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/kthread.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
74 #define CREATE_TRACE_POINTS
75 #include <trace/events/io_uring.h>
77 #include <uapi/linux/io_uring.h>
82 #define IORING_MAX_ENTRIES 32768
83 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
86 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
88 #define IORING_FILE_TABLE_SHIFT 9
89 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
90 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
91 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
94 u32 head ____cacheline_aligned_in_smp;
95 u32 tail ____cacheline_aligned_in_smp;
99 * This data is shared with the application through the mmap at offsets
100 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
102 * The offsets to the member fields are published through struct
103 * io_sqring_offsets when calling io_uring_setup.
107 * Head and tail offsets into the ring; the offsets need to be
108 * masked to get valid indices.
110 * The kernel controls head of the sq ring and the tail of the cq ring,
111 * and the application controls tail of the sq ring and the head of the
114 struct io_uring sq, cq;
116 * Bitmasks to apply to head and tail offsets (constant, equals
119 u32 sq_ring_mask, cq_ring_mask;
120 /* Ring sizes (constant, power of 2) */
121 u32 sq_ring_entries, cq_ring_entries;
123 * Number of invalid entries dropped by the kernel due to
124 * invalid index stored in array
126 * Written by the kernel, shouldn't be modified by the
127 * application (i.e. get number of "new events" by comparing to
130 * After a new SQ head value was read by the application this
131 * counter includes all submissions that were dropped reaching
132 * the new SQ head (and possibly more).
138 * Written by the kernel, shouldn't be modified by the
141 * The application needs a full memory barrier before checking
142 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
146 * Number of completion events lost because the queue was full;
147 * this should be avoided by the application by making sure
148 * there are not more requests pending than there is space in
149 * the completion queue.
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * As completion events come in out of order this counter is not
156 * ordered with any other data.
160 * Ring buffer of completion events.
162 * The kernel writes completion events fresh every time they are
163 * produced, so the application is allowed to modify pending
166 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
169 struct io_mapped_ubuf {
172 struct bio_vec *bvec;
173 unsigned int nr_bvecs;
176 struct fixed_file_table {
182 struct percpu_ref refs;
183 } ____cacheline_aligned_in_smp;
189 bool cq_overflow_flushed;
193 * Ring buffer of indices into array of io_uring_sqe, which is
194 * mmapped by the application using the IORING_OFF_SQES offset.
196 * This indirection could e.g. be used to assign fixed
197 * io_uring_sqe entries to operations and only submit them to
198 * the queue when needed.
200 * The kernel modifies neither the indices array nor the entries
204 unsigned cached_sq_head;
207 unsigned sq_thread_idle;
208 unsigned cached_sq_dropped;
209 atomic_t cached_cq_overflow;
210 struct io_uring_sqe *sq_sqes;
212 struct list_head defer_list;
213 struct list_head timeout_list;
214 struct list_head cq_overflow_list;
216 wait_queue_head_t inflight_wait;
217 } ____cacheline_aligned_in_smp;
219 struct io_rings *rings;
223 struct task_struct *sqo_thread; /* if using sq thread polling */
224 struct mm_struct *sqo_mm;
225 wait_queue_head_t sqo_wait;
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
232 struct fixed_file_table *file_table;
233 unsigned nr_user_files;
235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
239 struct user_struct *user;
241 const struct cred *creds;
243 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
244 struct completion *completions;
246 /* if all else fails... */
247 struct io_kiocb *fallback_req;
249 #if defined(CONFIG_UNIX)
250 struct socket *ring_sock;
254 unsigned cached_cq_tail;
257 atomic_t cq_timeouts;
258 struct wait_queue_head cq_wait;
259 struct fasync_struct *cq_fasync;
260 struct eventfd_ctx *cq_ev_fd;
261 } ____cacheline_aligned_in_smp;
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 hlist_head *cancel_hash;
279 unsigned cancel_hash_bits;
281 spinlock_t inflight_lock;
282 struct list_head inflight_list;
283 } ____cacheline_aligned_in_smp;
287 * First field must be the file pointer in all the
288 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
290 struct io_poll_iocb {
293 struct wait_queue_head *head;
299 struct wait_queue_entry wait;
302 struct io_timeout_data {
303 struct io_kiocb *req;
304 struct hrtimer timer;
305 struct timespec64 ts;
306 enum hrtimer_mode mode;
312 struct sockaddr __user *addr;
313 int __user *addr_len;
337 /* NOTE: kiocb has the file as the first member, so don't do it here */
345 struct sockaddr __user *addr;
351 struct user_msghdr __user *msg;
355 struct io_async_connect {
356 struct sockaddr_storage address;
359 struct io_async_msghdr {
360 struct iovec fast_iov[UIO_FASTIOV];
362 struct sockaddr __user *uaddr;
367 struct iovec fast_iov[UIO_FASTIOV];
373 struct io_async_ctx {
375 struct io_async_rw rw;
376 struct io_async_msghdr msg;
377 struct io_async_connect connect;
378 struct io_timeout_data timeout;
383 * NOTE! Each of the iocb union members has the file pointer
384 * as the first entry in their struct definition. So you can
385 * access the file pointer through any of the sub-structs,
386 * or directly as just 'ki_filp' in this struct.
392 struct io_poll_iocb poll;
393 struct io_accept accept;
395 struct io_cancel cancel;
396 struct io_timeout timeout;
397 struct io_connect connect;
398 struct io_sr_msg sr_msg;
401 struct io_async_ctx *io;
402 struct file *ring_file;
406 bool needs_fixed_file;
409 struct io_ring_ctx *ctx;
411 struct list_head list;
412 struct hlist_node hash_node;
414 struct list_head link_list;
417 #define REQ_F_NOWAIT 1 /* must not punt to workers */
418 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
419 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
420 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
421 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
422 #define REQ_F_IO_DRAINED 32 /* drain done */
423 #define REQ_F_LINK 64 /* linked sqes */
424 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
425 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
426 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
427 #define REQ_F_TIMEOUT 1024 /* timeout request */
428 #define REQ_F_ISREG 2048 /* regular file */
429 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
430 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
431 #define REQ_F_INFLIGHT 16384 /* on inflight list */
432 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
433 #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */
438 struct list_head inflight_entry;
440 struct io_wq_work work;
443 #define IO_PLUG_THRESHOLD 2
444 #define IO_IOPOLL_BATCH 8
446 struct io_submit_state {
447 struct blk_plug plug;
450 * io_kiocb alloc cache
452 void *reqs[IO_IOPOLL_BATCH];
453 unsigned int free_reqs;
454 unsigned int cur_req;
457 * File reference cache
461 unsigned int has_refs;
462 unsigned int used_refs;
463 unsigned int ios_left;
466 static void io_wq_submit_work(struct io_wq_work **workptr);
467 static void io_cqring_fill_event(struct io_kiocb *req, long res);
468 static void __io_free_req(struct io_kiocb *req);
469 static void io_put_req(struct io_kiocb *req);
470 static void io_double_put_req(struct io_kiocb *req);
471 static void __io_double_put_req(struct io_kiocb *req);
472 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
473 static void io_queue_linked_timeout(struct io_kiocb *req);
475 static struct kmem_cache *req_cachep;
477 static const struct file_operations io_uring_fops;
479 struct sock *io_uring_get_socket(struct file *file)
481 #if defined(CONFIG_UNIX)
482 if (file->f_op == &io_uring_fops) {
483 struct io_ring_ctx *ctx = file->private_data;
485 return ctx->ring_sock->sk;
490 EXPORT_SYMBOL(io_uring_get_socket);
492 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
494 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
496 complete(&ctx->completions[0]);
499 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
501 struct io_ring_ctx *ctx;
504 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
508 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
509 if (!ctx->fallback_req)
512 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
513 if (!ctx->completions)
517 * Use 5 bits less than the max cq entries, that should give us around
518 * 32 entries per hash list if totally full and uniformly spread.
520 hash_bits = ilog2(p->cq_entries);
524 ctx->cancel_hash_bits = hash_bits;
525 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
527 if (!ctx->cancel_hash)
529 __hash_init(ctx->cancel_hash, 1U << hash_bits);
531 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
532 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
535 ctx->flags = p->flags;
536 init_waitqueue_head(&ctx->cq_wait);
537 INIT_LIST_HEAD(&ctx->cq_overflow_list);
538 init_completion(&ctx->completions[0]);
539 init_completion(&ctx->completions[1]);
540 mutex_init(&ctx->uring_lock);
541 init_waitqueue_head(&ctx->wait);
542 spin_lock_init(&ctx->completion_lock);
543 INIT_LIST_HEAD(&ctx->poll_list);
544 INIT_LIST_HEAD(&ctx->defer_list);
545 INIT_LIST_HEAD(&ctx->timeout_list);
546 init_waitqueue_head(&ctx->inflight_wait);
547 spin_lock_init(&ctx->inflight_lock);
548 INIT_LIST_HEAD(&ctx->inflight_list);
551 if (ctx->fallback_req)
552 kmem_cache_free(req_cachep, ctx->fallback_req);
553 kfree(ctx->completions);
554 kfree(ctx->cancel_hash);
559 static inline bool __req_need_defer(struct io_kiocb *req)
561 struct io_ring_ctx *ctx = req->ctx;
563 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
564 + atomic_read(&ctx->cached_cq_overflow);
567 static inline bool req_need_defer(struct io_kiocb *req)
569 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
570 return __req_need_defer(req);
575 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
577 struct io_kiocb *req;
579 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
580 if (req && !req_need_defer(req)) {
581 list_del_init(&req->list);
588 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
590 struct io_kiocb *req;
592 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
594 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
596 if (!__req_need_defer(req)) {
597 list_del_init(&req->list);
605 static void __io_commit_cqring(struct io_ring_ctx *ctx)
607 struct io_rings *rings = ctx->rings;
609 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
610 /* order cqe stores with ring update */
611 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
613 if (wq_has_sleeper(&ctx->cq_wait)) {
614 wake_up_interruptible(&ctx->cq_wait);
615 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
620 static inline bool io_req_needs_user(struct io_kiocb *req)
622 return !(req->opcode == IORING_OP_READ_FIXED ||
623 req->opcode == IORING_OP_WRITE_FIXED);
626 static inline bool io_prep_async_work(struct io_kiocb *req,
627 struct io_kiocb **link)
629 bool do_hashed = false;
631 switch (req->opcode) {
632 case IORING_OP_WRITEV:
633 case IORING_OP_WRITE_FIXED:
634 /* only regular files should be hashed for writes */
635 if (req->flags & REQ_F_ISREG)
638 case IORING_OP_READV:
639 case IORING_OP_READ_FIXED:
640 case IORING_OP_SENDMSG:
641 case IORING_OP_RECVMSG:
642 case IORING_OP_ACCEPT:
643 case IORING_OP_POLL_ADD:
644 case IORING_OP_CONNECT:
646 * We know REQ_F_ISREG is not set on some of these
647 * opcodes, but this enables us to keep the check in
650 if (!(req->flags & REQ_F_ISREG))
651 req->work.flags |= IO_WQ_WORK_UNBOUND;
654 if (io_req_needs_user(req))
655 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
657 *link = io_prep_linked_timeout(req);
661 static inline void io_queue_async_work(struct io_kiocb *req)
663 struct io_ring_ctx *ctx = req->ctx;
664 struct io_kiocb *link;
667 do_hashed = io_prep_async_work(req, &link);
669 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
672 io_wq_enqueue(ctx->io_wq, &req->work);
674 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
675 file_inode(req->file));
679 io_queue_linked_timeout(link);
682 static void io_kill_timeout(struct io_kiocb *req)
686 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
688 atomic_inc(&req->ctx->cq_timeouts);
689 list_del_init(&req->list);
690 io_cqring_fill_event(req, 0);
695 static void io_kill_timeouts(struct io_ring_ctx *ctx)
697 struct io_kiocb *req, *tmp;
699 spin_lock_irq(&ctx->completion_lock);
700 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
701 io_kill_timeout(req);
702 spin_unlock_irq(&ctx->completion_lock);
705 static void io_commit_cqring(struct io_ring_ctx *ctx)
707 struct io_kiocb *req;
709 while ((req = io_get_timeout_req(ctx)) != NULL)
710 io_kill_timeout(req);
712 __io_commit_cqring(ctx);
714 while ((req = io_get_deferred_req(ctx)) != NULL) {
715 req->flags |= REQ_F_IO_DRAINED;
716 io_queue_async_work(req);
720 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
722 struct io_rings *rings = ctx->rings;
725 tail = ctx->cached_cq_tail;
727 * writes to the cq entry need to come after reading head; the
728 * control dependency is enough as we're using WRITE_ONCE to
731 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
734 ctx->cached_cq_tail++;
735 return &rings->cqes[tail & ctx->cq_mask];
738 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
740 if (waitqueue_active(&ctx->wait))
742 if (waitqueue_active(&ctx->sqo_wait))
743 wake_up(&ctx->sqo_wait);
745 eventfd_signal(ctx->cq_ev_fd, 1);
748 /* Returns true if there are no backlogged entries after the flush */
749 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
751 struct io_rings *rings = ctx->rings;
752 struct io_uring_cqe *cqe;
753 struct io_kiocb *req;
758 if (list_empty_careful(&ctx->cq_overflow_list))
760 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
761 rings->cq_ring_entries))
765 spin_lock_irqsave(&ctx->completion_lock, flags);
767 /* if force is set, the ring is going away. always drop after that */
769 ctx->cq_overflow_flushed = true;
772 while (!list_empty(&ctx->cq_overflow_list)) {
773 cqe = io_get_cqring(ctx);
777 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
779 list_move(&req->list, &list);
781 WRITE_ONCE(cqe->user_data, req->user_data);
782 WRITE_ONCE(cqe->res, req->result);
783 WRITE_ONCE(cqe->flags, 0);
785 WRITE_ONCE(ctx->rings->cq_overflow,
786 atomic_inc_return(&ctx->cached_cq_overflow));
790 io_commit_cqring(ctx);
791 spin_unlock_irqrestore(&ctx->completion_lock, flags);
792 io_cqring_ev_posted(ctx);
794 while (!list_empty(&list)) {
795 req = list_first_entry(&list, struct io_kiocb, list);
796 list_del(&req->list);
803 static void io_cqring_fill_event(struct io_kiocb *req, long res)
805 struct io_ring_ctx *ctx = req->ctx;
806 struct io_uring_cqe *cqe;
808 trace_io_uring_complete(ctx, req->user_data, res);
811 * If we can't get a cq entry, userspace overflowed the
812 * submission (by quite a lot). Increment the overflow count in
815 cqe = io_get_cqring(ctx);
817 WRITE_ONCE(cqe->user_data, req->user_data);
818 WRITE_ONCE(cqe->res, res);
819 WRITE_ONCE(cqe->flags, 0);
820 } else if (ctx->cq_overflow_flushed) {
821 WRITE_ONCE(ctx->rings->cq_overflow,
822 atomic_inc_return(&ctx->cached_cq_overflow));
824 refcount_inc(&req->refs);
826 list_add_tail(&req->list, &ctx->cq_overflow_list);
830 static void io_cqring_add_event(struct io_kiocb *req, long res)
832 struct io_ring_ctx *ctx = req->ctx;
835 spin_lock_irqsave(&ctx->completion_lock, flags);
836 io_cqring_fill_event(req, res);
837 io_commit_cqring(ctx);
838 spin_unlock_irqrestore(&ctx->completion_lock, flags);
840 io_cqring_ev_posted(ctx);
843 static inline bool io_is_fallback_req(struct io_kiocb *req)
845 return req == (struct io_kiocb *)
846 ((unsigned long) req->ctx->fallback_req & ~1UL);
849 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
851 struct io_kiocb *req;
853 req = ctx->fallback_req;
854 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
860 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
861 struct io_submit_state *state)
863 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
864 struct io_kiocb *req;
866 if (!percpu_ref_tryget(&ctx->refs))
870 req = kmem_cache_alloc(req_cachep, gfp);
873 } else if (!state->free_reqs) {
877 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
878 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
881 * Bulk alloc is all-or-nothing. If we fail to get a batch,
882 * retry single alloc to be on the safe side.
