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];
585 /* one is dropped after submission, the other at completion */
586 refcount_set(&req->refs, 2);
589 io_ring_drop_ctx_refs(ctx, 1);
593 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
596 kmem_cache_free_bulk(req_cachep, *nr, reqs);
597 io_ring_drop_ctx_refs(ctx, *nr);
602 static void io_free_req(struct io_kiocb *req)
604 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
606 io_ring_drop_ctx_refs(req->ctx, 1);
607 kmem_cache_free(req_cachep, req);
610 static void io_put_req(struct io_kiocb *req)
612 if (refcount_dec_and_test(&req->refs))
617 * Find and free completed poll iocbs
619 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
620 struct list_head *done)
622 void *reqs[IO_IOPOLL_BATCH];
623 struct io_kiocb *req;
627 while (!list_empty(done)) {
628 req = list_first_entry(done, struct io_kiocb, list);
629 list_del(&req->list);
631 io_cqring_fill_event(ctx, req->user_data, req->error);
634 if (refcount_dec_and_test(&req->refs)) {
635 /* If we're not using fixed files, we have to pair the
636 * completion part with the file put. Use regular
637 * completions for those, only batch free for fixed
640 if (req->flags & REQ_F_FIXED_FILE) {
641 reqs[to_free++] = req;
642 if (to_free == ARRAY_SIZE(reqs))
643 io_free_req_many(ctx, reqs, &to_free);
650 io_commit_cqring(ctx);
651 io_free_req_many(ctx, reqs, &to_free);
654 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
657 struct io_kiocb *req, *tmp;
663 * Only spin for completions if we don't have multiple devices hanging
664 * off our complete list, and we're under the requested amount.
666 spin = !ctx->poll_multi_file && *nr_events < min;
669 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
670 struct kiocb *kiocb = &req->rw;
673 * Move completed entries to our local list. If we find a
674 * request that requires polling, break out and complete
675 * the done list first, if we have entries there.
677 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
678 list_move_tail(&req->list, &done);
681 if (!list_empty(&done))
684 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
693 if (!list_empty(&done))
694 io_iopoll_complete(ctx, nr_events, &done);
700 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
701 * non-spinning poll check - we'll still enter the driver poll loop, but only
702 * as a non-spinning completion check.
704 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
707 while (!list_empty(&ctx->poll_list)) {
710 ret = io_do_iopoll(ctx, nr_events, min);
713 if (!min || *nr_events >= min)
721 * We can't just wait for polled events to come to us, we have to actively
722 * find and complete them.
724 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
726 if (!(ctx->flags & IORING_SETUP_IOPOLL))
729 mutex_lock(&ctx->uring_lock);
730 while (!list_empty(&ctx->poll_list)) {
731 unsigned int nr_events = 0;
733 io_iopoll_getevents(ctx, &nr_events, 1);
735 mutex_unlock(&ctx->uring_lock);
738 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
746 if (*nr_events < min)
747 tmin = min - *nr_events;
749 ret = io_iopoll_getevents(ctx, nr_events, tmin);
753 } while (min && !*nr_events && !need_resched());
758 static void kiocb_end_write(struct kiocb *kiocb)
760 if (kiocb->ki_flags & IOCB_WRITE) {
761 struct inode *inode = file_inode(kiocb->ki_filp);
764 * Tell lockdep we inherited freeze protection from submission
767 if (S_ISREG(inode->i_mode))
768 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
769 file_end_write(kiocb->ki_filp);
773 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
775 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
777 kiocb_end_write(kiocb);
779 io_cqring_add_event(req->ctx, req->user_data, res);
783 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
785 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
787 kiocb_end_write(kiocb);
791 req->flags |= REQ_F_IOPOLL_COMPLETED;
795 * After the iocb has been issued, it's safe to be found on the poll list.
796 * Adding the kiocb to the list AFTER submission ensures that we don't
797 * find it from a io_iopoll_getevents() thread before the issuer is done
798 * accessing the kiocb cookie.
800 static void io_iopoll_req_issued(struct io_kiocb *req)
802 struct io_ring_ctx *ctx = req->ctx;
805 * Track whether we have multiple files in our lists. This will impact
806 * how we do polling eventually, not spinning if we're on potentially
809 if (list_empty(&ctx->poll_list)) {
810 ctx->poll_multi_file = false;
811 } else if (!ctx->poll_multi_file) {
812 struct io_kiocb *list_req;
814 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
816 if (list_req->rw.ki_filp != req->rw.ki_filp)
817 ctx->poll_multi_file = true;
821 * For fast devices, IO may have already completed. If it has, add
822 * it to the front so we find it first.
824 if (req->flags & REQ_F_IOPOLL_COMPLETED)
825 list_add(&req->list, &ctx->poll_list);
827 list_add_tail(&req->list, &ctx->poll_list);
830 static void io_file_put(struct io_submit_state *state)
833 int diff = state->has_refs - state->used_refs;
836 fput_many(state->file, diff);
842 * Get as many references to a file as we have IOs left in this submission,
843 * assuming most submissions are for one file, or at least that each file
844 * has more than one submission.
846 static struct file *io_file_get(struct io_submit_state *state, int fd)
852 if (state->fd == fd) {
859 state->file = fget_many(fd, state->ios_left);
864 state->has_refs = state->ios_left;
865 state->used_refs = 1;
871 * If we tracked the file through the SCM inflight mechanism, we could support
872 * any file. For now, just ensure that anything potentially problematic is done
875 static bool io_file_supports_async(struct file *file)
877 umode_t mode = file_inode(file)->i_mode;
879 if (S_ISBLK(mode) || S_ISCHR(mode))
881 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
887 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
890 const struct io_uring_sqe *sqe = s->sqe;
891 struct io_ring_ctx *ctx = req->ctx;
892 struct kiocb *kiocb = &req->rw;
899 if (force_nonblock && !io_file_supports_async(req->file))
900 force_nonblock = false;
902 kiocb->ki_pos = READ_ONCE(sqe->off);
903 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
904 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
906 ioprio = READ_ONCE(sqe->ioprio);
908 ret = ioprio_check_cap(ioprio);
912 kiocb->ki_ioprio = ioprio;
914 kiocb->ki_ioprio = get_current_ioprio();
916 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
920 /* don't allow async punt if RWF_NOWAIT was requested */
921 if (kiocb->ki_flags & IOCB_NOWAIT)
922 req->flags |= REQ_F_NOWAIT;
925 kiocb->ki_flags |= IOCB_NOWAIT;
927 if (ctx->flags & IORING_SETUP_IOPOLL) {
928 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
929 !kiocb->ki_filp->f_op->iopoll)
933 kiocb->ki_flags |= IOCB_HIPRI;
934 kiocb->ki_complete = io_complete_rw_iopoll;
936 if (kiocb->ki_flags & IOCB_HIPRI)
938 kiocb->ki_complete = io_complete_rw;
943 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
949 case -ERESTARTNOINTR:
950 case -ERESTARTNOHAND:
951 case -ERESTART_RESTARTBLOCK:
953 * We can't just restart the syscall, since previously
954 * submitted sqes may already be in progress. Just fail this
960 kiocb->ki_complete(kiocb, ret, 0);
964 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
965 const struct io_uring_sqe *sqe,
966 struct iov_iter *iter)
968 size_t len = READ_ONCE(sqe->len);
969 struct io_mapped_ubuf *imu;
970 unsigned index, buf_index;
974 /* attempt to use fixed buffers without having provided iovecs */
975 if (unlikely(!ctx->user_bufs))
978 buf_index = READ_ONCE(sqe->buf_index);
979 if (unlikely(buf_index >= ctx->nr_user_bufs))
982 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
983 imu = &ctx->user_bufs[index];
984 buf_addr = READ_ONCE(sqe->addr);
987 if (buf_addr + len < buf_addr)
989 /* not inside the mapped region */
990 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
994 * May not be a start of buffer, set size appropriately
995 * and advance us to the beginning.
