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;
335 struct io_async_connect {
336 struct sockaddr_storage address;
339 struct io_async_msghdr {
340 struct iovec fast_iov[UIO_FASTIOV];
342 struct sockaddr __user *uaddr;
347 struct iovec fast_iov[UIO_FASTIOV];
353 struct io_async_ctx {
354 struct io_uring_sqe sqe;
356 struct io_async_rw rw;
357 struct io_async_msghdr msg;
358 struct io_async_connect connect;
359 struct io_timeout_data timeout;
364 * NOTE! Each of the iocb union members has the file pointer
365 * as the first entry in their struct definition. So you can
366 * access the file pointer through any of the sub-structs,
367 * or directly as just 'ki_filp' in this struct.
373 struct io_poll_iocb poll;
374 struct io_accept accept;
376 struct io_cancel cancel;
377 struct io_timeout timeout;
380 const struct io_uring_sqe *sqe;
381 struct io_async_ctx *io;
382 struct file *ring_file;
386 bool needs_fixed_file;
389 struct io_ring_ctx *ctx;
391 struct list_head list;
392 struct hlist_node hash_node;
394 struct list_head link_list;
397 #define REQ_F_NOWAIT 1 /* must not punt to workers */
398 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
399 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
400 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
401 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
402 #define REQ_F_IO_DRAINED 32 /* drain done */
403 #define REQ_F_LINK 64 /* linked sqes */
404 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
405 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
406 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
407 #define REQ_F_TIMEOUT 1024 /* timeout request */
408 #define REQ_F_ISREG 2048 /* regular file */
409 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
410 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
411 #define REQ_F_INFLIGHT 16384 /* on inflight list */
412 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
413 #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */
414 #define REQ_F_PREPPED 131072 /* request already opcode prepared */
419 struct list_head inflight_entry;
421 struct io_wq_work work;
424 #define IO_PLUG_THRESHOLD 2
425 #define IO_IOPOLL_BATCH 8
427 struct io_submit_state {
428 struct blk_plug plug;
431 * io_kiocb alloc cache
433 void *reqs[IO_IOPOLL_BATCH];
434 unsigned int free_reqs;
435 unsigned int cur_req;
438 * File reference cache
442 unsigned int has_refs;
443 unsigned int used_refs;
444 unsigned int ios_left;
447 static void io_wq_submit_work(struct io_wq_work **workptr);
448 static void io_cqring_fill_event(struct io_kiocb *req, long res);
449 static void __io_free_req(struct io_kiocb *req);
450 static void io_put_req(struct io_kiocb *req);
451 static void io_double_put_req(struct io_kiocb *req);
452 static void __io_double_put_req(struct io_kiocb *req);
453 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
454 static void io_queue_linked_timeout(struct io_kiocb *req);
456 static struct kmem_cache *req_cachep;
458 static const struct file_operations io_uring_fops;
460 struct sock *io_uring_get_socket(struct file *file)
462 #if defined(CONFIG_UNIX)
463 if (file->f_op == &io_uring_fops) {
464 struct io_ring_ctx *ctx = file->private_data;
466 return ctx->ring_sock->sk;
471 EXPORT_SYMBOL(io_uring_get_socket);
473 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
475 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
477 complete(&ctx->completions[0]);
480 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
482 struct io_ring_ctx *ctx;
485 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
489 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
490 if (!ctx->fallback_req)
493 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
494 if (!ctx->completions)
498 * Use 5 bits less than the max cq entries, that should give us around
499 * 32 entries per hash list if totally full and uniformly spread.