884 if (unlikely(ret <= 0)) {
885 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
890 state->free_reqs = ret - 1;
892 req = state->reqs[0];
894 req = state->reqs[state->cur_req];
901 req->ring_file = NULL;
905 /* one is dropped after submission, the other at completion */
906 refcount_set(&req->refs, 2);
908 INIT_IO_WORK(&req->work, io_wq_submit_work);
911 req = io_get_fallback_req(ctx);
914 percpu_ref_put(&ctx->refs);
918 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
921 kmem_cache_free_bulk(req_cachep, *nr, reqs);
922 percpu_ref_put_many(&ctx->refs, *nr);
927 static void __io_free_req(struct io_kiocb *req)
929 struct io_ring_ctx *ctx = req->ctx;
933 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
935 if (req->flags & REQ_F_INFLIGHT) {
938 spin_lock_irqsave(&ctx->inflight_lock, flags);
939 list_del(&req->inflight_entry);
940 if (waitqueue_active(&ctx->inflight_wait))
941 wake_up(&ctx->inflight_wait);
942 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
944 percpu_ref_put(&ctx->refs);
945 if (likely(!io_is_fallback_req(req)))
946 kmem_cache_free(req_cachep, req);
948 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
951 static bool io_link_cancel_timeout(struct io_kiocb *req)
953 struct io_ring_ctx *ctx = req->ctx;
956 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
958 io_cqring_fill_event(req, -ECANCELED);
959 io_commit_cqring(ctx);
960 req->flags &= ~REQ_F_LINK;
968 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
970 struct io_ring_ctx *ctx = req->ctx;
971 bool wake_ev = false;
973 /* Already got next link */
974 if (req->flags & REQ_F_LINK_NEXT)
978 * The list should never be empty when we are called here. But could
979 * potentially happen if the chain is messed up, check to be on the
982 while (!list_empty(&req->link_list)) {
983 struct io_kiocb *nxt = list_first_entry(&req->link_list,
984 struct io_kiocb, link_list);
986 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
987 (nxt->flags & REQ_F_TIMEOUT))) {
988 list_del_init(&nxt->link_list);
989 wake_ev |= io_link_cancel_timeout(nxt);
990 req->flags &= ~REQ_F_LINK_TIMEOUT;
994 list_del_init(&req->link_list);
995 if (!list_empty(&nxt->link_list))
996 nxt->flags |= REQ_F_LINK;
1001 req->flags |= REQ_F_LINK_NEXT;
1003 io_cqring_ev_posted(ctx);
1007 * Called if REQ_F_LINK is set, and we fail the head request
1009 static void io_fail_links(struct io_kiocb *req)
1011 struct io_ring_ctx *ctx = req->ctx;
1012 unsigned long flags;
1014 spin_lock_irqsave(&ctx->completion_lock, flags);
1016 while (!list_empty(&req->link_list)) {
1017 struct io_kiocb *link = list_first_entry(&req->link_list,
1018 struct io_kiocb, link_list);
1020 list_del_init(&link->link_list);
1021 trace_io_uring_fail_link(req, link);
1023 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1024 link->opcode == IORING_OP_LINK_TIMEOUT) {
1025 io_link_cancel_timeout(link);
1027 io_cqring_fill_event(link, -ECANCELED);
1028 __io_double_put_req(link);
1030 req->flags &= ~REQ_F_LINK_TIMEOUT;
1033 io_commit_cqring(ctx);
1034 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1035 io_cqring_ev_posted(ctx);
1038 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1040 if (likely(!(req->flags & REQ_F_LINK)))
1044 * If LINK is set, we have dependent requests in this chain. If we
1045 * didn't fail this request, queue the first one up, moving any other
1046 * dependencies to the next request. In case of failure, fail the rest
1049 if (req->flags & REQ_F_FAIL_LINK) {
1051 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1052 REQ_F_LINK_TIMEOUT) {
1053 struct io_ring_ctx *ctx = req->ctx;
1054 unsigned long flags;
1057 * If this is a timeout link, we could be racing with the
1058 * timeout timer. Grab the completion lock for this case to
1059 * protect against that.
1061 spin_lock_irqsave(&ctx->completion_lock, flags);
1062 io_req_link_next(req, nxt);
1063 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1065 io_req_link_next(req, nxt);
1069 static void io_free_req(struct io_kiocb *req)
1071 struct io_kiocb *nxt = NULL;
1073 io_req_find_next(req, &nxt);
1077 io_queue_async_work(nxt);
1081 * Drop reference to request, return next in chain (if there is one) if this
1082 * was the last reference to this request.
1084 __attribute__((nonnull))
1085 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1087 io_req_find_next(req, nxtptr);
1089 if (refcount_dec_and_test(&req->refs))
1093 static void io_put_req(struct io_kiocb *req)
1095 if (refcount_dec_and_test(&req->refs))
1100 * Must only be used if we don't need to care about links, usually from
1101 * within the completion handling itself.
1103 static void __io_double_put_req(struct io_kiocb *req)
1105 /* drop both submit and complete references */
1106 if (refcount_sub_and_test(2, &req->refs))
1110 static void io_double_put_req(struct io_kiocb *req)
1112 /* drop both submit and complete references */
1113 if (refcount_sub_and_test(2, &req->refs))
1117 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1119 struct io_rings *rings = ctx->rings;
1122 * noflush == true is from the waitqueue handler, just ensure we wake
1123 * up the task, and the next invocation will flush the entries. We
1124 * cannot safely to it from here.
1126 if (noflush && !list_empty(&ctx->cq_overflow_list))
1129 io_cqring_overflow_flush(ctx, false);
1131 /* See comment at the top of this file */
1133 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1136 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1138 struct io_rings *rings = ctx->rings;
1140 /* make sure SQ entry isn't read before tail */
1141 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1145 * Find and free completed poll iocbs
1147 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1148 struct list_head *done)
1150 void *reqs[IO_IOPOLL_BATCH];
1151 struct io_kiocb *req;
1155 while (!list_empty(done)) {
1156 req = list_first_entry(done, struct io_kiocb, list);
1157 list_del(&req->list);
1159 io_cqring_fill_event(req, req->result);
1162 if (refcount_dec_and_test(&req->refs)) {
1163 /* If we're not using fixed files, we have to pair the
1164 * completion part with the file put. Use regular
1165 * completions for those, only batch free for fixed
1166 * file and non-linked commands.
1168 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1169 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1171 reqs[to_free++] = req;
1172 if (to_free == ARRAY_SIZE(reqs))
1173 io_free_req_many(ctx, reqs, &to_free);
1180 io_commit_cqring(ctx);
1181 io_free_req_many(ctx, reqs, &to_free);
1184 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1187 struct io_kiocb *req, *tmp;
1193 * Only spin for completions if we don't have multiple devices hanging
1194 * off our complete list, and we're under the requested amount.
1196 spin = !ctx->poll_multi_file && *nr_events < min;
1199 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1200 struct kiocb *kiocb = &req->rw.kiocb;
1203 * Move completed entries to our local list. If we find a
1204 * request that requires polling, break out and complete
1205 * the done list first, if we have entries there.
1207 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1208 list_move_tail(&req->list, &done);
1211 if (!list_empty(&done))
1214 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1223 if (!list_empty(&done))
1224 io_iopoll_complete(ctx, nr_events, &done);
1230 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1231 * non-spinning poll check - we'll still enter the driver poll loop, but only
1232 * as a non-spinning completion check.
1234 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1237 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1240 ret = io_do_iopoll(ctx, nr_events, min);
1243 if (!min || *nr_events >= min)
1251 * We can't just wait for polled events to come to us, we have to actively
1252 * find and complete them.
1254 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1256 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1259 mutex_lock(&ctx->uring_lock);
1260 while (!list_empty(&ctx->poll_list)) {
1261 unsigned int nr_events = 0;
1263 io_iopoll_getevents(ctx, &nr_events, 1);
1266 * Ensure we allow local-to-the-cpu processing to take place,
1267 * in this case we need to ensure that we reap all events.
1271 mutex_unlock(&ctx->uring_lock);
1274 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1277 int iters = 0, ret = 0;
1283 * Don't enter poll loop if we already have events pending.
1284 * If we do, we can potentially be spinning for commands that
1285 * already triggered a CQE (eg in error).
1287 if (io_cqring_events(ctx, false))
1291 * If a submit got punted to a workqueue, we can have the
1292 * application entering polling for a command before it gets
1293 * issued. That app will hold the uring_lock for the duration
1294 * of the poll right here, so we need to take a breather every
1295 * now and then to ensure that the issue has a chance to add
1296 * the poll to the issued list. Otherwise we can spin here
1297 * forever, while the workqueue is stuck trying to acquire the
1300 if (!(++iters & 7)) {
1301 mutex_unlock(&ctx->uring_lock);
1302 mutex_lock(&ctx->uring_lock);
1305 if (*nr_events < min)
1306 tmin = min - *nr_events;
1308 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1312 } while (min && !*nr_events && !need_resched());
1317 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1323 * We disallow the app entering submit/complete with polling, but we
1324 * still need to lock the ring to prevent racing with polled issue
1325 * that got punted to a workqueue.
1327 mutex_lock(&ctx->uring_lock);
1328 ret = __io_iopoll_check(ctx, nr_events, min);
1329 mutex_unlock(&ctx->uring_lock);
1333 static void kiocb_end_write(struct io_kiocb *req)
1336 * Tell lockdep we inherited freeze protection from submission
1339 if (req->flags & REQ_F_ISREG) {
1340 struct inode *inode = file_inode(req->file);
1342 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1344 file_end_write(req->file);
1347 static inline void req_set_fail_links(struct io_kiocb *req)
1349 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1350 req->flags |= REQ_F_FAIL_LINK;
1353 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1355 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1357 if (kiocb->ki_flags & IOCB_WRITE)
1358 kiocb_end_write(req);
1360 if (res != req->result)
1361 req_set_fail_links(req);
1362 io_cqring_add_event(req, res);
1365 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1367 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1369 io_complete_rw_common(kiocb, res);
1373 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1375 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1376 struct io_kiocb *nxt = NULL;
1378 io_complete_rw_common(kiocb, res);
1379 io_put_req_find_next(req, &nxt);
1384 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1386 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1388 if (kiocb->ki_flags & IOCB_WRITE)
1389 kiocb_end_write(req);
1391 if (res != req->result)
1392 req_set_fail_links(req);
1395 req->flags |= REQ_F_IOPOLL_COMPLETED;
1399 * After the iocb has been issued, it's safe to be found on the poll list.
1400 * Adding the kiocb to the list AFTER submission ensures that we don't
1401 * find it from a io_iopoll_getevents() thread before the issuer is done
1402 * accessing the kiocb cookie.
1404 static void io_iopoll_req_issued(struct io_kiocb *req)
1406 struct io_ring_ctx *ctx = req->ctx;
1409 * Track whether we have multiple files in our lists. This will impact
1410 * how we do polling eventually, not spinning if we're on potentially
1411 * different devices.
1413 if (list_empty(&ctx->poll_list)) {
1414 ctx->poll_multi_file = false;
1415 } else if (!ctx->poll_multi_file) {
1416 struct io_kiocb *list_req;
1418 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1420 if (list_req->file != req->file)
1421 ctx->poll_multi_file = true;
1425 * For fast devices, IO may have already completed. If it has, add
1426 * it to the front so we find it first.
1428 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1429 list_add(&req->list, &ctx->poll_list);
1431 list_add_tail(&req->list, &ctx->poll_list);
1434 static void io_file_put(struct io_submit_state *state)
1437 int diff = state->has_refs - state->used_refs;
1440 fput_many(state->file, diff);
1446 * Get as many references to a file as we have IOs left in this submission,
1447 * assuming most submissions are for one file, or at least that each file
1448 * has more than one submission.