997 offset = buf_addr - imu->ubuf;
998 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1000 iov_iter_advance(iter, offset);
1004 static int io_import_iovec(struct io_ring_ctx *ctx, int rw,
1005 const struct sqe_submit *s, struct iovec **iovec,
1006 struct iov_iter *iter)
1008 const struct io_uring_sqe *sqe = s->sqe;
1009 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1010 size_t sqe_len = READ_ONCE(sqe->len);
1014 * We're reading ->opcode for the second time, but the first read
1015 * doesn't care whether it's _FIXED or not, so it doesn't matter
1016 * whether ->opcode changes concurrently. The first read does care
1017 * about whether it is a READ or a WRITE, so we don't trust this read
1018 * for that purpose and instead let the caller pass in the read/write
1021 opcode = READ_ONCE(sqe->opcode);
1022 if (opcode == IORING_OP_READ_FIXED ||
1023 opcode == IORING_OP_WRITE_FIXED) {
1024 int ret = io_import_fixed(ctx, rw, sqe, iter);
1032 #ifdef CONFIG_COMPAT
1034 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1038 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1042 * Make a note of the last file/offset/direction we punted to async
1043 * context. We'll use this information to see if we can piggy back a
1044 * sequential request onto the previous one, if it's still hasn't been
1045 * completed by the async worker.
1047 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1049 struct async_list *async_list = &req->ctx->pending_async[rw];
1050 struct kiocb *kiocb = &req->rw;
1051 struct file *filp = kiocb->ki_filp;
1052 off_t io_end = kiocb->ki_pos + len;
1054 if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) {
1055 unsigned long max_pages;
1057 /* Use 8x RA size as a decent limiter for both reads/writes */
1058 max_pages = filp->f_ra.ra_pages;
1060 max_pages = VM_READAHEAD_PAGES;
1063 /* If max pages are exceeded, reset the state */
1065 if (async_list->io_pages + len <= max_pages) {
1066 req->flags |= REQ_F_SEQ_PREV;
1067 async_list->io_pages += len;
1070 async_list->io_pages = 0;
1074 /* New file? Reset state. */
1075 if (async_list->file != filp) {
1076 async_list->io_pages = 0;
1077 async_list->file = filp;
1079 async_list->io_end = io_end;
1082 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1083 bool force_nonblock)
1085 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1086 struct kiocb *kiocb = &req->rw;
1087 struct iov_iter iter;
1092 ret = io_prep_rw(req, s, force_nonblock);
1095 file = kiocb->ki_filp;
1097 if (unlikely(!(file->f_mode & FMODE_READ)))
1099 if (unlikely(!file->f_op->read_iter))
1102 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1106 iov_count = iov_iter_count(&iter);
1107 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1111 /* Catch -EAGAIN return for forced non-blocking submission */
1112 ret2 = call_read_iter(file, kiocb, &iter);
1113 if (!force_nonblock || ret2 != -EAGAIN) {
1114 io_rw_done(kiocb, ret2);
1117 * If ->needs_lock is true, we're already in async
1121 io_async_list_note(READ, req, iov_count);
1129 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1130 bool force_nonblock)
1132 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1133 struct kiocb *kiocb = &req->rw;
1134 struct iov_iter iter;
1139 ret = io_prep_rw(req, s, force_nonblock);
1143 file = kiocb->ki_filp;
1144 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1146 if (unlikely(!file->f_op->write_iter))
1149 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1153 iov_count = iov_iter_count(&iter);
1156 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1157 /* If ->needs_lock is true, we're already in async context. */
1159 io_async_list_note(WRITE, req, iov_count);
1163 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1168 * Open-code file_start_write here to grab freeze protection,
1169 * which will be released by another thread in
1170 * io_complete_rw(). Fool lockdep by telling it the lock got
1171 * released so that it doesn't complain about the held lock when
1172 * we return to userspace.
1174 if (S_ISREG(file_inode(file)->i_mode)) {
1175 __sb_start_write(file_inode(file)->i_sb,
1176 SB_FREEZE_WRITE, true);
1177 __sb_writers_release(file_inode(file)->i_sb,
1180 kiocb->ki_flags |= IOCB_WRITE;
1182 ret2 = call_write_iter(file, kiocb, &iter);
1183 if (!force_nonblock || ret2 != -EAGAIN) {
1184 io_rw_done(kiocb, ret2);
1187 * If ->needs_lock is true, we're already in async
1191 io_async_list_note(WRITE, req, iov_count);
1201 * IORING_OP_NOP just posts a completion event, nothing else.
