1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 4096
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp;
83 u32 tail ____cacheline_aligned_in_smp;
87 * This data is shared with the application through the mmap at offset
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head and the application controls tail.
102 * Bitmask to apply to head and tail offsets (constant, equals
106 /* Ring size (constant, power of 2) */
109 * Number of invalid entries dropped by the kernel due to
110 * invalid index stored in array
112 * Written by the kernel, shouldn't be modified by the
113 * application (i.e. get number of "new events" by comparing to
116 * After a new SQ head value was read by the application this
117 * counter includes all submissions that were dropped reaching
118 * the new SQ head (and possibly more).
124 * Written by the kernel, shouldn't be modified by the
127 * The application needs a full memory barrier before checking
128 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
132 * Ring buffer of indices into array of io_uring_sqe, which is
133 * mmapped by the application using the IORING_OFF_SQES offset.
135 * This indirection could e.g. be used to assign fixed
136 * io_uring_sqe entries to operations and only submit them to
137 * the queue when needed.
139 * The kernel modifies neither the indices array nor the entries
146 * This data is shared with the application through the mmap at offset
147 * IORING_OFF_CQ_RING.
149 * The offsets to the member fields are published through struct
150 * io_cqring_offsets when calling io_uring_setup.
154 * Head and tail offsets into the ring; the offsets need to be
155 * masked to get valid indices.
157 * The application controls head and the kernel tail.
161 * Bitmask to apply to head and tail offsets (constant, equals
165 /* Ring size (constant, power of 2) */
168 * Number of completion events lost because the queue was full;
169 * this should be avoided by the application by making sure
170 * there are not more requests pending thatn there is space in
171 * the completion queue.
173 * Written by the kernel, shouldn't be modified by the
174 * application (i.e. get number of "new events" by comparing to
177 * As completion events come in out of order this counter is not
178 * ordered with any other data.
182 * Ring buffer of completion events.
184 * The kernel writes completion events fresh every time they are
185 * produced, so the application is allowed to modify pending
188 struct io_uring_cqe cqes[];
191 struct io_mapped_ubuf {
194 struct bio_vec *bvec;
195 unsigned int nr_bvecs;
201 struct list_head list;
210 struct percpu_ref refs;
211 } ____cacheline_aligned_in_smp;
219 struct io_sq_ring *sq_ring;
220 unsigned cached_sq_head;
223 unsigned sq_thread_idle;
224 struct io_uring_sqe *sq_sqes;
226 struct list_head defer_list;
227 } ____cacheline_aligned_in_smp;
230 struct workqueue_struct *sqo_wq;
231 struct task_struct *sqo_thread; /* if using sq thread polling */
232 struct mm_struct *sqo_mm;
233 wait_queue_head_t sqo_wait;
234 struct completion sqo_thread_started;
238 struct io_cq_ring *cq_ring;
239 unsigned cached_cq_tail;
242 struct wait_queue_head cq_wait;
243 struct fasync_struct *cq_fasync;
244 struct eventfd_ctx *cq_ev_fd;
245 } ____cacheline_aligned_in_smp;
248 * If used, fixed file set. Writers must ensure that ->refs is dead,
249 * readers must ensure that ->refs is alive as long as the file* is
250 * used. Only updated through io_uring_register(2).
252 struct file **user_files;
253 unsigned nr_user_files;
255 /* if used, fixed mapped user buffers */
256 unsigned nr_user_bufs;
257 struct io_mapped_ubuf *user_bufs;
259 struct user_struct *user;
261 struct completion ctx_done;
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
269 spinlock_t completion_lock;
270 bool poll_multi_file;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list;
278 struct list_head cancel_list;
279 } ____cacheline_aligned_in_smp;
281 struct async_list pending_async[2];
283 #if defined(CONFIG_UNIX)
284 struct socket *ring_sock;
289 const struct io_uring_sqe *sqe;
290 unsigned short index;
293 bool needs_fixed_file;
297 * First field must be the file pointer in all the
298 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
300 struct io_poll_iocb {
302 struct wait_queue_head *head;
306 struct wait_queue_entry wait;
310 * NOTE! Each of the iocb union members has the file pointer
311 * as the first entry in their struct definition. So you can
312 * access the file pointer through any of the sub-structs,
313 * or directly as just 'ki_filp' in this struct.
319 struct io_poll_iocb poll;
322 struct sqe_submit submit;
324 struct io_ring_ctx *ctx;
325 struct list_head list;
326 struct list_head link_list;
329 #define REQ_F_NOWAIT 1 /* must not punt to workers */
330 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
331 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
332 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
333 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
334 #define REQ_F_IO_DRAINED 32 /* drain done */
335 #define REQ_F_LINK 64 /* linked sqes */
336 #define REQ_F_LINK_DONE 128 /* linked sqes done */
337 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
342 struct work_struct work;
345 #define IO_PLUG_THRESHOLD 2
346 #define IO_IOPOLL_BATCH 8
348 struct io_submit_state {
349 struct blk_plug plug;
352 * io_kiocb alloc cache
354 void *reqs[IO_IOPOLL_BATCH];
355 unsigned int free_reqs;
356 unsigned int cur_req;
359 * File reference cache
363 unsigned int has_refs;
364 unsigned int used_refs;
365 unsigned int ios_left;
368 static void io_sq_wq_submit_work(struct work_struct *work);
370 static struct kmem_cache *req_cachep;
372 static const struct file_operations io_uring_fops;
374 struct sock *io_uring_get_socket(struct file *file)
376 #if defined(CONFIG_UNIX)
377 if (file->f_op == &io_uring_fops) {
378 struct io_ring_ctx *ctx = file->private_data;
380 return ctx->ring_sock->sk;
385 EXPORT_SYMBOL(io_uring_get_socket);
387 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
389 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
391 complete(&ctx->ctx_done);
394 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
396 struct io_ring_ctx *ctx;
399 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
403 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
404 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
409 ctx->flags = p->flags;
410 init_waitqueue_head(&ctx->cq_wait);
411 init_completion(&ctx->ctx_done);
412 init_completion(&ctx->sqo_thread_started);
413 mutex_init(&ctx->uring_lock);
414 init_waitqueue_head(&ctx->wait);
415 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
416 spin_lock_init(&ctx->pending_async[i].lock);
417 INIT_LIST_HEAD(&ctx->pending_async[i].list);
418 atomic_set(&ctx->pending_async[i].cnt, 0);
420 spin_lock_init(&ctx->completion_lock);
421 INIT_LIST_HEAD(&ctx->poll_list);
422 INIT_LIST_HEAD(&ctx->cancel_list);
423 INIT_LIST_HEAD(&ctx->defer_list);
427 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
428 struct io_kiocb *req)
430 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
433 return req->sequence != ctx->cached_cq_tail + ctx->sq_ring->dropped;
436 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
438 struct io_kiocb *req;
440 if (list_empty(&ctx->defer_list))
443 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
444 if (!io_sequence_defer(ctx, req)) {
445 list_del_init(&req->list);
452 static void __io_commit_cqring(struct io_ring_ctx *ctx)
454 struct io_cq_ring *ring = ctx->cq_ring;
456 if (ctx->cached_cq_tail != READ_ONCE(ring->r.tail)) {
457 /* order cqe stores with ring update */
458 smp_store_release(&ring->r.tail, ctx->cached_cq_tail);
460 if (wq_has_sleeper(&ctx->cq_wait)) {
461 wake_up_interruptible(&ctx->cq_wait);
462 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
467 static void io_commit_cqring(struct io_ring_ctx *ctx)
469 struct io_kiocb *req;
471 __io_commit_cqring(ctx);
473 while ((req = io_get_deferred_req(ctx)) != NULL) {
474 req->flags |= REQ_F_IO_DRAINED;
475 queue_work(ctx->sqo_wq, &req->work);
479 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
481 struct io_cq_ring *ring = ctx->cq_ring;
484 tail = ctx->cached_cq_tail;
486 * writes to the cq entry need to come after reading head; the
487 * control dependency is enough as we're using WRITE_ONCE to
490 if (tail - READ_ONCE(ring->r.head) == ring->ring_entries)
493 ctx->cached_cq_tail++;
494 return &ring->cqes[tail & ctx->cq_mask];
497 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
500 struct io_uring_cqe *cqe;
503 * If we can't get a cq entry, userspace overflowed the
504 * submission (by quite a lot). Increment the overflow count in
507 cqe = io_get_cqring(ctx);
509 WRITE_ONCE(cqe->user_data, ki_user_data);
510 WRITE_ONCE(cqe->res, res);
511 WRITE_ONCE(cqe->flags, 0);
513 unsigned overflow = READ_ONCE(ctx->cq_ring->overflow);
515 WRITE_ONCE(ctx->cq_ring->overflow, overflow + 1);
519 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
521 if (waitqueue_active(&ctx->wait))
523 if (waitqueue_active(&ctx->sqo_wait))
524 wake_up(&ctx->sqo_wait);
526 eventfd_signal(ctx->cq_ev_fd, 1);
529 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
534 spin_lock_irqsave(&ctx->completion_lock, flags);
535 io_cqring_fill_event(ctx, user_data, res);
536 io_commit_cqring(ctx);
537 spin_unlock_irqrestore(&ctx->completion_lock, flags);
539 io_cqring_ev_posted(ctx);
542 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
544 percpu_ref_put_many(&ctx->refs, refs);
546 if (waitqueue_active(&ctx->wait))
550 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
551 struct io_submit_state *state)
553 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
554 struct io_kiocb *req;
556 if (!percpu_ref_tryget(&ctx->refs))
560 req = kmem_cache_alloc(req_cachep, gfp);
563 } else if (!state->free_reqs) {
567 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
568 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
571 * Bulk alloc is all-or-nothing. If we fail to get a batch,
572 * retry single alloc to be on the safe side.