1450 static struct file *io_file_get(struct io_submit_state *state, int fd)
1456 if (state->fd == fd) {
1463 state->file = fget_many(fd, state->ios_left);
1468 state->has_refs = state->ios_left;
1469 state->used_refs = 1;
1475 * If we tracked the file through the SCM inflight mechanism, we could support
1476 * any file. For now, just ensure that anything potentially problematic is done
1479 static bool io_file_supports_async(struct file *file)
1481 umode_t mode = file_inode(file)->i_mode;
1483 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1485 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1491 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1492 bool force_nonblock)
1494 struct io_ring_ctx *ctx = req->ctx;
1495 struct kiocb *kiocb = &req->rw.kiocb;
1502 if (S_ISREG(file_inode(req->file)->i_mode))
1503 req->flags |= REQ_F_ISREG;
1505 kiocb->ki_pos = READ_ONCE(sqe->off);
1506 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1507 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1509 ioprio = READ_ONCE(sqe->ioprio);
1511 ret = ioprio_check_cap(ioprio);
1515 kiocb->ki_ioprio = ioprio;
1517 kiocb->ki_ioprio = get_current_ioprio();
1519 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1523 /* don't allow async punt if RWF_NOWAIT was requested */
1524 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1525 (req->file->f_flags & O_NONBLOCK))
1526 req->flags |= REQ_F_NOWAIT;
1529 kiocb->ki_flags |= IOCB_NOWAIT;
1531 if (ctx->flags & IORING_SETUP_IOPOLL) {
1532 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1533 !kiocb->ki_filp->f_op->iopoll)
1536 kiocb->ki_flags |= IOCB_HIPRI;
1537 kiocb->ki_complete = io_complete_rw_iopoll;
1540 if (kiocb->ki_flags & IOCB_HIPRI)
1542 kiocb->ki_complete = io_complete_rw;
1545 req->rw.addr = READ_ONCE(sqe->addr);
1546 req->rw.len = READ_ONCE(sqe->len);
1547 /* we own ->private, reuse it for the buffer index */
1548 req->rw.kiocb.private = (void *) (unsigned long)
1549 READ_ONCE(sqe->buf_index);
1553 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1559 case -ERESTARTNOINTR:
1560 case -ERESTARTNOHAND:
1561 case -ERESTART_RESTARTBLOCK:
1563 * We can't just restart the syscall, since previously
1564 * submitted sqes may already be in progress. Just fail this
1570 kiocb->ki_complete(kiocb, ret, 0);
1574 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1577 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1578 *nxt = __io_complete_rw(kiocb, ret);
1580 io_rw_done(kiocb, ret);
1583 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
1584 struct iov_iter *iter)
1586 struct io_ring_ctx *ctx = req->ctx;
1587 size_t len = req->rw.len;
1588 struct io_mapped_ubuf *imu;
1589 unsigned index, buf_index;
1593 /* attempt to use fixed buffers without having provided iovecs */
1594 if (unlikely(!ctx->user_bufs))
1597 buf_index = (unsigned long) req->rw.kiocb.private;
1598 if (unlikely(buf_index >= ctx->nr_user_bufs))
1601 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1602 imu = &ctx->user_bufs[index];
1603 buf_addr = req->rw.addr;
1606 if (buf_addr + len < buf_addr)
1608 /* not inside the mapped region */
1609 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1613 * May not be a start of buffer, set size appropriately
1614 * and advance us to the beginning.
1616 offset = buf_addr - imu->ubuf;
1617 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1621 * Don't use iov_iter_advance() here, as it's really slow for
1622 * using the latter parts of a big fixed buffer - it iterates
1623 * over each segment manually. We can cheat a bit here, because
1626 * 1) it's a BVEC iter, we set it up
1627 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1628 * first and last bvec
1630 * So just find our index, and adjust the iterator afterwards.
1631 * If the offset is within the first bvec (or the whole first
1632 * bvec, just use iov_iter_advance(). This makes it easier
1633 * since we can just skip the first segment, which may not
1634 * be PAGE_SIZE aligned.
1636 const struct bio_vec *bvec = imu->bvec;
1638 if (offset <= bvec->bv_len) {
1639 iov_iter_advance(iter, offset);
1641 unsigned long seg_skip;
1643 /* skip first vec */
1644 offset -= bvec->bv_len;
1645 seg_skip = 1 + (offset >> PAGE_SHIFT);
1647 iter->bvec = bvec + seg_skip;
1648 iter->nr_segs -= seg_skip;
1649 iter->count -= bvec->bv_len + offset;
1650 iter->iov_offset = offset & ~PAGE_MASK;
1657 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1658 struct iovec **iovec, struct iov_iter *iter)
1660 void __user *buf = u64_to_user_ptr(req->rw.addr);
1661 size_t sqe_len = req->rw.len;
1664 opcode = req->opcode;
1665 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1667 return io_import_fixed(req, rw, iter);
1670 /* buffer index only valid with fixed read/write */
1671 if (req->rw.kiocb.private)
1675 struct io_async_rw *iorw = &req->io->rw;
1678 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1679 if (iorw->iov == iorw->fast_iov)
1687 #ifdef CONFIG_COMPAT
1688 if (req->ctx->compat)
1689 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1693 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1697 * For files that don't have ->read_iter() and ->write_iter(), handle them
1698 * by looping over ->read() or ->write() manually.
1700 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1701 struct iov_iter *iter)
1706 * Don't support polled IO through this interface, and we can't
1707 * support non-blocking either. For the latter, this just causes
1708 * the kiocb to be handled from an async context.
1710 if (kiocb->ki_flags & IOCB_HIPRI)
1712 if (kiocb->ki_flags & IOCB_NOWAIT)
1715 while (iov_iter_count(iter)) {
1719 if (!iov_iter_is_bvec(iter)) {
1720 iovec = iov_iter_iovec(iter);
1722 /* fixed buffers import bvec */
1723 iovec.iov_base = kmap(iter->bvec->bv_page)
1725 iovec.iov_len = min(iter->count,
1726 iter->bvec->bv_len - iter->iov_offset);
1730 nr = file->f_op->read(file, iovec.iov_base,
1731 iovec.iov_len, &kiocb->ki_pos);
1733 nr = file->f_op->write(file, iovec.iov_base,
1734 iovec.iov_len, &kiocb->ki_pos);
1737 if (iov_iter_is_bvec(iter))
1738 kunmap(iter->bvec->bv_page);
1746 if (nr != iovec.iov_len)
1748 iov_iter_advance(iter, nr);
1754 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
1755 struct iovec *iovec, struct iovec *fast_iov,
1756 struct iov_iter *iter)
1758 req->io->rw.nr_segs = iter->nr_segs;
1759 req->io->rw.size = io_size;
1760 req->io->rw.iov = iovec;
1761 if (!req->io->rw.iov) {
1762 req->io->rw.iov = req->io->rw.fast_iov;
1763 memcpy(req->io->rw.iov, fast_iov,
1764 sizeof(struct iovec) * iter->nr_segs);
1768 static int io_alloc_async_ctx(struct io_kiocb *req)
1770 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1771 return req->io == NULL;
1774 static void io_rw_async(struct io_wq_work **workptr)
1776 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1777 struct iovec *iov = NULL;
1779 if (req->io->rw.iov != req->io->rw.fast_iov)
1780 iov = req->io->rw.iov;
1781 io_wq_submit_work(workptr);
1785 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
1786 struct iovec *iovec, struct iovec *fast_iov,
1787 struct iov_iter *iter)
1789 if (!req->io && io_alloc_async_ctx(req))
1792 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
1793 req->work.func = io_rw_async;
1797 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1798 bool force_nonblock)
1800 struct io_async_ctx *io;
1801 struct iov_iter iter;
1804 ret = io_prep_rw(req, sqe, force_nonblock);
1808 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1815 io->rw.iov = io->rw.fast_iov;
1817 ret = io_import_iovec(READ, req, &io->rw.iov, &iter);
1822 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
1826 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1827 bool force_nonblock)
1829 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1830 struct kiocb *kiocb = &req->rw.kiocb;
1831 struct iov_iter iter;
1833 ssize_t io_size, ret;
1835 ret = io_import_iovec(READ, req, &iovec, &iter);
1839 /* Ensure we clear previously set non-block flag */
1840 if (!force_nonblock)
1841 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
1844 if (req->flags & REQ_F_LINK)
1845 req->result = io_size;
1848 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1849 * we know to async punt it even if it was opened O_NONBLOCK
1851 if (force_nonblock && !io_file_supports_async(req->file)) {
1852 req->flags |= REQ_F_MUST_PUNT;
1856 iov_count = iov_iter_count(&iter);
1857 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
1861 if (req->file->f_op->read_iter)
1862 ret2 = call_read_iter(req->file, kiocb, &iter);
1864 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
1867 * In case of a short read, punt to async. This can happen
1868 * if we have data partially cached. Alternatively we can
1869 * return the short read, in which case the application will
1870 * need to issue another SQE and wait for it. That SQE will
1871 * need async punt anyway, so it's more efficient to do it
1874 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1875 (req->flags & REQ_F_ISREG) &&
1876 ret2 > 0 && ret2 < io_size)
1878 /* Catch -EAGAIN return for forced non-blocking submission */
1879 if (!force_nonblock || ret2 != -EAGAIN) {
1880 kiocb_done(kiocb, ret2, nxt, req->in_async);
1883 ret = io_setup_async_rw(req, io_size, iovec,
1884 inline_vecs, &iter);
1891 if (!io_wq_current_is_worker())
1896 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1897 bool force_nonblock)
1899 struct io_async_ctx *io;
1900 struct iov_iter iter;
1903 ret = io_prep_rw(req, sqe, force_nonblock);
1907 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1914 io->rw.iov = io->rw.fast_iov;
1916 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter);
1921 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
1925 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1926 bool force_nonblock)
1928 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1929 struct kiocb *kiocb = &req->rw.kiocb;
1930 struct iov_iter iter;
1932 ssize_t ret, io_size;
1934 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1938 /* Ensure we clear previously set non-block flag */
1939 if (!force_nonblock)
1940 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
1943 if (req->flags & REQ_F_LINK)
1944 req->result = io_size;
1947 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1948 * we know to async punt it even if it was opened O_NONBLOCK
1950 if (force_nonblock && !io_file_supports_async(req->file)) {
1951 req->flags |= REQ_F_MUST_PUNT;
1955 /* file path doesn't support NOWAIT for non-direct_IO */
1956 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
1957 (req->flags & REQ_F_ISREG))
1960 iov_count = iov_iter_count(&iter);
1961 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
1966 * Open-code file_start_write here to grab freeze protection,
1967 * which will be released by another thread in
1968 * io_complete_rw(). Fool lockdep by telling it the lock got
1969 * released so that it doesn't complain about the held lock when
1970 * we return to userspace.
1972 if (req->flags & REQ_F_ISREG) {
1973 __sb_start_write(file_inode(req->file)->i_sb,
1974 SB_FREEZE_WRITE, true);
1975 __sb_writers_release(file_inode(req->file)->i_sb,
1978 kiocb->ki_flags |= IOCB_WRITE;
1980 if (req->file->f_op->write_iter)
1981 ret2 = call_write_iter(req->file, kiocb, &iter);
1983 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
1984 if (!force_nonblock || ret2 != -EAGAIN) {
1985 kiocb_done(kiocb, ret2, nxt, req->in_async);
1988 ret = io_setup_async_rw(req, io_size, iovec,
1989 inline_vecs, &iter);
1996 if (!io_wq_current_is_worker())
2002 * IORING_OP_NOP just posts a completion event, nothing else.
