1591 ret = io_fsync(req, s->sqe, force_nonblock);
1593 case IORING_OP_POLL_ADD:
1594 ret = io_poll_add(req, s->sqe);
1596 case IORING_OP_POLL_REMOVE:
1597 ret = io_poll_remove(req, s->sqe);
1599 case IORING_OP_SYNC_FILE_RANGE:
1600 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1610 if (ctx->flags & IORING_SETUP_IOPOLL) {
1611 if (req->error == -EAGAIN)
1614 /* workqueue context doesn't hold uring_lock, grab it now */
1616 mutex_lock(&ctx->uring_lock);
1617 io_iopoll_req_issued(req);
1619 mutex_unlock(&ctx->uring_lock);
1625 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1626 const struct io_uring_sqe *sqe)
1628 switch (sqe->opcode) {
1629 case IORING_OP_READV:
1630 case IORING_OP_READ_FIXED:
1631 return &ctx->pending_async[READ];
1632 case IORING_OP_WRITEV:
1633 case IORING_OP_WRITE_FIXED:
1634 return &ctx->pending_async[WRITE];
1640 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1642 u8 opcode = READ_ONCE(sqe->opcode);
1644 return !(opcode == IORING_OP_READ_FIXED ||
1645 opcode == IORING_OP_WRITE_FIXED);
1648 static void io_sq_wq_submit_work(struct work_struct *work)
1650 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1651 struct io_ring_ctx *ctx = req->ctx;
1652 struct mm_struct *cur_mm = NULL;
1653 struct async_list *async_list;
1654 LIST_HEAD(req_list);
1655 mm_segment_t old_fs;
1658 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1661 struct sqe_submit *s = &req->submit;
1662 const struct io_uring_sqe *sqe = s->sqe;
1664 /* Ensure we clear previously set non-block flag */
1665 req->rw.ki_flags &= ~IOCB_NOWAIT;
1668 if (io_sqe_needs_user(sqe) && !cur_mm) {
1669 if (!mmget_not_zero(ctx->sqo_mm)) {
1672 cur_mm = ctx->sqo_mm;
1680 s->has_user = cur_mm != NULL;
1681 s->needs_lock = true;
1683 ret = __io_submit_sqe(ctx, req, s, false);
1685 * We can get EAGAIN for polled IO even though
1686 * we're forcing a sync submission from here,
1687 * since we can't wait for request slots on the
1696 /* drop submission reference */
1700 io_cqring_add_event(ctx, sqe->user_data, ret);
1704 /* async context always use a copy of the sqe */
1709 if (!list_empty(&req_list)) {
1710 req = list_first_entry(&req_list, struct io_kiocb,
1712 list_del(&req->list);
1715 if (list_empty(&async_list->list))
1719 spin_lock(&async_list->lock);
1720 if (list_empty(&async_list->list)) {
1721 spin_unlock(&async_list->lock);
1724 list_splice_init(&async_list->list, &req_list);
1725 spin_unlock(&async_list->lock);
1727 req = list_first_entry(&req_list, struct io_kiocb, list);
1728 list_del(&req->list);
1732 * Rare case of racing with a submitter. If we find the count has
1733 * dropped to zero AND we have pending work items, then restart
1734 * the processing. This is a tiny race window.
1737 ret = atomic_dec_return(&async_list->cnt);
1738 while (!ret && !list_empty(&async_list->list)) {
1739 spin_lock(&async_list->lock);
1740 atomic_inc(&async_list->cnt);
1741 list_splice_init(&async_list->list, &req_list);
1742 spin_unlock(&async_list->lock);
1744 if (!list_empty(&req_list)) {
1745 req = list_first_entry(&req_list,
1746 struct io_kiocb, list);
1747 list_del(&req->list);
1750 ret = atomic_dec_return(&async_list->cnt);
1762 * See if we can piggy back onto previously submitted work, that is still
1763 * running. We currently only allow this if the new request is sequential
1764 * to the previous one we punted.