574 if (unlikely(ret <= 0)) {
575 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
580 state->free_reqs = ret - 1;
582 req = state->reqs[0];
584 req = state->reqs[state->cur_req];
592 /* one is dropped after submission, the other at completion */
593 refcount_set(&req->refs, 2);
597 io_ring_drop_ctx_refs(ctx, 1);
601 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
604 kmem_cache_free_bulk(req_cachep, *nr, reqs);
605 io_ring_drop_ctx_refs(ctx, *nr);
610 static void __io_free_req(struct io_kiocb *req)
612 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
614 io_ring_drop_ctx_refs(req->ctx, 1);
615 kmem_cache_free(req_cachep, req);
618 static void io_req_link_next(struct io_kiocb *req)
620 struct io_kiocb *nxt;
623 * The list should never be empty when we are called here. But could
624 * potentially happen if the chain is messed up, check to be on the
627 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
629 list_del(&nxt->list);
630 if (!list_empty(&req->link_list)) {
631 INIT_LIST_HEAD(&nxt->link_list);
632 list_splice(&req->link_list, &nxt->link_list);
633 nxt->flags |= REQ_F_LINK;
636 nxt->flags |= REQ_F_LINK_DONE;
637 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
638 queue_work(req->ctx->sqo_wq, &nxt->work);
643 * Called if REQ_F_LINK is set, and we fail the head request
645 static void io_fail_links(struct io_kiocb *req)
647 struct io_kiocb *link;
649 while (!list_empty(&req->link_list)) {
650 link = list_first_entry(&req->link_list, struct io_kiocb, list);
651 list_del(&link->list);
653 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
658 static void io_free_req(struct io_kiocb *req)
661 * If LINK is set, we have dependent requests in this chain. If we
662 * didn't fail this request, queue the first one up, moving any other
663 * dependencies to the next request. In case of failure, fail the rest
666 if (req->flags & REQ_F_LINK) {
667 if (req->flags & REQ_F_FAIL_LINK)
670 io_req_link_next(req);
676 static void io_put_req(struct io_kiocb *req)
678 if (refcount_dec_and_test(&req->refs))
682 static unsigned io_cqring_events(struct io_cq_ring *ring)
684 /* See comment at the top of this file */
686 return READ_ONCE(ring->r.tail) - READ_ONCE(ring->r.head);
690 * Find and free completed poll iocbs
692 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
693 struct list_head *done)
695 void *reqs[IO_IOPOLL_BATCH];
696 struct io_kiocb *req;
700 while (!list_empty(done)) {
701 req = list_first_entry(done, struct io_kiocb, list);
702 list_del(&req->list);
704 io_cqring_fill_event(ctx, req->user_data, req->result);
707 if (refcount_dec_and_test(&req->refs)) {
708 /* If we're not using fixed files, we have to pair the
709 * completion part with the file put. Use regular
710 * completions for those, only batch free for fixed
711 * file and non-linked commands.
713 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
715 reqs[to_free++] = req;
716 if (to_free == ARRAY_SIZE(reqs))
717 io_free_req_many(ctx, reqs, &to_free);
724 io_commit_cqring(ctx);
725 io_free_req_many(ctx, reqs, &to_free);
728 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
731 struct io_kiocb *req, *tmp;
737 * Only spin for completions if we don't have multiple devices hanging
738 * off our complete list, and we're under the requested amount.
740 spin = !ctx->poll_multi_file && *nr_events < min;
743 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
744 struct kiocb *kiocb = &req->rw;
747 * Move completed entries to our local list. If we find a
748 * request that requires polling, break out and complete
749 * the done list first, if we have entries there.
751 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
752 list_move_tail(&req->list, &done);
755 if (!list_empty(&done))
758 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
767 if (!list_empty(&done))
768 io_iopoll_complete(ctx, nr_events, &done);
774 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
775 * non-spinning poll check - we'll still enter the driver poll loop, but only
776 * as a non-spinning completion check.
778 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
781 while (!list_empty(&ctx->poll_list)) {
784 ret = io_do_iopoll(ctx, nr_events, min);
787 if (!min || *nr_events >= min)
795 * We can't just wait for polled events to come to us, we have to actively
796 * find and complete them.
798 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
800 if (!(ctx->flags & IORING_SETUP_IOPOLL))
803 mutex_lock(&ctx->uring_lock);
804 while (!list_empty(&ctx->poll_list)) {
805 unsigned int nr_events = 0;
807 io_iopoll_getevents(ctx, &nr_events, 1);
809 mutex_unlock(&ctx->uring_lock);
812 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
818 * We disallow the app entering submit/complete with polling, but we
819 * still need to lock the ring to prevent racing with polled issue
820 * that got punted to a workqueue.
822 mutex_lock(&ctx->uring_lock);
829 * Don't enter poll loop if we already have events pending.
830 * If we do, we can potentially be spinning for commands that
831 * already triggered a CQE (eg in error).
833 if (io_cqring_events(ctx->cq_ring))
837 * If a submit got punted to a workqueue, we can have the
838 * application entering polling for a command before it gets
839 * issued. That app will hold the uring_lock for the duration
840 * of the poll right here, so we need to take a breather every
841 * now and then to ensure that the issue has a chance to add
842 * the poll to the issued list. Otherwise we can spin here
843 * forever, while the workqueue is stuck trying to acquire the
846 if (!(++iters & 7)) {
847 mutex_unlock(&ctx->uring_lock);
848 mutex_lock(&ctx->uring_lock);
851 if (*nr_events < min)
852 tmin = min - *nr_events;
854 ret = io_iopoll_getevents(ctx, nr_events, tmin);
858 } while (min && !*nr_events && !need_resched());
860 mutex_unlock(&ctx->uring_lock);
864 static void kiocb_end_write(struct kiocb *kiocb)
866 if (kiocb->ki_flags & IOCB_WRITE) {
867 struct inode *inode = file_inode(kiocb->ki_filp);
870 * Tell lockdep we inherited freeze protection from submission
873 if (S_ISREG(inode->i_mode))
874 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
875 file_end_write(kiocb->ki_filp);
879 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
881 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
883 kiocb_end_write(kiocb);
885 if ((req->flags & REQ_F_LINK) && res != req->result)
886 req->flags |= REQ_F_FAIL_LINK;
887 io_cqring_add_event(req->ctx, req->user_data, res);
891 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
893 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
895 kiocb_end_write(kiocb);
897 if ((req->flags & REQ_F_LINK) && res != req->result)
898 req->flags |= REQ_F_FAIL_LINK;
901 req->flags |= REQ_F_IOPOLL_COMPLETED;
905 * After the iocb has been issued, it's safe to be found on the poll list.
906 * Adding the kiocb to the list AFTER submission ensures that we don't
907 * find it from a io_iopoll_getevents() thread before the issuer is done
908 * accessing the kiocb cookie.
910 static void io_iopoll_req_issued(struct io_kiocb *req)
912 struct io_ring_ctx *ctx = req->ctx;
915 * Track whether we have multiple files in our lists. This will impact
916 * how we do polling eventually, not spinning if we're on potentially
919 if (list_empty(&ctx->poll_list)) {
920 ctx->poll_multi_file = false;
921 } else if (!ctx->poll_multi_file) {
922 struct io_kiocb *list_req;
924 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
926 if (list_req->rw.ki_filp != req->rw.ki_filp)
927 ctx->poll_multi_file = true;
931 * For fast devices, IO may have already completed. If it has, add
932 * it to the front so we find it first.
934 if (req->flags & REQ_F_IOPOLL_COMPLETED)
935 list_add(&req->list, &ctx->poll_list);
937 list_add_tail(&req->list, &ctx->poll_list);
940 static void io_file_put(struct io_submit_state *state)
943 int diff = state->has_refs - state->used_refs;
946 fput_many(state->file, diff);
952 * Get as many references to a file as we have IOs left in this submission,
953 * assuming most submissions are for one file, or at least that each file
954 * has more than one submission.