501 hash_bits = ilog2(p->cq_entries);
505 ctx->cancel_hash_bits = hash_bits;
506 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
508 if (!ctx->cancel_hash)
510 __hash_init(ctx->cancel_hash, 1U << hash_bits);
512 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
513 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
516 ctx->flags = p->flags;
517 init_waitqueue_head(&ctx->cq_wait);
518 INIT_LIST_HEAD(&ctx->cq_overflow_list);
519 init_completion(&ctx->completions[0]);
520 init_completion(&ctx->completions[1]);
521 mutex_init(&ctx->uring_lock);
522 init_waitqueue_head(&ctx->wait);
523 spin_lock_init(&ctx->completion_lock);
524 INIT_LIST_HEAD(&ctx->poll_list);
525 INIT_LIST_HEAD(&ctx->defer_list);
526 INIT_LIST_HEAD(&ctx->timeout_list);
527 init_waitqueue_head(&ctx->inflight_wait);
528 spin_lock_init(&ctx->inflight_lock);
529 INIT_LIST_HEAD(&ctx->inflight_list);
532 if (ctx->fallback_req)
533 kmem_cache_free(req_cachep, ctx->fallback_req);
534 kfree(ctx->completions);
535 kfree(ctx->cancel_hash);
540 static inline bool __req_need_defer(struct io_kiocb *req)
542 struct io_ring_ctx *ctx = req->ctx;
544 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
545 + atomic_read(&ctx->cached_cq_overflow);
548 static inline bool req_need_defer(struct io_kiocb *req)
550 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
551 return __req_need_defer(req);
556 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
558 struct io_kiocb *req;
560 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
561 if (req && !req_need_defer(req)) {
562 list_del_init(&req->list);
569 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
571 struct io_kiocb *req;
573 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
575 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
577 if (!__req_need_defer(req)) {
578 list_del_init(&req->list);
586 static void __io_commit_cqring(struct io_ring_ctx *ctx)
588 struct io_rings *rings = ctx->rings;
590 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
591 /* order cqe stores with ring update */
592 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
594 if (wq_has_sleeper(&ctx->cq_wait)) {
595 wake_up_interruptible(&ctx->cq_wait);
596 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
601 static inline bool io_req_needs_user(struct io_kiocb *req)
603 return !(req->opcode == IORING_OP_READ_FIXED ||
604 req->opcode == IORING_OP_WRITE_FIXED);
607 static inline bool io_prep_async_work(struct io_kiocb *req,
608 struct io_kiocb **link)
610 bool do_hashed = false;
613 switch (req->opcode) {
614 case IORING_OP_WRITEV:
615 case IORING_OP_WRITE_FIXED:
616 /* only regular files should be hashed for writes */
617 if (req->flags & REQ_F_ISREG)
620 case IORING_OP_READV:
621 case IORING_OP_READ_FIXED:
622 case IORING_OP_SENDMSG:
623 case IORING_OP_RECVMSG:
624 case IORING_OP_ACCEPT:
625 case IORING_OP_POLL_ADD:
626 case IORING_OP_CONNECT:
628 * We know REQ_F_ISREG is not set on some of these
629 * opcodes, but this enables us to keep the check in
632 if (!(req->flags & REQ_F_ISREG))
633 req->work.flags |= IO_WQ_WORK_UNBOUND;
636 if (io_req_needs_user(req))
637 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
640 *link = io_prep_linked_timeout(req);
644 static inline void io_queue_async_work(struct io_kiocb *req)
646 struct io_ring_ctx *ctx = req->ctx;
647 struct io_kiocb *link;
650 do_hashed = io_prep_async_work(req, &link);
652 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
655 io_wq_enqueue(ctx->io_wq, &req->work);
657 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
658 file_inode(req->file));
662 io_queue_linked_timeout(link);
665 static void io_kill_timeout(struct io_kiocb *req)
669 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
671 atomic_inc(&req->ctx->cq_timeouts);
672 list_del_init(&req->list);
673 io_cqring_fill_event(req, 0);
678 static void io_kill_timeouts(struct io_ring_ctx *ctx)
680 struct io_kiocb *req, *tmp;
682 spin_lock_irq(&ctx->completion_lock);