1766 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
1772 if (!(req->flags & REQ_F_SEQ_PREV))
1774 if (!atomic_read(&list->cnt))
1778 spin_lock(&list->lock);
1779 list_add_tail(&req->list, &list->list);
1780 if (!atomic_read(&list->cnt)) {
1781 list_del_init(&req->list);
1784 spin_unlock(&list->lock);
1788 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
1790 int op = READ_ONCE(sqe->opcode);
1794 case IORING_OP_POLL_REMOVE:
1801 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
1802 struct io_submit_state *state, struct io_kiocb *req)
1807 flags = READ_ONCE(s->sqe->flags);
1808 fd = READ_ONCE(s->sqe->fd);
1810 if (flags & IOSQE_IO_DRAIN) {
1811 req->flags |= REQ_F_IO_DRAIN;
1812 req->sequence = ctx->cached_sq_head - 1;
1815 if (!io_op_needs_file(s->sqe)) {
1820 if (flags & IOSQE_FIXED_FILE) {
1821 if (unlikely(!ctx->user_files ||
1822 (unsigned) fd >= ctx->nr_user_files))
1824 req->file = ctx->user_files[fd];
1825 req->flags |= REQ_F_FIXED_FILE;
1827 if (s->needs_fixed_file)
1829 req->file = io_file_get(state, fd);
1830 if (unlikely(!req->file))
1837 static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
1838 struct io_submit_state *state)
1840 struct io_kiocb *req;
1843 /* enforce forwards compatibility on users */
1844 if (unlikely(s->sqe->flags & ~(IOSQE_FIXED_FILE | IOSQE_IO_DRAIN)))
1847 req = io_get_req(ctx, state);
1851 ret = io_req_set_file(ctx, s, state, req);
1855 ret = io_req_defer(ctx, req, s->sqe);
1857 if (ret == -EIOCBQUEUED)
1862 ret = __io_submit_sqe(ctx, req, s, true);
1863 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
1864 struct io_uring_sqe *sqe_copy;
1866 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1868 struct async_list *list;
1870 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
1873 memcpy(&req->submit, s, sizeof(*s));
1874 list = io_async_list_from_sqe(ctx, s->sqe);
1875 if (!io_add_to_prev_work(list, req)) {
1877 atomic_inc(&list->cnt);
1878 INIT_WORK(&req->work, io_sq_wq_submit_work);
1879 queue_work(ctx->sqo_wq, &req->work);
1883 * Queued up for async execution, worker will release
1884 * submit reference when the iocb is actually
1892 /* drop submission reference */
1895 /* and drop final reference, if we failed */
1903 * Batched submission is done, ensure local IO is flushed out.
1905 static void io_submit_state_end(struct io_submit_state *state)
1907 blk_finish_plug(&state->plug);
1909 if (state->free_reqs)
1910 kmem_cache_free_bulk(req_cachep, state->free_reqs,
1911 &state->reqs[state->cur_req]);
1915 * Start submission side cache.
1917 static void io_submit_state_start(struct io_submit_state *state,
1918 struct io_ring_ctx *ctx, unsigned max_ios)
1920 blk_start_plug(&state->plug);
1921 state->free_reqs = 0;
1923 state->ios_left = max_ios;
1926 static void io_commit_sqring(struct io_ring_ctx *ctx)
1928 struct io_sq_ring *ring = ctx->sq_ring;
1930 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
1932 * Ensure any loads from the SQEs are done at this point,
1933 * since once we write the new head, the application could
1934 * write new data to them.
1936 smp_store_release(&ring->r.head, ctx->cached_sq_head);
1941 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1942 * that is mapped by userspace. This means that care needs to be taken to
1943 * ensure that reads are stable, as we cannot rely on userspace always
1944 * being a good citizen. If members of the sqe are validated and then later
1945 * used, it's important that those reads are done through READ_ONCE() to
1946 * prevent a re-load down the line.
1948 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
1950 struct io_sq_ring *ring = ctx->sq_ring;
1954 * The cached sq head (or cq tail) serves two purposes:
1956 * 1) allows us to batch the cost of updating the user visible
1958 * 2) allows the kernel side to track the head on its own, even
1959 * though the application is the one updating it.