956 static struct file *io_file_get(struct io_submit_state *state, int fd)
962 if (state->fd == fd) {
969 state->file = fget_many(fd, state->ios_left);
974 state->has_refs = state->ios_left;
975 state->used_refs = 1;
981 * If we tracked the file through the SCM inflight mechanism, we could support
982 * any file. For now, just ensure that anything potentially problematic is done
985 static bool io_file_supports_async(struct file *file)
987 umode_t mode = file_inode(file)->i_mode;
989 if (S_ISBLK(mode) || S_ISCHR(mode))
991 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
997 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
1000 const struct io_uring_sqe *sqe = s->sqe;
1001 struct io_ring_ctx *ctx = req->ctx;
1002 struct kiocb *kiocb = &req->rw;
1009 if (force_nonblock && !io_file_supports_async(req->file))
1010 force_nonblock = false;
1012 kiocb->ki_pos = READ_ONCE(sqe->off);
1013 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1014 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1016 ioprio = READ_ONCE(sqe->ioprio);
1018 ret = ioprio_check_cap(ioprio);
1022 kiocb->ki_ioprio = ioprio;
1024 kiocb->ki_ioprio = get_current_ioprio();
1026 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1030 /* don't allow async punt if RWF_NOWAIT was requested */
1031 if (kiocb->ki_flags & IOCB_NOWAIT)
1032 req->flags |= REQ_F_NOWAIT;
1035 kiocb->ki_flags |= IOCB_NOWAIT;
1037 if (ctx->flags & IORING_SETUP_IOPOLL) {
1038 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1039 !kiocb->ki_filp->f_op->iopoll)
1042 kiocb->ki_flags |= IOCB_HIPRI;
1043 kiocb->ki_complete = io_complete_rw_iopoll;
1045 if (kiocb->ki_flags & IOCB_HIPRI)
1047 kiocb->ki_complete = io_complete_rw;
1052 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1058 case -ERESTARTNOINTR:
1059 case -ERESTARTNOHAND:
1060 case -ERESTART_RESTARTBLOCK:
1062 * We can't just restart the syscall, since previously
1063 * submitted sqes may already be in progress. Just fail this
1069 kiocb->ki_complete(kiocb, ret, 0);
1073 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1074 const struct io_uring_sqe *sqe,
1075 struct iov_iter *iter)
1077 size_t len = READ_ONCE(sqe->len);
1078 struct io_mapped_ubuf *imu;
1079 unsigned index, buf_index;
1083 /* attempt to use fixed buffers without having provided iovecs */
1084 if (unlikely(!ctx->user_bufs))
1087 buf_index = READ_ONCE(sqe->buf_index);
1088 if (unlikely(buf_index >= ctx->nr_user_bufs))
1091 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1092 imu = &ctx->user_bufs[index];
1093 buf_addr = READ_ONCE(sqe->addr);
1096 if (buf_addr + len < buf_addr)
1098 /* not inside the mapped region */
1099 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1103 * May not be a start of buffer, set size appropriately
1104 * and advance us to the beginning.
1106 offset = buf_addr - imu->ubuf;
1107 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1111 * Don't use iov_iter_advance() here, as it's really slow for
1112 * using the latter parts of a big fixed buffer - it iterates
1113 * over each segment manually. We can cheat a bit here, because
1116 * 1) it's a BVEC iter, we set it up
1117 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1118 * first and last bvec
1120 * So just find our index, and adjust the iterator afterwards.
1121 * If the offset is within the first bvec (or the whole first
1122 * bvec, just use iov_iter_advance(). This makes it easier
1123 * since we can just skip the first segment, which may not
1124 * be PAGE_SIZE aligned.
1126 const struct bio_vec *bvec = imu->bvec;
1128 if (offset <= bvec->bv_len) {
1129 iov_iter_advance(iter, offset);
1131 unsigned long seg_skip;
1133 /* skip first vec */
1134 offset -= bvec->bv_len;
1135 seg_skip = 1 + (offset >> PAGE_SHIFT);
1137 iter->bvec = bvec + seg_skip;
1138 iter->nr_segs -= seg_skip;
1139 iter->count -= bvec->bv_len + offset;
1140 iter->iov_offset = offset & ~PAGE_MASK;
1147 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1148 const struct sqe_submit *s, struct iovec **iovec,
1149 struct iov_iter *iter)
1151 const struct io_uring_sqe *sqe = s->sqe;
1152 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1153 size_t sqe_len = READ_ONCE(sqe->len);
1157 * We're reading ->opcode for the second time, but the first read
1158 * doesn't care whether it's _FIXED or not, so it doesn't matter
1159 * whether ->opcode changes concurrently. The first read does care
1160 * about whether it is a READ or a WRITE, so we don't trust this read
1161 * for that purpose and instead let the caller pass in the read/write
1164 opcode = READ_ONCE(sqe->opcode);
1165 if (opcode == IORING_OP_READ_FIXED ||
1166 opcode == IORING_OP_WRITE_FIXED) {
1167 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1175 #ifdef CONFIG_COMPAT
1177 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1181 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1185 * Make a note of the last file/offset/direction we punted to async
1186 * context. We'll use this information to see if we can piggy back a
1187 * sequential request onto the previous one, if it's still hasn't been
1188 * completed by the async worker.
1190 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1192 struct async_list *async_list = &req->ctx->pending_async[rw];
1193 struct kiocb *kiocb = &req->rw;
1194 struct file *filp = kiocb->ki_filp;
1195 off_t io_end = kiocb->ki_pos + len;
1197 if (filp == async_list->file && kiocb->ki_pos == async_list->io_end) {
1198 unsigned long max_bytes;
1200 /* Use 8x RA size as a decent limiter for both reads/writes */
1201 max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3);
1203 max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3);
1205 /* If max len are exceeded, reset the state */
1206 if (async_list->io_len + len <= max_bytes) {
1207 req->flags |= REQ_F_SEQ_PREV;
1208 async_list->io_len += len;
1211 async_list->io_len = 0;
1215 /* New file? Reset state. */
1216 if (async_list->file != filp) {
1217 async_list->io_len = 0;
1218 async_list->file = filp;
1220 async_list->io_end = io_end;
1223 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1224 bool force_nonblock)
1226 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1227 struct kiocb *kiocb = &req->rw;
1228 struct iov_iter iter;
1231 ssize_t read_size, ret;
1233 ret = io_prep_rw(req, s, force_nonblock);
1236 file = kiocb->ki_filp;
1238 if (unlikely(!(file->f_mode & FMODE_READ)))
1240 if (unlikely(!file->f_op->read_iter))
1243 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1248 if (req->flags & REQ_F_LINK)
1249 req->result = read_size;
1251 iov_count = iov_iter_count(&iter);
1252 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1256 ret2 = call_read_iter(file, kiocb, &iter);
1258 * In case of a short read, punt to async. This can happen
1259 * if we have data partially cached. Alternatively we can
1260 * return the short read, in which case the application will
1261 * need to issue another SQE and wait for it. That SQE will
1262 * need async punt anyway, so it's more efficient to do it
1265 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1267 /* Catch -EAGAIN return for forced non-blocking submission */
1268 if (!force_nonblock || ret2 != -EAGAIN) {
1269 io_rw_done(kiocb, ret2);
1272 * If ->needs_lock is true, we're already in async
1276 io_async_list_note(READ, req, iov_count);
1284 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1285 bool force_nonblock)
1287 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1288 struct kiocb *kiocb = &req->rw;
1289 struct iov_iter iter;
1294 ret = io_prep_rw(req, s, force_nonblock);
1298 file = kiocb->ki_filp;
1299 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1301 if (unlikely(!file->f_op->write_iter))
1304 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1308 if (req->flags & REQ_F_LINK)
1311 iov_count = iov_iter_count(&iter);
1314 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1315 /* If ->needs_lock is true, we're already in async context. */
1317 io_async_list_note(WRITE, req, iov_count);
1321 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1326 * Open-code file_start_write here to grab freeze protection,
1327 * which will be released by another thread in
1328 * io_complete_rw(). Fool lockdep by telling it the lock got
1329 * released so that it doesn't complain about the held lock when
1330 * we return to userspace.
1332 if (S_ISREG(file_inode(file)->i_mode)) {
1333 __sb_start_write(file_inode(file)->i_sb,
1334 SB_FREEZE_WRITE, true);
1335 __sb_writers_release(file_inode(file)->i_sb,
1338 kiocb->ki_flags |= IOCB_WRITE;
1340 ret2 = call_write_iter(file, kiocb, &iter);
1341 if (!force_nonblock || ret2 != -EAGAIN) {
1342 io_rw_done(kiocb, ret2);
1345 * If ->needs_lock is true, we're already in async
1349 io_async_list_note(WRITE, req, iov_count);
1359 * IORING_OP_NOP just posts a completion event, nothing else.