1203 static int io_nop(struct io_kiocb *req, u64 user_data)
1205 struct io_ring_ctx *ctx = req->ctx;
1208 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1211 io_cqring_add_event(ctx, user_data, err);
1216 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1218 struct io_ring_ctx *ctx = req->ctx;
1223 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1225 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1231 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1232 bool force_nonblock)
1234 loff_t sqe_off = READ_ONCE(sqe->off);
1235 loff_t sqe_len = READ_ONCE(sqe->len);
1236 loff_t end = sqe_off + sqe_len;
1237 unsigned fsync_flags;
1240 fsync_flags = READ_ONCE(sqe->fsync_flags);
1241 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1244 ret = io_prep_fsync(req, sqe);
1248 /* fsync always requires a blocking context */
1252 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1253 end > 0 ? end : LLONG_MAX,
1254 fsync_flags & IORING_FSYNC_DATASYNC);
1256 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1261 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1263 struct io_ring_ctx *ctx = req->ctx;
1269 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1271 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1277 static int io_sync_file_range(struct io_kiocb *req,
1278 const struct io_uring_sqe *sqe,
1279 bool force_nonblock)
1286 ret = io_prep_sfr(req, sqe);
1290 /* sync_file_range always requires a blocking context */
1294 sqe_off = READ_ONCE(sqe->off);
1295 sqe_len = READ_ONCE(sqe->len);
1296 flags = READ_ONCE(sqe->sync_range_flags);
1298 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1300 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1305 static void io_poll_remove_one(struct io_kiocb *req)
1307 struct io_poll_iocb *poll = &req->poll;
1309 spin_lock(&poll->head->lock);
1310 WRITE_ONCE(poll->canceled, true);
1311 if (!list_empty(&poll->wait.entry)) {
1312 list_del_init(&poll->wait.entry);
1313 queue_work(req->ctx->sqo_wq, &req->work);
1315 spin_unlock(&poll->head->lock);
1317 list_del_init(&req->list);
1320 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1322 struct io_kiocb *req;
1324 spin_lock_irq(&ctx->completion_lock);
1325 while (!list_empty(&ctx->cancel_list)) {
1326 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1327 io_poll_remove_one(req);
1329 spin_unlock_irq(&ctx->completion_lock);
1333 * Find a running poll command that matches one specified in sqe->addr,
1334 * and remove it if found.
1336 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1338 struct io_ring_ctx *ctx = req->ctx;
1339 struct io_kiocb *poll_req, *next;
1342 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1344 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1348 spin_lock_irq(&ctx->completion_lock);
1349 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1350 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1351 io_poll_remove_one(poll_req);
1356 spin_unlock_irq(&ctx->completion_lock);
1358 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1363 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1366 req->poll.done = true;
1367 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1368 io_commit_cqring(ctx);
1371 static void io_poll_complete_work(struct work_struct *work)
1373 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1374 struct io_poll_iocb *poll = &req->poll;
1375 struct poll_table_struct pt = { ._key = poll->events };
1376 struct io_ring_ctx *ctx = req->ctx;
1379 if (!READ_ONCE(poll->canceled))
1380 mask = vfs_poll(poll->file, &pt) & poll->events;
1383 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1384 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1385 * synchronize with them. In the cancellation case the list_del_init
1386 * itself is not actually needed, but harmless so we keep it in to
1387 * avoid further branches in the fast path.
1389 spin_lock_irq(&ctx->completion_lock);
1390 if (!mask && !READ_ONCE(poll->canceled)) {
1391 add_wait_queue(poll->head, &poll->wait);
1392 spin_unlock_irq(&ctx->completion_lock);
1395 list_del_init(&req->list);
1396 io_poll_complete(ctx, req, mask);
1397 spin_unlock_irq(&ctx->completion_lock);
1399 io_cqring_ev_posted(ctx);
1403 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1406 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1408 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1409 struct io_ring_ctx *ctx = req->ctx;
1410 __poll_t mask = key_to_poll(key);
1411 unsigned long flags;
1413 /* for instances that support it check for an event match first: */
1414 if (mask && !(mask & poll->events))
1417 list_del_init(&poll->wait.entry);
1419 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1420 list_del(&req->list);
1421 io_poll_complete(ctx, req, mask);
1422 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1424 io_cqring_ev_posted(ctx);
1427 queue_work(ctx->sqo_wq, &req->work);
1433 struct io_poll_table {
1434 struct poll_table_struct pt;
1435 struct io_kiocb *req;
1439 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1440 struct poll_table_struct *p)
1442 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1444 if (unlikely(pt->req->poll.head)) {
1445 pt->error = -EINVAL;
1450 pt->req->poll.head = head;
1451 add_wait_queue(head, &pt->req->poll.wait);
1454 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1456 struct io_poll_iocb *poll = &req->poll;
1457 struct io_ring_ctx *ctx = req->ctx;
1458 struct io_poll_table ipt;
1459 bool cancel = false;
1463 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1465 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1470 INIT_WORK(&req->work, io_poll_complete_work);
1471 events = READ_ONCE(sqe->poll_events);
1472 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1476 poll->canceled = false;
1478 ipt.pt._qproc = io_poll_queue_proc;
1479 ipt.pt._key = poll->events;
1481 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1483 /* initialized the list so that we can do list_empty checks */
1484 INIT_LIST_HEAD(&poll->wait.entry);
1485 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1487 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1489 spin_lock_irq(&ctx->completion_lock);
1490 if (likely(poll->head)) {
1491 spin_lock(&poll->head->lock);
1492 if (unlikely(list_empty(&poll->wait.entry))) {
1498 if (mask || ipt.error)
1499 list_del_init(&poll->wait.entry);
1501 WRITE_ONCE(poll->canceled, true);
1502 else if (!poll->done) /* actually waiting for an event */
1503 list_add_tail(&req->list, &ctx->cancel_list);
1504 spin_unlock(&poll->head->lock);
1506 if (mask) { /* no async, we'd stolen it */
1508 io_poll_complete(ctx, req, mask);
1510 spin_unlock_irq(&ctx->completion_lock);
1513 io_cqring_ev_posted(ctx);
1519 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1520 const struct io_uring_sqe *sqe)
1522 struct io_uring_sqe *sqe_copy;
1524 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1527 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1531 spin_lock_irq(&ctx->completion_lock);