683 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
684 io_kill_timeout(req);
685 spin_unlock_irq(&ctx->completion_lock);
688 static void io_commit_cqring(struct io_ring_ctx *ctx)
690 struct io_kiocb *req;
692 while ((req = io_get_timeout_req(ctx)) != NULL)
693 io_kill_timeout(req);
695 __io_commit_cqring(ctx);
697 while ((req = io_get_deferred_req(ctx)) != NULL) {
698 req->flags |= REQ_F_IO_DRAINED;
699 io_queue_async_work(req);
703 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
705 struct io_rings *rings = ctx->rings;
708 tail = ctx->cached_cq_tail;
710 * writes to the cq entry need to come after reading head; the
711 * control dependency is enough as we're using WRITE_ONCE to
714 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
717 ctx->cached_cq_tail++;
718 return &rings->cqes[tail & ctx->cq_mask];
721 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
723 if (waitqueue_active(&ctx->wait))
725 if (waitqueue_active(&ctx->sqo_wait))
726 wake_up(&ctx->sqo_wait);
728 eventfd_signal(ctx->cq_ev_fd, 1);
731 /* Returns true if there are no backlogged entries after the flush */
732 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
734 struct io_rings *rings = ctx->rings;
735 struct io_uring_cqe *cqe;
736 struct io_kiocb *req;
741 if (list_empty_careful(&ctx->cq_overflow_list))
743 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
744 rings->cq_ring_entries))
748 spin_lock_irqsave(&ctx->completion_lock, flags);
750 /* if force is set, the ring is going away. always drop after that */
752 ctx->cq_overflow_flushed = true;
755 while (!list_empty(&ctx->cq_overflow_list)) {
756 cqe = io_get_cqring(ctx);
760 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
762 list_move(&req->list, &list);
764 WRITE_ONCE(cqe->user_data, req->user_data);
765 WRITE_ONCE(cqe->res, req->result);
766 WRITE_ONCE(cqe->flags, 0);
768 WRITE_ONCE(ctx->rings->cq_overflow,
769 atomic_inc_return(&ctx->cached_cq_overflow));
773 io_commit_cqring(ctx);
774 spin_unlock_irqrestore(&ctx->completion_lock, flags);
775 io_cqring_ev_posted(ctx);
777 while (!list_empty(&list)) {
778 req = list_first_entry(&list, struct io_kiocb, list);
779 list_del(&req->list);
786 static void io_cqring_fill_event(struct io_kiocb *req, long res)
788 struct io_ring_ctx *ctx = req->ctx;
789 struct io_uring_cqe *cqe;
791 trace_io_uring_complete(ctx, req->user_data, res);
794 * If we can't get a cq entry, userspace overflowed the
795 * submission (by quite a lot). Increment the overflow count in
798 cqe = io_get_cqring(ctx);
800 WRITE_ONCE(cqe->user_data, req->user_data);
801 WRITE_ONCE(cqe->res, res);
802 WRITE_ONCE(cqe->flags, 0);
803 } else if (ctx->cq_overflow_flushed) {
804 WRITE_ONCE(ctx->rings->cq_overflow,
805 atomic_inc_return(&ctx->cached_cq_overflow));
807 refcount_inc(&req->refs);
809 list_add_tail(&req->list, &ctx->cq_overflow_list);
813 static void io_cqring_add_event(struct io_kiocb *req, long res)
815 struct io_ring_ctx *ctx = req->ctx;
818 spin_lock_irqsave(&ctx->completion_lock, flags);
819 io_cqring_fill_event(req, res);
820 io_commit_cqring(ctx);
821 spin_unlock_irqrestore(&ctx->completion_lock, flags);
823 io_cqring_ev_posted(ctx);
826 static inline bool io_is_fallback_req(struct io_kiocb *req)
828 return req == (struct io_kiocb *)
829 ((unsigned long) req->ctx->fallback_req & ~1UL);
832 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
834 struct io_kiocb *req;
836 req = ctx->fallback_req;
837 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
843 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
844 struct io_submit_state *state)
846 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
847 struct io_kiocb *req;
849 if (!percpu_ref_tryget(&ctx->refs))
853 req = kmem_cache_alloc(req_cachep, gfp);
856 } else if (!state->free_reqs) {
860 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
861 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
864 * Bulk alloc is all-or-nothing. If we fail to get a batch,
865 * retry single alloc to be on the safe side.