2004 static int io_nop(struct io_kiocb *req)
2006 struct io_ring_ctx *ctx = req->ctx;
2008 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2011 io_cqring_add_event(req, 0);
2016 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2018 struct io_ring_ctx *ctx = req->ctx;
2023 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2025 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2028 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2029 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2032 req->sync.off = READ_ONCE(sqe->off);
2033 req->sync.len = READ_ONCE(sqe->len);
2037 static bool io_req_cancelled(struct io_kiocb *req)
2039 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2040 req_set_fail_links(req);
2041 io_cqring_add_event(req, -ECANCELED);
2049 static void io_fsync_finish(struct io_wq_work **workptr)
2051 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2052 loff_t end = req->sync.off + req->sync.len;
2053 struct io_kiocb *nxt = NULL;
2056 if (io_req_cancelled(req))
2059 ret = vfs_fsync_range(req->file, req->sync.off,
2060 end > 0 ? end : LLONG_MAX,
2061 req->sync.flags & IORING_FSYNC_DATASYNC);
2063 req_set_fail_links(req);
2064 io_cqring_add_event(req, ret);
2065 io_put_req_find_next(req, &nxt);
2067 *workptr = &nxt->work;
2070 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2071 bool force_nonblock)
2073 struct io_wq_work *work, *old_work;
2075 /* fsync always requires a blocking context */
2076 if (force_nonblock) {
2078 req->work.func = io_fsync_finish;
2082 work = old_work = &req->work;
2083 io_fsync_finish(&work);
2084 if (work && work != old_work)
2085 *nxt = container_of(work, struct io_kiocb, work);
2089 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2091 struct io_ring_ctx *ctx = req->ctx;
2096 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2098 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2101 req->sync.off = READ_ONCE(sqe->off);
2102 req->sync.len = READ_ONCE(sqe->len);
2103 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2107 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2109 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2110 struct io_kiocb *nxt = NULL;
2113 if (io_req_cancelled(req))
2116 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
2119 req_set_fail_links(req);
2120 io_cqring_add_event(req, ret);
2121 io_put_req_find_next(req, &nxt);
2123 *workptr = &nxt->work;
2126 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2127 bool force_nonblock)
2129 struct io_wq_work *work, *old_work;
2131 /* sync_file_range always requires a blocking context */
2132 if (force_nonblock) {
2134 req->work.func = io_sync_file_range_finish;
2138 work = old_work = &req->work;
2139 io_sync_file_range_finish(&work);
2140 if (work && work != old_work)
2141 *nxt = container_of(work, struct io_kiocb, work);
2145 #if defined(CONFIG_NET)
2146 static void io_sendrecv_async(struct io_wq_work **workptr)
2148 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2149 struct iovec *iov = NULL;
2151 if (req->io->rw.iov != req->io->rw.fast_iov)
2152 iov = req->io->msg.iov;
2153 io_wq_submit_work(workptr);
2158 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2160 #if defined(CONFIG_NET)
2161 struct io_sr_msg *sr = &req->sr_msg;
2162 struct io_async_ctx *io = req->io;
2164 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2165 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2170 io->msg.iov = io->msg.fast_iov;
2171 return sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2178 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2179 bool force_nonblock)
2181 #if defined(CONFIG_NET)
2182 struct io_async_msghdr *kmsg = NULL;
2183 struct socket *sock;
2186 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2189 sock = sock_from_file(req->file, &ret);
2191 struct io_async_ctx io;
2192 struct sockaddr_storage addr;
2196 kmsg = &req->io->msg;
2197 kmsg->msg.msg_name = &addr;
2198 /* if iov is set, it's allocated already */
2200 kmsg->iov = kmsg->fast_iov;
2201 kmsg->msg.msg_iter.iov = kmsg->iov;
2203 struct io_sr_msg *sr = &req->sr_msg;
2206 kmsg->msg.msg_name = &addr;
2208 io.msg.iov = io.msg.fast_iov;
2209 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
2210 sr->msg_flags, &io.msg.iov);
2215 flags = req->sr_msg.msg_flags;
2216 if (flags & MSG_DONTWAIT)
2217 req->flags |= REQ_F_NOWAIT;
2218 else if (force_nonblock)
2219 flags |= MSG_DONTWAIT;
2221 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
2222 if (force_nonblock && ret == -EAGAIN) {
2225 if (io_alloc_async_ctx(req))
2227 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2228 req->work.func = io_sendrecv_async;
2231 if (ret == -ERESTARTSYS)
2235 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2237 io_cqring_add_event(req, ret);
2239 req_set_fail_links(req);
2240 io_put_req_find_next(req, nxt);
2247 static int io_recvmsg_prep(struct io_kiocb *req,
2248 const struct io_uring_sqe *sqe)
2250 #if defined(CONFIG_NET)
2251 struct io_sr_msg *sr = &req->sr_msg;
2252 struct io_async_ctx *io = req->io;
2254 sr->msg_flags = READ_ONCE(sqe->msg_flags);
2255 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
2260 io->msg.iov = io->msg.fast_iov;
2261 return recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
2262 &io->msg.uaddr, &io->msg.iov);
2268 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2269 bool force_nonblock)
2271 #if defined(CONFIG_NET)
2272 struct io_async_msghdr *kmsg = NULL;
2273 struct socket *sock;
2276 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2279 sock = sock_from_file(req->file, &ret);
2281 struct io_async_ctx io;
2282 struct sockaddr_storage addr;
2286 kmsg = &req->io->msg;
2287 kmsg->msg.msg_name = &addr;
2288 /* if iov is set, it's allocated already */
2290 kmsg->iov = kmsg->fast_iov;
2291 kmsg->msg.msg_iter.iov = kmsg->iov;
2293 struct io_sr_msg *sr = &req->sr_msg;
2296 kmsg->msg.msg_name = &addr;
2298 io.msg.iov = io.msg.fast_iov;
2299 ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg,
2300 sr->msg_flags, &io.msg.uaddr,
2306 flags = req->sr_msg.msg_flags;
2307 if (flags & MSG_DONTWAIT)
2308 req->flags |= REQ_F_NOWAIT;
2309 else if (force_nonblock)
2310 flags |= MSG_DONTWAIT;
2312 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
2313 kmsg->uaddr, flags);
2314 if (force_nonblock && ret == -EAGAIN) {
2317 if (io_alloc_async_ctx(req))
2319 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2320 req->work.func = io_sendrecv_async;
2323 if (ret == -ERESTARTSYS)
2327 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2329 io_cqring_add_event(req, ret);
2331 req_set_fail_links(req);
2332 io_put_req_find_next(req, nxt);
2339 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2341 #if defined(CONFIG_NET)
2342 struct io_accept *accept = &req->accept;
2344 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2346 if (sqe->ioprio || sqe->len || sqe->buf_index)
2349 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
2350 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2351 accept->flags = READ_ONCE(sqe->accept_flags);
2358 #if defined(CONFIG_NET)
2359 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2360 bool force_nonblock)
2362 struct io_accept *accept = &req->accept;
2363 unsigned file_flags;
2366 file_flags = force_nonblock ? O_NONBLOCK : 0;
2367 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
2368 accept->addr_len, accept->flags);
2369 if (ret == -EAGAIN && force_nonblock)
2371 if (ret == -ERESTARTSYS)
2374 req_set_fail_links(req);
2375 io_cqring_add_event(req, ret);
2376 io_put_req_find_next(req, nxt);
2380 static void io_accept_finish(struct io_wq_work **workptr)
2382 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2383 struct io_kiocb *nxt = NULL;
2385 if (io_req_cancelled(req))
2387 __io_accept(req, &nxt, false);
2389 *workptr = &nxt->work;
2393 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2394 bool force_nonblock)
2396 #if defined(CONFIG_NET)
2399 ret = __io_accept(req, nxt, force_nonblock);
2400 if (ret == -EAGAIN && force_nonblock) {
2401 req->work.func = io_accept_finish;
2402 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2412 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2414 #if defined(CONFIG_NET)
2415 struct io_connect *conn = &req->connect;
2416 struct io_async_ctx *io = req->io;
2418 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2420 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2423 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
2424 conn->addr_len = READ_ONCE(sqe->addr2);
2429 return move_addr_to_kernel(conn->addr, conn->addr_len,
2430 &io->connect.address);
2436 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
2437 bool force_nonblock)
2439 #if defined(CONFIG_NET)
2440 struct io_async_ctx __io, *io;
2441 unsigned file_flags;
2447 ret = move_addr_to_kernel(req->connect.addr,
2448 req->connect.addr_len,
2449 &__io.connect.address);
2455 file_flags = force_nonblock ? O_NONBLOCK : 0;
2457 ret = __sys_connect_file(req->file, &io->connect.address,
2458 req->connect.addr_len, file_flags);
2459 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2462 if (io_alloc_async_ctx(req)) {
2466 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
2469 if (ret == -ERESTARTSYS)
2473 req_set_fail_links(req);
2474 io_cqring_add_event(req, ret);
2475 io_put_req_find_next(req, nxt);
2482 static void io_poll_remove_one(struct io_kiocb *req)
2484 struct io_poll_iocb *poll = &req->poll;
2486 spin_lock(&poll->head->lock);
2487 WRITE_ONCE(poll->canceled, true);
2488 if (!list_empty(&poll->wait.entry)) {
2489 list_del_init(&poll->wait.entry);
2490 io_queue_async_work(req);
2492 spin_unlock(&poll->head->lock);
2493 hash_del(&req->hash_node);
2496 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2498 struct hlist_node *tmp;
2499 struct io_kiocb *req;
2502 spin_lock_irq(&ctx->completion_lock);
2503 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2504 struct hlist_head *list;
2506 list = &ctx->cancel_hash[i];
2507 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2508 io_poll_remove_one(req);
2510 spin_unlock_irq(&ctx->completion_lock);
2513 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2515 struct hlist_head *list;
2516 struct io_kiocb *req;
2518 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2519 hlist_for_each_entry(req, list, hash_node) {
2520 if (sqe_addr == req->user_data) {
2521 io_poll_remove_one(req);
2529 static int io_poll_remove_prep(struct io_kiocb *req,
2530 const struct io_uring_sqe *sqe)
2532 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2534 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2538 req->poll.addr = READ_ONCE(sqe->addr);
2543 * Find a running poll command that matches one specified in sqe->addr,
2544 * and remove it if found.
2546 static int io_poll_remove(struct io_kiocb *req)
2548 struct io_ring_ctx *ctx = req->ctx;
2552 addr = req->poll.addr;
2553 spin_lock_irq(&ctx->completion_lock);
2554 ret = io_poll_cancel(ctx, addr);
2555 spin_unlock_irq(&ctx->completion_lock);
2557 io_cqring_add_event(req, ret);
2559 req_set_fail_links(req);
2564 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2566 struct io_ring_ctx *ctx = req->ctx;
2568 req->poll.done = true;
2570 io_cqring_fill_event(req, error);
2572 io_cqring_fill_event(req, mangle_poll(mask));
2573 io_commit_cqring(ctx);
2576 static void io_poll_complete_work(struct io_wq_work **workptr)
2578 struct io_wq_work *work = *workptr;
2579 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2580 struct io_poll_iocb *poll = &req->poll;
2581 struct poll_table_struct pt = { ._key = poll->events };
2582 struct io_ring_ctx *ctx = req->ctx;
2583 struct io_kiocb *nxt = NULL;
2587 if (work->flags & IO_WQ_WORK_CANCEL) {
2588 WRITE_ONCE(poll->canceled, true);
2590 } else if (READ_ONCE(poll->canceled)) {
2594 if (ret != -ECANCELED)
2595 mask = vfs_poll(poll->file, &pt) & poll->events;
2598 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2599 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2600 * synchronize with them. In the cancellation case the list_del_init
2601 * itself is not actually needed, but harmless so we keep it in to
2602 * avoid further branches in the fast path.
2604 spin_lock_irq(&ctx->completion_lock);
2605 if (!mask && ret != -ECANCELED) {
2606 add_wait_queue(poll->head, &poll->wait);
2607 spin_unlock_irq(&ctx->completion_lock);
2610 hash_del(&req->hash_node);
2611 io_poll_complete(req, mask, ret);
2612 spin_unlock_irq(&ctx->completion_lock);
2614 io_cqring_ev_posted(ctx);
2617 req_set_fail_links(req);
2618 io_put_req_find_next(req, &nxt);
2620 *workptr = &nxt->work;
2623 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2626 struct io_poll_iocb *poll = wait->private;
2627 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2628 struct io_ring_ctx *ctx = req->ctx;
2629 __poll_t mask = key_to_poll(key);
2630 unsigned long flags;
2632 /* for instances that support it check for an event match first: */
2633 if (mask && !(mask & poll->events))
2636 list_del_init(&poll->wait.entry);
2639 * Run completion inline if we can. We're using trylock here because
2640 * we are violating the completion_lock -> poll wq lock ordering.
2641 * If we have a link timeout we're going to need the completion_lock
2642 * for finalizing the request, mark us as having grabbed that already.
2644 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2645 hash_del(&req->hash_node);
2646 io_poll_complete(req, mask, 0);
2647 req->flags |= REQ_F_COMP_LOCKED;
2649 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2651 io_cqring_ev_posted(ctx);
2653 io_queue_async_work(req);
2659 struct io_poll_table {
2660 struct poll_table_struct pt;
2661 struct io_kiocb *req;
2665 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2666 struct poll_table_struct *p)
2668 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2670 if (unlikely(pt->req->poll.head)) {
2671 pt->error = -EINVAL;
2676 pt->req->poll.head = head;
2677 add_wait_queue(head, &pt->req->poll.wait);
2680 static void io_poll_req_insert(struct io_kiocb *req)
2682 struct io_ring_ctx *ctx = req->ctx;
2683 struct hlist_head *list;
2685 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2686 hlist_add_head(&req->hash_node, list);
2689 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2691 struct io_poll_iocb *poll = &req->poll;
2694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2696 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2701 events = READ_ONCE(sqe->poll_events);
2702 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2706 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
2708 struct io_poll_iocb *poll = &req->poll;
2709 struct io_ring_ctx *ctx = req->ctx;
2710 struct io_poll_table ipt;
2711 bool cancel = false;
2714 INIT_IO_WORK(&req->work, io_poll_complete_work);
2715 INIT_HLIST_NODE(&req->hash_node);
2719 poll->canceled = false;
2721 ipt.pt._qproc = io_poll_queue_proc;
2722 ipt.pt._key = poll->events;
2724 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2726 /* initialized the list so that we can do list_empty checks */
2727 INIT_LIST_HEAD(&poll->wait.entry);
2728 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
2729 poll->wait.private = poll;
2731 INIT_LIST_HEAD(&req->list);
2733 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2735 spin_lock_irq(&ctx->completion_lock);
2736 if (likely(poll->head)) {
2737 spin_lock(&poll->head->lock);
2738 if (unlikely(list_empty(&poll->wait.entry))) {
2744 if (mask || ipt.error)
2745 list_del_init(&poll->wait.entry);
2747 WRITE_ONCE(poll->canceled, true);
2748 else if (!poll->done) /* actually waiting for an event */
2749 io_poll_req_insert(req);
2750 spin_unlock(&poll->head->lock);
2752 if (mask) { /* no async, we'd stolen it */
2754 io_poll_complete(req, mask, 0);
2756 spin_unlock_irq(&ctx->completion_lock);
2759 io_cqring_ev_posted(ctx);
2760 io_put_req_find_next(req, nxt);
2765 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2767 struct io_timeout_data *data = container_of(timer,
2768 struct io_timeout_data, timer);
2769 struct io_kiocb *req = data->req;
2770 struct io_ring_ctx *ctx = req->ctx;
2771 unsigned long flags;
2773 atomic_inc(&ctx->cq_timeouts);
2775 spin_lock_irqsave(&ctx->completion_lock, flags);
2777 * We could be racing with timeout deletion. If the list is empty,
2778 * then timeout lookup already found it and will be handling it.
2780 if (!list_empty(&req->list)) {
2781 struct io_kiocb *prev;
2784 * Adjust the reqs sequence before the current one because it
2785 * will consume a slot in the cq_ring and the cq_tail
2786 * pointer will be increased, otherwise other timeout reqs may
2787 * return in advance without waiting for enough wait_nr.