1961 head = ctx->cached_sq_head;
1962 /* make sure SQ entry isn't read before tail */
1963 if (head == smp_load_acquire(&ring->r.tail))
1966 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
1967 if (head < ctx->sq_entries) {
1969 s->sqe = &ctx->sq_sqes[head];
1970 ctx->cached_sq_head++;
1974 /* drop invalid entries */
1975 ctx->cached_sq_head++;
1980 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
1981 unsigned int nr, bool has_user, bool mm_fault)
1983 struct io_submit_state state, *statep = NULL;
1984 int ret, i, submitted = 0;
1986 if (nr > IO_PLUG_THRESHOLD) {
1987 io_submit_state_start(&state, ctx, nr);
1991 for (i = 0; i < nr; i++) {
1992 if (unlikely(mm_fault)) {
1995 sqes[i].has_user = has_user;
1996 sqes[i].needs_lock = true;
1997 sqes[i].needs_fixed_file = true;
1998 ret = io_submit_sqe(ctx, &sqes[i], statep);
2005 io_cqring_add_event(ctx, sqes[i].sqe->user_data, ret);
2009 io_submit_state_end(&state);
2014 static int io_sq_thread(void *data)
2016 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2017 struct io_ring_ctx *ctx = data;
2018 struct mm_struct *cur_mm = NULL;
2019 mm_segment_t old_fs;
2022 unsigned long timeout;
2027 timeout = inflight = 0;
2028 while (!kthread_should_park()) {
2029 bool all_fixed, mm_fault = false;
2033 unsigned nr_events = 0;
2035 if (ctx->flags & IORING_SETUP_IOPOLL) {
2037 * We disallow the app entering submit/complete
2038 * with polling, but we still need to lock the
2039 * ring to prevent racing with polled issue
2040 * that got punted to a workqueue.
2042 mutex_lock(&ctx->uring_lock);
2043 io_iopoll_check(ctx, &nr_events, 0);
2044 mutex_unlock(&ctx->uring_lock);
2047 * Normal IO, just pretend everything completed.
2048 * We don't have to poll completions for that.
2050 nr_events = inflight;
2053 inflight -= nr_events;
2055 timeout = jiffies + ctx->sq_thread_idle;
2058 if (!io_get_sqring(ctx, &sqes[0])) {
2060 * We're polling. If we're within the defined idle
2061 * period, then let us spin without work before going
2064 if (inflight || !time_after(jiffies, timeout)) {
2070 * Drop cur_mm before scheduling, we can't hold it for
2071 * long periods (or over schedule()). Do this before
2072 * adding ourselves to the waitqueue, as the unuse/drop
2081 prepare_to_wait(&ctx->sqo_wait, &wait,
2082 TASK_INTERRUPTIBLE);
2084 /* Tell userspace we may need a wakeup call */
2085 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
2086 /* make sure to read SQ tail after writing flags */
2089 if (!io_get_sqring(ctx, &sqes[0])) {
2090 if (kthread_should_park()) {
2091 finish_wait(&ctx->sqo_wait, &wait);
2094 if (signal_pending(current))
2095 flush_signals(current);
2097 finish_wait(&ctx->sqo_wait, &wait);
2099 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2102 finish_wait(&ctx->sqo_wait, &wait);
2104 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2110 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2114 if (i == ARRAY_SIZE(sqes))
2116 } while (io_get_sqring(ctx, &sqes[i]));
2118 /* Unless all new commands are FIXED regions, grab mm */
2119 if (!all_fixed && !cur_mm) {
2120 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2122 use_mm(ctx->sqo_mm);
2123 cur_mm = ctx->sqo_mm;
2127 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2130 /* Commit SQ ring head once we've consumed all SQEs */
2131 io_commit_sqring(ctx);
2145 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2147 struct io_submit_state state, *statep = NULL;
2150 if (to_submit > IO_PLUG_THRESHOLD) {
2151 io_submit_state_start(&state, ctx, to_submit);
2155 for (i = 0; i < to_submit; i++) {
2156 struct sqe_submit s;
2159 if (!io_get_sqring(ctx, &s))
2163 s.needs_lock = false;
2164 s.needs_fixed_file = false;
2167 ret = io_submit_sqe(ctx, &s, statep);
2169 io_cqring_add_event(ctx, s.sqe->user_data, ret);
2171 io_commit_sqring(ctx);
2174 io_submit_state_end(statep);
2179 static unsigned io_cqring_events(struct io_cq_ring *ring)
2181 /* See comment at the top of this file */
2183 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
2187 * Wait until events become available, if we don't already have some. The
2188 * application must reap them itself, as they reside on the shared cq ring.