867 if (unlikely(ret <= 0)) {
868 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
873 state->free_reqs = ret - 1;
875 req = state->reqs[0];
877 req = state->reqs[state->cur_req];
884 req->ring_file = NULL;
888 /* one is dropped after submission, the other at completion */
889 refcount_set(&req->refs, 2);
891 INIT_IO_WORK(&req->work, io_wq_submit_work);
894 req = io_get_fallback_req(ctx);
897 percpu_ref_put(&ctx->refs);
901 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
904 kmem_cache_free_bulk(req_cachep, *nr, reqs);
905 percpu_ref_put_many(&ctx->refs, *nr);
910 static void __io_free_req(struct io_kiocb *req)
912 struct io_ring_ctx *ctx = req->ctx;
916 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
918 if (req->flags & REQ_F_INFLIGHT) {
921 spin_lock_irqsave(&ctx->inflight_lock, flags);
922 list_del(&req->inflight_entry);
923 if (waitqueue_active(&ctx->inflight_wait))
924 wake_up(&ctx->inflight_wait);
925 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
927 percpu_ref_put(&ctx->refs);
928 if (likely(!io_is_fallback_req(req)))
929 kmem_cache_free(req_cachep, req);
931 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
934 static bool io_link_cancel_timeout(struct io_kiocb *req)
936 struct io_ring_ctx *ctx = req->ctx;
939 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
941 io_cqring_fill_event(req, -ECANCELED);
942 io_commit_cqring(ctx);
943 req->flags &= ~REQ_F_LINK;
951 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
953 struct io_ring_ctx *ctx = req->ctx;
954 bool wake_ev = false;
956 /* Already got next link */
957 if (req->flags & REQ_F_LINK_NEXT)
961 * The list should never be empty when we are called here. But could
962 * potentially happen if the chain is messed up, check to be on the
965 while (!list_empty(&req->link_list)) {
966 struct io_kiocb *nxt = list_first_entry(&req->link_list,
967 struct io_kiocb, link_list);
969 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
970 (nxt->flags & REQ_F_TIMEOUT))) {
971 list_del_init(&nxt->link_list);
972 wake_ev |= io_link_cancel_timeout(nxt);
973 req->flags &= ~REQ_F_LINK_TIMEOUT;
977 list_del_init(&req->link_list);
978 if (!list_empty(&nxt->link_list))
979 nxt->flags |= REQ_F_LINK;
984 req->flags |= REQ_F_LINK_NEXT;
986 io_cqring_ev_posted(ctx);
990 * Called if REQ_F_LINK is set, and we fail the head request
992 static void io_fail_links(struct io_kiocb *req)
994 struct io_ring_ctx *ctx = req->ctx;
997 spin_lock_irqsave(&ctx->completion_lock, flags);
999 while (!list_empty(&req->link_list)) {
1000 struct io_kiocb *link = list_first_entry(&req->link_list,
1001 struct io_kiocb, link_list);
1003 list_del_init(&link->link_list);
1004 trace_io_uring_fail_link(req, link);
1006 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1007 link->opcode == IORING_OP_LINK_TIMEOUT) {
1008 io_link_cancel_timeout(link);
1010 io_cqring_fill_event(link, -ECANCELED);
1011 __io_double_put_req(link);
1013 req->flags &= ~REQ_F_LINK_TIMEOUT;
1016 io_commit_cqring(ctx);
1017 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1018 io_cqring_ev_posted(ctx);
1021 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1023 if (likely(!(req->flags & REQ_F_LINK)))
1027 * If LINK is set, we have dependent requests in this chain. If we
1028 * didn't fail this request, queue the first one up, moving any other
1029 * dependencies to the next request. In case of failure, fail the rest
1032 if (req->flags & REQ_F_FAIL_LINK) {
1034 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1035 REQ_F_LINK_TIMEOUT) {
1036 struct io_ring_ctx *ctx = req->ctx;
1037 unsigned long flags;
1040 * If this is a timeout link, we could be racing with the
1041 * timeout timer. Grab the completion lock for this case to
1042 * protect against that.