2790 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2792 list_del_init(&req->list);
2795 io_cqring_fill_event(req, -ETIME);
2796 io_commit_cqring(ctx);
2797 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2799 io_cqring_ev_posted(ctx);
2800 req_set_fail_links(req);
2802 return HRTIMER_NORESTART;
2805 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2807 struct io_kiocb *req;
2810 list_for_each_entry(req, &ctx->timeout_list, list) {
2811 if (user_data == req->user_data) {
2812 list_del_init(&req->list);
2821 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2825 req_set_fail_links(req);
2826 io_cqring_fill_event(req, -ECANCELED);
2831 static int io_timeout_remove_prep(struct io_kiocb *req,
2832 const struct io_uring_sqe *sqe)
2834 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2836 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2839 req->timeout.addr = READ_ONCE(sqe->addr);
2840 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
2841 if (req->timeout.flags)
2848 * Remove or update an existing timeout command
2850 static int io_timeout_remove(struct io_kiocb *req)
2852 struct io_ring_ctx *ctx = req->ctx;
2855 spin_lock_irq(&ctx->completion_lock);
2856 ret = io_timeout_cancel(ctx, req->timeout.addr);
2858 io_cqring_fill_event(req, ret);
2859 io_commit_cqring(ctx);
2860 spin_unlock_irq(&ctx->completion_lock);
2861 io_cqring_ev_posted(ctx);
2863 req_set_fail_links(req);
2868 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2869 bool is_timeout_link)
2871 struct io_timeout_data *data;
2874 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2876 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2878 if (sqe->off && is_timeout_link)
2880 flags = READ_ONCE(sqe->timeout_flags);
2881 if (flags & ~IORING_TIMEOUT_ABS)
2884 req->timeout.count = READ_ONCE(sqe->off);
2886 if (!req->io && io_alloc_async_ctx(req))
2889 data = &req->io->timeout;
2891 req->flags |= REQ_F_TIMEOUT;
2893 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2896 if (flags & IORING_TIMEOUT_ABS)
2897 data->mode = HRTIMER_MODE_ABS;
2899 data->mode = HRTIMER_MODE_REL;
2901 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2905 static int io_timeout(struct io_kiocb *req)
2908 struct io_ring_ctx *ctx = req->ctx;
2909 struct io_timeout_data *data;
2910 struct list_head *entry;
2913 data = &req->io->timeout;
2916 * sqe->off holds how many events that need to occur for this
2917 * timeout event to be satisfied. If it isn't set, then this is
2918 * a pure timeout request, sequence isn't used.
2920 count = req->timeout.count;
2922 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2923 spin_lock_irq(&ctx->completion_lock);
2924 entry = ctx->timeout_list.prev;
2928 req->sequence = ctx->cached_sq_head + count - 1;
2929 data->seq_offset = count;
2932 * Insertion sort, ensuring the first entry in the list is always
2933 * the one we need first.
2935 spin_lock_irq(&ctx->completion_lock);
2936 list_for_each_prev(entry, &ctx->timeout_list) {
2937 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2938 unsigned nxt_sq_head;
2939 long long tmp, tmp_nxt;
2940 u32 nxt_offset = nxt->io->timeout.seq_offset;
2942 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2946 * Since cached_sq_head + count - 1 can overflow, use type long
2949 tmp = (long long)ctx->cached_sq_head + count - 1;
2950 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2951 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2954 * cached_sq_head may overflow, and it will never overflow twice
2955 * once there is some timeout req still be valid.
2957 if (ctx->cached_sq_head < nxt_sq_head)
2964 * Sequence of reqs after the insert one and itself should
2965 * be adjusted because each timeout req consumes a slot.
2970 req->sequence -= span;
2972 list_add(&req->list, entry);
2973 data->timer.function = io_timeout_fn;
2974 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2975 spin_unlock_irq(&ctx->completion_lock);
2979 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2981 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2983 return req->user_data == (unsigned long) data;
2986 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2988 enum io_wq_cancel cancel_ret;
2991 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2992 switch (cancel_ret) {
2993 case IO_WQ_CANCEL_OK:
2996 case IO_WQ_CANCEL_RUNNING:
2999 case IO_WQ_CANCEL_NOTFOUND:
3007 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3008 struct io_kiocb *req, __u64 sqe_addr,
3009 struct io_kiocb **nxt, int success_ret)
3011 unsigned long flags;
3014 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3015 if (ret != -ENOENT) {
3016 spin_lock_irqsave(&ctx->completion_lock, flags);
3020 spin_lock_irqsave(&ctx->completion_lock, flags);
3021 ret = io_timeout_cancel(ctx, sqe_addr);
3024 ret = io_poll_cancel(ctx, sqe_addr);
3028 io_cqring_fill_event(req, ret);
3029 io_commit_cqring(ctx);
3030 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3031 io_cqring_ev_posted(ctx);
3034 req_set_fail_links(req);
3035 io_put_req_find_next(req, nxt);
3038 static int io_async_cancel_prep(struct io_kiocb *req,
3039 const struct io_uring_sqe *sqe)
3041 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3043 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
3047 req->cancel.addr = READ_ONCE(sqe->addr);
3051 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
3053 struct io_ring_ctx *ctx = req->ctx;
3055 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
3059 static int io_req_defer_prep(struct io_kiocb *req,
3060 const struct io_uring_sqe *sqe)
3064 switch (req->opcode) {
3067 case IORING_OP_READV:
3068 case IORING_OP_READ_FIXED:
3069 ret = io_read_prep(req, sqe, true);
3071 case IORING_OP_WRITEV:
3072 case IORING_OP_WRITE_FIXED:
3073 ret = io_write_prep(req, sqe, true);
3075 case IORING_OP_POLL_ADD:
3076 ret = io_poll_add_prep(req, sqe);
3078 case IORING_OP_POLL_REMOVE:
3079 ret = io_poll_remove_prep(req, sqe);
3081 case IORING_OP_FSYNC:
3082 ret = io_prep_fsync(req, sqe);
3084 case IORING_OP_SYNC_FILE_RANGE:
3085 ret = io_prep_sfr(req, sqe);
3087 case IORING_OP_SENDMSG:
3088 ret = io_sendmsg_prep(req, sqe);
3090 case IORING_OP_RECVMSG:
3091 ret = io_recvmsg_prep(req, sqe);
3093 case IORING_OP_CONNECT:
3094 ret = io_connect_prep(req, sqe);
3096 case IORING_OP_TIMEOUT:
3097 ret = io_timeout_prep(req, sqe, false);
3099 case IORING_OP_TIMEOUT_REMOVE:
3100 ret = io_timeout_remove_prep(req, sqe);
3102 case IORING_OP_ASYNC_CANCEL:
3103 ret = io_async_cancel_prep(req, sqe);
3105 case IORING_OP_LINK_TIMEOUT:
3106 ret = io_timeout_prep(req, sqe, true);
3108 case IORING_OP_ACCEPT:
3109 ret = io_accept_prep(req, sqe);
3112 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
3121 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3123 struct io_ring_ctx *ctx = req->ctx;
3126 /* Still need defer if there is pending req in defer list. */
3127 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
3130 if (!req->io && io_alloc_async_ctx(req))
3133 ret = io_req_defer_prep(req, sqe);
3137 spin_lock_irq(&ctx->completion_lock);
3138 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
3139 spin_unlock_irq(&ctx->completion_lock);
3143 trace_io_uring_defer(ctx, req, req->user_data);
3144 list_add_tail(&req->list, &ctx->defer_list);
3145 spin_unlock_irq(&ctx->completion_lock);
3146 return -EIOCBQUEUED;
3149 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3150 struct io_kiocb **nxt, bool force_nonblock)
3152 struct io_ring_ctx *ctx = req->ctx;
3155 switch (req->opcode) {
3159 case IORING_OP_READV:
3160 case IORING_OP_READ_FIXED:
3162 ret = io_read_prep(req, sqe, force_nonblock);
3166 ret = io_read(req, nxt, force_nonblock);
3168 case IORING_OP_WRITEV:
3169 case IORING_OP_WRITE_FIXED:
3171 ret = io_write_prep(req, sqe, force_nonblock);
3175 ret = io_write(req, nxt, force_nonblock);
3177 case IORING_OP_FSYNC:
3179 ret = io_prep_fsync(req, sqe);
3183 ret = io_fsync(req, nxt, force_nonblock);
3185 case IORING_OP_POLL_ADD:
3187 ret = io_poll_add_prep(req, sqe);
3191 ret = io_poll_add(req, nxt);
3193 case IORING_OP_POLL_REMOVE:
3195 ret = io_poll_remove_prep(req, sqe);
3199 ret = io_poll_remove(req);
3201 case IORING_OP_SYNC_FILE_RANGE:
3203 ret = io_prep_sfr(req, sqe);
3207 ret = io_sync_file_range(req, nxt, force_nonblock);
3209 case IORING_OP_SENDMSG:
3211 ret = io_sendmsg_prep(req, sqe);
3215 ret = io_sendmsg(req, nxt, force_nonblock);
3217 case IORING_OP_RECVMSG:
3219 ret = io_recvmsg_prep(req, sqe);
3223 ret = io_recvmsg(req, nxt, force_nonblock);
3225 case IORING_OP_TIMEOUT:
3227 ret = io_timeout_prep(req, sqe, false);
3231 ret = io_timeout(req);
3233 case IORING_OP_TIMEOUT_REMOVE:
3235 ret = io_timeout_remove_prep(req, sqe);
3239 ret = io_timeout_remove(req);
3241 case IORING_OP_ACCEPT:
3243 ret = io_accept_prep(req, sqe);
3247 ret = io_accept(req, nxt, force_nonblock);
3249 case IORING_OP_CONNECT:
3251 ret = io_connect_prep(req, sqe);
3255 ret = io_connect(req, nxt, force_nonblock);
3257 case IORING_OP_ASYNC_CANCEL:
3259 ret = io_async_cancel_prep(req, sqe);
3263 ret = io_async_cancel(req, nxt);
3273 if (ctx->flags & IORING_SETUP_IOPOLL) {
3274 if (req->result == -EAGAIN)
3277 io_iopoll_req_issued(req);
3283 static void io_link_work_cb(struct io_wq_work **workptr)
3285 struct io_wq_work *work = *workptr;
3286 struct io_kiocb *link = work->data;
3288 io_queue_linked_timeout(link);
3289 work->func = io_wq_submit_work;
3292 static void io_wq_submit_work(struct io_wq_work **workptr)
3294 struct io_wq_work *work = *workptr;
3295 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3296 struct io_kiocb *nxt = NULL;
3299 if (work->flags & IO_WQ_WORK_CANCEL)
3303 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3304 req->in_async = true;
3306 ret = io_issue_sqe(req, NULL, &nxt, false);
3308 * We can get EAGAIN for polled IO even though we're
3309 * forcing a sync submission from here, since we can't
3310 * wait for request slots on the block side.
3318 /* drop submission reference */
3322 req_set_fail_links(req);
3323 io_cqring_add_event(req, ret);
3327 /* if a dependent link is ready, pass it back */
3329 struct io_kiocb *link;
3331 io_prep_async_work(nxt, &link);
3332 *workptr = &nxt->work;
3334 nxt->work.flags |= IO_WQ_WORK_CB;
3335 nxt->work.func = io_link_work_cb;
3336 nxt->work.data = link;
3341 static bool io_req_op_valid(int op)
3343 return op >= IORING_OP_NOP && op < IORING_OP_LAST;
3346 static int io_req_needs_file(struct io_kiocb *req)
3348 switch (req->opcode) {
3350 case IORING_OP_POLL_REMOVE:
3351 case IORING_OP_TIMEOUT:
3352 case IORING_OP_TIMEOUT_REMOVE:
3353 case IORING_OP_ASYNC_CANCEL:
3354 case IORING_OP_LINK_TIMEOUT:
3357 if (io_req_op_valid(req->opcode))
3363 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3366 struct fixed_file_table *table;
3368 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3369 return table->files[index & IORING_FILE_TABLE_MASK];
3372 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
3373 const struct io_uring_sqe *sqe)
3375 struct io_ring_ctx *ctx = req->ctx;
3379 flags = READ_ONCE(sqe->flags);
3380 fd = READ_ONCE(sqe->fd);
3382 if (flags & IOSQE_IO_DRAIN)
3383 req->flags |= REQ_F_IO_DRAIN;
3385 ret = io_req_needs_file(req);
3389 if (flags & IOSQE_FIXED_FILE) {
3390 if (unlikely(!ctx->file_table ||
3391 (unsigned) fd >= ctx->nr_user_files))
3393 fd = array_index_nospec(fd, ctx->nr_user_files);
3394 req->file = io_file_from_index(ctx, fd);
3397 req->flags |= REQ_F_FIXED_FILE;
3399 if (req->needs_fixed_file)
3401 trace_io_uring_file_get(ctx, fd);
3402 req->file = io_file_get(state, fd);
3403 if (unlikely(!req->file))
3410 static int io_grab_files(struct io_kiocb *req)
3413 struct io_ring_ctx *ctx = req->ctx;
3416 spin_lock_irq(&ctx->inflight_lock);
3418 * We use the f_ops->flush() handler to ensure that we can flush
3419 * out work accessing these files if the fd is closed. Check if
3420 * the fd has changed since we started down this path, and disallow
3421 * this operation if it has.
3423 if (fcheck(req->ring_fd) == req->ring_file) {
3424 list_add(&req->inflight_entry, &ctx->inflight_list);
3425 req->flags |= REQ_F_INFLIGHT;
3426 req->work.files = current->files;
3429 spin_unlock_irq(&ctx->inflight_lock);
3435 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3437 struct io_timeout_data *data = container_of(timer,
3438 struct io_timeout_data, timer);
3439 struct io_kiocb *req = data->req;
3440 struct io_ring_ctx *ctx = req->ctx;
3441 struct io_kiocb *prev = NULL;
3442 unsigned long flags;
3444 spin_lock_irqsave(&ctx->completion_lock, flags);
3447 * We don't expect the list to be empty, that will only happen if we
3448 * race with the completion of the linked work.