2190 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2191 const sigset_t __user *sig, size_t sigsz)
2193 struct io_cq_ring *ring = ctx->cq_ring;
2194 sigset_t ksigmask, sigsaved;
2197 if (io_cqring_events(ring) >= min_events)
2201 #ifdef CONFIG_COMPAT
2202 if (in_compat_syscall())
2203 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2204 &ksigmask, &sigsaved, sigsz);
2207 ret = set_user_sigmask(sig, &ksigmask,
2214 ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events);
2215 if (ret == -ERESTARTSYS)
2219 restore_user_sigmask(sig, &sigsaved);
2221 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
2224 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2226 #if defined(CONFIG_UNIX)
2227 if (ctx->ring_sock) {
2228 struct sock *sock = ctx->ring_sock->sk;
2229 struct sk_buff *skb;
2231 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2237 for (i = 0; i < ctx->nr_user_files; i++)
2238 fput(ctx->user_files[i]);
2242 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2244 if (!ctx->user_files)
2247 __io_sqe_files_unregister(ctx);
2248 kfree(ctx->user_files);
2249 ctx->user_files = NULL;
2250 ctx->nr_user_files = 0;
2254 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2256 if (ctx->sqo_thread) {
2258 * The park is a bit of a work-around, without it we get
2259 * warning spews on shutdown with SQPOLL set and affinity
2260 * set to a single CPU.
2262 kthread_park(ctx->sqo_thread);
2263 kthread_stop(ctx->sqo_thread);
2264 ctx->sqo_thread = NULL;
2268 static void io_finish_async(struct io_ring_ctx *ctx)
2270 io_sq_thread_stop(ctx);
2273 destroy_workqueue(ctx->sqo_wq);
2278 #if defined(CONFIG_UNIX)
2279 static void io_destruct_skb(struct sk_buff *skb)
2281 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2283 io_finish_async(ctx);
2284 unix_destruct_scm(skb);
2288 * Ensure the UNIX gc is aware of our file set, so we are certain that
2289 * the io_uring can be safely unregistered on process exit, even if we have
2290 * loops in the file referencing.
2292 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2294 struct sock *sk = ctx->ring_sock->sk;
2295 struct scm_fp_list *fpl;
2296 struct sk_buff *skb;
2299 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2300 unsigned long inflight = ctx->user->unix_inflight + nr;
2302 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2306 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2310 skb = alloc_skb(0, GFP_KERNEL);
2317 skb->destructor = io_destruct_skb;
2319 fpl->user = get_uid(ctx->user);
2320 for (i = 0; i < nr; i++) {
2321 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2322 unix_inflight(fpl->user, fpl->fp[i]);
2325 fpl->max = fpl->count = nr;
2326 UNIXCB(skb).fp = fpl;
2327 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2328 skb_queue_head(&sk->sk_receive_queue, skb);
2330 for (i = 0; i < nr; i++)
2337 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2338 * causes regular reference counting to break down. We rely on the UNIX
2339 * garbage collection to take care of this problem for us.
2341 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2343 unsigned left, total;
2347 left = ctx->nr_user_files;
2349 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2351 ret = __io_sqe_files_scm(ctx, this_files, total);
2355 total += this_files;
2361 while (total < ctx->nr_user_files) {
2362 fput(ctx->user_files[total]);
2369 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2375 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2378 __s32 __user *fds = (__s32 __user *) arg;
2382 if (ctx->user_files)
2386 if (nr_args > IORING_MAX_FIXED_FILES)
2389 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2390 if (!ctx->user_files)
2393 for (i = 0; i < nr_args; i++) {
2395 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2398 ctx->user_files[i] = fget(fd);
2401 if (!ctx->user_files[i])
2404 * Don't allow io_uring instances to be registered. If UNIX
2405 * isn't enabled, then this causes a reference cycle and this
2406 * instance can never get freed. If UNIX is enabled we'll
2407 * handle it just fine, but there's still no point in allowing
2408 * a ring fd as it doesn't support regular read/write anyway.
