1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/workqueue.h>
60 #include <linux/kthread.h>
61 #include <linux/blkdev.h>
62 #include <linux/bvec.h>
63 #include <linux/net.h>
65 #include <net/af_unix.h>
67 #include <linux/anon_inodes.h>
68 #include <linux/sched/mm.h>
69 #include <linux/uaccess.h>
70 #include <linux/nospec.h>
71 #include <linux/sizes.h>
72 #include <linux/hugetlb.h>
74 #include <uapi/linux/io_uring.h>
78 #define IORING_MAX_ENTRIES 32768
79 #define IORING_MAX_FIXED_FILES 1024
82 u32 head ____cacheline_aligned_in_smp;
83 u32 tail ____cacheline_aligned_in_smp;
87 * This data is shared with the application through the mmap at offsets
88 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
90 * The offsets to the member fields are published through struct
91 * io_sqring_offsets when calling io_uring_setup.
95 * Head and tail offsets into the ring; the offsets need to be
96 * masked to get valid indices.
98 * The kernel controls head of the sq ring and the tail of the cq ring,
99 * and the application controls tail of the sq ring and the head of the
102 struct io_uring sq, cq;
104 * Bitmasks to apply to head and tail offsets (constant, equals
107 u32 sq_ring_mask, cq_ring_mask;
108 /* Ring sizes (constant, power of 2) */
109 u32 sq_ring_entries, cq_ring_entries;
111 * Number of invalid entries dropped by the kernel due to
112 * invalid index stored in array
114 * Written by the kernel, shouldn't be modified by the
115 * application (i.e. get number of "new events" by comparing to
118 * After a new SQ head value was read by the application this
119 * counter includes all submissions that were dropped reaching
120 * the new SQ head (and possibly more).
126 * Written by the kernel, shouldn't be modified by the
129 * The application needs a full memory barrier before checking
130 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
134 * Number of completion events lost because the queue was full;
135 * this should be avoided by the application by making sure
136 * there are not more requests pending thatn there is space in
137 * the completion queue.
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * As completion events come in out of order this counter is not
144 * ordered with any other data.
148 * Ring buffer of completion events.
150 * The kernel writes completion events fresh every time they are
151 * produced, so the application is allowed to modify pending
154 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
157 struct io_mapped_ubuf {
160 struct bio_vec *bvec;
161 unsigned int nr_bvecs;
167 struct list_head list;
176 struct percpu_ref refs;
177 } ____cacheline_aligned_in_smp;
185 * Ring buffer of indices into array of io_uring_sqe, which is
186 * mmapped by the application using the IORING_OFF_SQES offset.
188 * This indirection could e.g. be used to assign fixed
189 * io_uring_sqe entries to operations and only submit them to
190 * the queue when needed.
192 * The kernel modifies neither the indices array nor the entries
196 unsigned cached_sq_head;
199 unsigned sq_thread_idle;
200 struct io_uring_sqe *sq_sqes;
202 struct list_head defer_list;
203 } ____cacheline_aligned_in_smp;
206 struct workqueue_struct *sqo_wq[2];
207 struct task_struct *sqo_thread; /* if using sq thread polling */
208 struct mm_struct *sqo_mm;
209 wait_queue_head_t sqo_wait;
210 struct completion sqo_thread_started;
213 unsigned cached_cq_tail;
216 struct wait_queue_head cq_wait;
217 struct fasync_struct *cq_fasync;
218 struct eventfd_ctx *cq_ev_fd;
219 } ____cacheline_aligned_in_smp;
221 struct io_rings *rings;
224 * If used, fixed file set. Writers must ensure that ->refs is dead,
225 * readers must ensure that ->refs is alive as long as the file* is
226 * used. Only updated through io_uring_register(2).
228 struct file **user_files;
229 unsigned nr_user_files;
231 /* if used, fixed mapped user buffers */
232 unsigned nr_user_bufs;
233 struct io_mapped_ubuf *user_bufs;
235 struct user_struct *user;
237 struct completion ctx_done;
240 struct mutex uring_lock;
241 wait_queue_head_t wait;
242 } ____cacheline_aligned_in_smp;
245 spinlock_t completion_lock;
246 bool poll_multi_file;
248 * ->poll_list is protected by the ctx->uring_lock for
249 * io_uring instances that don't use IORING_SETUP_SQPOLL.
250 * For SQPOLL, only the single threaded io_sq_thread() will
251 * manipulate the list, hence no extra locking is needed there.
253 struct list_head poll_list;
254 struct list_head cancel_list;
255 } ____cacheline_aligned_in_smp;
257 struct async_list pending_async[2];
259 #if defined(CONFIG_UNIX)
260 struct socket *ring_sock;
265 const struct io_uring_sqe *sqe;
266 unsigned short index;
270 bool needs_fixed_file;
274 * First field must be the file pointer in all the
275 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
277 struct io_poll_iocb {
279 struct wait_queue_head *head;
283 struct wait_queue_entry wait;
287 * NOTE! Each of the iocb union members has the file pointer
288 * as the first entry in their struct definition. So you can
289 * access the file pointer through any of the sub-structs,
290 * or directly as just 'ki_filp' in this struct.
296 struct io_poll_iocb poll;
299 struct sqe_submit submit;
301 struct io_ring_ctx *ctx;
302 struct list_head list;
303 struct list_head link_list;
306 #define REQ_F_NOWAIT 1 /* must not punt to workers */
307 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
308 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
309 #define REQ_F_SEQ_PREV 8 /* sequential with previous */
310 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
311 #define REQ_F_IO_DRAINED 32 /* drain done */
312 #define REQ_F_LINK 64 /* linked sqes */
313 #define REQ_F_LINK_DONE 128 /* linked sqes done */
314 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
315 #define REQ_F_SHADOW_DRAIN 512 /* link-drain shadow req */
320 struct work_struct work;
323 #define IO_PLUG_THRESHOLD 2
324 #define IO_IOPOLL_BATCH 8
326 struct io_submit_state {
327 struct blk_plug plug;
330 * io_kiocb alloc cache
332 void *reqs[IO_IOPOLL_BATCH];
333 unsigned int free_reqs;
334 unsigned int cur_req;
337 * File reference cache
341 unsigned int has_refs;
342 unsigned int used_refs;
343 unsigned int ios_left;
346 static void io_sq_wq_submit_work(struct work_struct *work);
347 static void __io_free_req(struct io_kiocb *req);
349 static struct kmem_cache *req_cachep;
351 static const struct file_operations io_uring_fops;
353 struct sock *io_uring_get_socket(struct file *file)
355 #if defined(CONFIG_UNIX)
356 if (file->f_op == &io_uring_fops) {
357 struct io_ring_ctx *ctx = file->private_data;
359 return ctx->ring_sock->sk;
364 EXPORT_SYMBOL(io_uring_get_socket);
366 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
368 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
370 complete(&ctx->ctx_done);
373 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
375 struct io_ring_ctx *ctx;
378 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
382 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
383 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
388 ctx->flags = p->flags;
389 init_waitqueue_head(&ctx->cq_wait);
390 init_completion(&ctx->ctx_done);
391 init_completion(&ctx->sqo_thread_started);
392 mutex_init(&ctx->uring_lock);
393 init_waitqueue_head(&ctx->wait);
394 for (i = 0; i < ARRAY_SIZE(ctx->pending_async); i++) {
395 spin_lock_init(&ctx->pending_async[i].lock);
396 INIT_LIST_HEAD(&ctx->pending_async[i].list);
397 atomic_set(&ctx->pending_async[i].cnt, 0);
399 spin_lock_init(&ctx->completion_lock);
400 INIT_LIST_HEAD(&ctx->poll_list);
401 INIT_LIST_HEAD(&ctx->cancel_list);
402 INIT_LIST_HEAD(&ctx->defer_list);
406 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
407 struct io_kiocb *req)
409 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) != REQ_F_IO_DRAIN)
412 return req->sequence != ctx->cached_cq_tail + ctx->rings->sq_dropped;
415 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
417 struct io_kiocb *req;
419 if (list_empty(&ctx->defer_list))
422 req = list_first_entry(&ctx->defer_list, struct io_kiocb, list);
423 if (!io_sequence_defer(ctx, req)) {
424 list_del_init(&req->list);
431 static void __io_commit_cqring(struct io_ring_ctx *ctx)
433 struct io_rings *rings = ctx->rings;
435 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
436 /* order cqe stores with ring update */
437 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
439 if (wq_has_sleeper(&ctx->cq_wait)) {
440 wake_up_interruptible(&ctx->cq_wait);
441 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
446 static inline void io_queue_async_work(struct io_ring_ctx *ctx,
447 struct io_kiocb *req)
451 if (req->submit.sqe) {
452 switch (req->submit.sqe->opcode) {
453 case IORING_OP_WRITEV:
454 case IORING_OP_WRITE_FIXED:
455 rw = !(req->rw.ki_flags & IOCB_DIRECT);
460 queue_work(ctx->sqo_wq[rw], &req->work);
463 static void io_commit_cqring(struct io_ring_ctx *ctx)
465 struct io_kiocb *req;
467 __io_commit_cqring(ctx);
469 while ((req = io_get_deferred_req(ctx)) != NULL) {
470 if (req->flags & REQ_F_SHADOW_DRAIN) {
471 /* Just for drain, free it. */
475 req->flags |= REQ_F_IO_DRAINED;
476 io_queue_async_work(ctx, req);
480 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
482 struct io_rings *rings = ctx->rings;
485 tail = ctx->cached_cq_tail;
487 * writes to the cq entry need to come after reading head; the
488 * control dependency is enough as we're using WRITE_ONCE to
491 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
494 ctx->cached_cq_tail++;
495 return &rings->cqes[tail & ctx->cq_mask];
498 static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
501 struct io_uring_cqe *cqe;
504 * If we can't get a cq entry, userspace overflowed the
505 * submission (by quite a lot). Increment the overflow count in
508 cqe = io_get_cqring(ctx);
510 WRITE_ONCE(cqe->user_data, ki_user_data);
511 WRITE_ONCE(cqe->res, res);
512 WRITE_ONCE(cqe->flags, 0);
514 unsigned overflow = READ_ONCE(ctx->rings->cq_overflow);
516 WRITE_ONCE(ctx->rings->cq_overflow, overflow + 1);
520 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
522 if (waitqueue_active(&ctx->wait))
524 if (waitqueue_active(&ctx->sqo_wait))
525 wake_up(&ctx->sqo_wait);
527 eventfd_signal(ctx->cq_ev_fd, 1);
530 static void io_cqring_add_event(struct io_ring_ctx *ctx, u64 user_data,
535 spin_lock_irqsave(&ctx->completion_lock, flags);
536 io_cqring_fill_event(ctx, user_data, res);
537 io_commit_cqring(ctx);
538 spin_unlock_irqrestore(&ctx->completion_lock, flags);
540 io_cqring_ev_posted(ctx);
543 static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
545 percpu_ref_put_many(&ctx->refs, refs);
547 if (waitqueue_active(&ctx->wait))
551 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
552 struct io_submit_state *state)
554 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
555 struct io_kiocb *req;
557 if (!percpu_ref_tryget(&ctx->refs))
561 req = kmem_cache_alloc(req_cachep, gfp);
564 } else if (!state->free_reqs) {
568 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
569 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
572 * Bulk alloc is all-or-nothing. If we fail to get a batch,
573 * retry single alloc to be on the safe side.