3450 if (!list_empty(&req->link_list)) {
3451 prev = list_entry(req->link_list.prev, struct io_kiocb,
3453 if (refcount_inc_not_zero(&prev->refs)) {
3454 list_del_init(&req->link_list);
3455 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3460 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3463 req_set_fail_links(prev);
3464 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3468 io_cqring_add_event(req, -ETIME);
3471 return HRTIMER_NORESTART;
3474 static void io_queue_linked_timeout(struct io_kiocb *req)
3476 struct io_ring_ctx *ctx = req->ctx;
3479 * If the list is now empty, then our linked request finished before
3480 * we got a chance to setup the timer
3482 spin_lock_irq(&ctx->completion_lock);
3483 if (!list_empty(&req->link_list)) {
3484 struct io_timeout_data *data = &req->io->timeout;
3486 data->timer.function = io_link_timeout_fn;
3487 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3490 spin_unlock_irq(&ctx->completion_lock);
3492 /* drop submission reference */
3496 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3498 struct io_kiocb *nxt;
3500 if (!(req->flags & REQ_F_LINK))
3503 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3505 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
3508 req->flags |= REQ_F_LINK_TIMEOUT;
3512 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3514 struct io_kiocb *linked_timeout;
3515 struct io_kiocb *nxt = NULL;
3519 linked_timeout = io_prep_linked_timeout(req);
3521 ret = io_issue_sqe(req, sqe, &nxt, true);
3524 * We async punt it if the file wasn't marked NOWAIT, or if the file
3525 * doesn't support non-blocking read/write attempts
3527 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3528 (req->flags & REQ_F_MUST_PUNT))) {
3529 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3530 ret = io_grab_files(req);
3536 * Queued up for async execution, worker will release
3537 * submit reference when the iocb is actually submitted.
3539 io_queue_async_work(req);
3544 /* drop submission reference */
3547 if (linked_timeout) {
3549 io_queue_linked_timeout(linked_timeout);
3551 io_put_req(linked_timeout);
3554 /* and drop final reference, if we failed */
3556 io_cqring_add_event(req, ret);
3557 req_set_fail_links(req);
3568 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3572 if (unlikely(req->ctx->drain_next)) {
3573 req->flags |= REQ_F_IO_DRAIN;
3574 req->ctx->drain_next = false;
3576 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3578 ret = io_req_defer(req, sqe);
3580 if (ret != -EIOCBQUEUED) {
3581 io_cqring_add_event(req, ret);
3582 req_set_fail_links(req);
3583 io_double_put_req(req);
3586 __io_queue_sqe(req, sqe);
3589 static inline void io_queue_link_head(struct io_kiocb *req)
3591 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3592 io_cqring_add_event(req, -ECANCELED);
3593 io_double_put_req(req);
3595 io_queue_sqe(req, NULL);
3598 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3601 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
3602 struct io_submit_state *state, struct io_kiocb **link)
3604 struct io_ring_ctx *ctx = req->ctx;
3607 /* enforce forwards compatibility on users */
3608 if (unlikely(sqe->flags & ~SQE_VALID_FLAGS)) {
3613 ret = io_req_set_file(state, req, sqe);
3614 if (unlikely(ret)) {
3616 io_cqring_add_event(req, ret);
3617 io_double_put_req(req);
3622 * If we already have a head request, queue this one for async
3623 * submittal once the head completes. If we don't have a head but
3624 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3625 * submitted sync once the chain is complete. If none of those
3626 * conditions are true (normal request), then just queue it.
3629 struct io_kiocb *prev = *link;
3631 if (sqe->flags & IOSQE_IO_DRAIN)
3632 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3634 if (sqe->flags & IOSQE_IO_HARDLINK)
3635 req->flags |= REQ_F_HARDLINK;
3637 if (io_alloc_async_ctx(req)) {
3642 ret = io_req_defer_prep(req, sqe);
3644 /* fail even hard links since we don't submit */
3645 prev->flags |= REQ_F_FAIL_LINK;
3648 trace_io_uring_link(ctx, req, prev);
3649 list_add_tail(&req->link_list, &prev->link_list);
3650 } else if (sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
3651 req->flags |= REQ_F_LINK;
3652 if (sqe->flags & IOSQE_IO_HARDLINK)
3653 req->flags |= REQ_F_HARDLINK;
3655 INIT_LIST_HEAD(&req->link_list);
3656 ret = io_req_defer_prep(req, sqe);
3658 req->flags |= REQ_F_FAIL_LINK;
3661 io_queue_sqe(req, sqe);
3668 * Batched submission is done, ensure local IO is flushed out.
3670 static void io_submit_state_end(struct io_submit_state *state)
3672 blk_finish_plug(&state->plug);
3674 if (state->free_reqs)
3675 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3676 &state->reqs[state->cur_req]);
3680 * Start submission side cache.
3682 static void io_submit_state_start(struct io_submit_state *state,
3683 unsigned int max_ios)
3685 blk_start_plug(&state->plug);
3686 state->free_reqs = 0;
3688 state->ios_left = max_ios;
3691 static void io_commit_sqring(struct io_ring_ctx *ctx)
3693 struct io_rings *rings = ctx->rings;
3695 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3697 * Ensure any loads from the SQEs are done at this point,
3698 * since once we write the new head, the application could
3699 * write new data to them.
3701 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3706 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
3707 * that is mapped by userspace. This means that care needs to be taken to
3708 * ensure that reads are stable, as we cannot rely on userspace always
3709 * being a good citizen. If members of the sqe are validated and then later
3710 * used, it's important that those reads are done through READ_ONCE() to
3711 * prevent a re-load down the line.
3713 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
3714 const struct io_uring_sqe **sqe_ptr)
3716 struct io_rings *rings = ctx->rings;
3717 u32 *sq_array = ctx->sq_array;
3721 * The cached sq head (or cq tail) serves two purposes:
3723 * 1) allows us to batch the cost of updating the user visible
3725 * 2) allows the kernel side to track the head on its own, even
3726 * though the application is the one updating it.
3728 head = ctx->cached_sq_head;
3729 /* make sure SQ entry isn't read before tail */
3730 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3733 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3734 if (likely(head < ctx->sq_entries)) {
3736 * All io need record the previous position, if LINK vs DARIN,
3737 * it can be used to mark the position of the first IO in the
3740 req->sequence = ctx->cached_sq_head;
3741 *sqe_ptr = &ctx->sq_sqes[head];
3742 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
3743 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
3744 ctx->cached_sq_head++;
3748 /* drop invalid entries */
3749 ctx->cached_sq_head++;
3750 ctx->cached_sq_dropped++;
3751 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3755 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3756 struct file *ring_file, int ring_fd,
3757 struct mm_struct **mm, bool async)
3759 struct io_submit_state state, *statep = NULL;
3760 struct io_kiocb *link = NULL;
3761 int i, submitted = 0;
3762 bool mm_fault = false;
3764 /* if we have a backlog and couldn't flush it all, return BUSY */
3765 if (!list_empty(&ctx->cq_overflow_list) &&
3766 !io_cqring_overflow_flush(ctx, false))
3769 if (nr > IO_PLUG_THRESHOLD) {
3770 io_submit_state_start(&state, nr);
3774 for (i = 0; i < nr; i++) {
3775 const struct io_uring_sqe *sqe;
3776 struct io_kiocb *req;
3777 unsigned int sqe_flags;
3779 req = io_get_req(ctx, statep);
3780 if (unlikely(!req)) {
3782 submitted = -EAGAIN;
3785 if (!io_get_sqring(ctx, req, &sqe)) {
3790 if (io_req_needs_user(req) && !*mm) {
3791 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3793 use_mm(ctx->sqo_mm);
3799 sqe_flags = sqe->flags;
3801 req->ring_file = ring_file;
3802 req->ring_fd = ring_fd;
3803 req->has_user = *mm != NULL;
3804 req->in_async = async;
3805 req->needs_fixed_file = async;
3806 trace_io_uring_submit_sqe(ctx, req->user_data, true, async);
3807 if (!io_submit_sqe(req, sqe, statep, &link))
3810 * If previous wasn't linked and we have a linked command,
3811 * that's the end of the chain. Submit the previous link.
3813 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) && link) {
3814 io_queue_link_head(link);
3820 io_queue_link_head(link);
3822 io_submit_state_end(&state);
3824 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3825 io_commit_sqring(ctx);
3830 static int io_sq_thread(void *data)
3832 struct io_ring_ctx *ctx = data;
3833 struct mm_struct *cur_mm = NULL;
3834 const struct cred *old_cred;
3835 mm_segment_t old_fs;
3838 unsigned long timeout;
3841 complete(&ctx->completions[1]);
3845 old_cred = override_creds(ctx->creds);
3847 ret = timeout = inflight = 0;
3848 while (!kthread_should_park()) {
3849 unsigned int to_submit;
3852 unsigned nr_events = 0;
3854 if (ctx->flags & IORING_SETUP_IOPOLL) {
3856 * inflight is the count of the maximum possible
3857 * entries we submitted, but it can be smaller
3858 * if we dropped some of them. If we don't have
3859 * poll entries available, then we know that we
3860 * have nothing left to poll for. Reset the
3861 * inflight count to zero in that case.
3863 mutex_lock(&ctx->uring_lock);
3864 if (!list_empty(&ctx->poll_list))
3865 __io_iopoll_check(ctx, &nr_events, 0);
3868 mutex_unlock(&ctx->uring_lock);
3871 * Normal IO, just pretend everything completed.
3872 * We don't have to poll completions for that.
3874 nr_events = inflight;
3877 inflight -= nr_events;
3879 timeout = jiffies + ctx->sq_thread_idle;
3882 to_submit = io_sqring_entries(ctx);
3885 * If submit got -EBUSY, flag us as needing the application
3886 * to enter the kernel to reap and flush events.
3888 if (!to_submit || ret == -EBUSY) {
3890 * We're polling. If we're within the defined idle
3891 * period, then let us spin without work before going
3892 * to sleep. The exception is if we got EBUSY doing
3893 * more IO, we should wait for the application to
3894 * reap events and wake us up.
3897 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3903 * Drop cur_mm before scheduling, we can't hold it for
3904 * long periods (or over schedule()). Do this before
3905 * adding ourselves to the waitqueue, as the unuse/drop
3914 prepare_to_wait(&ctx->sqo_wait, &wait,
3915 TASK_INTERRUPTIBLE);
3917 /* Tell userspace we may need a wakeup call */
3918 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3919 /* make sure to read SQ tail after writing flags */
3922 to_submit = io_sqring_entries(ctx);
3923 if (!to_submit || ret == -EBUSY) {
3924 if (kthread_should_park()) {
3925 finish_wait(&ctx->sqo_wait, &wait);
3928 if (signal_pending(current))
3929 flush_signals(current);
3931 finish_wait(&ctx->sqo_wait, &wait);
3933 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3936 finish_wait(&ctx->sqo_wait, &wait);
3938 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3941 to_submit = min(to_submit, ctx->sq_entries);
3942 mutex_lock(&ctx->uring_lock);
3943 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3944 mutex_unlock(&ctx->uring_lock);
3954 revert_creds(old_cred);
3961 struct io_wait_queue {
3962 struct wait_queue_entry wq;
3963 struct io_ring_ctx *ctx;
3965 unsigned nr_timeouts;
3968 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3970 struct io_ring_ctx *ctx = iowq->ctx;
3973 * Wake up if we have enough events, or if a timeout occurred since we
3974 * started waiting. For timeouts, we always want to return to userspace,
3975 * regardless of event count.
3977 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3978 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3981 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3982 int wake_flags, void *key)
3984 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3987 /* use noflush == true, as we can't safely rely on locking context */
3988 if (!io_should_wake(iowq, true))
3991 return autoremove_wake_function(curr, mode, wake_flags, key);
3995 * Wait until events become available, if we don't already have some. The
3996 * application must reap them itself, as they reside on the shared cq ring.
3998 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3999 const sigset_t __user *sig, size_t sigsz)
4001 struct io_wait_queue iowq = {
4004 .func = io_wake_function,
4005 .entry = LIST_HEAD_INIT(iowq.wq.entry),
4008 .to_wait = min_events,
4010 struct io_rings *rings = ctx->rings;
4013 if (io_cqring_events(ctx, false) >= min_events)
4017 #ifdef CONFIG_COMPAT
4018 if (in_compat_syscall())
4019 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
4023 ret = set_user_sigmask(sig, sigsz);
4029 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
4030 trace_io_uring_cqring_wait(ctx, min_events);
4032 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
4033 TASK_INTERRUPTIBLE);
4034 if (io_should_wake(&iowq, false))
4037 if (signal_pending(current)) {
4042 finish_wait(&ctx->wait, &iowq.wq);
4044 restore_saved_sigmask_unless(ret == -EINTR);
4046 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
4049 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
4051 #if defined(CONFIG_UNIX)
4052 if (ctx->ring_sock) {
4053 struct sock *sock = ctx->ring_sock->sk;
4054 struct sk_buff *skb;
4056 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
4062 for (i = 0; i < ctx->nr_user_files; i++) {
4065 file = io_file_from_index(ctx, i);
4072 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
4074 unsigned nr_tables, i;
4076 if (!ctx->file_table)
4079 __io_sqe_files_unregister(ctx);
4080 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
4081 for (i = 0; i < nr_tables; i++)
4082 kfree(ctx->file_table[i].files);
4083 kfree(ctx->file_table);
4084 ctx->file_table = NULL;
4085 ctx->nr_user_files = 0;
4089 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
4091 if (ctx->sqo_thread) {
4092 wait_for_completion(&ctx->completions[1]);
4094 * The park is a bit of a work-around, without it we get
4095 * warning spews on shutdown with SQPOLL set and affinity
4096 * set to a single CPU.