2410 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2411 fput(ctx->user_files[i]);
2414 ctx->nr_user_files++;
2419 for (i = 0; i < ctx->nr_user_files; i++)
2420 fput(ctx->user_files[i]);
2422 kfree(ctx->user_files);
2423 ctx->user_files = NULL;
2424 ctx->nr_user_files = 0;
2428 ret = io_sqe_files_scm(ctx);
2430 io_sqe_files_unregister(ctx);
2435 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2436 struct io_uring_params *p)
2440 init_waitqueue_head(&ctx->sqo_wait);
2441 mmgrab(current->mm);
2442 ctx->sqo_mm = current->mm;
2444 if (ctx->flags & IORING_SETUP_SQPOLL) {
2446 if (!capable(CAP_SYS_ADMIN))
2449 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2450 if (!ctx->sq_thread_idle)
2451 ctx->sq_thread_idle = HZ;
2453 if (p->flags & IORING_SETUP_SQ_AFF) {
2454 int cpu = p->sq_thread_cpu;
2457 if (cpu >= nr_cpu_ids)
2459 if (!cpu_online(cpu))
2462 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2466 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2469 if (IS_ERR(ctx->sqo_thread)) {
2470 ret = PTR_ERR(ctx->sqo_thread);
2471 ctx->sqo_thread = NULL;
2474 wake_up_process(ctx->sqo_thread);
2475 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2476 /* Can't have SQ_AFF without SQPOLL */
2481 /* Do QD, or 2 * CPUS, whatever is smallest */
2482 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
2483 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2491 io_sq_thread_stop(ctx);
2492 mmdrop(ctx->sqo_mm);
2497 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2499 atomic_long_sub(nr_pages, &user->locked_vm);
2502 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2504 unsigned long page_limit, cur_pages, new_pages;
2506 /* Don't allow more pages than we can safely lock */
2507 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2510 cur_pages = atomic_long_read(&user->locked_vm);
2511 new_pages = cur_pages + nr_pages;
2512 if (new_pages > page_limit)
2514 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2515 new_pages) != cur_pages);
2520 static void io_mem_free(void *ptr)
2527 page = virt_to_head_page(ptr);
2528 if (put_page_testzero(page))
2529 free_compound_page(page);
2532 static void *io_mem_alloc(size_t size)
2534 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2537 return (void *) __get_free_pages(gfp_flags, get_order(size));
2540 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2542 struct io_sq_ring *sq_ring;
2543 struct io_cq_ring *cq_ring;
2546 bytes = struct_size(sq_ring, array, sq_entries);
2547 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
2548 bytes += struct_size(cq_ring, cqes, cq_entries);
2550 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
2553 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
2557 if (!ctx->user_bufs)
2560 for (i = 0; i < ctx->nr_user_bufs; i++) {
2561 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2563 for (j = 0; j < imu->nr_bvecs; j++)
2564 put_page(imu->bvec[j].bv_page);
2566 if (ctx->account_mem)
2567 io_unaccount_mem(ctx->user, imu->nr_bvecs);
2572 kfree(ctx->user_bufs);
2573 ctx->user_bufs = NULL;
2574 ctx->nr_user_bufs = 0;
2578 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
2579 void __user *arg, unsigned index)
2581 struct iovec __user *src;
2583 #ifdef CONFIG_COMPAT
2585 struct compat_iovec __user *ciovs;
2586 struct compat_iovec ciov;
2588 ciovs = (struct compat_iovec __user *) arg;
2589 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
2592 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
2593 dst->iov_len = ciov.iov_len;
2597 src = (struct iovec __user *) arg;
2598 if (copy_from_user(dst, &src[index], sizeof(*dst)))
2603 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
2606 struct vm_area_struct **vmas = NULL;
2607 struct page **pages = NULL;
2608 int i, j, got_pages = 0;
2613 if (!nr_args || nr_args > UIO_MAXIOV)
2616 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
2618 if (!ctx->user_bufs)
2621 for (i = 0; i < nr_args; i++) {
2622 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2623 unsigned long off, start, end, ubuf;
2628 ret = io_copy_iov(ctx, &iov, arg, i);
2633 * Don't impose further limits on the size and buffer
2634 * constraints here, we'll -EINVAL later when IO is
2635 * submitted if they are wrong.