575 if (unlikely(ret <= 0)) {
576 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
581 state->free_reqs = ret - 1;
583 req = state->reqs[0];
585 req = state->reqs[state->cur_req];
593 /* one is dropped after submission, the other at completion */
594 refcount_set(&req->refs, 2);
598 io_ring_drop_ctx_refs(ctx, 1);
602 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
605 kmem_cache_free_bulk(req_cachep, *nr, reqs);
606 io_ring_drop_ctx_refs(ctx, *nr);
611 static void __io_free_req(struct io_kiocb *req)
613 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
615 io_ring_drop_ctx_refs(req->ctx, 1);
616 kmem_cache_free(req_cachep, req);
619 static void io_req_link_next(struct io_kiocb *req)
621 struct io_kiocb *nxt;
624 * The list should never be empty when we are called here. But could
625 * potentially happen if the chain is messed up, check to be on the
628 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
630 list_del(&nxt->list);
631 if (!list_empty(&req->link_list)) {
632 INIT_LIST_HEAD(&nxt->link_list);
633 list_splice(&req->link_list, &nxt->link_list);
634 nxt->flags |= REQ_F_LINK;
637 nxt->flags |= REQ_F_LINK_DONE;
638 INIT_WORK(&nxt->work, io_sq_wq_submit_work);
639 io_queue_async_work(req->ctx, nxt);
644 * Called if REQ_F_LINK is set, and we fail the head request
646 static void io_fail_links(struct io_kiocb *req)
648 struct io_kiocb *link;
650 while (!list_empty(&req->link_list)) {
651 link = list_first_entry(&req->link_list, struct io_kiocb, list);
652 list_del(&link->list);
654 io_cqring_add_event(req->ctx, link->user_data, -ECANCELED);
659 static void io_free_req(struct io_kiocb *req)
662 * If LINK is set, we have dependent requests in this chain. If we
663 * didn't fail this request, queue the first one up, moving any other
664 * dependencies to the next request. In case of failure, fail the rest
667 if (req->flags & REQ_F_LINK) {
668 if (req->flags & REQ_F_FAIL_LINK)
671 io_req_link_next(req);
677 static void io_put_req(struct io_kiocb *req)
679 if (refcount_dec_and_test(&req->refs))
683 static unsigned io_cqring_events(struct io_rings *rings)
685 /* See comment at the top of this file */
687 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
691 * Find and free completed poll iocbs
693 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
694 struct list_head *done)
696 void *reqs[IO_IOPOLL_BATCH];
697 struct io_kiocb *req;
701 while (!list_empty(done)) {
702 req = list_first_entry(done, struct io_kiocb, list);
703 list_del(&req->list);
705 io_cqring_fill_event(ctx, req->user_data, req->result);
708 if (refcount_dec_and_test(&req->refs)) {
709 /* If we're not using fixed files, we have to pair the
710 * completion part with the file put. Use regular
711 * completions for those, only batch free for fixed
712 * file and non-linked commands.
714 if ((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
716 reqs[to_free++] = req;
717 if (to_free == ARRAY_SIZE(reqs))
718 io_free_req_many(ctx, reqs, &to_free);
725 io_commit_cqring(ctx);
726 io_free_req_many(ctx, reqs, &to_free);
729 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
732 struct io_kiocb *req, *tmp;
738 * Only spin for completions if we don't have multiple devices hanging
739 * off our complete list, and we're under the requested amount.
741 spin = !ctx->poll_multi_file && *nr_events < min;
744 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
745 struct kiocb *kiocb = &req->rw;
748 * Move completed entries to our local list. If we find a
749 * request that requires polling, break out and complete
750 * the done list first, if we have entries there.
752 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
753 list_move_tail(&req->list, &done);
756 if (!list_empty(&done))
759 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
768 if (!list_empty(&done))
769 io_iopoll_complete(ctx, nr_events, &done);
775 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
776 * non-spinning poll check - we'll still enter the driver poll loop, but only
777 * as a non-spinning completion check.
779 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
782 while (!list_empty(&ctx->poll_list) && !need_resched()) {
785 ret = io_do_iopoll(ctx, nr_events, min);
788 if (!min || *nr_events >= min)
796 * We can't just wait for polled events to come to us, we have to actively
797 * find and complete them.
799 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
801 if (!(ctx->flags & IORING_SETUP_IOPOLL))
804 mutex_lock(&ctx->uring_lock);
805 while (!list_empty(&ctx->poll_list)) {
806 unsigned int nr_events = 0;
808 io_iopoll_getevents(ctx, &nr_events, 1);
811 * Ensure we allow local-to-the-cpu processing to take place,
812 * in this case we need to ensure that we reap all events.
816 mutex_unlock(&ctx->uring_lock);
819 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
825 * We disallow the app entering submit/complete with polling, but we
826 * still need to lock the ring to prevent racing with polled issue
827 * that got punted to a workqueue.
829 mutex_lock(&ctx->uring_lock);
836 * Don't enter poll loop if we already have events pending.
837 * If we do, we can potentially be spinning for commands that
838 * already triggered a CQE (eg in error).
840 if (io_cqring_events(ctx->rings))
844 * If a submit got punted to a workqueue, we can have the
845 * application entering polling for a command before it gets
846 * issued. That app will hold the uring_lock for the duration
847 * of the poll right here, so we need to take a breather every
848 * now and then to ensure that the issue has a chance to add
849 * the poll to the issued list. Otherwise we can spin here
850 * forever, while the workqueue is stuck trying to acquire the
853 if (!(++iters & 7)) {
854 mutex_unlock(&ctx->uring_lock);
855 mutex_lock(&ctx->uring_lock);
858 if (*nr_events < min)
859 tmin = min - *nr_events;
861 ret = io_iopoll_getevents(ctx, nr_events, tmin);
865 } while (min && !*nr_events && !need_resched());
867 mutex_unlock(&ctx->uring_lock);
871 static void kiocb_end_write(struct kiocb *kiocb)
873 if (kiocb->ki_flags & IOCB_WRITE) {
874 struct inode *inode = file_inode(kiocb->ki_filp);
877 * Tell lockdep we inherited freeze protection from submission
880 if (S_ISREG(inode->i_mode))
881 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
882 file_end_write(kiocb->ki_filp);
886 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
888 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
890 kiocb_end_write(kiocb);
892 if ((req->flags & REQ_F_LINK) && res != req->result)
893 req->flags |= REQ_F_FAIL_LINK;
894 io_cqring_add_event(req->ctx, req->user_data, res);
898 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
900 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
902 kiocb_end_write(kiocb);
904 if ((req->flags & REQ_F_LINK) && res != req->result)
905 req->flags |= REQ_F_FAIL_LINK;
908 req->flags |= REQ_F_IOPOLL_COMPLETED;
912 * After the iocb has been issued, it's safe to be found on the poll list.
913 * Adding the kiocb to the list AFTER submission ensures that we don't
914 * find it from a io_iopoll_getevents() thread before the issuer is done
915 * accessing the kiocb cookie.
917 static void io_iopoll_req_issued(struct io_kiocb *req)
919 struct io_ring_ctx *ctx = req->ctx;
922 * Track whether we have multiple files in our lists. This will impact
923 * how we do polling eventually, not spinning if we're on potentially
926 if (list_empty(&ctx->poll_list)) {
927 ctx->poll_multi_file = false;
928 } else if (!ctx->poll_multi_file) {
929 struct io_kiocb *list_req;
931 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
933 if (list_req->rw.ki_filp != req->rw.ki_filp)
934 ctx->poll_multi_file = true;
938 * For fast devices, IO may have already completed. If it has, add
939 * it to the front so we find it first.
941 if (req->flags & REQ_F_IOPOLL_COMPLETED)
942 list_add(&req->list, &ctx->poll_list);
944 list_add_tail(&req->list, &ctx->poll_list);
947 static void io_file_put(struct io_submit_state *state)
950 int diff = state->has_refs - state->used_refs;
953 fput_many(state->file, diff);
959 * Get as many references to a file as we have IOs left in this submission,
960 * assuming most submissions are for one file, or at least that each file
961 * has more than one submission.