4098 kthread_park(ctx->sqo_thread);
4099 kthread_stop(ctx->sqo_thread);
4100 ctx->sqo_thread = NULL;
4104 static void io_finish_async(struct io_ring_ctx *ctx)
4106 io_sq_thread_stop(ctx);
4109 io_wq_destroy(ctx->io_wq);
4114 #if defined(CONFIG_UNIX)
4115 static void io_destruct_skb(struct sk_buff *skb)
4117 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
4120 io_wq_flush(ctx->io_wq);
4122 unix_destruct_scm(skb);
4126 * Ensure the UNIX gc is aware of our file set, so we are certain that
4127 * the io_uring can be safely unregistered on process exit, even if we have
4128 * loops in the file referencing.
4130 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
4132 struct sock *sk = ctx->ring_sock->sk;
4133 struct scm_fp_list *fpl;
4134 struct sk_buff *skb;
4137 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
4138 unsigned long inflight = ctx->user->unix_inflight + nr;
4140 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
4144 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
4148 skb = alloc_skb(0, GFP_KERNEL);
4157 fpl->user = get_uid(ctx->user);
4158 for (i = 0; i < nr; i++) {
4159 struct file *file = io_file_from_index(ctx, i + offset);
4163 fpl->fp[nr_files] = get_file(file);
4164 unix_inflight(fpl->user, fpl->fp[nr_files]);
4169 fpl->max = SCM_MAX_FD;
4170 fpl->count = nr_files;
4171 UNIXCB(skb).fp = fpl;
4172 skb->destructor = io_destruct_skb;
4173 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
4174 skb_queue_head(&sk->sk_receive_queue, skb);
4176 for (i = 0; i < nr_files; i++)
4187 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
4188 * causes regular reference counting to break down. We rely on the UNIX
4189 * garbage collection to take care of this problem for us.
4191 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4193 unsigned left, total;
4197 left = ctx->nr_user_files;
4199 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
4201 ret = __io_sqe_files_scm(ctx, this_files, total);
4205 total += this_files;
4211 while (total < ctx->nr_user_files) {
4212 struct file *file = io_file_from_index(ctx, total);
4222 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4228 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
4233 for (i = 0; i < nr_tables; i++) {
4234 struct fixed_file_table *table = &ctx->file_table[i];
4235 unsigned this_files;
4237 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
4238 table->files = kcalloc(this_files, sizeof(struct file *),
4242 nr_files -= this_files;
4248 for (i = 0; i < nr_tables; i++) {
4249 struct fixed_file_table *table = &ctx->file_table[i];
4250 kfree(table->files);
4255 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
4258 __s32 __user *fds = (__s32 __user *) arg;
4263 if (ctx->file_table)
4267 if (nr_args > IORING_MAX_FIXED_FILES)
4270 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
4271 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
4273 if (!ctx->file_table)
4276 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
4277 kfree(ctx->file_table);
4278 ctx->file_table = NULL;
4282 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
4283 struct fixed_file_table *table;
4287 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4289 /* allow sparse sets */
4295 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4296 index = i & IORING_FILE_TABLE_MASK;
4297 table->files[index] = fget(fd);
4300 if (!table->files[index])
4303 * Don't allow io_uring instances to be registered. If UNIX
4304 * isn't enabled, then this causes a reference cycle and this
4305 * instance can never get freed. If UNIX is enabled we'll
4306 * handle it just fine, but there's still no point in allowing
4307 * a ring fd as it doesn't support regular read/write anyway.
4309 if (table->files[index]->f_op == &io_uring_fops) {
4310 fput(table->files[index]);
4317 for (i = 0; i < ctx->nr_user_files; i++) {
4320 file = io_file_from_index(ctx, i);
4324 for (i = 0; i < nr_tables; i++)
4325 kfree(ctx->file_table[i].files);
4327 kfree(ctx->file_table);
4328 ctx->file_table = NULL;
4329 ctx->nr_user_files = 0;
4333 ret = io_sqe_files_scm(ctx);
4335 io_sqe_files_unregister(ctx);
4340 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4342 #if defined(CONFIG_UNIX)
4343 struct file *file = io_file_from_index(ctx, index);
4344 struct sock *sock = ctx->ring_sock->sk;
4345 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4346 struct sk_buff *skb;
4349 __skb_queue_head_init(&list);
4352 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4353 * remove this entry and rearrange the file array.
4355 skb = skb_dequeue(head);
4357 struct scm_fp_list *fp;
4359 fp = UNIXCB(skb).fp;
4360 for (i = 0; i < fp->count; i++) {
4363 if (fp->fp[i] != file)
4366 unix_notinflight(fp->user, fp->fp[i]);
4367 left = fp->count - 1 - i;
4369 memmove(&fp->fp[i], &fp->fp[i + 1],
4370 left * sizeof(struct file *));
4377 __skb_queue_tail(&list, skb);
4387 __skb_queue_tail(&list, skb);
4389 skb = skb_dequeue(head);
4392 if (skb_peek(&list)) {
4393 spin_lock_irq(&head->lock);
4394 while ((skb = __skb_dequeue(&list)) != NULL)
4395 __skb_queue_tail(head, skb);
4396 spin_unlock_irq(&head->lock);
4399 fput(io_file_from_index(ctx, index));
4403 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4406 #if defined(CONFIG_UNIX)
4407 struct sock *sock = ctx->ring_sock->sk;
4408 struct sk_buff_head *head = &sock->sk_receive_queue;
4409 struct sk_buff *skb;
4412 * See if we can merge this file into an existing skb SCM_RIGHTS
4413 * file set. If there's no room, fall back to allocating a new skb
4414 * and filling it in.
4416 spin_lock_irq(&head->lock);
4417 skb = skb_peek(head);
4419 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4421 if (fpl->count < SCM_MAX_FD) {
4422 __skb_unlink(skb, head);
4423 spin_unlock_irq(&head->lock);
4424 fpl->fp[fpl->count] = get_file(file);
4425 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4427 spin_lock_irq(&head->lock);
4428 __skb_queue_head(head, skb);
4433 spin_unlock_irq(&head->lock);
4440 return __io_sqe_files_scm(ctx, 1, index);
4446 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4449 struct io_uring_files_update up;
4454 if (!ctx->file_table)
4458 if (copy_from_user(&up, arg, sizeof(up)))
4460 if (check_add_overflow(up.offset, nr_args, &done))
4462 if (done > ctx->nr_user_files)
4466 fds = (__s32 __user *) up.fds;
4468 struct fixed_file_table *table;
4472 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4476 i = array_index_nospec(up.offset, ctx->nr_user_files);
4477 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4478 index = i & IORING_FILE_TABLE_MASK;
4479 if (table->files[index]) {
4480 io_sqe_file_unregister(ctx, i);
4481 table->files[index] = NULL;
4492 * Don't allow io_uring instances to be registered. If
4493 * UNIX isn't enabled, then this causes a reference
4494 * cycle and this instance can never get freed. If UNIX
4495 * is enabled we'll handle it just fine, but there's
4496 * still no point in allowing a ring fd as it doesn't
4497 * support regular read/write anyway.
4499 if (file->f_op == &io_uring_fops) {
4504 table->files[index] = file;
4505 err = io_sqe_file_register(ctx, file, i);
4514 return done ? done : err;
4517 static void io_put_work(struct io_wq_work *work)
4519 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4524 static void io_get_work(struct io_wq_work *work)
4526 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4528 refcount_inc(&req->refs);
4531 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4532 struct io_uring_params *p)
4534 struct io_wq_data data;
4535 unsigned concurrency;
4538 init_waitqueue_head(&ctx->sqo_wait);
4539 mmgrab(current->mm);
4540 ctx->sqo_mm = current->mm;
4542 if (ctx->flags & IORING_SETUP_SQPOLL) {
4544 if (!capable(CAP_SYS_ADMIN))
4547 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4548 if (!ctx->sq_thread_idle)
4549 ctx->sq_thread_idle = HZ;
4551 if (p->flags & IORING_SETUP_SQ_AFF) {
4552 int cpu = p->sq_thread_cpu;
4555 if (cpu >= nr_cpu_ids)
4557 if (!cpu_online(cpu))
4560 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4564 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4567 if (IS_ERR(ctx->sqo_thread)) {
4568 ret = PTR_ERR(ctx->sqo_thread);
4569 ctx->sqo_thread = NULL;
4572 wake_up_process(ctx->sqo_thread);
4573 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4574 /* Can't have SQ_AFF without SQPOLL */
4579 data.mm = ctx->sqo_mm;
4580 data.user = ctx->user;
4581 data.creds = ctx->creds;
4582 data.get_work = io_get_work;
4583 data.put_work = io_put_work;
4585 /* Do QD, or 4 * CPUS, whatever is smallest */
4586 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4587 ctx->io_wq = io_wq_create(concurrency, &data);
4588 if (IS_ERR(ctx->io_wq)) {
4589 ret = PTR_ERR(ctx->io_wq);
4596 io_finish_async(ctx);
4597 mmdrop(ctx->sqo_mm);
4602 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4604 atomic_long_sub(nr_pages, &user->locked_vm);
4607 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4609 unsigned long page_limit, cur_pages, new_pages;
4611 /* Don't allow more pages than we can safely lock */
4612 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4615 cur_pages = atomic_long_read(&user->locked_vm);
4616 new_pages = cur_pages + nr_pages;
4617 if (new_pages > page_limit)
4619 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4620 new_pages) != cur_pages);
4625 static void io_mem_free(void *ptr)
4632 page = virt_to_head_page(ptr);
4633 if (put_page_testzero(page))
4634 free_compound_page(page);
4637 static void *io_mem_alloc(size_t size)
4639 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4642 return (void *) __get_free_pages(gfp_flags, get_order(size));
4645 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4648 struct io_rings *rings;
4649 size_t off, sq_array_size;
4651 off = struct_size(rings, cqes, cq_entries);
4652 if (off == SIZE_MAX)
4656 off = ALIGN(off, SMP_CACHE_BYTES);
4661 sq_array_size = array_size(sizeof(u32), sq_entries);
4662 if (sq_array_size == SIZE_MAX)
4665 if (check_add_overflow(off, sq_array_size, &off))
4674 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4678 pages = (size_t)1 << get_order(
4679 rings_size(sq_entries, cq_entries, NULL));
4680 pages += (size_t)1 << get_order(
4681 array_size(sizeof(struct io_uring_sqe), sq_entries));
4686 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4690 if (!ctx->user_bufs)
4693 for (i = 0; i < ctx->nr_user_bufs; i++) {
4694 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4696 for (j = 0; j < imu->nr_bvecs; j++)
4697 put_user_page(imu->bvec[j].bv_page);
4699 if (ctx->account_mem)
4700 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4705 kfree(ctx->user_bufs);
4706 ctx->user_bufs = NULL;
4707 ctx->nr_user_bufs = 0;
4711 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4712 void __user *arg, unsigned index)
4714 struct iovec __user *src;
4716 #ifdef CONFIG_COMPAT
4718 struct compat_iovec __user *ciovs;
4719 struct compat_iovec ciov;
4721 ciovs = (struct compat_iovec __user *) arg;
4722 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4725 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
4726 dst->iov_len = ciov.iov_len;
4730 src = (struct iovec __user *) arg;
4731 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4736 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4739 struct vm_area_struct **vmas = NULL;
4740 struct page **pages = NULL;
4741 int i, j, got_pages = 0;
4746 if (!nr_args || nr_args > UIO_MAXIOV)
4749 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4751 if (!ctx->user_bufs)
4754 for (i = 0; i < nr_args; i++) {
4755 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4756 unsigned long off, start, end, ubuf;
4761 ret = io_copy_iov(ctx, &iov, arg, i);
4766 * Don't impose further limits on the size and buffer
4767 * constraints here, we'll -EINVAL later when IO is
4768 * submitted if they are wrong.