2638 if (!iov.iov_base || !iov.iov_len)
2641 /* arbitrary limit, but we need something */
2642 if (iov.iov_len > SZ_1G)
2645 ubuf = (unsigned long) iov.iov_base;
2646 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2647 start = ubuf >> PAGE_SHIFT;
2648 nr_pages = end - start;
2650 if (ctx->account_mem) {
2651 ret = io_account_mem(ctx->user, nr_pages);
2657 if (!pages || nr_pages > got_pages) {
2660 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
2662 vmas = kvmalloc_array(nr_pages,
2663 sizeof(struct vm_area_struct *),
2665 if (!pages || !vmas) {
2667 if (ctx->account_mem)
2668 io_unaccount_mem(ctx->user, nr_pages);
2671 got_pages = nr_pages;
2674 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
2678 if (ctx->account_mem)
2679 io_unaccount_mem(ctx->user, nr_pages);
2684 down_read(¤t->mm->mmap_sem);
2685 pret = get_user_pages(ubuf, nr_pages,
2686 FOLL_WRITE | FOLL_LONGTERM,
2688 if (pret == nr_pages) {
2689 /* don't support file backed memory */
2690 for (j = 0; j < nr_pages; j++) {
2691 struct vm_area_struct *vma = vmas[j];
2694 !is_file_hugepages(vma->vm_file)) {
2700 ret = pret < 0 ? pret : -EFAULT;
2702 up_read(¤t->mm->mmap_sem);
2705 * if we did partial map, or found file backed vmas,
2706 * release any pages we did get
2709 for (j = 0; j < pret; j++)
2712 if (ctx->account_mem)
2713 io_unaccount_mem(ctx->user, nr_pages);
2718 off = ubuf & ~PAGE_MASK;
2720 for (j = 0; j < nr_pages; j++) {
2723 vec_len = min_t(size_t, size, PAGE_SIZE - off);
2724 imu->bvec[j].bv_page = pages[j];
2725 imu->bvec[j].bv_len = vec_len;
2726 imu->bvec[j].bv_offset = off;
2730 /* store original address for later verification */
2732 imu->len = iov.iov_len;
2733 imu->nr_bvecs = nr_pages;
2735 ctx->nr_user_bufs++;
2743 io_sqe_buffer_unregister(ctx);
2747 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
2749 __s32 __user *fds = arg;
2755 if (copy_from_user(&fd, fds, sizeof(*fds)))
2758 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
2759 if (IS_ERR(ctx->cq_ev_fd)) {
2760 int ret = PTR_ERR(ctx->cq_ev_fd);
2761 ctx->cq_ev_fd = NULL;
2768 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2770 if (ctx->cq_ev_fd) {
2771 eventfd_ctx_put(ctx->cq_ev_fd);
2772 ctx->cq_ev_fd = NULL;
2779 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
2781 io_finish_async(ctx);
2783 mmdrop(ctx->sqo_mm);
2785 io_iopoll_reap_events(ctx);
2786 io_sqe_buffer_unregister(ctx);
2787 io_sqe_files_unregister(ctx);
2788 io_eventfd_unregister(ctx);
2790 #if defined(CONFIG_UNIX)
2792 sock_release(ctx->ring_sock);
2795 io_mem_free(ctx->sq_ring);
2796 io_mem_free(ctx->sq_sqes);
2797 io_mem_free(ctx->cq_ring);
2799 percpu_ref_exit(&ctx->refs);
2800 if (ctx->account_mem)
2801 io_unaccount_mem(ctx->user,
2802 ring_pages(ctx->sq_entries, ctx->cq_entries));
2803 free_uid(ctx->user);
2807 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2809 struct io_ring_ctx *ctx = file->private_data;
2812 poll_wait(file, &ctx->cq_wait, wait);
2814 * synchronizes with barrier from wq_has_sleeper call in
2818 if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
2819 ctx->sq_ring->ring_entries)
2820 mask |= EPOLLOUT | EPOLLWRNORM;
2821 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
2822 mask |= EPOLLIN | EPOLLRDNORM;
2827 static int io_uring_fasync(int fd, struct file *file, int on)
2829 struct io_ring_ctx *ctx = file->private_data;
2831 return fasync_helper(fd, file, on, &ctx->cq_fasync);
2834 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2836 mutex_lock(&ctx->uring_lock);
2837 percpu_ref_kill(&ctx->refs);
2838 mutex_unlock(&ctx->uring_lock);
2840 io_poll_remove_all(ctx);
2841 io_iopoll_reap_events(ctx);
2842 wait_for_completion(&ctx->ctx_done);
2843 io_ring_ctx_free(ctx);
2846 static int io_uring_release(struct inode *inode, struct file *file)
2848 struct io_ring_ctx *ctx = file->private_data;
2850 file->private_data = NULL;
2851 io_ring_ctx_wait_and_kill(ctx);