1532 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1533 spin_unlock_irq(&ctx->completion_lock);
1538 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1539 req->submit.sqe = sqe_copy;
1541 INIT_WORK(&req->work, io_sq_wq_submit_work);
1542 list_add_tail(&req->list, &ctx->defer_list);
1543 spin_unlock_irq(&ctx->completion_lock);
1544 return -EIOCBQUEUED;
1547 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1548 const struct sqe_submit *s, bool force_nonblock)
1552 if (unlikely(s->index >= ctx->sq_entries))
1554 req->user_data = READ_ONCE(s->sqe->user_data);
1556 opcode = READ_ONCE(s->sqe->opcode);
1559 ret = io_nop(req, req->user_data);
1561 case IORING_OP_READV:
1562 if (unlikely(s->sqe->buf_index))
1564 ret = io_read(req, s, force_nonblock);
1566 case IORING_OP_WRITEV:
1567 if (unlikely(s->sqe->buf_index))
1569 ret = io_write(req, s, force_nonblock);
1571 case IORING_OP_READ_FIXED:
1572 ret = io_read(req, s, force_nonblock);
1574 case IORING_OP_WRITE_FIXED:
1575 ret = io_write(req, s, force_nonblock);
1577 case IORING_OP_FSYNC:
1578 ret = io_fsync(req, s->sqe, force_nonblock);
1580 case IORING_OP_POLL_ADD:
1581 ret = io_poll_add(req, s->sqe);
1583 case IORING_OP_POLL_REMOVE:
1584 ret = io_poll_remove(req, s->sqe);
1586 case IORING_OP_SYNC_FILE_RANGE:
1587 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1597 if (ctx->flags & IORING_SETUP_IOPOLL) {
1598 if (req->error == -EAGAIN)
1601 /* workqueue context doesn't hold uring_lock, grab it now */
1603 mutex_lock(&ctx->uring_lock);
1604 io_iopoll_req_issued(req);
1606 mutex_unlock(&ctx->uring_lock);
1612 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1613 const struct io_uring_sqe *sqe)
1615 switch (sqe->opcode) {
1616 case IORING_OP_READV:
1617 case IORING_OP_READ_FIXED:
1618 return &ctx->pending_async[READ];
1619 case IORING_OP_WRITEV:
1620 case IORING_OP_WRITE_FIXED:
1621 return &ctx->pending_async[WRITE];
1627 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1629 u8 opcode = READ_ONCE(sqe->opcode);
1631 return !(opcode == IORING_OP_READ_FIXED ||
1632 opcode == IORING_OP_WRITE_FIXED);
1635 static void io_sq_wq_submit_work(struct work_struct *work)
1637 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1638 struct io_ring_ctx *ctx = req->ctx;
1639 struct mm_struct *cur_mm = NULL;
1640 struct async_list *async_list;
1641 LIST_HEAD(req_list);
1642 mm_segment_t old_fs;
1645 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1648 struct sqe_submit *s = &req->submit;
1649 const struct io_uring_sqe *sqe = s->sqe;
1651 /* Ensure we clear previously set non-block flag */
1652 req->rw.ki_flags &= ~IOCB_NOWAIT;
1655 if (io_sqe_needs_user(sqe) && !cur_mm) {
1656 if (!mmget_not_zero(ctx->sqo_mm)) {
1659 cur_mm = ctx->sqo_mm;
1667 s->has_user = cur_mm != NULL;
1668 s->needs_lock = true;
1670 ret = __io_submit_sqe(ctx, req, s, false);
1672 * We can get EAGAIN for polled IO even though
1673 * we're forcing a sync submission from here,
1674 * since we can't wait for request slots on the
1683 /* drop submission reference */
1687 io_cqring_add_event(ctx, sqe->user_data, ret);
1691 /* async context always use a copy of the sqe */
1696 if (!list_empty(&req_list)) {
1697 req = list_first_entry(&req_list, struct io_kiocb,
1699 list_del(&req->list);
1702 if (list_empty(&async_list->list))
1706 spin_lock(&async_list->lock);
1707 if (list_empty(&async_list->list)) {
1708 spin_unlock(&async_list->lock);
1711 list_splice_init(&async_list->list, &req_list);
1712 spin_unlock(&async_list->lock);
1714 req = list_first_entry(&req_list, struct io_kiocb, list);
1715 list_del(&req->list);
1719 * Rare case of racing with a submitter. If we find the count has
1720 * dropped to zero AND we have pending work items, then restart
1721 * the processing. This is a tiny race window.
1724 ret = atomic_dec_return(&async_list->cnt);
1725 while (!ret && !list_empty(&async_list->list)) {
1726 spin_lock(&async_list->lock);
1727 atomic_inc(&async_list->cnt);
1728 list_splice_init(&async_list->list, &req_list);
1729 spin_unlock(&async_list->lock);
1731 if (!list_empty(&req_list)) {
1732 req = list_first_entry(&req_list,
1733 struct io_kiocb, list);
1734 list_del(&req->list);
1737 ret = atomic_dec_return(&async_list->cnt);
1749 * See if we can piggy back onto previously submitted work, that is still
1750 * running. We currently only allow this if the new request is sequential
1751 * to the previous one we punted.
1753 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
1759 if (!(req->flags & REQ_F_SEQ_PREV))
1761 if (!atomic_read(&list->cnt))
1765 spin_lock(&list->lock);
1766 list_add_tail(&req->list, &list->list);
1767 if (!atomic_read(&list->cnt)) {
1768 list_del_init(&req->list);
1771 spin_unlock(&list->lock);
1775 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
1777 int op = READ_ONCE(sqe->opcode);
1781 case IORING_OP_POLL_REMOVE:
1788 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
1789 struct io_submit_state *state, struct io_kiocb *req)
1794 flags = READ_ONCE(s->sqe->flags);
1795 fd = READ_ONCE(s->sqe->fd);
1797 if (flags & IOSQE_IO_DRAIN) {
1798 req->flags |= REQ_F_IO_DRAIN;
1799 req->sequence = ctx->cached_sq_head - 1;
1802 if (!io_op_needs_file(s->sqe))
1805 if (flags & IOSQE_FIXED_FILE) {
1806 if (unlikely(!ctx->user_files ||
1807 (unsigned) fd >= ctx->nr_user_files))
1809 req->file = ctx->user_files[fd];
1810 req->flags |= REQ_F_FIXED_FILE;
1812 if (s->needs_fixed_file)
1814 req->file = io_file_get(state, fd);
1815 if (unlikely(!req->file))
1822 static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
1823 struct io_submit_state *state)
1825 struct io_kiocb *req;
1828 /* enforce forwards compatibility on users */
1829 if (unlikely(s->sqe->flags & ~(IOSQE_FIXED_FILE | IOSQE_IO_DRAIN)))
1832 req = io_get_req(ctx, state);
1836 ret = io_req_set_file(ctx, s, state, req);
1840 ret = io_req_defer(ctx, req, s->sqe);
1842 if (ret == -EIOCBQUEUED)
1847 ret = __io_submit_sqe(ctx, req, s, true);
1848 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
1849 struct io_uring_sqe *sqe_copy;
1851 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1853 struct async_list *list;
1855 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
1858 memcpy(&req->submit, s, sizeof(*s));
1859 list = io_async_list_from_sqe(ctx, s->sqe);
1860 if (!io_add_to_prev_work(list, req)) {
1862 atomic_inc(&list->cnt);
1863 INIT_WORK(&req->work, io_sq_wq_submit_work);
1864 queue_work(ctx->sqo_wq, &req->work);
1868 * Queued up for async execution, worker will release
1869 * submit reference when the iocb is actually
1877 /* drop submission reference */
1880 /* and drop final reference, if we failed */
1888 * Batched submission is done, ensure local IO is flushed out.
1890 static void io_submit_state_end(struct io_submit_state *state)
1892 blk_finish_plug(&state->plug);
1894 if (state->free_reqs)
1895 kmem_cache_free_bulk(req_cachep, state->free_reqs,
1896 &state->reqs[state->cur_req]);
1900 * Start submission side cache.