963 static struct file *io_file_get(struct io_submit_state *state, int fd)
969 if (state->fd == fd) {
976 state->file = fget_many(fd, state->ios_left);
981 state->has_refs = state->ios_left;
982 state->used_refs = 1;
988 * If we tracked the file through the SCM inflight mechanism, we could support
989 * any file. For now, just ensure that anything potentially problematic is done
992 static bool io_file_supports_async(struct file *file)
994 umode_t mode = file_inode(file)->i_mode;
996 if (S_ISBLK(mode) || S_ISCHR(mode))
998 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1004 static int io_prep_rw(struct io_kiocb *req, const struct sqe_submit *s,
1005 bool force_nonblock)
1007 const struct io_uring_sqe *sqe = s->sqe;
1008 struct io_ring_ctx *ctx = req->ctx;
1009 struct kiocb *kiocb = &req->rw;
1016 if (force_nonblock && !io_file_supports_async(req->file))
1017 force_nonblock = false;
1019 kiocb->ki_pos = READ_ONCE(sqe->off);
1020 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1021 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1023 ioprio = READ_ONCE(sqe->ioprio);
1025 ret = ioprio_check_cap(ioprio);
1029 kiocb->ki_ioprio = ioprio;
1031 kiocb->ki_ioprio = get_current_ioprio();
1033 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1037 /* don't allow async punt if RWF_NOWAIT was requested */
1038 if (kiocb->ki_flags & IOCB_NOWAIT)
1039 req->flags |= REQ_F_NOWAIT;
1042 kiocb->ki_flags |= IOCB_NOWAIT;
1044 if (ctx->flags & IORING_SETUP_IOPOLL) {
1045 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1046 !kiocb->ki_filp->f_op->iopoll)
1049 kiocb->ki_flags |= IOCB_HIPRI;
1050 kiocb->ki_complete = io_complete_rw_iopoll;
1052 if (kiocb->ki_flags & IOCB_HIPRI)
1054 kiocb->ki_complete = io_complete_rw;
1059 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1065 case -ERESTARTNOINTR:
1066 case -ERESTARTNOHAND:
1067 case -ERESTART_RESTARTBLOCK:
1069 * We can't just restart the syscall, since previously
1070 * submitted sqes may already be in progress. Just fail this
1076 kiocb->ki_complete(kiocb, ret, 0);
1080 static int io_import_fixed(struct io_ring_ctx *ctx, int rw,
1081 const struct io_uring_sqe *sqe,
1082 struct iov_iter *iter)
1084 size_t len = READ_ONCE(sqe->len);
1085 struct io_mapped_ubuf *imu;
1086 unsigned index, buf_index;
1090 /* attempt to use fixed buffers without having provided iovecs */
1091 if (unlikely(!ctx->user_bufs))
1094 buf_index = READ_ONCE(sqe->buf_index);
1095 if (unlikely(buf_index >= ctx->nr_user_bufs))
1098 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1099 imu = &ctx->user_bufs[index];
1100 buf_addr = READ_ONCE(sqe->addr);
1103 if (buf_addr + len < buf_addr)
1105 /* not inside the mapped region */
1106 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1110 * May not be a start of buffer, set size appropriately
1111 * and advance us to the beginning.
1113 offset = buf_addr - imu->ubuf;
1114 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1118 * Don't use iov_iter_advance() here, as it's really slow for
1119 * using the latter parts of a big fixed buffer - it iterates
1120 * over each segment manually. We can cheat a bit here, because
1123 * 1) it's a BVEC iter, we set it up
1124 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1125 * first and last bvec
1127 * So just find our index, and adjust the iterator afterwards.
1128 * If the offset is within the first bvec (or the whole first
1129 * bvec, just use iov_iter_advance(). This makes it easier
1130 * since we can just skip the first segment, which may not
1131 * be PAGE_SIZE aligned.
1133 const struct bio_vec *bvec = imu->bvec;
1135 if (offset <= bvec->bv_len) {
1136 iov_iter_advance(iter, offset);
1138 unsigned long seg_skip;
1140 /* skip first vec */
1141 offset -= bvec->bv_len;
1142 seg_skip = 1 + (offset >> PAGE_SHIFT);
1144 iter->bvec = bvec + seg_skip;
1145 iter->nr_segs -= seg_skip;
1146 iter->count -= bvec->bv_len + offset;
1147 iter->iov_offset = offset & ~PAGE_MASK;
1154 static ssize_t io_import_iovec(struct io_ring_ctx *ctx, int rw,
1155 const struct sqe_submit *s, struct iovec **iovec,
1156 struct iov_iter *iter)
1158 const struct io_uring_sqe *sqe = s->sqe;
1159 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1160 size_t sqe_len = READ_ONCE(sqe->len);
1164 * We're reading ->opcode for the second time, but the first read
1165 * doesn't care whether it's _FIXED or not, so it doesn't matter
1166 * whether ->opcode changes concurrently. The first read does care
1167 * about whether it is a READ or a WRITE, so we don't trust this read
1168 * for that purpose and instead let the caller pass in the read/write
1171 opcode = READ_ONCE(sqe->opcode);
1172 if (opcode == IORING_OP_READ_FIXED ||
1173 opcode == IORING_OP_WRITE_FIXED) {
1174 ssize_t ret = io_import_fixed(ctx, rw, sqe, iter);
1182 #ifdef CONFIG_COMPAT
1184 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1188 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1191 static inline bool io_should_merge(struct async_list *al, struct kiocb *kiocb)
1193 if (al->file == kiocb->ki_filp) {
1197 * Allow merging if we're anywhere in the range of the same
1198 * page. Generally this happens for sub-page reads or writes,
1199 * and it's beneficial to allow the first worker to bring the
1200 * page in and the piggy backed work can then work on the
1203 start = al->io_start & PAGE_MASK;
1204 end = (al->io_start + al->io_len + PAGE_SIZE - 1) & PAGE_MASK;
1205 if (kiocb->ki_pos >= start && kiocb->ki_pos <= end)
1214 * Make a note of the last file/offset/direction we punted to async
1215 * context. We'll use this information to see if we can piggy back a
1216 * sequential request onto the previous one, if it's still hasn't been
1217 * completed by the async worker.
1219 static void io_async_list_note(int rw, struct io_kiocb *req, size_t len)
1221 struct async_list *async_list = &req->ctx->pending_async[rw];
1222 struct kiocb *kiocb = &req->rw;
1223 struct file *filp = kiocb->ki_filp;
1225 if (io_should_merge(async_list, kiocb)) {
1226 unsigned long max_bytes;
1228 /* Use 8x RA size as a decent limiter for both reads/writes */
1229 max_bytes = filp->f_ra.ra_pages << (PAGE_SHIFT + 3);
1231 max_bytes = VM_READAHEAD_PAGES << (PAGE_SHIFT + 3);
1233 /* If max len are exceeded, reset the state */
1234 if (async_list->io_len + len <= max_bytes) {
1235 req->flags |= REQ_F_SEQ_PREV;
1236 async_list->io_len += len;
1238 async_list->file = NULL;
1242 /* New file? Reset state. */
1243 if (async_list->file != filp) {
1244 async_list->io_start = kiocb->ki_pos;
1245 async_list->io_len = len;
1246 async_list->file = filp;
1250 static int io_read(struct io_kiocb *req, const struct sqe_submit *s,
1251 bool force_nonblock)
1253 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1254 struct kiocb *kiocb = &req->rw;
1255 struct iov_iter iter;
1258 ssize_t read_size, ret;
1260 ret = io_prep_rw(req, s, force_nonblock);
1263 file = kiocb->ki_filp;
1265 if (unlikely(!(file->f_mode & FMODE_READ)))
1267 if (unlikely(!file->f_op->read_iter))
1270 ret = io_import_iovec(req->ctx, READ, s, &iovec, &iter);
1275 if (req->flags & REQ_F_LINK)
1276 req->result = read_size;
1278 iov_count = iov_iter_count(&iter);
1279 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1283 ret2 = call_read_iter(file, kiocb, &iter);
1285 * In case of a short read, punt to async. This can happen
1286 * if we have data partially cached. Alternatively we can
1287 * return the short read, in which case the application will
1288 * need to issue another SQE and wait for it. That SQE will
1289 * need async punt anyway, so it's more efficient to do it
1292 if (force_nonblock && ret2 > 0 && ret2 < read_size)
1294 /* Catch -EAGAIN return for forced non-blocking submission */
1295 if (!force_nonblock || ret2 != -EAGAIN) {
1296 io_rw_done(kiocb, ret2);
1299 * If ->needs_lock is true, we're already in async
1303 io_async_list_note(READ, req, iov_count);
1311 static int io_write(struct io_kiocb *req, const struct sqe_submit *s,
1312 bool force_nonblock)
1314 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1315 struct kiocb *kiocb = &req->rw;
1316 struct iov_iter iter;
1321 ret = io_prep_rw(req, s, force_nonblock);
1325 file = kiocb->ki_filp;
1326 if (unlikely(!(file->f_mode & FMODE_WRITE)))
1328 if (unlikely(!file->f_op->write_iter))
1331 ret = io_import_iovec(req->ctx, WRITE, s, &iovec, &iter);
1335 if (req->flags & REQ_F_LINK)
1338 iov_count = iov_iter_count(&iter);
1341 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT)) {
1342 /* If ->needs_lock is true, we're already in async context. */
1344 io_async_list_note(WRITE, req, iov_count);
1348 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1353 * Open-code file_start_write here to grab freeze protection,
1354 * which will be released by another thread in
1355 * io_complete_rw(). Fool lockdep by telling it the lock got
1356 * released so that it doesn't complain about the held lock when
1357 * we return to userspace.
1359 if (S_ISREG(file_inode(file)->i_mode)) {
1360 __sb_start_write(file_inode(file)->i_sb,
1361 SB_FREEZE_WRITE, true);
1362 __sb_writers_release(file_inode(file)->i_sb,
1365 kiocb->ki_flags |= IOCB_WRITE;
1367 ret2 = call_write_iter(file, kiocb, &iter);
1368 if (!force_nonblock || ret2 != -EAGAIN) {
1369 io_rw_done(kiocb, ret2);
1372 * If ->needs_lock is true, we're already in async
1376 io_async_list_note(WRITE, req, iov_count);
1386 * IORING_OP_NOP just posts a completion event, nothing else.