2855 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2857 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
2858 unsigned long sz = vma->vm_end - vma->vm_start;
2859 struct io_ring_ctx *ctx = file->private_data;
2865 case IORING_OFF_SQ_RING:
2868 case IORING_OFF_SQES:
2871 case IORING_OFF_CQ_RING:
2878 page = virt_to_head_page(ptr);
2879 if (sz > (PAGE_SIZE << compound_order(page)))
2882 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2883 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2886 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2887 u32, min_complete, u32, flags, const sigset_t __user *, sig,
2890 struct io_ring_ctx *ctx;
2895 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
2903 if (f.file->f_op != &io_uring_fops)
2907 ctx = f.file->private_data;
2908 if (!percpu_ref_tryget(&ctx->refs))
2912 * For SQ polling, the thread will do all submissions and completions.
2913 * Just return the requested submit count, and wake the thread if
2916 if (ctx->flags & IORING_SETUP_SQPOLL) {
2917 if (flags & IORING_ENTER_SQ_WAKEUP)
2918 wake_up(&ctx->sqo_wait);
2919 submitted = to_submit;
2925 to_submit = min(to_submit, ctx->sq_entries);
2927 mutex_lock(&ctx->uring_lock);
2928 submitted = io_ring_submit(ctx, to_submit);
2929 mutex_unlock(&ctx->uring_lock);
2931 if (flags & IORING_ENTER_GETEVENTS) {
2932 unsigned nr_events = 0;
2934 min_complete = min(min_complete, ctx->cq_entries);
2936 if (ctx->flags & IORING_SETUP_IOPOLL) {
2937 mutex_lock(&ctx->uring_lock);
2938 ret = io_iopoll_check(ctx, &nr_events, min_complete);
2939 mutex_unlock(&ctx->uring_lock);
2941 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
2946 io_ring_drop_ctx_refs(ctx, 1);
2949 return submitted ? submitted : ret;
2952 static const struct file_operations io_uring_fops = {
2953 .release = io_uring_release,
2954 .mmap = io_uring_mmap,
2955 .poll = io_uring_poll,
2956 .fasync = io_uring_fasync,
2959 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
2960 struct io_uring_params *p)
2962 struct io_sq_ring *sq_ring;
2963 struct io_cq_ring *cq_ring;
2966 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
2970 ctx->sq_ring = sq_ring;
2971 sq_ring->ring_mask = p->sq_entries - 1;
2972 sq_ring->ring_entries = p->sq_entries;
2973 ctx->sq_mask = sq_ring->ring_mask;
2974 ctx->sq_entries = sq_ring->ring_entries;
2976 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
2977 if (size == SIZE_MAX)
2980 ctx->sq_sqes = io_mem_alloc(size);
2984 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
2988 ctx->cq_ring = cq_ring;
2989 cq_ring->ring_mask = p->cq_entries - 1;
2990 cq_ring->ring_entries = p->cq_entries;
2991 ctx->cq_mask = cq_ring->ring_mask;
2992 ctx->cq_entries = cq_ring->ring_entries;
2997 * Allocate an anonymous fd, this is what constitutes the application
2998 * visible backing of an io_uring instance. The application mmaps this
2999 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3000 * we have to tie this fd to a socket for file garbage collection purposes.
3002 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3007 #if defined(CONFIG_UNIX)
3008 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3014 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3018 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3019 O_RDWR | O_CLOEXEC);
3022 ret = PTR_ERR(file);
3026 #if defined(CONFIG_UNIX)
3027 ctx->ring_sock->file = file;
3028 ctx->ring_sock->sk->sk_user_data = ctx;
3030 fd_install(ret, file);
3033 #if defined(CONFIG_UNIX)
3034 sock_release(ctx->ring_sock);
3035 ctx->ring_sock = NULL;
3040 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3042 struct user_struct *user = NULL;
3043 struct io_ring_ctx *ctx;
3047 if (!entries || entries > IORING_MAX_ENTRIES)
3051 * Use twice as many entries for the CQ ring. It's possible for the
3052 * application to drive a higher depth than the size of the SQ ring,
3053 * since the sqes are only used at submission time. This allows for
3054 * some flexibility in overcommitting a bit.