1361 static int io_nop(struct io_kiocb *req, u64 user_data)
1363 struct io_ring_ctx *ctx = req->ctx;
1366 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1369 io_cqring_add_event(ctx, user_data, err);
1374 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1376 struct io_ring_ctx *ctx = req->ctx;
1381 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1383 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1389 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1390 bool force_nonblock)
1392 loff_t sqe_off = READ_ONCE(sqe->off);
1393 loff_t sqe_len = READ_ONCE(sqe->len);
1394 loff_t end = sqe_off + sqe_len;
1395 unsigned fsync_flags;
1398 fsync_flags = READ_ONCE(sqe->fsync_flags);
1399 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1402 ret = io_prep_fsync(req, sqe);
1406 /* fsync always requires a blocking context */
1410 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1411 end > 0 ? end : LLONG_MAX,
1412 fsync_flags & IORING_FSYNC_DATASYNC);
1414 if (ret < 0 && (req->flags & REQ_F_LINK))
1415 req->flags |= REQ_F_FAIL_LINK;
1416 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1421 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1423 struct io_ring_ctx *ctx = req->ctx;
1429 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1431 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1437 static int io_sync_file_range(struct io_kiocb *req,
1438 const struct io_uring_sqe *sqe,
1439 bool force_nonblock)
1446 ret = io_prep_sfr(req, sqe);
1450 /* sync_file_range always requires a blocking context */
1454 sqe_off = READ_ONCE(sqe->off);
1455 sqe_len = READ_ONCE(sqe->len);
1456 flags = READ_ONCE(sqe->sync_range_flags);
1458 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1460 if (ret < 0 && (req->flags & REQ_F_LINK))
1461 req->flags |= REQ_F_FAIL_LINK;
1462 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1467 #if defined(CONFIG_NET)
1468 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1469 bool force_nonblock,
1470 long (*fn)(struct socket *, struct user_msghdr __user *,
1473 struct socket *sock;
1476 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1479 sock = sock_from_file(req->file, &ret);
1481 struct user_msghdr __user *msg;
1484 flags = READ_ONCE(sqe->msg_flags);
1485 if (flags & MSG_DONTWAIT)
1486 req->flags |= REQ_F_NOWAIT;
1487 else if (force_nonblock)
1488 flags |= MSG_DONTWAIT;
1490 msg = (struct user_msghdr __user *) (unsigned long)
1491 READ_ONCE(sqe->addr);
1493 ret = fn(sock, msg, flags);
1494 if (force_nonblock && ret == -EAGAIN)
1498 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1504 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1505 bool force_nonblock)
1507 #if defined(CONFIG_NET)
1508 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1514 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1515 bool force_nonblock)
1517 #if defined(CONFIG_NET)
1518 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1524 static void io_poll_remove_one(struct io_kiocb *req)
1526 struct io_poll_iocb *poll = &req->poll;
1528 spin_lock(&poll->head->lock);
1529 WRITE_ONCE(poll->canceled, true);
1530 if (!list_empty(&poll->wait.entry)) {
1531 list_del_init(&poll->wait.entry);
1532 queue_work(req->ctx->sqo_wq, &req->work);
1534 spin_unlock(&poll->head->lock);
1536 list_del_init(&req->list);
1539 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1541 struct io_kiocb *req;
1543 spin_lock_irq(&ctx->completion_lock);
1544 while (!list_empty(&ctx->cancel_list)) {
1545 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1546 io_poll_remove_one(req);
1548 spin_unlock_irq(&ctx->completion_lock);
1552 * Find a running poll command that matches one specified in sqe->addr,
1553 * and remove it if found.
1555 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1557 struct io_ring_ctx *ctx = req->ctx;
1558 struct io_kiocb *poll_req, *next;
1561 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1563 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1567 spin_lock_irq(&ctx->completion_lock);
1568 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1569 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1570 io_poll_remove_one(poll_req);
1575 spin_unlock_irq(&ctx->completion_lock);
1577 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1582 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1585 req->poll.done = true;
1586 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1587 io_commit_cqring(ctx);
1590 static void io_poll_complete_work(struct work_struct *work)
1592 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1593 struct io_poll_iocb *poll = &req->poll;
1594 struct poll_table_struct pt = { ._key = poll->events };
1595 struct io_ring_ctx *ctx = req->ctx;
1598 if (!READ_ONCE(poll->canceled))
1599 mask = vfs_poll(poll->file, &pt) & poll->events;
1602 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1603 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1604 * synchronize with them. In the cancellation case the list_del_init
1605 * itself is not actually needed, but harmless so we keep it in to
1606 * avoid further branches in the fast path.
1608 spin_lock_irq(&ctx->completion_lock);
1609 if (!mask && !READ_ONCE(poll->canceled)) {
1610 add_wait_queue(poll->head, &poll->wait);
1611 spin_unlock_irq(&ctx->completion_lock);
1614 list_del_init(&req->list);
1615 io_poll_complete(ctx, req, mask);
1616 spin_unlock_irq(&ctx->completion_lock);
1618 io_cqring_ev_posted(ctx);
1622 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1625 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1627 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1628 struct io_ring_ctx *ctx = req->ctx;
1629 __poll_t mask = key_to_poll(key);
1630 unsigned long flags;
1632 /* for instances that support it check for an event match first: */
1633 if (mask && !(mask & poll->events))
1636 list_del_init(&poll->wait.entry);
1638 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1639 list_del(&req->list);
1640 io_poll_complete(ctx, req, mask);
1641 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1643 io_cqring_ev_posted(ctx);
1646 queue_work(ctx->sqo_wq, &req->work);
1652 struct io_poll_table {
1653 struct poll_table_struct pt;
1654 struct io_kiocb *req;
1658 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1659 struct poll_table_struct *p)
1661 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1663 if (unlikely(pt->req->poll.head)) {
1664 pt->error = -EINVAL;
1669 pt->req->poll.head = head;
1670 add_wait_queue(head, &pt->req->poll.wait);
1673 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1675 struct io_poll_iocb *poll = &req->poll;
1676 struct io_ring_ctx *ctx = req->ctx;
1677 struct io_poll_table ipt;
1678 bool cancel = false;
1682 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1684 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1689 INIT_WORK(&req->work, io_poll_complete_work);
1690 events = READ_ONCE(sqe->poll_events);
1691 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1695 poll->canceled = false;
1697 ipt.pt._qproc = io_poll_queue_proc;
1698 ipt.pt._key = poll->events;
1700 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1702 /* initialized the list so that we can do list_empty checks */
1703 INIT_LIST_HEAD(&poll->wait.entry);
1704 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1706 INIT_LIST_HEAD(&req->list);
1708 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1710 spin_lock_irq(&ctx->completion_lock);
1711 if (likely(poll->head)) {
1712 spin_lock(&poll->head->lock);
1713 if (unlikely(list_empty(&poll->wait.entry))) {
1719 if (mask || ipt.error)
1720 list_del_init(&poll->wait.entry);
1722 WRITE_ONCE(poll->canceled, true);
1723 else if (!poll->done) /* actually waiting for an event */
1724 list_add_tail(&req->list, &ctx->cancel_list);
1725 spin_unlock(&poll->head->lock);
1727 if (mask) { /* no async, we'd stolen it */
1729 io_poll_complete(ctx, req, mask);
1731 spin_unlock_irq(&ctx->completion_lock);
1734 io_cqring_ev_posted(ctx);
1740 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1741 const struct io_uring_sqe *sqe)
1743 struct io_uring_sqe *sqe_copy;
1745 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1748 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1752 spin_lock_irq(&ctx->completion_lock);
1753 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1754 spin_unlock_irq(&ctx->completion_lock);
1759 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1760 req->submit.sqe = sqe_copy;
1762 INIT_WORK(&req->work, io_sq_wq_submit_work);
1763 list_add_tail(&req->list, &ctx->defer_list);
1764 spin_unlock_irq(&ctx->completion_lock);
1765 return -EIOCBQUEUED;
1768 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1769 const struct sqe_submit *s, bool force_nonblock)
1773 req->user_data = READ_ONCE(s->sqe->user_data);
1775 if (unlikely(s->index >= ctx->sq_entries))
1778 opcode = READ_ONCE(s->sqe->opcode);
1781 ret = io_nop(req, req->user_data);