1388 static int io_nop(struct io_kiocb *req, u64 user_data)
1390 struct io_ring_ctx *ctx = req->ctx;
1393 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1396 io_cqring_add_event(ctx, user_data, err);
1401 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1403 struct io_ring_ctx *ctx = req->ctx;
1408 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1410 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1416 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1417 bool force_nonblock)
1419 loff_t sqe_off = READ_ONCE(sqe->off);
1420 loff_t sqe_len = READ_ONCE(sqe->len);
1421 loff_t end = sqe_off + sqe_len;
1422 unsigned fsync_flags;
1425 fsync_flags = READ_ONCE(sqe->fsync_flags);
1426 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1429 ret = io_prep_fsync(req, sqe);
1433 /* fsync always requires a blocking context */
1437 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1438 end > 0 ? end : LLONG_MAX,
1439 fsync_flags & IORING_FSYNC_DATASYNC);
1441 if (ret < 0 && (req->flags & REQ_F_LINK))
1442 req->flags |= REQ_F_FAIL_LINK;
1443 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1448 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1450 struct io_ring_ctx *ctx = req->ctx;
1456 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1458 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1464 static int io_sync_file_range(struct io_kiocb *req,
1465 const struct io_uring_sqe *sqe,
1466 bool force_nonblock)
1473 ret = io_prep_sfr(req, sqe);
1477 /* sync_file_range always requires a blocking context */
1481 sqe_off = READ_ONCE(sqe->off);
1482 sqe_len = READ_ONCE(sqe->len);
1483 flags = READ_ONCE(sqe->sync_range_flags);
1485 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1487 if (ret < 0 && (req->flags & REQ_F_LINK))
1488 req->flags |= REQ_F_FAIL_LINK;
1489 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1494 #if defined(CONFIG_NET)
1495 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1496 bool force_nonblock,
1497 long (*fn)(struct socket *, struct user_msghdr __user *,
1500 struct socket *sock;
1503 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1506 sock = sock_from_file(req->file, &ret);
1508 struct user_msghdr __user *msg;
1511 flags = READ_ONCE(sqe->msg_flags);
1512 if (flags & MSG_DONTWAIT)
1513 req->flags |= REQ_F_NOWAIT;
1514 else if (force_nonblock)
1515 flags |= MSG_DONTWAIT;
1517 msg = (struct user_msghdr __user *) (unsigned long)
1518 READ_ONCE(sqe->addr);
1520 ret = fn(sock, msg, flags);
1521 if (force_nonblock && ret == -EAGAIN)
1525 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1531 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1532 bool force_nonblock)
1534 #if defined(CONFIG_NET)
1535 return io_send_recvmsg(req, sqe, force_nonblock, __sys_sendmsg_sock);
1541 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1542 bool force_nonblock)
1544 #if defined(CONFIG_NET)
1545 return io_send_recvmsg(req, sqe, force_nonblock, __sys_recvmsg_sock);
1551 static void io_poll_remove_one(struct io_kiocb *req)
1553 struct io_poll_iocb *poll = &req->poll;
1555 spin_lock(&poll->head->lock);
1556 WRITE_ONCE(poll->canceled, true);
1557 if (!list_empty(&poll->wait.entry)) {
1558 list_del_init(&poll->wait.entry);
1559 io_queue_async_work(req->ctx, req);
1561 spin_unlock(&poll->head->lock);
1563 list_del_init(&req->list);
1566 static void io_poll_remove_all(struct io_ring_ctx *ctx)
1568 struct io_kiocb *req;
1570 spin_lock_irq(&ctx->completion_lock);
1571 while (!list_empty(&ctx->cancel_list)) {
1572 req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
1573 io_poll_remove_one(req);
1575 spin_unlock_irq(&ctx->completion_lock);
1579 * Find a running poll command that matches one specified in sqe->addr,
1580 * and remove it if found.
1582 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1584 struct io_ring_ctx *ctx = req->ctx;
1585 struct io_kiocb *poll_req, *next;
1588 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1590 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
1594 spin_lock_irq(&ctx->completion_lock);
1595 list_for_each_entry_safe(poll_req, next, &ctx->cancel_list, list) {
1596 if (READ_ONCE(sqe->addr) == poll_req->user_data) {
1597 io_poll_remove_one(poll_req);
1602 spin_unlock_irq(&ctx->completion_lock);
1604 io_cqring_add_event(req->ctx, sqe->user_data, ret);
1609 static void io_poll_complete(struct io_ring_ctx *ctx, struct io_kiocb *req,
1612 req->poll.done = true;
1613 io_cqring_fill_event(ctx, req->user_data, mangle_poll(mask));
1614 io_commit_cqring(ctx);
1617 static void io_poll_complete_work(struct work_struct *work)
1619 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1620 struct io_poll_iocb *poll = &req->poll;
1621 struct poll_table_struct pt = { ._key = poll->events };
1622 struct io_ring_ctx *ctx = req->ctx;
1625 if (!READ_ONCE(poll->canceled))
1626 mask = vfs_poll(poll->file, &pt) & poll->events;
1629 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1630 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
1631 * synchronize with them. In the cancellation case the list_del_init
1632 * itself is not actually needed, but harmless so we keep it in to
1633 * avoid further branches in the fast path.
1635 spin_lock_irq(&ctx->completion_lock);
1636 if (!mask && !READ_ONCE(poll->canceled)) {
1637 add_wait_queue(poll->head, &poll->wait);
1638 spin_unlock_irq(&ctx->completion_lock);
1641 list_del_init(&req->list);
1642 io_poll_complete(ctx, req, mask);
1643 spin_unlock_irq(&ctx->completion_lock);
1645 io_cqring_ev_posted(ctx);
1649 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1652 struct io_poll_iocb *poll = container_of(wait, struct io_poll_iocb,
1654 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
1655 struct io_ring_ctx *ctx = req->ctx;
1656 __poll_t mask = key_to_poll(key);
1657 unsigned long flags;
1659 /* for instances that support it check for an event match first: */
1660 if (mask && !(mask & poll->events))
1663 list_del_init(&poll->wait.entry);
1665 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
1666 list_del(&req->list);
1667 io_poll_complete(ctx, req, mask);
1668 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1670 io_cqring_ev_posted(ctx);
1673 io_queue_async_work(ctx, req);
1679 struct io_poll_table {
1680 struct poll_table_struct pt;
1681 struct io_kiocb *req;
1685 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1686 struct poll_table_struct *p)
1688 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
1690 if (unlikely(pt->req->poll.head)) {
1691 pt->error = -EINVAL;
1696 pt->req->poll.head = head;
1697 add_wait_queue(head, &pt->req->poll.wait);
1700 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1702 struct io_poll_iocb *poll = &req->poll;
1703 struct io_ring_ctx *ctx = req->ctx;
1704 struct io_poll_table ipt;
1705 bool cancel = false;
1709 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
1711 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
1716 req->submit.sqe = NULL;
1717 INIT_WORK(&req->work, io_poll_complete_work);
1718 events = READ_ONCE(sqe->poll_events);
1719 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
1723 poll->canceled = false;
1725 ipt.pt._qproc = io_poll_queue_proc;
1726 ipt.pt._key = poll->events;
1728 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1730 /* initialized the list so that we can do list_empty checks */
1731 INIT_LIST_HEAD(&poll->wait.entry);
1732 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
1734 INIT_LIST_HEAD(&req->list);
1736 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
1738 spin_lock_irq(&ctx->completion_lock);
1739 if (likely(poll->head)) {
1740 spin_lock(&poll->head->lock);
1741 if (unlikely(list_empty(&poll->wait.entry))) {
1747 if (mask || ipt.error)
1748 list_del_init(&poll->wait.entry);
1750 WRITE_ONCE(poll->canceled, true);
1751 else if (!poll->done) /* actually waiting for an event */
1752 list_add_tail(&req->list, &ctx->cancel_list);
1753 spin_unlock(&poll->head->lock);
1755 if (mask) { /* no async, we'd stolen it */
1757 io_poll_complete(ctx, req, mask);
1759 spin_unlock_irq(&ctx->completion_lock);
1762 io_cqring_ev_posted(ctx);
1768 static int io_req_defer(struct io_ring_ctx *ctx, struct io_kiocb *req,
1769 const struct io_uring_sqe *sqe)
1771 struct io_uring_sqe *sqe_copy;
1773 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list))
1776 sqe_copy = kmalloc(sizeof(*sqe_copy), GFP_KERNEL);
1780 spin_lock_irq(&ctx->completion_lock);
1781 if (!io_sequence_defer(ctx, req) && list_empty(&ctx->defer_list)) {
1782 spin_unlock_irq(&ctx->completion_lock);
1787 memcpy(sqe_copy, sqe, sizeof(*sqe_copy));
1788 req->submit.sqe = sqe_copy;
1790 INIT_WORK(&req->work, io_sq_wq_submit_work);
1791 list_add_tail(&req->list, &ctx->defer_list);
1792 spin_unlock_irq(&ctx->completion_lock);
1793 return -EIOCBQUEUED;
1796 static int __io_submit_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
1797 const struct sqe_submit *s, bool force_nonblock)
1801 req->user_data = READ_ONCE(s->sqe->user_data);
1803 if (unlikely(s->index >= ctx->sq_entries))
1806 opcode = READ_ONCE(s->sqe->opcode);
1809 ret = io_nop(req, req->user_data);
1811 case IORING_OP_READV:
1812 if (unlikely(s->sqe->buf_index))
1814 ret = io_read(req, s, force_nonblock);
1816 case IORING_OP_WRITEV:
1817 if (unlikely(s->sqe->buf_index))
1819 ret = io_write(req, s, force_nonblock);
1821 case IORING_OP_READ_FIXED:
1822 ret = io_read(req, s, force_nonblock);
1824 case IORING_OP_WRITE_FIXED:
1825 ret = io_write(req, s, force_nonblock);
1827 case IORING_OP_FSYNC:
1828 ret = io_fsync(req, s->sqe, force_nonblock);
1830 case IORING_OP_POLL_ADD:
1831 ret = io_poll_add(req, s->sqe);
1833 case IORING_OP_POLL_REMOVE:
1834 ret = io_poll_remove(req, s->sqe);
1836 case IORING_OP_SYNC_FILE_RANGE:
1837 ret = io_sync_file_range(req, s->sqe, force_nonblock);
1839 case IORING_OP_SENDMSG:
1840 ret = io_sendmsg(req, s->sqe, force_nonblock);
1842 case IORING_OP_RECVMSG:
1843 ret = io_recvmsg(req, s->sqe, force_nonblock);
1853 if (ctx->flags & IORING_SETUP_IOPOLL) {
1854 if (req->result == -EAGAIN)
1857 /* workqueue context doesn't hold uring_lock, grab it now */
1859 mutex_lock(&ctx->uring_lock);
1860 io_iopoll_req_issued(req);
1862 mutex_unlock(&ctx->uring_lock);
1868 static struct async_list *io_async_list_from_sqe(struct io_ring_ctx *ctx,
1869 const struct io_uring_sqe *sqe)
1871 switch (sqe->opcode) {
1872 case IORING_OP_READV:
1873 case IORING_OP_READ_FIXED:
1874 return &ctx->pending_async[READ];
1875 case IORING_OP_WRITEV:
1876 case IORING_OP_WRITE_FIXED:
1877 return &ctx->pending_async[WRITE];
1883 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
1885 u8 opcode = READ_ONCE(sqe->opcode);
1887 return !(opcode == IORING_OP_READ_FIXED ||
1888 opcode == IORING_OP_WRITE_FIXED);
1891 static void io_sq_wq_submit_work(struct work_struct *work)
1893 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
1894 struct io_ring_ctx *ctx = req->ctx;
1895 struct mm_struct *cur_mm = NULL;
1896 struct async_list *async_list;
1897 LIST_HEAD(req_list);
1898 mm_segment_t old_fs;
1901 async_list = io_async_list_from_sqe(ctx, req->submit.sqe);
1904 struct sqe_submit *s = &req->submit;
1905 const struct io_uring_sqe *sqe = s->sqe;
1906 unsigned int flags = req->flags;
1908 /* Ensure we clear previously set non-block flag */
1909 req->rw.ki_flags &= ~IOCB_NOWAIT;
1912 if (io_sqe_needs_user(sqe) && !cur_mm) {
1913 if (!mmget_not_zero(ctx->sqo_mm)) {
1916 cur_mm = ctx->sqo_mm;
1924 s->has_user = cur_mm != NULL;
1925 s->needs_lock = true;
1927 ret = __io_submit_sqe(ctx, req, s, false);
1929 * We can get EAGAIN for polled IO even though
1930 * we're forcing a sync submission from here,
1931 * since we can't wait for request slots on the
1940 /* drop submission reference */
1944 io_cqring_add_event(ctx, sqe->user_data, ret);
1948 /* async context always use a copy of the sqe */
1951 /* req from defer and link list needn't decrease async cnt */
1952 if (flags & (REQ_F_IO_DRAINED | REQ_F_LINK_DONE))
1957 if (!list_empty(&req_list)) {
1958 req = list_first_entry(&req_list, struct io_kiocb,
1960 list_del(&req->list);
1963 if (list_empty(&async_list->list))
1967 spin_lock(&async_list->lock);
1968 if (list_empty(&async_list->list)) {
1969 spin_unlock(&async_list->lock);
1972 list_splice_init(&async_list->list, &req_list);
1973 spin_unlock(&async_list->lock);
1975 req = list_first_entry(&req_list, struct io_kiocb, list);
1976 list_del(&req->list);
1980 * Rare case of racing with a submitter. If we find the count has
1981 * dropped to zero AND we have pending work items, then restart
1982 * the processing. This is a tiny race window.