3056 p->sq_entries = roundup_pow_of_two(entries);
3057 p->cq_entries = 2 * p->sq_entries;
3059 user = get_uid(current_user());
3060 account_mem = !capable(CAP_IPC_LOCK);
3063 ret = io_account_mem(user,
3064 ring_pages(p->sq_entries, p->cq_entries));
3071 ctx = io_ring_ctx_alloc(p);
3074 io_unaccount_mem(user, ring_pages(p->sq_entries,
3079 ctx->compat = in_compat_syscall();
3080 ctx->account_mem = account_mem;
3083 ret = io_allocate_scq_urings(ctx, p);
3087 ret = io_sq_offload_start(ctx, p);
3091 ret = io_uring_get_fd(ctx);
3095 memset(&p->sq_off, 0, sizeof(p->sq_off));
3096 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
3097 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
3098 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
3099 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
3100 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
3101 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
3102 p->sq_off.array = offsetof(struct io_sq_ring, array);
3104 memset(&p->cq_off, 0, sizeof(p->cq_off));
3105 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
3106 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
3107 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
3108 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
3109 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
3110 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
3113 io_ring_ctx_wait_and_kill(ctx);
3118 * Sets up an aio uring context, and returns the fd. Applications asks for a
3119 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3120 * params structure passed in.
3122 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3124 struct io_uring_params p;
3128 if (copy_from_user(&p, params, sizeof(p)))
3130 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3135 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3136 IORING_SETUP_SQ_AFF))
3139 ret = io_uring_create(entries, &p);
3143 if (copy_to_user(params, &p, sizeof(p)))
3149 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3150 struct io_uring_params __user *, params)
3152 return io_uring_setup(entries, params);
3155 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3156 void __user *arg, unsigned nr_args)
3157 __releases(ctx->uring_lock)
3158 __acquires(ctx->uring_lock)
3163 * We're inside the ring mutex, if the ref is already dying, then
3164 * someone else killed the ctx or is already going through
3165 * io_uring_register().
3167 if (percpu_ref_is_dying(&ctx->refs))
3170 percpu_ref_kill(&ctx->refs);
3173 * Drop uring mutex before waiting for references to exit. If another
3174 * thread is currently inside io_uring_enter() it might need to grab
3175 * the uring_lock to make progress. If we hold it here across the drain
3176 * wait, then we can deadlock. It's safe to drop the mutex here, since
3177 * no new references will come in after we've killed the percpu ref.
3179 mutex_unlock(&ctx->uring_lock);
3180 wait_for_completion(&ctx->ctx_done);
3181 mutex_lock(&ctx->uring_lock);
3184 case IORING_REGISTER_BUFFERS:
3185 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3187 case IORING_UNREGISTER_BUFFERS:
3191 ret = io_sqe_buffer_unregister(ctx);
3193 case IORING_REGISTER_FILES:
3194 ret = io_sqe_files_register(ctx, arg, nr_args);
3196 case IORING_UNREGISTER_FILES:
3200 ret = io_sqe_files_unregister(ctx);
3202 case IORING_REGISTER_EVENTFD:
3206 ret = io_eventfd_register(ctx, arg);
3208 case IORING_UNREGISTER_EVENTFD:
3212 ret = io_eventfd_unregister(ctx);
3219 /* bring the ctx back to life */
3220 reinit_completion(&ctx->ctx_done);
3221 percpu_ref_reinit(&ctx->refs);
3225 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3226 void __user *, arg, unsigned int, nr_args)
3228 struct io_ring_ctx *ctx;
3237 if (f.file->f_op != &io_uring_fops)
3240 ctx = f.file->private_data;
3242 mutex_lock(&ctx->uring_lock);
3243 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3244 mutex_unlock(&ctx->uring_lock);
3250 static int __init io_uring_init(void)
3252 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3255 __initcall(io_uring_init);