1902 static void io_submit_state_start(struct io_submit_state *state,
1903 struct io_ring_ctx *ctx, unsigned max_ios)
1905 blk_start_plug(&state->plug);
1906 state->free_reqs = 0;
1908 state->ios_left = max_ios;
1911 static void io_commit_sqring(struct io_ring_ctx *ctx)
1913 struct io_sq_ring *ring = ctx->sq_ring;
1915 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
1917 * Ensure any loads from the SQEs are done at this point,
1918 * since once we write the new head, the application could
1919 * write new data to them.
1921 smp_store_release(&ring->r.head, ctx->cached_sq_head);
1926 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
1927 * that is mapped by userspace. This means that care needs to be taken to
1928 * ensure that reads are stable, as we cannot rely on userspace always
1929 * being a good citizen. If members of the sqe are validated and then later
1930 * used, it's important that those reads are done through READ_ONCE() to
1931 * prevent a re-load down the line.
1933 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
1935 struct io_sq_ring *ring = ctx->sq_ring;
1939 * The cached sq head (or cq tail) serves two purposes:
1941 * 1) allows us to batch the cost of updating the user visible
1943 * 2) allows the kernel side to track the head on its own, even
1944 * though the application is the one updating it.
1946 head = ctx->cached_sq_head;
1947 /* make sure SQ entry isn't read before tail */
1948 if (head == smp_load_acquire(&ring->r.tail))
1951 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
1952 if (head < ctx->sq_entries) {
1954 s->sqe = &ctx->sq_sqes[head];
1955 ctx->cached_sq_head++;
1959 /* drop invalid entries */
1960 ctx->cached_sq_head++;
1965 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
1966 unsigned int nr, bool has_user, bool mm_fault)
1968 struct io_submit_state state, *statep = NULL;
1969 int ret, i, submitted = 0;
1971 if (nr > IO_PLUG_THRESHOLD) {
1972 io_submit_state_start(&state, ctx, nr);
1976 for (i = 0; i < nr; i++) {
1977 if (unlikely(mm_fault)) {
1980 sqes[i].has_user = has_user;
1981 sqes[i].needs_lock = true;
1982 sqes[i].needs_fixed_file = true;
1983 ret = io_submit_sqe(ctx, &sqes[i], statep);
1990 io_cqring_add_event(ctx, sqes[i].sqe->user_data, ret);
1994 io_submit_state_end(&state);
1999 static int io_sq_thread(void *data)
2001 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2002 struct io_ring_ctx *ctx = data;
2003 struct mm_struct *cur_mm = NULL;
2004 mm_segment_t old_fs;
2007 unsigned long timeout;
2012 timeout = inflight = 0;
2013 while (!kthread_should_park()) {
2014 bool all_fixed, mm_fault = false;
2018 unsigned nr_events = 0;
2020 if (ctx->flags & IORING_SETUP_IOPOLL) {
2022 * We disallow the app entering submit/complete
2023 * with polling, but we still need to lock the
2024 * ring to prevent racing with polled issue
2025 * that got punted to a workqueue.
2027 mutex_lock(&ctx->uring_lock);
2028 io_iopoll_check(ctx, &nr_events, 0);
2029 mutex_unlock(&ctx->uring_lock);
2032 * Normal IO, just pretend everything completed.
2033 * We don't have to poll completions for that.
2035 nr_events = inflight;
2038 inflight -= nr_events;
2040 timeout = jiffies + ctx->sq_thread_idle;
2043 if (!io_get_sqring(ctx, &sqes[0])) {
2045 * We're polling. If we're within the defined idle
2046 * period, then let us spin without work before going
2049 if (inflight || !time_after(jiffies, timeout)) {
2055 * Drop cur_mm before scheduling, we can't hold it for
2056 * long periods (or over schedule()). Do this before
2057 * adding ourselves to the waitqueue, as the unuse/drop
2066 prepare_to_wait(&ctx->sqo_wait, &wait,
2067 TASK_INTERRUPTIBLE);
2069 /* Tell userspace we may need a wakeup call */
2070 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
2071 /* make sure to read SQ tail after writing flags */
2074 if (!io_get_sqring(ctx, &sqes[0])) {
2075 if (kthread_should_park()) {
2076 finish_wait(&ctx->sqo_wait, &wait);
2079 if (signal_pending(current))
2080 flush_signals(current);
2082 finish_wait(&ctx->sqo_wait, &wait);
2084 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2087 finish_wait(&ctx->sqo_wait, &wait);
2089 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2095 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2099 if (i == ARRAY_SIZE(sqes))
2101 } while (io_get_sqring(ctx, &sqes[i]));
2103 /* Unless all new commands are FIXED regions, grab mm */
2104 if (!all_fixed && !cur_mm) {
2105 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2107 use_mm(ctx->sqo_mm);
2108 cur_mm = ctx->sqo_mm;
2112 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2115 /* Commit SQ ring head once we've consumed all SQEs */
2116 io_commit_sqring(ctx);
2130 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2132 struct io_submit_state state, *statep = NULL;
2135 if (to_submit > IO_PLUG_THRESHOLD) {
2136 io_submit_state_start(&state, ctx, to_submit);
2140 for (i = 0; i < to_submit; i++) {
2141 struct sqe_submit s;
2144 if (!io_get_sqring(ctx, &s))
2148 s.needs_lock = false;
2149 s.needs_fixed_file = false;
2152 ret = io_submit_sqe(ctx, &s, statep);
2154 io_cqring_add_event(ctx, s.sqe->user_data, ret);
2156 io_commit_sqring(ctx);
2159 io_submit_state_end(statep);
2164 static unsigned io_cqring_events(struct io_cq_ring *ring)
2166 /* See comment at the top of this file */
2168 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
2172 * Wait until events become available, if we don't already have some. The
2173 * application must reap them itself, as they reside on the shared cq ring.
2175 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2176 const sigset_t __user *sig, size_t sigsz)
2178 struct io_cq_ring *ring = ctx->cq_ring;
2179 sigset_t ksigmask, sigsaved;
2182 if (io_cqring_events(ring) >= min_events)
2186 #ifdef CONFIG_COMPAT
2187 if (in_compat_syscall())
2188 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2189 &ksigmask, &sigsaved, sigsz);
2192 ret = set_user_sigmask(sig, &ksigmask,
2199 ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events);
2202 restore_user_sigmask(sig, &sigsaved, ret == -ERESTARTSYS);
2204 if (ret == -ERESTARTSYS)
2207 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
2210 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2212 #if defined(CONFIG_UNIX)
2213 if (ctx->ring_sock) {
2214 struct sock *sock = ctx->ring_sock->sk;
2215 struct sk_buff *skb;
2217 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2223 for (i = 0; i < ctx->nr_user_files; i++)
2224 fput(ctx->user_files[i]);
2228 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2230 if (!ctx->user_files)
2233 __io_sqe_files_unregister(ctx);
2234 kfree(ctx->user_files);
2235 ctx->user_files = NULL;
2236 ctx->nr_user_files = 0;
2240 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2242 if (ctx->sqo_thread) {
2244 * The park is a bit of a work-around, without it we get
2245 * warning spews on shutdown with SQPOLL set and affinity
2246 * set to a single CPU.