1985 ret = atomic_dec_return(&async_list->cnt);
1986 while (!ret && !list_empty(&async_list->list)) {
1987 spin_lock(&async_list->lock);
1988 atomic_inc(&async_list->cnt);
1989 list_splice_init(&async_list->list, &req_list);
1990 spin_unlock(&async_list->lock);
1992 if (!list_empty(&req_list)) {
1993 req = list_first_entry(&req_list,
1994 struct io_kiocb, list);
1995 list_del(&req->list);
1998 ret = atomic_dec_return(&async_list->cnt);
2011 * See if we can piggy back onto previously submitted work, that is still
2012 * running. We currently only allow this if the new request is sequential
2013 * to the previous one we punted.
2015 static bool io_add_to_prev_work(struct async_list *list, struct io_kiocb *req)
2021 if (!(req->flags & REQ_F_SEQ_PREV))
2023 if (!atomic_read(&list->cnt))
2027 spin_lock(&list->lock);
2028 list_add_tail(&req->list, &list->list);
2030 * Ensure we see a simultaneous modification from io_sq_wq_submit_work()
2033 if (!atomic_read(&list->cnt)) {
2034 list_del_init(&req->list);
2037 spin_unlock(&list->lock);
2041 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2043 int op = READ_ONCE(sqe->opcode);
2047 case IORING_OP_POLL_REMOVE:
2054 static int io_req_set_file(struct io_ring_ctx *ctx, const struct sqe_submit *s,
2055 struct io_submit_state *state, struct io_kiocb *req)
2060 flags = READ_ONCE(s->sqe->flags);
2061 fd = READ_ONCE(s->sqe->fd);
2063 if (flags & IOSQE_IO_DRAIN)
2064 req->flags |= REQ_F_IO_DRAIN;
2066 * All io need record the previous position, if LINK vs DARIN,
2067 * it can be used to mark the position of the first IO in the
2070 req->sequence = s->sequence;
2072 if (!io_op_needs_file(s->sqe))
2075 if (flags & IOSQE_FIXED_FILE) {
2076 if (unlikely(!ctx->user_files ||
2077 (unsigned) fd >= ctx->nr_user_files))
2079 req->file = ctx->user_files[fd];
2080 req->flags |= REQ_F_FIXED_FILE;
2082 if (s->needs_fixed_file)
2084 req->file = io_file_get(state, fd);
2085 if (unlikely(!req->file))
2092 static int __io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2093 struct sqe_submit *s, bool force_nonblock)
2097 ret = __io_submit_sqe(ctx, req, s, force_nonblock);
2098 if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
2099 struct io_uring_sqe *sqe_copy;
2101 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2103 struct async_list *list;
2106 memcpy(&req->submit, s, sizeof(*s));
2107 list = io_async_list_from_sqe(ctx, s->sqe);
2108 if (!io_add_to_prev_work(list, req)) {
2110 atomic_inc(&list->cnt);
2111 INIT_WORK(&req->work, io_sq_wq_submit_work);
2112 io_queue_async_work(ctx, req);
2116 * Queued up for async execution, worker will release
2117 * submit reference when the iocb is actually submitted.
2123 /* drop submission reference */
2126 /* and drop final reference, if we failed */
2128 io_cqring_add_event(ctx, req->user_data, ret);
2129 if (req->flags & REQ_F_LINK)
2130 req->flags |= REQ_F_FAIL_LINK;
2137 static int io_queue_sqe(struct io_ring_ctx *ctx, struct io_kiocb *req,
2138 struct sqe_submit *s, bool force_nonblock)
2142 ret = io_req_defer(ctx, req, s->sqe);
2144 if (ret != -EIOCBQUEUED) {
2146 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2151 return __io_queue_sqe(ctx, req, s, force_nonblock);
2154 static int io_queue_link_head(struct io_ring_ctx *ctx, struct io_kiocb *req,
2155 struct sqe_submit *s, struct io_kiocb *shadow,
2156 bool force_nonblock)
2159 int need_submit = false;
2162 return io_queue_sqe(ctx, req, s, force_nonblock);
2165 * Mark the first IO in link list as DRAIN, let all the following
2166 * IOs enter the defer list. all IO needs to be completed before link
2169 req->flags |= REQ_F_IO_DRAIN;
2170 ret = io_req_defer(ctx, req, s->sqe);
2172 if (ret != -EIOCBQUEUED) {
2174 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2179 * If ret == 0 means that all IOs in front of link io are
2180 * running done. let's queue link head.
2185 /* Insert shadow req to defer_list, blocking next IOs */
2186 spin_lock_irq(&ctx->completion_lock);
2187 list_add_tail(&shadow->list, &ctx->defer_list);
2188 spin_unlock_irq(&ctx->completion_lock);
2191 return __io_queue_sqe(ctx, req, s, force_nonblock);
2196 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
2198 static void io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
2199 struct io_submit_state *state, struct io_kiocb **link,
2200 bool force_nonblock)
2202 struct io_uring_sqe *sqe_copy;
2203 struct io_kiocb *req;
2206 /* enforce forwards compatibility on users */
2207 if (unlikely(s->sqe->flags & ~SQE_VALID_FLAGS)) {
2212 req = io_get_req(ctx, state);
2213 if (unlikely(!req)) {
2218 ret = io_req_set_file(ctx, s, state, req);
2219 if (unlikely(ret)) {
2223 io_cqring_add_event(ctx, s->sqe->user_data, ret);
2228 * If we already have a head request, queue this one for async
2229 * submittal once the head completes. If we don't have a head but
2230 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
2231 * submitted sync once the chain is complete. If none of those
2232 * conditions are true (normal request), then just queue it.
2235 struct io_kiocb *prev = *link;
2237 sqe_copy = kmemdup(s->sqe, sizeof(*sqe_copy), GFP_KERNEL);
2244 memcpy(&req->submit, s, sizeof(*s));
2245 list_add_tail(&req->list, &prev->link_list);
2246 } else if (s->sqe->flags & IOSQE_IO_LINK) {
2247 req->flags |= REQ_F_LINK;
2249 memcpy(&req->submit, s, sizeof(*s));
2250 INIT_LIST_HEAD(&req->link_list);
2253 io_queue_sqe(ctx, req, s, force_nonblock);
2258 * Batched submission is done, ensure local IO is flushed out.
2260 static void io_submit_state_end(struct io_submit_state *state)
2262 blk_finish_plug(&state->plug);
2264 if (state->free_reqs)
2265 kmem_cache_free_bulk(req_cachep, state->free_reqs,
2266 &state->reqs[state->cur_req]);
2270 * Start submission side cache.
2272 static void io_submit_state_start(struct io_submit_state *state,
2273 struct io_ring_ctx *ctx, unsigned max_ios)
2275 blk_start_plug(&state->plug);
2276 state->free_reqs = 0;
2278 state->ios_left = max_ios;
2281 static void io_commit_sqring(struct io_ring_ctx *ctx)
2283 struct io_rings *rings = ctx->rings;
2285 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
2287 * Ensure any loads from the SQEs are done at this point,
2288 * since once we write the new head, the application could
2289 * write new data to them.
2291 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
2296 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
2297 * that is mapped by userspace. This means that care needs to be taken to
2298 * ensure that reads are stable, as we cannot rely on userspace always
2299 * being a good citizen. If members of the sqe are validated and then later
2300 * used, it's important that those reads are done through READ_ONCE() to
2301 * prevent a re-load down the line.