1783 case IORING_OP_READV:
1784 if (unlikely(s->sqe->buf_index))
1786 ret = io_read(req, s, force_nonblock);
1788 case IORING_OP_WRITEV:
1789 if (unlikely(s->sqe->buf_index))
1791 ret = io_write(req, s, force_nonblock);
1793 case IORING_OP_READ_FIXED:
1794 ret = io_read(req, s, force_nonblock);
1796 case IORING_OP_WRITE_FIXED:
1797 ret = io_write(req, s, force_nonblock);
1799 case IORING_OP_FSYNC:
1800 ret = io_fsync(req, s->sqe, force_nonblock);
1802 case IORING_OP_POLL_ADD:
1803 ret = io_poll_add(req, s->sqe);
1805 case IORING_OP_POLL_REMOVE:
1806 ret = io_poll_remove(req, s->sqe);
1808 case IORING_OP_SYNC_FILE_RANGE:
1809 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1811 case IORING_OP_SENDMSG:
1812 ret = io_sendmsg(req, s->sqe, force_nonblock);
1814 case IORING_OP_RECVMSG:
1815 ret = io_recvmsg(req, s->sqe, force_nonblock);
1825 if (ctx->flags & IORING_SETUP_IOPOLL) {
1826 if (req->result == -EAGAIN)
1829 /* workqueue context doesn't hold uring_lock, grab it now */
1831 mutex_lock(&ctx->uring_lock);
1832 io_iopoll_req_issued(req);
1834 mutex_unlock(&ctx->uring_lock);
1840 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1841 const struct io_uring_sqe *sqe)
1843 switch (sqe->opcode) {
1844 case IORING_OP_READV:
1845 case IORING_OP_READ_FIXED:
1846 return &ctx->pending_async[READ];
1847 case IORING_OP_WRITEV:
1848 case IORING_OP_WRITE_FIXED:
1849 return &ctx->pending_async[WRITE];
1855 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1857 u8 opcode = READ_ONCE(sqe->opcode);
1859 return !(opcode == IORING_OP_READ_FIXED ||
1860 opcode == IORING_OP_WRITE_FIXED);
1863 static void io_sq_wq_submit_work(struct work_struct *work)
1865 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1866 struct io_ring_ctx *ctx = req->ctx;
1867 struct mm_struct *cur_mm = NULL;
1868 struct async_list *async_list;
1869 LIST_HEAD(req_list);
1870 mm_segment_t old_fs;
1873 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1876 struct sqe_submit *s = &req->submit;
1877 const struct io_uring_sqe *sqe = s->sqe;
1878 unsigned int flags = req->flags;
1880 /* Ensure we clear previously set non-block flag */
1881 req->rw.ki_flags &= ~IOCB_NOWAIT;
1884 if (io_sqe_needs_user(sqe) && !cur_mm) {
1885 if (!mmget_not_zero(ctx->sqo_mm)) {
1888 cur_mm = ctx->sqo_mm;
1896 s->has_user = cur_mm != NULL;
1897 s->needs_lock = true;
1899 ret = __io_submit_sqe(ctx, req, s, false);
1901 * We can get EAGAIN for polled IO even though
1902 * we're forcing a sync submission from here,
1903 * since we can't wait for request slots on the
1912 /* drop submission reference */
1916 io_cqring_add_event(ctx, sqe->user_data, ret);
1920 /* async context always use a copy of the sqe */
1923 /* req from defer and link list needn't decrease async cnt */
1924 if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE))
1929 if (!list_empty(&req_list)) {
1930 req = list_first_entry(&req_list, struct io_kiocb,
1932 list_del(&req->list);
1935 if (list_empty(&async_list->list))
1939 spin_lock(&async_list->lock);
1940 if (list_empty(&async_list->list)) {
1941 spin_unlock(&async_list->lock);
1944 list_splice_init(&async_list->list, &req_list);
1945 spin_unlock(&async_list->lock);
1947 req = list_first_entry(&req_list, struct io_kiocb, list);
1948 list_del(&req->list);
1952 * Rare case of racing with a submitter. If we find the count has
1953 * dropped to zero AND we have pending work items, then restart
1954 * the processing. This is a tiny race window.
1957 ret = atomic_dec_return(&async_list->cnt);
1958 while (!ret && !list_empty(&async_list->list)) {
1959 spin_lock(&async_list->lock);
1960 atomic_inc(&async_list->cnt);
1961 list_splice_init(&async_list->list, &req_list);
1962 spin_unlock(&async_list->lock);
1964 if (!list_empty(&req_list)) {
1965 req = list_first_entry(&req_list,
1966 struct io_kiocb, list);
1967 list_del(&req->list);
1970 ret = atomic_dec_return(&async_list->cnt);
1983 * See if we can piggy back onto previously submitted work, that is still
1984 * running. We currently only allow this if the new request is sequential
1985 * to the previous one we punted.
1987 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
1993 if (!(req->flags & REQ_F_SEQ_PREV))
1995 if (!atomic_read(&list->cnt))
1999 spin_lock(&list->lock);
2000 list_add_tail(&req->list, &list->list);
2002 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
2005 if (!atomic_read(&list->cnt)) {
2006 list_del_init(&req->list);
2009 spin_unlock(&list->lock);
2013 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2015 int op = READ_ONCE(sqe->opcode);
2019 case IORING_OP_POLL_REMOVE:
2026 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2027 struct io_submit_state *state, struct io_kiocb *req)
2032 flags = READ_ONCE(s->sqe->flags);
2033 fd = READ_ONCE(s->sqe->fd);
2035 if (flags & IOSQE_IO_DRAIN) {
2036 req->flags |= REQ_F_IO_DRAIN;
2037 req->sequence = ctx->cached_sq_head - 1;
2040 if (!io_op_needs_file(s->sqe))
2043 if (flags & IOSQE_FIXED_FILE) {
2044 if (unlikely(!ctx->user_files ||
2045 (unsigned) fd >= ctx->nr_user_files))
2047 req->file = ctx->user_files[fd];
2048 req->flags |= REQ_F_FIXED_FILE;
2050 if (s->needs_fixed_file)
2052 req->file = io_file_get(state, fd);
2053 if (unlikely(!req->file))
2060 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2061 struct sqe_submit *s)
2065 ret = io_req_defer(ctx, req, s->sqe);
2067 if (ret != -EIOCBQUEUED) {
2069 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2074 ret = __io_submit_sqe(ctx, req, s, true);
2075 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
2076 struct io_uring_sqe *sqe_copy;
2078 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
2080 struct async_list *list;
2082 memcpy(sqe_copy, s->sqe, sizeof(*sqe_copy));
2085 memcpy(&req->submit, s, sizeof(*s));
2086 list = io_async_list_from_sqe(ctx, s->sqe);
2087 if (!io_add_to_prev_work(list, req)) {
2089 atomic_inc(&list->cnt);
2090 INIT_WORK(&req->work, io_sq_wq_submit_work);
2091 queue_work(ctx->sqo_wq, &req->work);
2095 * Queued up for async execution, worker will release
2096 * submit reference when the iocb is actually submitted.
2102 /* drop submission reference */
2105 /* and drop final reference, if we failed */
2107 io_cqring_add_event(ctx, req->user_data, ret);
2108 if (req->flags & REQ_F_LINK)
2109 req->flags |= REQ_F_FAIL_LINK;
2116 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2118 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2119 struct io_submit_state *state, struct io_kiocb **link)
2121 struct io_uring_sqe *sqe_copy;
2122 struct io_kiocb *req;
2125 /* enforce forwards compatibility on users */
2126 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2131 req = io_get_req(ctx, state);
2132 if (unlikely(!req)) {
2137 ret = io_req_set_file(ctx, s, state, req);
2138 if (unlikely(ret)) {
2142 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2147 * If we already have a head request, queue this one for async
2148 * submittal once the head completes. If we don't have a head but
2149 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2150 * submitted sync once the chain is complete. If none of those
2151 * conditions are true (normal request), then just queue it.
2154 struct io_kiocb *prev = *link;
2156 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2163 memcpy(&req->submit, s, sizeof(*s));
2164 list_add_tail(&req->list, &prev->link_list);
2165 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2166 req->flags |= REQ_F_LINK;
2168 memcpy(&req->submit, s, sizeof(*s));
2169 INIT_LIST_HEAD(&req->link_list);
2172 io_queue_sqe(ctx, req, s);
2177 * Batched submission is done, ensure local IO is flushed out.
2179 static void io_submit_state_end(struct io_submit_state *state)
2181 blk_finish_plug(&state->plug);
2183 if (state->free_reqs)
2184 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2185 &state->reqs[state->cur_req]);
2189 * Start submission side cache.
2191 static void io_submit_state_start(struct io_submit_state *state,
2192 struct io_ring_ctx *ctx, unsigned max_ios)
2194 blk_start_plug(&state->plug);
2195 state->free_reqs = 0;
2197 state->ios_left = max_ios;
2200 static void io_commit_sqring(struct io_ring_ctx *ctx)
2202 struct io_sq_ring *ring = ctx->sq_ring;
2204 if (ctx->cached_sq_head != READ_ONCE(ring->r.head)) {
2206 * Ensure any loads from the SQEs are done at this point,
2207 * since once we write the new head, the application could
2208 * write new data to them.
2210 smp_store_release(&ring->r.head, ctx->cached_sq_head);
2215 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2216 * that is mapped by userspace. This means that care needs to be taken to
2217 * ensure that reads are stable, as we cannot rely on userspace always
2218 * being a good citizen. If members of the sqe are validated and then later
2219 * used, it's important that those reads are done through READ_ONCE() to
2220 * prevent a re-load down the line.