2248 kthread_park(ctx->sqo_thread);
2249 kthread_stop(ctx->sqo_thread);
2250 ctx->sqo_thread = NULL;
2254 static void io_finish_async(struct io_ring_ctx *ctx)
2256 io_sq_thread_stop(ctx);
2259 destroy_workqueue(ctx->sqo_wq);
2264 #if defined(CONFIG_UNIX)
2265 static void io_destruct_skb(struct sk_buff *skb)
2267 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2269 io_finish_async(ctx);
2270 unix_destruct_scm(skb);
2274 * Ensure the UNIX gc is aware of our file set, so we are certain that
2275 * the io_uring can be safely unregistered on process exit, even if we have
2276 * loops in the file referencing.
2278 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2280 struct sock *sk = ctx->ring_sock->sk;
2281 struct scm_fp_list *fpl;
2282 struct sk_buff *skb;
2285 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2286 unsigned long inflight = ctx->user->unix_inflight + nr;
2288 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2292 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2296 skb = alloc_skb(0, GFP_KERNEL);
2303 skb->destructor = io_destruct_skb;
2305 fpl->user = get_uid(ctx->user);
2306 for (i = 0; i < nr; i++) {
2307 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2308 unix_inflight(fpl->user, fpl->fp[i]);
2311 fpl->max = fpl->count = nr;
2312 UNIXCB(skb).fp = fpl;
2313 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2314 skb_queue_head(&sk->sk_receive_queue, skb);
2316 for (i = 0; i < nr; i++)
2323 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2324 * causes regular reference counting to break down. We rely on the UNIX
2325 * garbage collection to take care of this problem for us.
2327 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2329 unsigned left, total;
2333 left = ctx->nr_user_files;
2335 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2337 ret = __io_sqe_files_scm(ctx, this_files, total);
2341 total += this_files;
2347 while (total < ctx->nr_user_files) {
2348 fput(ctx->user_files[total]);
2355 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2361 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2364 __s32 __user *fds = (__s32 __user *) arg;
2368 if (ctx->user_files)
2372 if (nr_args > IORING_MAX_FIXED_FILES)
2375 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2376 if (!ctx->user_files)
2379 for (i = 0; i < nr_args; i++) {
2381 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2384 ctx->user_files[i] = fget(fd);
2387 if (!ctx->user_files[i])
2390 * Don't allow io_uring instances to be registered. If UNIX
2391 * isn't enabled, then this causes a reference cycle and this
2392 * instance can never get freed. If UNIX is enabled we'll
2393 * handle it just fine, but there's still no point in allowing
2394 * a ring fd as it doesn't support regular read/write anyway.
2396 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2397 fput(ctx->user_files[i]);
2400 ctx->nr_user_files++;
2405 for (i = 0; i < ctx->nr_user_files; i++)
2406 fput(ctx->user_files[i]);
2408 kfree(ctx->user_files);
2409 ctx->user_files = NULL;
2410 ctx->nr_user_files = 0;
2414 ret = io_sqe_files_scm(ctx);
2416 io_sqe_files_unregister(ctx);
2421 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2422 struct io_uring_params *p)
2426 init_waitqueue_head(&ctx->sqo_wait);
2427 mmgrab(current->mm);
2428 ctx->sqo_mm = current->mm;
2430 if (ctx->flags & IORING_SETUP_SQPOLL) {
2432 if (!capable(CAP_SYS_ADMIN))
2435 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2436 if (!ctx->sq_thread_idle)
2437 ctx->sq_thread_idle = HZ;
2439 if (p->flags & IORING_SETUP_SQ_AFF) {
2440 int cpu = p->sq_thread_cpu;
2443 if (cpu >= nr_cpu_ids)
2445 if (!cpu_online(cpu))
2448 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2452 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2455 if (IS_ERR(ctx->sqo_thread)) {
2456 ret = PTR_ERR(ctx->sqo_thread);
2457 ctx->sqo_thread = NULL;
2460 wake_up_process(ctx->sqo_thread);
2461 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2462 /* Can't have SQ_AFF without SQPOLL */
2467 /* Do QD, or 2 * CPUS, whatever is smallest */
2468 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
2469 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2477 io_sq_thread_stop(ctx);
2478 mmdrop(ctx->sqo_mm);
2483 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2485 atomic_long_sub(nr_pages, &user->locked_vm);
2488 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2490 unsigned long page_limit, cur_pages, new_pages;
2492 /* Don't allow more pages than we can safely lock */
2493 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2496 cur_pages = atomic_long_read(&user->locked_vm);
2497 new_pages = cur_pages + nr_pages;
2498 if (new_pages > page_limit)
2500 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2501 new_pages) != cur_pages);
2506 static void io_mem_free(void *ptr)
2513 page = virt_to_head_page(ptr);
2514 if (put_page_testzero(page))
2515 free_compound_page(page);
2518 static void *io_mem_alloc(size_t size)
2520 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2523 return (void *) __get_free_pages(gfp_flags, get_order(size));
2526 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2528 struct io_sq_ring *sq_ring;
2529 struct io_cq_ring *cq_ring;
2532 bytes = struct_size(sq_ring, array, sq_entries);
2533 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
2534 bytes += struct_size(cq_ring, cqes, cq_entries);
2536 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
2539 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
2543 if (!ctx->user_bufs)
2546 for (i = 0; i < ctx->nr_user_bufs; i++) {
2547 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2549 for (j = 0; j < imu->nr_bvecs; j++)
2550 put_page(imu->bvec[j].bv_page);
2552 if (ctx->account_mem)
2553 io_unaccount_mem(ctx->user, imu->nr_bvecs);
2558 kfree(ctx->user_bufs);
2559 ctx->user_bufs = NULL;
2560 ctx->nr_user_bufs = 0;
2564 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
2565 void __user *arg, unsigned index)
2567 struct iovec __user *src;
2569 #ifdef CONFIG_COMPAT
2571 struct compat_iovec __user *ciovs;
2572 struct compat_iovec ciov;
2574 ciovs = (struct compat_iovec __user *) arg;
2575 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
2578 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
2579 dst->iov_len = ciov.iov_len;
2583 src = (struct iovec __user *) arg;
2584 if (copy_from_user(dst, &src[index], sizeof(*dst)))
2589 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
2592 struct vm_area_struct **vmas = NULL;
2593 struct page **pages = NULL;
2594 int i, j, got_pages = 0;
2599 if (!nr_args || nr_args > UIO_MAXIOV)
2602 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
2604 if (!ctx->user_bufs)
2607 for (i = 0; i < nr_args; i++) {
2608 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2609 unsigned long off, start, end, ubuf;
2614 ret = io_copy_iov(ctx, &iov, arg, i);
2619 * Don't impose further limits on the size and buffer
2620 * constraints here, we'll -EINVAL later when IO is
2621 * submitted if they are wrong.