2303 static bool io_get_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
2305 struct io_rings *rings = ctx->rings;
2306 u32 *sq_array = ctx->sq_array;
2310 * The cached sq head (or cq tail) serves two purposes:
2312 * 1) allows us to batch the cost of updating the user visible
2314 * 2) allows the kernel side to track the head on its own, even
2315 * though the application is the one updating it.
2317 head = ctx->cached_sq_head;
2318 /* make sure SQ entry isn't read before tail */
2319 if (head == smp_load_acquire(&rings->sq.tail))
2322 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
2323 if (head < ctx->sq_entries) {
2325 s->sqe = &ctx->sq_sqes[head];
2326 s->sequence = ctx->cached_sq_head;
2327 ctx->cached_sq_head++;
2331 /* drop invalid entries */
2332 ctx->cached_sq_head++;
2333 rings->sq_dropped++;
2337 static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
2338 unsigned int nr, bool has_user, bool mm_fault)
2340 struct io_submit_state state, *statep = NULL;
2341 struct io_kiocb *link = NULL;
2342 struct io_kiocb *shadow_req = NULL;
2343 bool prev_was_link = false;
2344 int i, submitted = 0;
2346 if (nr > IO_PLUG_THRESHOLD) {
2347 io_submit_state_start(&state, ctx, nr);
2351 for (i = 0; i < nr; i++) {
2353 * If previous wasn't linked and we have a linked command,
2354 * that's the end of the chain. Submit the previous link.
2356 if (!prev_was_link && link) {
2357 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2362 prev_was_link = (sqes[i].sqe->flags & IOSQE_IO_LINK) != 0;
2364 if (link && (sqes[i].sqe->flags & IOSQE_IO_DRAIN)) {
2366 shadow_req = io_get_req(ctx, NULL);
2367 if (unlikely(!shadow_req))
2369 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2370 refcount_dec(&shadow_req->refs);
2372 shadow_req->sequence = sqes[i].sequence;
2376 if (unlikely(mm_fault)) {
2377 io_cqring_add_event(ctx, sqes[i].sqe->user_data,
2380 sqes[i].has_user = has_user;
2381 sqes[i].needs_lock = true;
2382 sqes[i].needs_fixed_file = true;
2383 io_submit_sqe(ctx, &sqes[i], statep, &link, true);
2389 io_queue_link_head(ctx, link, &link->submit, shadow_req, true);
2391 io_submit_state_end(&state);
2396 static int io_sq_thread(void *data)
2398 struct sqe_submit sqes[IO_IOPOLL_BATCH];
2399 struct io_ring_ctx *ctx = data;
2400 struct mm_struct *cur_mm = NULL;
2401 mm_segment_t old_fs;
2404 unsigned long timeout;
2406 complete(&ctx->sqo_thread_started);
2411 timeout = inflight = 0;
2412 while (!kthread_should_park()) {
2413 bool all_fixed, mm_fault = false;
2417 unsigned nr_events = 0;
2419 if (ctx->flags & IORING_SETUP_IOPOLL) {
2420 io_iopoll_check(ctx, &nr_events, 0);
2423 * Normal IO, just pretend everything completed.
2424 * We don't have to poll completions for that.
2426 nr_events = inflight;
2429 inflight -= nr_events;
2431 timeout = jiffies + ctx->sq_thread_idle;
2434 if (!io_get_sqring(ctx, &sqes[0])) {
2436 * We're polling. If we're within the defined idle
2437 * period, then let us spin without work before going
2440 if (inflight || !time_after(jiffies, timeout)) {
2446 * Drop cur_mm before scheduling, we can't hold it for
2447 * long periods (or over schedule()). Do this before
2448 * adding ourselves to the waitqueue, as the unuse/drop
2457 prepare_to_wait(&ctx->sqo_wait, &wait,
2458 TASK_INTERRUPTIBLE);
2460 /* Tell userspace we may need a wakeup call */
2461 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
2462 /* make sure to read SQ tail after writing flags */
2465 if (!io_get_sqring(ctx, &sqes[0])) {
2466 if (kthread_should_park()) {
2467 finish_wait(&ctx->sqo_wait, &wait);
2470 if (signal_pending(current))
2471 flush_signals(current);
2473 finish_wait(&ctx->sqo_wait, &wait);
2475 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2478 finish_wait(&ctx->sqo_wait, &wait);
2480 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
2486 if (all_fixed && io_sqe_needs_user(sqes[i].sqe))
2490 if (i == ARRAY_SIZE(sqes))
2492 } while (io_get_sqring(ctx, &sqes[i]));
2494 /* Unless all new commands are FIXED regions, grab mm */
2495 if (!all_fixed && !cur_mm) {
2496 mm_fault = !mmget_not_zero(ctx->sqo_mm);
2498 use_mm(ctx->sqo_mm);
2499 cur_mm = ctx->sqo_mm;
2503 inflight += io_submit_sqes(ctx, sqes, i, cur_mm != NULL,
2506 /* Commit SQ ring head once we've consumed all SQEs */
2507 io_commit_sqring(ctx);
2521 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit,
2522 bool block_for_last)
2524 struct io_submit_state state, *statep = NULL;
2525 struct io_kiocb *link = NULL;
2526 struct io_kiocb *shadow_req = NULL;
2527 bool prev_was_link = false;
2530 if (to_submit > IO_PLUG_THRESHOLD) {
2531 io_submit_state_start(&state, ctx, to_submit);
2535 for (i = 0; i < to_submit; i++) {
2536 bool force_nonblock = true;
2537 struct sqe_submit s;
2539 if (!io_get_sqring(ctx, &s))
2543 * If previous wasn't linked and we have a linked command,
2544 * that's the end of the chain. Submit the previous link.
2546 if (!prev_was_link && link) {
2547 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2552 prev_was_link = (s.sqe->flags & IOSQE_IO_LINK) != 0;
2554 if (link && (s.sqe->flags & IOSQE_IO_DRAIN)) {
2556 shadow_req = io_get_req(ctx, NULL);
2557 if (unlikely(!shadow_req))
2559 shadow_req->flags |= (REQ_F_IO_DRAIN | REQ_F_SHADOW_DRAIN);
2560 refcount_dec(&shadow_req->refs);
2562 shadow_req->sequence = s.sequence;
2567 s.needs_lock = false;
2568 s.needs_fixed_file = false;
2572 * The caller will block for events after submit, submit the
2573 * last IO non-blocking. This is either the only IO it's
2574 * submitting, or it already submitted the previous ones. This
2575 * improves performance by avoiding an async punt that we don't
2578 if (block_for_last && submit == to_submit)
2579 force_nonblock = false;
2581 io_submit_sqe(ctx, &s, statep, &link, force_nonblock);
2583 io_commit_sqring(ctx);
2586 io_queue_link_head(ctx, link, &link->submit, shadow_req,
2589 io_submit_state_end(statep);
2595 * Wait until events become available, if we don't already have some. The
2596 * application must reap them itself, as they reside on the shared cq ring.
2598 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
2599 const sigset_t __user *sig, size_t sigsz)
2601 struct io_rings *rings = ctx->rings;
2604 if (io_cqring_events(rings) >= min_events)
2608 #ifdef CONFIG_COMPAT
2609 if (in_compat_syscall())
2610 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
2614 ret = set_user_sigmask(sig, sigsz);
2620 ret = wait_event_interruptible(ctx->wait, io_cqring_events(rings) >= min_events);
2621 restore_saved_sigmask_unless(ret == -ERESTARTSYS);
2622 if (ret == -ERESTARTSYS)
2625 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
2628 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
2630 #if defined(CONFIG_UNIX)
2631 if (ctx->ring_sock) {
2632 struct sock *sock = ctx->ring_sock->sk;
2633 struct sk_buff *skb;
2635 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
2641 for (i = 0; i < ctx->nr_user_files; i++)
2642 fput(ctx->user_files[i]);
2646 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
2648 if (!ctx->user_files)
2651 __io_sqe_files_unregister(ctx);
2652 kfree(ctx->user_files);
2653 ctx->user_files = NULL;
2654 ctx->nr_user_files = 0;
2658 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
2660 if (ctx->sqo_thread) {
2661 wait_for_completion(&ctx->sqo_thread_started);
2663 * The park is a bit of a work-around, without it we get
2664 * warning spews on shutdown with SQPOLL set and affinity
2665 * set to a single CPU.
2667 kthread_park(ctx->sqo_thread);
2668 kthread_stop(ctx->sqo_thread);
2669 ctx->sqo_thread = NULL;
2673 static void io_finish_async(struct io_ring_ctx *ctx)
2677 io_sq_thread_stop(ctx);
2679 for (i = 0; i < ARRAY_SIZE(ctx->sqo_wq); i++) {
2680 if (ctx->sqo_wq[i]) {
2681 destroy_workqueue(ctx->sqo_wq[i]);
2682 ctx->sqo_wq[i] = NULL;
2687 #if defined(CONFIG_UNIX)
2688 static void io_destruct_skb(struct sk_buff *skb)
2690 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
2692 io_finish_async(ctx);
2693 unix_destruct_scm(skb);
2697 * Ensure the UNIX gc is aware of our file set, so we are certain that
2698 * the io_uring can be safely unregistered on process exit, even if we have
2699 * loops in the file referencing.
2701 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
2703 struct sock *sk = ctx->ring_sock->sk;
2704 struct scm_fp_list *fpl;
2705 struct sk_buff *skb;
2708 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
2709 unsigned long inflight = ctx->user->unix_inflight + nr;
2711 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
2715 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
2719 skb = alloc_skb(0, GFP_KERNEL);
2726 skb->destructor = io_destruct_skb;
2728 fpl->user = get_uid(ctx->user);
2729 for (i = 0; i < nr; i++) {
2730 fpl->fp[i] = get_file(ctx->user_files[i + offset]);
2731 unix_inflight(fpl->user, fpl->fp[i]);
2734 fpl->max = fpl->count = nr;
2735 UNIXCB(skb).fp = fpl;
2736 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
2737 skb_queue_head(&sk->sk_receive_queue, skb);
2739 for (i = 0; i < nr; i++)
2746 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
2747 * causes regular reference counting to break down. We rely on the UNIX
2748 * garbage collection to take care of this problem for us.