2222 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2224 struct io_sq_ring *ring = ctx->sq_ring;
2228 * The cached sq head (or cq tail) serves two purposes:
2230 * 1) allows us to batch the cost of updating the user visible
2232 * 2) allows the kernel side to track the head on its own, even
2233 * though the application is the one updating it.
2235 head = ctx->cached_sq_head;
2236 /* make sure SQ entry isn't read before tail */
2237 if (head == smp_load_acquire(&ring->r.tail))
2240 head = READ_ONCE(ring->array[head & ctx->sq_mask]);
2241 if (head < ctx->sq_entries) {
2243 s->sqe = &ctx->sq_sqes[head];
2244 ctx->cached_sq_head++;
2248 /* drop invalid entries */
2249 ctx->cached_sq_head++;
2254 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2255 unsigned int nr, bool has_user, bool mm_fault)
2257 struct io_submit_state state, *statep = NULL;
2258 struct io_kiocb *link = NULL;
2259 bool prev_was_link = false;
2260 int i, submitted = 0;
2262 if (nr > IO_PLUG_THRESHOLD) {
2263 io_submit_state_start(&state, ctx, nr);
2267 for (i = 0; i < nr; i++) {
2269 * If previous wasn't linked and we have a linked command,
2270 * that's the end of the chain. Submit the previous link.
2272 if (!prev_was_link && link) {
2273 io_queue_sqe(ctx, link, &link->submit);
2276 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2278 if (unlikely(mm_fault)) {
2279 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2282 sqes[i].has_user = has_user;
2283 sqes[i].needs_lock = true;
2284 sqes[i].needs_fixed_file = true;
2285 io_submit_sqe(ctx, &sqes[i], statep, &link);
2291 io_queue_sqe(ctx, link, &link->submit);
2293 io_submit_state_end(&state);
2298 static int io_sq_thread(void *data)
2300 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2301 struct io_ring_ctx *ctx = data;
2302 struct mm_struct *cur_mm = NULL;
2303 mm_segment_t old_fs;
2306 unsigned long timeout;
2308 complete(&ctx->sqo_thread_started);
2313 timeout = inflight = 0;
2314 while (!kthread_should_park()) {
2315 bool all_fixed, mm_fault = false;
2319 unsigned nr_events = 0;
2321 if (ctx->flags & IORING_SETUP_IOPOLL) {
2322 io_iopoll_check(ctx, &nr_events, 0);
2325 * Normal IO, just pretend everything completed.
2326 * We don't have to poll completions for that.
2328 nr_events = inflight;
2331 inflight -= nr_events;
2333 timeout = jiffies + ctx->sq_thread_idle;
2336 if (!io_get_sqring(ctx, &sqes[0])) {
2338 * We're polling. If we're within the defined idle
2339 * period, then let us spin without work before going
2342 if (inflight || !time_after(jiffies, timeout)) {
2348 * Drop cur_mm before scheduling, we can't hold it for
2349 * long periods (or over schedule()). Do this before
2350 * adding ourselves to the waitqueue, as the unuse/drop
2359 prepare_to_wait(&ctx->sqo_wait, &wait,
2360 TASK_INTERRUPTIBLE);
2362 /* Tell userspace we may need a wakeup call */
2363 ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
2364 /* make sure to read SQ tail after writing flags */
2367 if (!io_get_sqring(ctx, &sqes[0])) {
2368 if (kthread_should_park()) {
2369 finish_wait(&ctx->sqo_wait, &wait);
2372 if (signal_pending(current))
2373 flush_signals(current);
2375 finish_wait(&ctx->sqo_wait, &wait);
2377 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2380 finish_wait(&ctx->sqo_wait, &wait);
2382 ctx->sq_ring->flags &= ~IORING_SQ_NEED_WAKEUP;
2388 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2392 if (i == ARRAY_SIZE(sqes))
2394 } while (io_get_sqring(ctx, &sqes[i]));
2396 /* Unless all new commands are FIXED regions, grab mm */
2397 if (!all_fixed && !cur_mm) {
2398 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2400 use_mm(ctx->sqo_mm);
2401 cur_mm = ctx->sqo_mm;
2405 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2408 /* Commit SQ ring head once we've consumed all SQEs */
2409 io_commit_sqring(ctx);
2423 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
2425 struct io_submit_state state, *statep = NULL;
2426 struct io_kiocb *link = NULL;
2427 bool prev_was_link = false;
2430 if (to_submit > IO_PLUG_THRESHOLD) {
2431 io_submit_state_start(&state, ctx, to_submit);
2435 for (i = 0; i < to_submit; i++) {
2436 struct sqe_submit s;
2438 if (!io_get_sqring(ctx, &s))
2442 * If previous wasn't linked and we have a linked command,
2443 * that's the end of the chain. Submit the previous link.
2445 if (!prev_was_link && link) {
2446 io_queue_sqe(ctx, link, &link->submit);
2449 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2452 s.needs_lock = false;
2453 s.needs_fixed_file = false;
2455 io_submit_sqe(ctx, &s, statep, &link);
2457 io_commit_sqring(ctx);
2460 io_queue_sqe(ctx, link, &link->submit);
2462 io_submit_state_end(statep);
2468 * Wait until events become available, if we don't already have some. The
2469 * application must reap them itself, as they reside on the shared cq ring.
2471 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2472 const sigset_t __user *sig, size_t sigsz)
2474 struct io_cq_ring *ring = ctx->cq_ring;
2477 if (io_cqring_events(ring) >= min_events)
2481 #ifdef CONFIG_COMPAT
2482 if (in_compat_syscall())
2483 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2487 ret = set_user_sigmask(sig, sigsz);
2493 ret = wait_event_interruptible(ctx->wait, io_cqring_events(ring) >= min_events);
2494 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
2495 if (ret == -ERESTARTSYS)
2498 return READ_ONCE(ring->r.head) == READ_ONCE(ring->r.tail) ? ret : 0;
2501 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2503 #if defined(CONFIG_UNIX)
2504 if (ctx->ring_sock) {
2505 struct sock *sock = ctx->ring_sock->sk;
2506 struct sk_buff *skb;
2508 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2514 for (i = 0; i < ctx->nr_user_files; i++)
2515 fput(ctx->user_files[i]);
2519 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2521 if (!ctx->user_files)
2524 __io_sqe_files_unregister(ctx);
2525 kfree(ctx->user_files);
2526 ctx->user_files = NULL;
2527 ctx->nr_user_files = 0;
2531 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2533 if (ctx->sqo_thread) {
2534 wait_for_completion(&ctx->sqo_thread_started);
2536 * The park is a bit of a work-around, without it we get
2537 * warning spews on shutdown with SQPOLL set and affinity
2538 * set to a single CPU.
2540 kthread_park(ctx->sqo_thread);
2541 kthread_stop(ctx->sqo_thread);
2542 ctx->sqo_thread = NULL;
2546 static void io_finish_async(struct io_ring_ctx *ctx)
2548 io_sq_thread_stop(ctx);
2551 destroy_workqueue(ctx->sqo_wq);
2556 #if defined(CONFIG_UNIX)
2557 static void io_destruct_skb(struct sk_buff *skb)
2559 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2561 io_finish_async(ctx);
2562 unix_destruct_scm(skb);
2566 * Ensure the UNIX gc is aware of our file set, so we are certain that
2567 * the io_uring can be safely unregistered on process exit, even if we have
2568 * loops in the file referencing.
2570 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2572 struct sock *sk = ctx->ring_sock->sk;
2573 struct scm_fp_list *fpl;
2574 struct sk_buff *skb;
2577 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2578 unsigned long inflight = ctx->user->unix_inflight + nr;
2580 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2584 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2588 skb = alloc_skb(0, GFP_KERNEL);
2595 skb->destructor = io_destruct_skb;
2597 fpl->user = get_uid(ctx->user);
2598 for (i = 0; i < nr; i++) {
2599 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2600 unix_inflight(fpl->user, fpl->fp[i]);
2603 fpl->max = fpl->count = nr;
2604 UNIXCB(skb).fp = fpl;
2605 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2606 skb_queue_head(&sk->sk_receive_queue, skb);
2608 for (i = 0; i < nr; i++)
2615 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2616 * causes regular reference counting to break down. We rely on the UNIX
2617 * garbage collection to take care of this problem for us.