2624 if (!iov.iov_base || !iov.iov_len)
2627 /* arbitrary limit, but we need something */
2628 if (iov.iov_len > SZ_1G)
2631 ubuf = (unsigned long) iov.iov_base;
2632 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2633 start = ubuf >> PAGE_SHIFT;
2634 nr_pages = end - start;
2636 if (ctx->account_mem) {
2637 ret = io_account_mem(ctx->user, nr_pages);
2643 if (!pages || nr_pages > got_pages) {
2646 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
2648 vmas = kvmalloc_array(nr_pages,
2649 sizeof(struct vm_area_struct *),
2651 if (!pages || !vmas) {
2653 if (ctx->account_mem)
2654 io_unaccount_mem(ctx->user, nr_pages);
2657 got_pages = nr_pages;
2660 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
2664 if (ctx->account_mem)
2665 io_unaccount_mem(ctx->user, nr_pages);
2670 down_read(¤t->mm->mmap_sem);
2671 pret = get_user_pages(ubuf, nr_pages,
2672 FOLL_WRITE | FOLL_LONGTERM,
2674 if (pret == nr_pages) {
2675 /* don't support file backed memory */
2676 for (j = 0; j < nr_pages; j++) {
2677 struct vm_area_struct *vma = vmas[j];
2680 !is_file_hugepages(vma->vm_file)) {
2686 ret = pret < 0 ? pret : -EFAULT;
2688 up_read(¤t->mm->mmap_sem);
2691 * if we did partial map, or found file backed vmas,
2692 * release any pages we did get
2695 for (j = 0; j < pret; j++)
2698 if (ctx->account_mem)
2699 io_unaccount_mem(ctx->user, nr_pages);
2704 off = ubuf & ~PAGE_MASK;
2706 for (j = 0; j < nr_pages; j++) {
2709 vec_len = min_t(size_t, size, PAGE_SIZE - off);
2710 imu->bvec[j].bv_page = pages[j];
2711 imu->bvec[j].bv_len = vec_len;
2712 imu->bvec[j].bv_offset = off;
2716 /* store original address for later verification */
2718 imu->len = iov.iov_len;
2719 imu->nr_bvecs = nr_pages;
2721 ctx->nr_user_bufs++;
2729 io_sqe_buffer_unregister(ctx);
2733 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
2735 __s32 __user *fds = arg;
2741 if (copy_from_user(&fd, fds, sizeof(*fds)))
2744 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
2745 if (IS_ERR(ctx->cq_ev_fd)) {
2746 int ret = PTR_ERR(ctx->cq_ev_fd);
2747 ctx->cq_ev_fd = NULL;
2754 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
2756 if (ctx->cq_ev_fd) {
2757 eventfd_ctx_put(ctx->cq_ev_fd);
2758 ctx->cq_ev_fd = NULL;
2765 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
2767 io_finish_async(ctx);
2769 mmdrop(ctx->sqo_mm);
2771 io_iopoll_reap_events(ctx);
2772 io_sqe_buffer_unregister(ctx);
2773 io_sqe_files_unregister(ctx);
2774 io_eventfd_unregister(ctx);
2776 #if defined(CONFIG_UNIX)
2777 if (ctx->ring_sock) {
2778 ctx->ring_sock->file = NULL; /* so that iput() is called */
2779 sock_release(ctx->ring_sock);
2783 io_mem_free(ctx->sq_ring);
2784 io_mem_free(ctx->sq_sqes);
2785 io_mem_free(ctx->cq_ring);
2787 percpu_ref_exit(&ctx->refs);
2788 if (ctx->account_mem)
2789 io_unaccount_mem(ctx->user,
2790 ring_pages(ctx->sq_entries, ctx->cq_entries));
2791 free_uid(ctx->user);
2795 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
2797 struct io_ring_ctx *ctx = file->private_data;
2800 poll_wait(file, &ctx->cq_wait, wait);
2802 * synchronizes with barrier from wq_has_sleeper call in
2806 if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
2807 ctx->sq_ring->ring_entries)
2808 mask |= EPOLLOUT | EPOLLWRNORM;
2809 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
2810 mask |= EPOLLIN | EPOLLRDNORM;
2815 static int io_uring_fasync(int fd, struct file *file, int on)
2817 struct io_ring_ctx *ctx = file->private_data;
2819 return fasync_helper(fd, file, on, &ctx->cq_fasync);
2822 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
2824 mutex_lock(&ctx->uring_lock);
2825 percpu_ref_kill(&ctx->refs);
2826 mutex_unlock(&ctx->uring_lock);
2828 io_poll_remove_all(ctx);
2829 io_iopoll_reap_events(ctx);
2830 wait_for_completion(&ctx->ctx_done);
2831 io_ring_ctx_free(ctx);
2834 static int io_uring_release(struct inode *inode, struct file *file)
2836 struct io_ring_ctx *ctx = file->private_data;
2838 file->private_data = NULL;
2839 io_ring_ctx_wait_and_kill(ctx);
2843 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
2845 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
2846 unsigned long sz = vma->vm_end - vma->vm_start;
2847 struct io_ring_ctx *ctx = file->private_data;
2853 case IORING_OFF_SQ_RING:
2856 case IORING_OFF_SQES:
2859 case IORING_OFF_CQ_RING:
2866 page = virt_to_head_page(ptr);
2867 if (sz > (PAGE_SIZE << compound_order(page)))
2870 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
2871 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
2874 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
2875 u32, min_complete, u32, flags, const sigset_t __user *, sig,
2878 struct io_ring_ctx *ctx;
2883 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
2891 if (f.file->f_op != &io_uring_fops)
2895 ctx = f.file->private_data;
2896 if (!percpu_ref_tryget(&ctx->refs))
2900 * For SQ polling, the thread will do all submissions and completions.