2750 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2752 unsigned left, total;
2756 left = ctx->nr_user_files;
2758 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
2760 ret = __io_sqe_files_scm(ctx, this_files, total);
2764 total += this_files;
2770 while (total < ctx->nr_user_files) {
2771 fput(ctx->user_files[total]);
2778 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
2784 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
2787 __s32 __user *fds = (__s32 __user *) arg;
2791 if (ctx->user_files)
2795 if (nr_args > IORING_MAX_FIXED_FILES)
2798 ctx->user_files = kcalloc(nr_args, sizeof(struct file *), GFP_KERNEL);
2799 if (!ctx->user_files)
2802 for (i = 0; i < nr_args; i++) {
2804 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
2807 ctx->user_files[i] = fget(fd);
2810 if (!ctx->user_files[i])
2813 * Don't allow io_uring instances to be registered. If UNIX
2814 * isn't enabled, then this causes a reference cycle and this
2815 * instance can never get freed. If UNIX is enabled we'll
2816 * handle it just fine, but there's still no point in allowing
2817 * a ring fd as it doesn't support regular read/write anyway.
2819 if (ctx->user_files[i]->f_op == &io_uring_fops) {
2820 fput(ctx->user_files[i]);
2823 ctx->nr_user_files++;
2828 for (i = 0; i < ctx->nr_user_files; i++)
2829 fput(ctx->user_files[i]);
2831 kfree(ctx->user_files);
2832 ctx->user_files = NULL;
2833 ctx->nr_user_files = 0;
2837 ret = io_sqe_files_scm(ctx);
2839 io_sqe_files_unregister(ctx);
2844 static int io_sq_offload_start(struct io_ring_ctx *ctx,
2845 struct io_uring_params *p)
2849 init_waitqueue_head(&ctx->sqo_wait);
2850 mmgrab(current->mm);
2851 ctx->sqo_mm = current->mm;
2853 if (ctx->flags & IORING_SETUP_SQPOLL) {
2855 if (!capable(CAP_SYS_ADMIN))
2858 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
2859 if (!ctx->sq_thread_idle)
2860 ctx->sq_thread_idle = HZ;
2862 if (p->flags & IORING_SETUP_SQ_AFF) {
2863 int cpu = p->sq_thread_cpu;
2866 if (cpu >= nr_cpu_ids)
2868 if (!cpu_online(cpu))
2871 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
2875 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
2878 if (IS_ERR(ctx->sqo_thread)) {
2879 ret = PTR_ERR(ctx->sqo_thread);
2880 ctx->sqo_thread = NULL;
2883 wake_up_process(ctx->sqo_thread);
2884 } else if (p->flags & IORING_SETUP_SQ_AFF) {
2885 /* Can't have SQ_AFF without SQPOLL */
2890 /* Do QD, or 2 * CPUS, whatever is smallest */
2891 ctx->sqo_wq[0] = alloc_workqueue("io_ring-wq",
2892 WQ_UNBOUND | WQ_FREEZABLE,
2893 min(ctx->sq_entries - 1, 2 * num_online_cpus()));
2894 if (!ctx->sqo_wq[0]) {
2900 * This is for buffered writes, where we want to limit the parallelism
2901 * due to file locking in file systems. As "normal" buffered writes
2902 * should parellelize on writeout quite nicely, limit us to having 2
2903 * pending. This avoids massive contention on the inode when doing
2904 * buffered async writes.
2906 ctx->sqo_wq[1] = alloc_workqueue("io_ring-write-wq",
2907 WQ_UNBOUND | WQ_FREEZABLE, 2);
2908 if (!ctx->sqo_wq[1]) {
2915 io_finish_async(ctx);
2916 mmdrop(ctx->sqo_mm);
2921 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
2923 atomic_long_sub(nr_pages, &user->locked_vm);
2926 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
2928 unsigned long page_limit, cur_pages, new_pages;
2930 /* Don't allow more pages than we can safely lock */
2931 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
2934 cur_pages = atomic_long_read(&user->locked_vm);
2935 new_pages = cur_pages + nr_pages;
2936 if (new_pages > page_limit)
2938 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
2939 new_pages) != cur_pages);
2944 static void io_mem_free(void *ptr)
2951 page = virt_to_head_page(ptr);
2952 if (put_page_testzero(page))
2953 free_compound_page(page);
2956 static void *io_mem_alloc(size_t size)
2958 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
2961 return (void *) __get_free_pages(gfp_flags, get_order(size));
2964 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
2967 struct io_rings *rings;
2968 size_t off, sq_array_size;
2970 off = struct_size(rings, cqes, cq_entries);
2971 if (off == SIZE_MAX)
2975 off = ALIGN(off, SMP_CACHE_BYTES);
2980 sq_array_size = array_size(sizeof(u32), sq_entries);
2981 if (sq_array_size == SIZE_MAX)
2984 if (check_add_overflow(off, sq_array_size, &off))
2993 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
2997 pages = (size_t)1 << get_order(
2998 rings_size(sq_entries, cq_entries, NULL));
2999 pages += (size_t)1 << get_order(
3000 array_size(sizeof(struct io_uring_sqe), sq_entries));
3005 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
3009 if (!ctx->user_bufs)
3012 for (i = 0; i < ctx->nr_user_bufs; i++) {
3013 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3015 for (j = 0; j < imu->nr_bvecs; j++)
3016 put_user_page(imu->bvec[j].bv_page);
3018 if (ctx->account_mem)
3019 io_unaccount_mem(ctx->user, imu->nr_bvecs);
3024 kfree(ctx->user_bufs);
3025 ctx->user_bufs = NULL;
3026 ctx->nr_user_bufs = 0;
3030 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
3031 void __user *arg, unsigned index)
3033 struct iovec __user *src;
3035 #ifdef CONFIG_COMPAT
3037 struct compat_iovec __user *ciovs;
3038 struct compat_iovec ciov;
3040 ciovs = (struct compat_iovec __user *) arg;
3041 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
3044 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
3045 dst->iov_len = ciov.iov_len;
3049 src = (struct iovec __user *) arg;
3050 if (copy_from_user(dst, &src[index], sizeof(*dst)))
3055 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
3058 struct vm_area_struct **vmas = NULL;
3059 struct page **pages = NULL;
3060 int i, j, got_pages = 0;
3065 if (!nr_args || nr_args > UIO_MAXIOV)
3068 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
3070 if (!ctx->user_bufs)
3073 for (i = 0; i < nr_args; i++) {
3074 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
3075 unsigned long off, start, end, ubuf;
3080 ret = io_copy_iov(ctx, &iov, arg, i);
3085 * Don't impose further limits on the size and buffer
3086 * constraints here, we'll -EINVAL later when IO is
3087 * submitted if they are wrong.
3090 if (!iov.iov_base || !iov.iov_len)
3093 /* arbitrary limit, but we need something */
3094 if (iov.iov_len > SZ_1G)
3097 ubuf = (unsigned long) iov.iov_base;
3098 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
3099 start = ubuf >> PAGE_SHIFT;
3100 nr_pages = end - start;
3102 if (ctx->account_mem) {
3103 ret = io_account_mem(ctx->user, nr_pages);
3109 if (!pages || nr_pages > got_pages) {
3112 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
3114 vmas = kvmalloc_array(nr_pages,
3115 sizeof(struct vm_area_struct *),
3117 if (!pages || !vmas) {
3119 if (ctx->account_mem)
3120 io_unaccount_mem(ctx->user, nr_pages);
3123 got_pages = nr_pages;
3126 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
3130 if (ctx->account_mem)
3131 io_unaccount_mem(ctx->user, nr_pages);
3136 down_read(¤t->mm->mmap_sem);
3137 pret = get_user_pages(ubuf, nr_pages,
3138 FOLL_WRITE | FOLL_LONGTERM,
3140 if (pret == nr_pages) {
3141 /* don't support file backed memory */
3142 for (j = 0; j < nr_pages; j++) {
3143 struct vm_area_struct *vma = vmas[j];
3146 !is_file_hugepages(vma->vm_file)) {
3152 ret = pret < 0 ? pret : -EFAULT;
3154 up_read(¤t->mm->mmap_sem);
3157 * if we did partial map, or found file backed vmas,
3158 * release any pages we did get
3161 put_user_pages(pages, pret);
3162 if (ctx->account_mem)
3163 io_unaccount_mem(ctx->user, nr_pages);
3168 off = ubuf & ~PAGE_MASK;
3170 for (j = 0; j < nr_pages; j++) {
3173 vec_len = min_t(size_t, size, PAGE_SIZE - off);
3174 imu->bvec[j].bv_page = pages[j];
3175 imu->bvec[j].bv_len = vec_len;
3176 imu->bvec[j].bv_offset = off;
3180 /* store original address for later verification */
3182 imu->len = iov.iov_len;
3183 imu->nr_bvecs = nr_pages;
3185 ctx->nr_user_bufs++;
3193 io_sqe_buffer_unregister(ctx);
3197 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
3199 __s32 __user *fds = arg;
3205 if (copy_from_user(&fd, fds, sizeof(*fds)))
3208 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
3209 if (IS_ERR(ctx->cq_ev_fd)) {
3210 int ret = PTR_ERR(ctx->cq_ev_fd);
3211 ctx->cq_ev_fd = NULL;
3218 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
3220 if (ctx->cq_ev_fd) {
3221 eventfd_ctx_put(ctx->cq_ev_fd);
3222 ctx->cq_ev_fd = NULL;
3229 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
3231 io_finish_async(ctx);
3233 mmdrop(ctx->sqo_mm);
3235 io_iopoll_reap_events(ctx);
3236 io_sqe_buffer_unregister(ctx);
3237 io_sqe_files_unregister(ctx);
3238 io_eventfd_unregister(ctx);
3240 #if defined(CONFIG_UNIX)
3241 if (ctx->ring_sock) {
3242 ctx->ring_sock->file = NULL; /* so that iput() is called */
3243 sock_release(ctx->ring_sock);
3247 io_mem_free(ctx->rings);
3248 io_mem_free(ctx->sq_sqes);
3250 percpu_ref_exit(&ctx->refs);
3251 if (ctx->account_mem)
3252 io_unaccount_mem(ctx->user,
3253 ring_pages(ctx->sq_entries, ctx->cq_entries));
3254 free_uid(ctx->user);
3258 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
3260 struct io_ring_ctx *ctx = file->private_data;
3263 poll_wait(file, &ctx->cq_wait, wait);
3265 * synchronizes with barrier from wq_has_sleeper call in
3269 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
3270 ctx->rings->sq_ring_entries)
3271 mask |= EPOLLOUT | EPOLLWRNORM;
3272 if (READ_ONCE(ctx->rings->sq.head) != ctx->cached_cq_tail)
3273 mask |= EPOLLIN | EPOLLRDNORM;
3278 static int io_uring_fasync(int fd, struct file *file, int on)
3280 struct io_ring_ctx *ctx = file->private_data;
3282 return fasync_helper(fd, file, on, &ctx->cq_fasync);
3285 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
3287 mutex_lock(&ctx->uring_lock);
3288 percpu_ref_kill(&ctx->refs);
3289 mutex_unlock(&ctx->uring_lock);
3291 io_poll_remove_all(ctx);
3292 io_iopoll_reap_events(ctx);
3293 wait_for_completion(&ctx->ctx_done);
3294 io_ring_ctx_free(ctx);
3297 static int io_uring_release(struct inode *inode, struct file *file)
3299 struct io_ring_ctx *ctx = file->private_data;
3301 file->private_data = NULL;
3302 io_ring_ctx_wait_and_kill(ctx);
3306 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
3308 loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
3309 unsigned long sz = vma->vm_end - vma->vm_start;
3310 struct io_ring_ctx *ctx = file->private_data;
3316 case IORING_OFF_SQ_RING:
3317 case IORING_OFF_CQ_RING:
3320 case IORING_OFF_SQES:
3327 page = virt_to_head_page(ptr);
3328 if (sz > (PAGE_SIZE << compound_order(page)))
3331 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
3332 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
3335 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
3336 u32, min_complete, u32, flags, const sigset_t __user *, sig,
3339 struct io_ring_ctx *ctx;
3344 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
3352 if (f.file->f_op != &io_uring_fops)
3356 ctx = f.file->private_data;
3357 if (!percpu_ref_tryget(&ctx->refs))
3361 * For SQ polling, the thread will do all submissions and completions.