2619 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2621 unsigned left, total;
2625 left = ctx->nr_user_files;
2627 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2629 ret = __io_sqe_files_scm(ctx, this_files, total);
2633 total += this_files;
2639 while (total < ctx->nr_user_files) {
2640 fput(ctx->user_files[total]);
2647 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2653 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2656 __s32 __user *fds = (__s32 __user *) arg;
2660 if (ctx->user_files)
2664 if (nr_args > IORING_MAX_FIXED_FILES)
2667 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2668 if (!ctx->user_files)
2671 for (i = 0; i < nr_args; i++) {
2673 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2676 ctx->user_files[i] = fget(fd);
2679 if (!ctx->user_files[i])
2682 * Don't allow io_uring instances to be registered. If UNIX
2683 * isn't enabled, then this causes a reference cycle and this
2684 * instance can never get freed. If UNIX is enabled we'll
2685 * handle it just fine, but there's still no point in allowing
2686 * a ring fd as it doesn't support regular read/write anyway.
2688 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2689 fput(ctx->user_files[i]);
2692 ctx->nr_user_files++;
2697 for (i = 0; i < ctx->nr_user_files; i++)
2698 fput(ctx->user_files[i]);
2700 kfree(ctx->user_files);
2701 ctx->user_files = NULL;
2702 ctx->nr_user_files = 0;
2706 ret = io_sqe_files_scm(ctx);
2708 io_sqe_files_unregister(ctx);
2713 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2714 struct io_uring_params *p)
2718 init_waitqueue_head(&ctx->sqo_wait);
2719 mmgrab(current->mm);
2720 ctx->sqo_mm = current->mm;
2722 if (ctx->flags & IORING_SETUP_SQPOLL) {
2724 if (!capable(CAP_SYS_ADMIN))
2727 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2728 if (!ctx->sq_thread_idle)
2729 ctx->sq_thread_idle = HZ;
2731 if (p->flags & IORING_SETUP_SQ_AFF) {
2732 int cpu = p->sq_thread_cpu;
2735 if (cpu >= nr_cpu_ids)
2737 if (!cpu_online(cpu))
2740 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2744 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2747 if (IS_ERR(ctx->sqo_thread)) {
2748 ret = PTR_ERR(ctx->sqo_thread);
2749 ctx->sqo_thread = NULL;
2752 wake_up_process(ctx->sqo_thread);
2753 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2754 /* Can't have SQ_AFF without SQPOLL */
2759 /* Do QD, or 2 * CPUS, whatever is smallest */
2760 ctx->sqo_wq = alloc_workqueue("io_ring-wq", WQ_UNBOUND | WQ_FREEZABLE,
2761 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2769 io_sq_thread_stop(ctx);
2770 mmdrop(ctx->sqo_mm);
2775 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2777 atomic_long_sub(nr_pages, &user->locked_vm);
2780 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2782 unsigned long page_limit, cur_pages, new_pages;
2784 /* Don't allow more pages than we can safely lock */
2785 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2788 cur_pages = atomic_long_read(&user->locked_vm);
2789 new_pages = cur_pages + nr_pages;
2790 if (new_pages > page_limit)
2792 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2793 new_pages) != cur_pages);
2798 static void io_mem_free(void *ptr)
2805 page = virt_to_head_page(ptr);
2806 if (put_page_testzero(page))
2807 free_compound_page(page);
2810 static void *io_mem_alloc(size_t size)
2812 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2815 return (void *) __get_free_pages(gfp_flags, get_order(size));
2818 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2820 struct io_sq_ring *sq_ring;
2821 struct io_cq_ring *cq_ring;
2824 bytes = struct_size(sq_ring, array, sq_entries);
2825 bytes += array_size(sizeof(struct io_uring_sqe), sq_entries);
2826 bytes += struct_size(cq_ring, cqes, cq_entries);
2828 return (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
2831 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
2835 if (!ctx->user_bufs)
2838 for (i = 0; i < ctx->nr_user_bufs; i++) {
2839 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2841 for (j = 0; j < imu->nr_bvecs; j++)
2842 put_page(imu->bvec[j].bv_page);
2844 if (ctx->account_mem)
2845 io_unaccount_mem(ctx->user, imu->nr_bvecs);
2850 kfree(ctx->user_bufs);
2851 ctx->user_bufs = NULL;
2852 ctx->nr_user_bufs = 0;
2856 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
2857 void __user *arg, unsigned index)
2859 struct iovec __user *src;
2861 #ifdef CONFIG_COMPAT
2863 struct compat_iovec __user *ciovs;
2864 struct compat_iovec ciov;
2866 ciovs = (struct compat_iovec __user *) arg;
2867 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
2870 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
2871 dst->iov_len = ciov.iov_len;
2875 src = (struct iovec __user *) arg;
2876 if (copy_from_user(dst, &src[index], sizeof(*dst)))
2881 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
2884 struct vm_area_struct **vmas = NULL;
2885 struct page **pages = NULL;
2886 int i, j, got_pages = 0;
2891 if (!nr_args || nr_args > UIO_MAXIOV)
2894 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
2896 if (!ctx->user_bufs)
2899 for (i = 0; i < nr_args; i++) {
2900 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
2901 unsigned long off, start, end, ubuf;
2906 ret = io_copy_iov(ctx, &iov, arg, i);
2911 * Don't impose further limits on the size and buffer
2912 * constraints here, we'll -EINVAL later when IO is
2913 * submitted if they are wrong.
2916 if (!iov.iov_base || !iov.iov_len)
2919 /* arbitrary limit, but we need something */
2920 if (iov.iov_len > SZ_1G)
2923 ubuf = (unsigned long) iov.iov_base;
2924 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
2925 start = ubuf >> PAGE_SHIFT;
2926 nr_pages = end - start;
2928 if (ctx->account_mem) {
2929 ret = io_account_mem(ctx->user, nr_pages);
2935 if (!pages || nr_pages > got_pages) {
2938 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
2940 vmas = kvmalloc_array(nr_pages,
2941 sizeof(struct vm_area_struct *),
2943 if (!pages || !vmas) {
2945 if (ctx->account_mem)
2946 io_unaccount_mem(ctx->user, nr_pages);
2949 got_pages = nr_pages;
2952 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
2956 if (ctx->account_mem)
2957 io_unaccount_mem(ctx->user, nr_pages);
2962 down_read(¤t->mm->mmap_sem);
2963 pret = get_user_pages(ubuf, nr_pages,
2964 FOLL_WRITE | FOLL_LONGTERM,
2966 if (pret == nr_pages) {
2967 /* don't support file backed memory */
2968 for (j = 0; j < nr_pages; j++) {
2969 struct vm_area_struct *vma = vmas[j];
2972 !is_file_hugepages(vma->vm_file)) {
2978 ret = pret < 0 ? pret : -EFAULT;
2980 up_read(¤t->mm->mmap_sem);
2983 * if we did partial map, or found file backed vmas,
2984 * release any pages we did get
2987 for (j = 0; j < pret; j++)
2990 if (ctx->account_mem)
2991 io_unaccount_mem(ctx->user, nr_pages);
2996 off = ubuf & ~PAGE_MASK;
2998 for (j = 0; j < nr_pages; j++) {
3001 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3002 imu->bvec[j].bv_page = pages[j];
3003 imu->bvec[j].bv_len = vec_len;
3004 imu->bvec[j].bv_offset = off;
3008 /* store original address for later verification */
3010 imu->len = iov.iov_len;
3011 imu->nr_bvecs = nr_pages;
3013 ctx->nr_user_bufs++;
3021 io_sqe_buffer_unregister(ctx);
3025 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3027 __s32 __user *fds = arg;
3033 if (copy_from_user(&fd, fds, sizeof(*fds)))
3036 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3037 if (IS_ERR(ctx->cq_ev_fd)) {
3038 int ret = PTR_ERR(ctx->cq_ev_fd);
3039 ctx->cq_ev_fd = NULL;
3046 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3048 if (ctx->cq_ev_fd) {
3049 eventfd_ctx_put(ctx->cq_ev_fd);
3050 ctx->cq_ev_fd = NULL;
3057 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3059 io_finish_async(ctx);
3061 mmdrop(ctx->sqo_mm);
3063 io_iopoll_reap_events(ctx);
3064 io_sqe_buffer_unregister(ctx);
3065 io_sqe_files_unregister(ctx);
3066 io_eventfd_unregister(ctx);
3068 #if defined(CONFIG_UNIX)
3069 if (ctx->ring_sock) {
3070 ctx->ring_sock->file = NULL; /* so that iput() is called */
3071 sock_release(ctx->ring_sock);
3075 io_mem_free(ctx->sq_ring);
3076 io_mem_free(ctx->sq_sqes);
3077 io_mem_free(ctx->cq_ring);
3079 percpu_ref_exit(&ctx->refs);
3080 if (ctx->account_mem)
3081 io_unaccount_mem(ctx->user,
3082 ring_pages(ctx->sq_entries, ctx->cq_entries));
3083 free_uid(ctx->user);
3087 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3089 struct io_ring_ctx *ctx = file->private_data;
3092 poll_wait(file, &ctx->cq_wait, wait);
3094 * synchronizes with barrier from wq_has_sleeper call in
3098 if (READ_ONCE(ctx->sq_ring->r.tail) - ctx->cached_sq_head !=
3099 ctx->sq_ring->ring_entries)
3100 mask |= EPOLLOUT | EPOLLWRNORM;
3101 if (READ_ONCE(ctx->cq_ring->r.head) != ctx->cached_cq_tail)
3102 mask |= EPOLLIN | EPOLLRDNORM;
3107 static int io_uring_fasync(int fd, struct file *file, int on)
3109 struct io_ring_ctx *ctx = file->private_data;
3111 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3114 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3116 mutex_lock(&ctx->uring_lock);
3117 percpu_ref_kill(&ctx->refs);
3118 mutex_unlock(&ctx->uring_lock);
3120 io_poll_remove_all(ctx);
3121 io_iopoll_reap_events(ctx);
3122 wait_for_completion(&ctx->ctx_done);
3123 io_ring_ctx_free(ctx);
3126 static int io_uring_release(struct inode *inode, struct file *file)
3128 struct io_ring_ctx *ctx = file->private_data;
3130 file->private_data = NULL;
3131 io_ring_ctx_wait_and_kill(ctx);
3135 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3137 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3138 unsigned long sz = vma->vm_end - vma->vm_start;
3139 struct io_ring_ctx *ctx = file->private_data;
3145 case IORING_OFF_SQ_RING:
3148 case IORING_OFF_SQES:
3151 case IORING_OFF_CQ_RING:
3158 page = virt_to_head_page(ptr);
3159 if (sz > (PAGE_SIZE << compound_order(page)))
3162 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3163 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3166 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3167 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3170 struct io_ring_ctx *ctx;
3175 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3183 if (f.file->f_op != &io_uring_fops)
3187 ctx = f.file->private_data;
3188 if (!percpu_ref_tryget(&ctx->refs))
3192 * For SQ polling, the thread will do all submissions and completions.