2901 * Just return the requested submit count, and wake the thread if
2904 if (ctx->flags & IORING_SETUP_SQPOLL) {
2905 if (flags & IORING_ENTER_SQ_WAKEUP)
2906 wake_up(&ctx->sqo_wait);
2907 submitted = to_submit;
2913 to_submit = min(to_submit, ctx->sq_entries);
2915 mutex_lock(&ctx->uring_lock);
2916 submitted = io_ring_submit(ctx, to_submit);
2917 mutex_unlock(&ctx->uring_lock);
2919 if (flags & IORING_ENTER_GETEVENTS) {
2920 unsigned nr_events = 0;
2922 min_complete = min(min_complete, ctx->cq_entries);
2924 if (ctx->flags & IORING_SETUP_IOPOLL) {
2925 mutex_lock(&ctx->uring_lock);
2926 ret = io_iopoll_check(ctx, &nr_events, min_complete);
2927 mutex_unlock(&ctx->uring_lock);
2929 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
2934 io_ring_drop_ctx_refs(ctx, 1);
2937 return submitted ? submitted : ret;
2940 static const struct file_operations io_uring_fops = {
2941 .release = io_uring_release,
2942 .mmap = io_uring_mmap,
2943 .poll = io_uring_poll,
2944 .fasync = io_uring_fasync,
2947 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
2948 struct io_uring_params *p)
2950 struct io_sq_ring *sq_ring;
2951 struct io_cq_ring *cq_ring;
2954 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
2958 ctx->sq_ring = sq_ring;
2959 sq_ring->ring_mask = p->sq_entries - 1;
2960 sq_ring->ring_entries = p->sq_entries;
2961 ctx->sq_mask = sq_ring->ring_mask;
2962 ctx->sq_entries = sq_ring->ring_entries;
2964 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
2965 if (size == SIZE_MAX)
2968 ctx->sq_sqes = io_mem_alloc(size);
2972 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
2976 ctx->cq_ring = cq_ring;
2977 cq_ring->ring_mask = p->cq_entries - 1;
2978 cq_ring->ring_entries = p->cq_entries;
2979 ctx->cq_mask = cq_ring->ring_mask;
2980 ctx->cq_entries = cq_ring->ring_entries;
2985 * Allocate an anonymous fd, this is what constitutes the application
2986 * visible backing of an io_uring instance. The application mmaps this
2987 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
2988 * we have to tie this fd to a socket for file garbage collection purposes.
2990 static int io_uring_get_fd(struct io_ring_ctx *ctx)
2995 #if defined(CONFIG_UNIX)
2996 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3002 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3006 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3007 O_RDWR | O_CLOEXEC);
3010 ret = PTR_ERR(file);
3014 #if defined(CONFIG_UNIX)
3015 ctx->ring_sock->file = file;
3016 ctx->ring_sock->sk->sk_user_data = ctx;
3018 fd_install(ret, file);
3021 #if defined(CONFIG_UNIX)
3022 sock_release(ctx->ring_sock);
3023 ctx->ring_sock = NULL;
3028 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3030 struct user_struct *user = NULL;
3031 struct io_ring_ctx *ctx;
3035 if (!entries || entries > IORING_MAX_ENTRIES)
3039 * Use twice as many entries for the CQ ring. It's possible for the
3040 * application to drive a higher depth than the size of the SQ ring,
3041 * since the sqes are only used at submission time. This allows for
3042 * some flexibility in overcommitting a bit.
3044 p->sq_entries = roundup_pow_of_two(entries);
3045 p->cq_entries = 2 * p->sq_entries;
3047 user = get_uid(current_user());
3048 account_mem = !capable(CAP_IPC_LOCK);
3051 ret = io_account_mem(user,
3052 ring_pages(p->sq_entries, p->cq_entries));
3059 ctx = io_ring_ctx_alloc(p);
3062 io_unaccount_mem(user, ring_pages(p->sq_entries,
3067 ctx->compat = in_compat_syscall();
3068 ctx->account_mem = account_mem;
3071 ret = io_allocate_scq_urings(ctx, p);
3075 ret = io_sq_offload_start(ctx, p);
3079 ret = io_uring_get_fd(ctx);
3083 memset(&p->sq_off, 0, sizeof(p->sq_off));
3084 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
3085 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
3086 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
3087 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
3088 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
3089 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
3090 p->sq_off.array = offsetof(struct io_sq_ring, array);
3092 memset(&p->cq_off, 0, sizeof(p->cq_off));
3093 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
3094 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
3095 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
3096 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
3097 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
3098 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
3101 io_ring_ctx_wait_and_kill(ctx);
3106 * Sets up an aio uring context, and returns the fd. Applications asks for a
3107 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3108 * params structure passed in.
3110 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3112 struct io_uring_params p;
3116 if (copy_from_user(&p, params, sizeof(p)))
3118 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3123 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3124 IORING_SETUP_SQ_AFF))
3127 ret = io_uring_create(entries, &p);
3131 if (copy_to_user(params, &p, sizeof(p)))
3137 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3138 struct io_uring_params __user *, params)
3140 return io_uring_setup(entries, params);
3143 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3144 void __user *arg, unsigned nr_args)
3145 __releases(ctx->uring_lock)
3146 __acquires(ctx->uring_lock)
3151 * We're inside the ring mutex, if the ref is already dying, then
3152 * someone else killed the ctx or is already going through
3153 * io_uring_register().
3155 if (percpu_ref_is_dying(&ctx->refs))
3158 percpu_ref_kill(&ctx->refs);
3161 * Drop uring mutex before waiting for references to exit. If another
3162 * thread is currently inside io_uring_enter() it might need to grab
3163 * the uring_lock to make progress. If we hold it here across the drain
3164 * wait, then we can deadlock. It's safe to drop the mutex here, since
3165 * no new references will come in after we've killed the percpu ref.
3167 mutex_unlock(&ctx->uring_lock);
3168 wait_for_completion(&ctx->ctx_done);
3169 mutex_lock(&ctx->uring_lock);
3172 case IORING_REGISTER_BUFFERS:
3173 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3175 case IORING_UNREGISTER_BUFFERS:
3179 ret = io_sqe_buffer_unregister(ctx);
3181 case IORING_REGISTER_FILES:
3182 ret = io_sqe_files_register(ctx, arg, nr_args);
3184 case IORING_UNREGISTER_FILES:
3188 ret = io_sqe_files_unregister(ctx);
3190 case IORING_REGISTER_EVENTFD:
3194 ret = io_eventfd_register(ctx, arg);
3196 case IORING_UNREGISTER_EVENTFD:
3200 ret = io_eventfd_unregister(ctx);
3207 /* bring the ctx back to life */
3208 reinit_completion(&ctx->ctx_done);
3209 percpu_ref_reinit(&ctx->refs);
3213 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3214 void __user *, arg, unsigned int, nr_args)
3216 struct io_ring_ctx *ctx;
3225 if (f.file->f_op != &io_uring_fops)
3228 ctx = f.file->private_data;
3230 mutex_lock(&ctx->uring_lock);
3231 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3232 mutex_unlock(&ctx->uring_lock);
3238 static int __init io_uring_init(void)
3240 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3243 __initcall(io_uring_init);