3362 * Just return the requested submit count, and wake the thread if
3366 if (ctx->flags & IORING_SETUP_SQPOLL) {
3367 if (flags & IORING_ENTER_SQ_WAKEUP)
3368 wake_up(&ctx->sqo_wait);
3369 submitted = to_submit;
3370 } else if (to_submit) {
3371 bool block_for_last = false;
3373 to_submit = min(to_submit, ctx->sq_entries);
3376 * Allow last submission to block in a series, IFF the caller
3377 * asked to wait for events and we don't currently have
3378 * enough. This potentially avoids an async punt.
3380 if (to_submit == min_complete &&
3381 io_cqring_events(ctx->rings) < min_complete)
3382 block_for_last = true;
3384 mutex_lock(&ctx->uring_lock);
3385 submitted = io_ring_submit(ctx, to_submit, block_for_last);
3386 mutex_unlock(&ctx->uring_lock);
3388 if (flags & IORING_ENTER_GETEVENTS) {
3389 unsigned nr_events = 0;
3391 min_complete = min(min_complete, ctx->cq_entries);
3393 if (ctx->flags & IORING_SETUP_IOPOLL) {
3394 ret = io_iopoll_check(ctx, &nr_events, min_complete);
3396 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
3400 io_ring_drop_ctx_refs(ctx, 1);
3403 return submitted ? submitted : ret;
3406 static const struct file_operations io_uring_fops = {
3407 .release = io_uring_release,
3408 .mmap = io_uring_mmap,
3409 .poll = io_uring_poll,
3410 .fasync = io_uring_fasync,
3413 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
3414 struct io_uring_params *p)
3416 struct io_rings *rings;
3417 size_t size, sq_array_offset;
3419 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
3420 if (size == SIZE_MAX)
3423 rings = io_mem_alloc(size);
3428 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
3429 rings->sq_ring_mask = p->sq_entries - 1;
3430 rings->cq_ring_mask = p->cq_entries - 1;
3431 rings->sq_ring_entries = p->sq_entries;
3432 rings->cq_ring_entries = p->cq_entries;
3433 ctx->sq_mask = rings->sq_ring_mask;
3434 ctx->cq_mask = rings->cq_ring_mask;
3435 ctx->sq_entries = rings->sq_ring_entries;
3436 ctx->cq_entries = rings->cq_ring_entries;
3438 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
3439 if (size == SIZE_MAX)
3442 ctx->sq_sqes = io_mem_alloc(size);
3450 * Allocate an anonymous fd, this is what constitutes the application
3451 * visible backing of an io_uring instance. The application mmaps this
3452 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
3453 * we have to tie this fd to a socket for file garbage collection purposes.
3455 static int io_uring_get_fd(struct io_ring_ctx *ctx)
3460 #if defined(CONFIG_UNIX)
3461 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
3467 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
3471 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
3472 O_RDWR | O_CLOEXEC);
3475 ret = PTR_ERR(file);
3479 #if defined(CONFIG_UNIX)
3480 ctx->ring_sock->file = file;
3481 ctx->ring_sock->sk->sk_user_data = ctx;
3483 fd_install(ret, file);
3486 #if defined(CONFIG_UNIX)
3487 sock_release(ctx->ring_sock);
3488 ctx->ring_sock = NULL;
3493 static int io_uring_create(unsigned entries, struct io_uring_params *p)
3495 struct user_struct *user = NULL;
3496 struct io_ring_ctx *ctx;
3500 if (!entries || entries > IORING_MAX_ENTRIES)
3504 * Use twice as many entries for the CQ ring. It's possible for the
3505 * application to drive a higher depth than the size of the SQ ring,
3506 * since the sqes are only used at submission time. This allows for
3507 * some flexibility in overcommitting a bit.
3509 p->sq_entries = roundup_pow_of_two(entries);
3510 p->cq_entries = 2 * p->sq_entries;
3512 user = get_uid(current_user());
3513 account_mem = !capable(CAP_IPC_LOCK);
3516 ret = io_account_mem(user,
3517 ring_pages(p->sq_entries, p->cq_entries));
3524 ctx = io_ring_ctx_alloc(p);
3527 io_unaccount_mem(user, ring_pages(p->sq_entries,
3532 ctx->compat = in_compat_syscall();
3533 ctx->account_mem = account_mem;
3536 ret = io_allocate_scq_urings(ctx, p);
3540 ret = io_sq_offload_start(ctx, p);
3544 ret = io_uring_get_fd(ctx);
3548 memset(&p->sq_off, 0, sizeof(p->sq_off));
3549 p->sq_off.head = offsetof(struct io_rings, sq.head);
3550 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
3551 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
3552 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
3553 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
3554 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
3555 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
3557 memset(&p->cq_off, 0, sizeof(p->cq_off));
3558 p->cq_off.head = offsetof(struct io_rings, cq.head);
3559 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
3560 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
3561 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
3562 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
3563 p->cq_off.cqes = offsetof(struct io_rings, cqes);
3565 p->features = IORING_FEAT_SINGLE_MMAP;
3568 io_ring_ctx_wait_and_kill(ctx);
3573 * Sets up an aio uring context, and returns the fd. Applications asks for a
3574 * ring size, we return the actual sq/cq ring sizes (among other things) in the
3575 * params structure passed in.
3577 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
3579 struct io_uring_params p;
3583 if (copy_from_user(&p, params, sizeof(p)))
3585 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
3590 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
3591 IORING_SETUP_SQ_AFF))
3594 ret = io_uring_create(entries, &p);
3598 if (copy_to_user(params, &p, sizeof(p)))
3604 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
3605 struct io_uring_params __user *, params)
3607 return io_uring_setup(entries, params);
3610 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
3611 void __user *arg, unsigned nr_args)
3612 __releases(ctx->uring_lock)
3613 __acquires(ctx->uring_lock)
3618 * We're inside the ring mutex, if the ref is already dying, then
3619 * someone else killed the ctx or is already going through
3620 * io_uring_register().
3622 if (percpu_ref_is_dying(&ctx->refs))
3625 percpu_ref_kill(&ctx->refs);
3628 * Drop uring mutex before waiting for references to exit. If another
3629 * thread is currently inside io_uring_enter() it might need to grab
3630 * the uring_lock to make progress. If we hold it here across the drain
3631 * wait, then we can deadlock. It's safe to drop the mutex here, since
3632 * no new references will come in after we've killed the percpu ref.
3634 mutex_unlock(&ctx->uring_lock);
3635 wait_for_completion(&ctx->ctx_done);
3636 mutex_lock(&ctx->uring_lock);
3639 case IORING_REGISTER_BUFFERS:
3640 ret = io_sqe_buffer_register(ctx, arg, nr_args);
3642 case IORING_UNREGISTER_BUFFERS:
3646 ret = io_sqe_buffer_unregister(ctx);
3648 case IORING_REGISTER_FILES:
3649 ret = io_sqe_files_register(ctx, arg, nr_args);
3651 case IORING_UNREGISTER_FILES:
3655 ret = io_sqe_files_unregister(ctx);
3657 case IORING_REGISTER_EVENTFD:
3661 ret = io_eventfd_register(ctx, arg);
3663 case IORING_UNREGISTER_EVENTFD:
3667 ret = io_eventfd_unregister(ctx);
3674 /* bring the ctx back to life */
3675 reinit_completion(&ctx->ctx_done);
3676 percpu_ref_reinit(&ctx->refs);
3680 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
3681 void __user *, arg, unsigned int, nr_args)
3683 struct io_ring_ctx *ctx;
3692 if (f.file->f_op != &io_uring_fops)
3695 ctx = f.file->private_data;
3697 mutex_lock(&ctx->uring_lock);
3698 ret = __io_uring_register(ctx, opcode, arg, nr_args);
3699 mutex_unlock(&ctx->uring_lock);
3705 static int __init io_uring_init(void)
3707 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3710 __initcall(io_uring_init);