3193 * Just return the requested submit count, and wake the thread if
3196 if (ctx->flags & IORING_SETUP_SQPOLL) {
3197 if (flags & IORING_ENTER_SQ_WAKEUP)
3198 wake_up(&ctx->sqo_wait);
3199 submitted = to_submit;
3205 to_submit = min(to_submit, ctx->sq_entries);
3207 mutex_lock(&ctx->uring_lock);
3208 submitted = io_ring_submit(ctx, to_submit);
3209 mutex_unlock(&ctx->uring_lock);
3211 if (flags & IORING_ENTER_GETEVENTS) {
3212 unsigned nr_events = 0;
3214 min_complete = min(min_complete, ctx->cq_entries);
3216 if (ctx->flags & IORING_SETUP_IOPOLL) {
3217 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3219 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3224 io_ring_drop_ctx_refs(ctx, 1);
3227 return submitted ? submitted : ret;
3230 static const struct file_operations io_uring_fops = {
3231 .release = io_uring_release,
3232 .mmap = io_uring_mmap,
3233 .poll = io_uring_poll,
3234 .fasync = io_uring_fasync,
3237 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3238 struct io_uring_params *p)
3240 struct io_sq_ring *sq_ring;
3241 struct io_cq_ring *cq_ring;
3244 sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
3248 ctx->sq_ring = sq_ring;
3249 sq_ring->ring_mask = p->sq_entries - 1;
3250 sq_ring->ring_entries = p->sq_entries;
3251 ctx->sq_mask = sq_ring->ring_mask;
3252 ctx->sq_entries = sq_ring->ring_entries;
3254 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3255 if (size == SIZE_MAX)
3258 ctx->sq_sqes = io_mem_alloc(size);
3262 cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
3266 ctx->cq_ring = cq_ring;
3267 cq_ring->ring_mask = p->cq_entries - 1;
3268 cq_ring->ring_entries = p->cq_entries;
3269 ctx->cq_mask = cq_ring->ring_mask;
3270 ctx->cq_entries = cq_ring->ring_entries;
3275 * Allocate an anonymous fd, this is what constitutes the application
3276 * visible backing of an io_uring instance. The application mmaps this
3277 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3278 * we have to tie this fd to a socket for file garbage collection purposes.
3280 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3285 #if defined(CONFIG_UNIX)
3286 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3292 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3296 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3297 O_RDWR | O_CLOEXEC);
3300 ret = PTR_ERR(file);
3304 #if defined(CONFIG_UNIX)
3305 ctx->ring_sock->file = file;
3306 ctx->ring_sock->sk->sk_user_data = ctx;
3308 fd_install(ret, file);
3311 #if defined(CONFIG_UNIX)
3312 sock_release(ctx->ring_sock);
3313 ctx->ring_sock = NULL;
3318 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3320 struct user_struct *user = NULL;
3321 struct io_ring_ctx *ctx;
3325 if (!entries || entries > IORING_MAX_ENTRIES)
3329 * Use twice as many entries for the CQ ring. It's possible for the
3330 * application to drive a higher depth than the size of the SQ ring,
3331 * since the sqes are only used at submission time. This allows for
3332 * some flexibility in overcommitting a bit.
3334 p->sq_entries = roundup_pow_of_two(entries);
3335 p->cq_entries = 2 * p->sq_entries;
3337 user = get_uid(current_user());
3338 account_mem = !capable(CAP_IPC_LOCK);
3341 ret = io_account_mem(user,
3342 ring_pages(p->sq_entries, p->cq_entries));
3349 ctx = io_ring_ctx_alloc(p);
3352 io_unaccount_mem(user, ring_pages(p->sq_entries,
3357 ctx->compat = in_compat_syscall();
3358 ctx->account_mem = account_mem;
3361 ret = io_allocate_scq_urings(ctx, p);
3365 ret = io_sq_offload_start(ctx, p);
3369 ret = io_uring_get_fd(ctx);
3373 memset(&p->sq_off, 0, sizeof(p->sq_off));
3374 p->sq_off.head = offsetof(struct io_sq_ring, r.head);
3375 p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
3376 p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
3377 p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
3378 p->sq_off.flags = offsetof(struct io_sq_ring, flags);
3379 p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
3380 p->sq_off.array = offsetof(struct io_sq_ring, array);
3382 memset(&p->cq_off, 0, sizeof(p->cq_off));
3383 p->cq_off.head = offsetof(struct io_cq_ring, r.head);
3384 p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
3385 p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
3386 p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
3387 p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
3388 p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
3391 io_ring_ctx_wait_and_kill(ctx);
3396 * Sets up an aio uring context, and returns the fd. Applications asks for a
3397 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3398 * params structure passed in.
3400 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3402 struct io_uring_params p;
3406 if (copy_from_user(&p, params, sizeof(p)))
3408 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3413 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3414 IORING_SETUP_SQ_AFF))
3417 ret = io_uring_create(entries, &p);
3421 if (copy_to_user(params, &p, sizeof(p)))
3427 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3428 struct io_uring_params __user *, params)
3430 return io_uring_setup(entries, params);
3433 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3434 void __user *arg, unsigned nr_args)
3435 __releases(ctx->uring_lock)
3436 __acquires(ctx->uring_lock)
3441 * We're inside the ring mutex, if the ref is already dying, then
3442 * someone else killed the ctx or is already going through
3443 * io_uring_register().
3445 if (percpu_ref_is_dying(&ctx->refs))
3448 percpu_ref_kill(&ctx->refs);
3451 * Drop uring mutex before waiting for references to exit. If another
3452 * thread is currently inside io_uring_enter() it might need to grab
3453 * the uring_lock to make progress. If we hold it here across the drain
3454 * wait, then we can deadlock. It's safe to drop the mutex here, since
3455 * no new references will come in after we've killed the percpu ref.
3457 mutex_unlock(&ctx->uring_lock);
3458 wait_for_completion(&ctx->ctx_done);
3459 mutex_lock(&ctx->uring_lock);
3462 case IORING_REGISTER_BUFFERS:
3463 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3465 case IORING_UNREGISTER_BUFFERS:
3469 ret = io_sqe_buffer_unregister(ctx);
3471 case IORING_REGISTER_FILES:
3472 ret = io_sqe_files_register(ctx, arg, nr_args);
3474 case IORING_UNREGISTER_FILES:
3478 ret = io_sqe_files_unregister(ctx);
3480 case IORING_REGISTER_EVENTFD:
3484 ret = io_eventfd_register(ctx, arg);
3486 case IORING_UNREGISTER_EVENTFD:
3490 ret = io_eventfd_unregister(ctx);
3497 /* bring the ctx back to life */
3498 reinit_completion(&ctx->ctx_done);
3499 percpu_ref_reinit(&ctx->refs);
3503 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3504 void __user *, arg, unsigned int, nr_args)
3506 struct io_ring_ctx *ctx;
3515 if (f.file->f_op != &io_uring_fops)
3518 ctx = f.file->private_data;
3520 mutex_lock(&ctx->uring_lock);
3521 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3522 mutex_unlock(&ctx->uring_lock);
3528 static int __init io_uring_init(void)
3530 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3533 __initcall(io_uring_init);