4771 if (!iov.iov_base || !iov.iov_len)
4774 /* arbitrary limit, but we need something */
4775 if (iov.iov_len > SZ_1G)
4778 ubuf = (unsigned long) iov.iov_base;
4779 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4780 start = ubuf >> PAGE_SHIFT;
4781 nr_pages = end - start;
4783 if (ctx->account_mem) {
4784 ret = io_account_mem(ctx->user, nr_pages);
4790 if (!pages || nr_pages > got_pages) {
4793 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4795 vmas = kvmalloc_array(nr_pages,
4796 sizeof(struct vm_area_struct *),
4798 if (!pages || !vmas) {
4800 if (ctx->account_mem)
4801 io_unaccount_mem(ctx->user, nr_pages);
4804 got_pages = nr_pages;
4807 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4811 if (ctx->account_mem)
4812 io_unaccount_mem(ctx->user, nr_pages);
4817 down_read(¤t->mm->mmap_sem);
4818 pret = get_user_pages(ubuf, nr_pages,
4819 FOLL_WRITE | FOLL_LONGTERM,
4821 if (pret == nr_pages) {
4822 /* don't support file backed memory */
4823 for (j = 0; j < nr_pages; j++) {
4824 struct vm_area_struct *vma = vmas[j];
4827 !is_file_hugepages(vma->vm_file)) {
4833 ret = pret < 0 ? pret : -EFAULT;
4835 up_read(¤t->mm->mmap_sem);
4838 * if we did partial map, or found file backed vmas,
4839 * release any pages we did get
4842 put_user_pages(pages, pret);
4843 if (ctx->account_mem)
4844 io_unaccount_mem(ctx->user, nr_pages);
4849 off = ubuf & ~PAGE_MASK;
4851 for (j = 0; j < nr_pages; j++) {
4854 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4855 imu->bvec[j].bv_page = pages[j];
4856 imu->bvec[j].bv_len = vec_len;
4857 imu->bvec[j].bv_offset = off;
4861 /* store original address for later verification */
4863 imu->len = iov.iov_len;
4864 imu->nr_bvecs = nr_pages;
4866 ctx->nr_user_bufs++;
4874 io_sqe_buffer_unregister(ctx);
4878 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4880 __s32 __user *fds = arg;
4886 if (copy_from_user(&fd, fds, sizeof(*fds)))
4889 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4890 if (IS_ERR(ctx->cq_ev_fd)) {
4891 int ret = PTR_ERR(ctx->cq_ev_fd);
4892 ctx->cq_ev_fd = NULL;
4899 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4901 if (ctx->cq_ev_fd) {
4902 eventfd_ctx_put(ctx->cq_ev_fd);
4903 ctx->cq_ev_fd = NULL;
4910 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4912 io_finish_async(ctx);
4914 mmdrop(ctx->sqo_mm);
4916 io_iopoll_reap_events(ctx);
4917 io_sqe_buffer_unregister(ctx);
4918 io_sqe_files_unregister(ctx);
4919 io_eventfd_unregister(ctx);
4921 #if defined(CONFIG_UNIX)
4922 if (ctx->ring_sock) {
4923 ctx->ring_sock->file = NULL; /* so that iput() is called */
4924 sock_release(ctx->ring_sock);
4928 io_mem_free(ctx->rings);
4929 io_mem_free(ctx->sq_sqes);
4931 percpu_ref_exit(&ctx->refs);
4932 if (ctx->account_mem)
4933 io_unaccount_mem(ctx->user,
4934 ring_pages(ctx->sq_entries, ctx->cq_entries));
4935 free_uid(ctx->user);
4936 put_cred(ctx->creds);
4937 kfree(ctx->completions);
4938 kfree(ctx->cancel_hash);
4939 kmem_cache_free(req_cachep, ctx->fallback_req);
4943 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4945 struct io_ring_ctx *ctx = file->private_data;
4948 poll_wait(file, &ctx->cq_wait, wait);
4950 * synchronizes with barrier from wq_has_sleeper call in
4954 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4955 ctx->rings->sq_ring_entries)
4956 mask |= EPOLLOUT | EPOLLWRNORM;
4957 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4958 mask |= EPOLLIN | EPOLLRDNORM;
4963 static int io_uring_fasync(int fd, struct file *file, int on)
4965 struct io_ring_ctx *ctx = file->private_data;
4967 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4970 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4972 mutex_lock(&ctx->uring_lock);
4973 percpu_ref_kill(&ctx->refs);
4974 mutex_unlock(&ctx->uring_lock);
4976 io_kill_timeouts(ctx);
4977 io_poll_remove_all(ctx);
4980 io_wq_cancel_all(ctx->io_wq);
4982 io_iopoll_reap_events(ctx);
4983 /* if we failed setting up the ctx, we might not have any rings */
4985 io_cqring_overflow_flush(ctx, true);
4986 wait_for_completion(&ctx->completions[0]);
4987 io_ring_ctx_free(ctx);
4990 static int io_uring_release(struct inode *inode, struct file *file)
4992 struct io_ring_ctx *ctx = file->private_data;
4994 file->private_data = NULL;
4995 io_ring_ctx_wait_and_kill(ctx);
4999 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
5000 struct files_struct *files)
5002 struct io_kiocb *req;
5005 while (!list_empty_careful(&ctx->inflight_list)) {
5006 struct io_kiocb *cancel_req = NULL;
5008 spin_lock_irq(&ctx->inflight_lock);
5009 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
5010 if (req->work.files != files)
5012 /* req is being completed, ignore */
5013 if (!refcount_inc_not_zero(&req->refs))
5019 prepare_to_wait(&ctx->inflight_wait, &wait,
5020 TASK_UNINTERRUPTIBLE);
5021 spin_unlock_irq(&ctx->inflight_lock);
5023 /* We need to keep going until we don't find a matching req */
5027 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
5028 io_put_req(cancel_req);
5031 finish_wait(&ctx->inflight_wait, &wait);
5034 static int io_uring_flush(struct file *file, void *data)
5036 struct io_ring_ctx *ctx = file->private_data;
5038 io_uring_cancel_files(ctx, data);
5039 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
5040 io_cqring_overflow_flush(ctx, true);
5041 io_wq_cancel_all(ctx->io_wq);
5046 static void *io_uring_validate_mmap_request(struct file *file,
5047 loff_t pgoff, size_t sz)
5049 struct io_ring_ctx *ctx = file->private_data;
5050 loff_t offset = pgoff << PAGE_SHIFT;
5055 case IORING_OFF_SQ_RING:
5056 case IORING_OFF_CQ_RING:
5059 case IORING_OFF_SQES:
5063 return ERR_PTR(-EINVAL);
5066 page = virt_to_head_page(ptr);
5067 if (sz > page_size(page))
5068 return ERR_PTR(-EINVAL);
5075 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5077 size_t sz = vma->vm_end - vma->vm_start;
5081 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
5083 return PTR_ERR(ptr);
5085 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
5086 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
5089 #else /* !CONFIG_MMU */
5091 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5093 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
5096 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
5098 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
5101 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
5102 unsigned long addr, unsigned long len,
5103 unsigned long pgoff, unsigned long flags)
5107 ptr = io_uring_validate_mmap_request(file, pgoff, len);
5109 return PTR_ERR(ptr);
5111 return (unsigned long) ptr;
5114 #endif /* !CONFIG_MMU */
5116 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
5117 u32, min_complete, u32, flags, const sigset_t __user *, sig,
5120 struct io_ring_ctx *ctx;
5125 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
5133 if (f.file->f_op != &io_uring_fops)
5137 ctx = f.file->private_data;
5138 if (!percpu_ref_tryget(&ctx->refs))
5142 * For SQ polling, the thread will do all submissions and completions.
5143 * Just return the requested submit count, and wake the thread if
5147 if (ctx->flags & IORING_SETUP_SQPOLL) {
5148 if (!list_empty_careful(&ctx->cq_overflow_list))
5149 io_cqring_overflow_flush(ctx, false);
5150 if (flags & IORING_ENTER_SQ_WAKEUP)
5151 wake_up(&ctx->sqo_wait);
5152 submitted = to_submit;
5153 } else if (to_submit) {
5154 struct mm_struct *cur_mm;
5156 to_submit = min(to_submit, ctx->sq_entries);
5157 mutex_lock(&ctx->uring_lock);
5158 /* already have mm, so io_submit_sqes() won't try to grab it */
5159 cur_mm = ctx->sqo_mm;
5160 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
5162 mutex_unlock(&ctx->uring_lock);
5164 if (submitted != to_submit)
5167 if (flags & IORING_ENTER_GETEVENTS) {
5168 unsigned nr_events = 0;
5170 min_complete = min(min_complete, ctx->cq_entries);
5172 if (ctx->flags & IORING_SETUP_IOPOLL) {
5173 ret = io_iopoll_check(ctx, &nr_events, min_complete);
5175 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
5180 percpu_ref_put(&ctx->refs);
5183 return submitted ? submitted : ret;
5186 static const struct file_operations io_uring_fops = {
5187 .release = io_uring_release,
5188 .flush = io_uring_flush,
5189 .mmap = io_uring_mmap,
5191 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
5192 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
5194 .poll = io_uring_poll,
5195 .fasync = io_uring_fasync,
5198 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
5199 struct io_uring_params *p)
5201 struct io_rings *rings;
5202 size_t size, sq_array_offset;
5204 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
5205 if (size == SIZE_MAX)
5208 rings = io_mem_alloc(size);
5213 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
5214 rings->sq_ring_mask = p->sq_entries - 1;
5215 rings->cq_ring_mask = p->cq_entries - 1;
5216 rings->sq_ring_entries = p->sq_entries;
5217 rings->cq_ring_entries = p->cq_entries;
5218 ctx->sq_mask = rings->sq_ring_mask;
5219 ctx->cq_mask = rings->cq_ring_mask;
5220 ctx->sq_entries = rings->sq_ring_entries;
5221 ctx->cq_entries = rings->cq_ring_entries;
5223 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
5224 if (size == SIZE_MAX) {
5225 io_mem_free(ctx->rings);
5230 ctx->sq_sqes = io_mem_alloc(size);
5231 if (!ctx->sq_sqes) {
5232 io_mem_free(ctx->rings);
5241 * Allocate an anonymous fd, this is what constitutes the application
5242 * visible backing of an io_uring instance. The application mmaps this
5243 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
5244 * we have to tie this fd to a socket for file garbage collection purposes.
5246 static int io_uring_get_fd(struct io_ring_ctx *ctx)
5251 #if defined(CONFIG_UNIX)
5252 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
5258 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
5262 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
5263 O_RDWR | O_CLOEXEC);
5266 ret = PTR_ERR(file);
5270 #if defined(CONFIG_UNIX)
5271 ctx->ring_sock->file = file;
5272 ctx->ring_sock->sk->sk_user_data = ctx;
5274 fd_install(ret, file);
5277 #if defined(CONFIG_UNIX)
5278 sock_release(ctx->ring_sock);
5279 ctx->ring_sock = NULL;
5284 static int io_uring_create(unsigned entries, struct io_uring_params *p)
5286 struct user_struct *user = NULL;
5287 struct io_ring_ctx *ctx;
5291 if (!entries || entries > IORING_MAX_ENTRIES)
5295 * Use twice as many entries for the CQ ring. It's possible for the
5296 * application to drive a higher depth than the size of the SQ ring,
5297 * since the sqes are only used at submission time. This allows for
5298 * some flexibility in overcommitting a bit. If the application has
5299 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5300 * of CQ ring entries manually.
5302 p->sq_entries = roundup_pow_of_two(entries);
5303 if (p->flags & IORING_SETUP_CQSIZE) {
5305 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5306 * to a power-of-two, if it isn't already. We do NOT impose
5307 * any cq vs sq ring sizing.
5309 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5311 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5313 p->cq_entries = 2 * p->sq_entries;
5316 user = get_uid(current_user());
5317 account_mem = !capable(CAP_IPC_LOCK);
5320 ret = io_account_mem(user,
5321 ring_pages(p->sq_entries, p->cq_entries));
5328 ctx = io_ring_ctx_alloc(p);
5331 io_unaccount_mem(user, ring_pages(p->sq_entries,
5336 ctx->compat = in_compat_syscall();
5337 ctx->account_mem = account_mem;
5339 ctx->creds = get_current_cred();
5341 ret = io_allocate_scq_urings(ctx, p);
5345 ret = io_sq_offload_start(ctx, p);
5349 memset(&p->sq_off, 0, sizeof(p->sq_off));
5350 p->sq_off.head = offsetof(struct io_rings, sq.head);
5351 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5352 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5353 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5354 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5355 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5356 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5358 memset(&p->cq_off, 0, sizeof(p->cq_off));
5359 p->cq_off.head = offsetof(struct io_rings, cq.head);
5360 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5361 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5362 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5363 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5364 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5367 * Install ring fd as the very last thing, so we don't risk someone
5368 * having closed it before we finish setup
5370 ret = io_uring_get_fd(ctx);
5374 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5375 IORING_FEAT_SUBMIT_STABLE;
5376 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5379 io_ring_ctx_wait_and_kill(ctx);
5384 * Sets up an aio uring context, and returns the fd. Applications asks for a
5385 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5386 * params structure passed in.
5388 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5390 struct io_uring_params p;
5394 if (copy_from_user(&p, params, sizeof(p)))
5396 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5401 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5402 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5405 ret = io_uring_create(entries, &p);
5409 if (copy_to_user(params, &p, sizeof(p)))
5415 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5416 struct io_uring_params __user *, params)
5418 return io_uring_setup(entries, params);
5421 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5422 void __user *arg, unsigned nr_args)
5423 __releases(ctx->uring_lock)
5424 __acquires(ctx->uring_lock)
5429 * We're inside the ring mutex, if the ref is already dying, then
5430 * someone else killed the ctx or is already going through
5431 * io_uring_register().
5433 if (percpu_ref_is_dying(&ctx->refs))
5436 percpu_ref_kill(&ctx->refs);
5439 * Drop uring mutex before waiting for references to exit. If another
5440 * thread is currently inside io_uring_enter() it might need to grab
5441 * the uring_lock to make progress. If we hold it here across the drain
5442 * wait, then we can deadlock. It's safe to drop the mutex here, since
5443 * no new references will come in after we've killed the percpu ref.
5445 mutex_unlock(&ctx->uring_lock);
5446 wait_for_completion(&ctx->completions[0]);
5447 mutex_lock(&ctx->uring_lock);
5450 case IORING_REGISTER_BUFFERS:
5451 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5453 case IORING_UNREGISTER_BUFFERS:
5457 ret = io_sqe_buffer_unregister(ctx);
5459 case IORING_REGISTER_FILES:
5460 ret = io_sqe_files_register(ctx, arg, nr_args);
5462 case IORING_UNREGISTER_FILES:
5466 ret = io_sqe_files_unregister(ctx);
5468 case IORING_REGISTER_FILES_UPDATE:
5469 ret = io_sqe_files_update(ctx, arg, nr_args);
5471 case IORING_REGISTER_EVENTFD:
5475 ret = io_eventfd_register(ctx, arg);
5477 case IORING_UNREGISTER_EVENTFD:
5481 ret = io_eventfd_unregister(ctx);
5488 /* bring the ctx back to life */
5489 reinit_completion(&ctx->completions[0]);
5490 percpu_ref_reinit(&ctx->refs);
5494 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5495 void __user *, arg, unsigned int, nr_args)
5497 struct io_ring_ctx *ctx;
5506 if (f.file->f_op != &io_uring_fops)
5509 ctx = f.file->private_data;
5511 mutex_lock(&ctx->uring_lock);
5512 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5513 mutex_unlock(&ctx->uring_lock);
5514 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5515 ctx->cq_ev_fd != NULL, ret);
5521 static int __init io_uring_init(void)
5523 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5526 __initcall(io_uring_init);