1044 spin_lock_irqsave(&ctx->completion_lock, flags);
1045 io_req_link_next(req, nxt);
1046 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1048 io_req_link_next(req, nxt);
1052 static void io_free_req(struct io_kiocb *req)
1054 struct io_kiocb *nxt = NULL;
1056 io_req_find_next(req, &nxt);
1060 io_queue_async_work(nxt);
1064 * Drop reference to request, return next in chain (if there is one) if this
1065 * was the last reference to this request.
1067 __attribute__((nonnull))
1068 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1070 io_req_find_next(req, nxtptr);
1072 if (refcount_dec_and_test(&req->refs))
1076 static void io_put_req(struct io_kiocb *req)
1078 if (refcount_dec_and_test(&req->refs))
1083 * Must only be used if we don't need to care about links, usually from
1084 * within the completion handling itself.
1086 static void __io_double_put_req(struct io_kiocb *req)
1088 /* drop both submit and complete references */
1089 if (refcount_sub_and_test(2, &req->refs))
1093 static void io_double_put_req(struct io_kiocb *req)
1095 /* drop both submit and complete references */
1096 if (refcount_sub_and_test(2, &req->refs))
1100 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1102 struct io_rings *rings = ctx->rings;
1105 * noflush == true is from the waitqueue handler, just ensure we wake
1106 * up the task, and the next invocation will flush the entries. We
1107 * cannot safely to it from here.
1109 if (noflush && !list_empty(&ctx->cq_overflow_list))
1112 io_cqring_overflow_flush(ctx, false);
1114 /* See comment at the top of this file */
1116 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1119 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1121 struct io_rings *rings = ctx->rings;
1123 /* make sure SQ entry isn't read before tail */
1124 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1128 * Find and free completed poll iocbs
1130 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1131 struct list_head *done)
1133 void *reqs[IO_IOPOLL_BATCH];
1134 struct io_kiocb *req;
1138 while (!list_empty(done)) {
1139 req = list_first_entry(done, struct io_kiocb, list);
1140 list_del(&req->list);
1142 io_cqring_fill_event(req, req->result);
1145 if (refcount_dec_and_test(&req->refs)) {
1146 /* If we're not using fixed files, we have to pair the
1147 * completion part with the file put. Use regular
1148 * completions for those, only batch free for fixed
1149 * file and non-linked commands.
1151 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1152 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1154 reqs[to_free++] = req;
1155 if (to_free == ARRAY_SIZE(reqs))
1156 io_free_req_many(ctx, reqs, &to_free);
1163 io_commit_cqring(ctx);
1164 io_free_req_many(ctx, reqs, &to_free);
1167 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1170 struct io_kiocb *req, *tmp;
1176 * Only spin for completions if we don't have multiple devices hanging
1177 * off our complete list, and we're under the requested amount.
1179 spin = !ctx->poll_multi_file && *nr_events < min;
1182 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1183 struct kiocb *kiocb = &req->rw;
1186 * Move completed entries to our local list. If we find a
1187 * request that requires polling, break out and complete
1188 * the done list first, if we have entries there.
1190 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1191 list_move_tail(&req->list, &done);
1194 if (!list_empty(&done))
1197 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1206 if (!list_empty(&done))
1207 io_iopoll_complete(ctx, nr_events, &done);
1213 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1214 * non-spinning poll check - we'll still enter the driver poll loop, but only
1215 * as a non-spinning completion check.
1217 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1220 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1223 ret = io_do_iopoll(ctx, nr_events, min);
1226 if (!min || *nr_events >= min)
1234 * We can't just wait for polled events to come to us, we have to actively
1235 * find and complete them.
1237 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1239 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1242 mutex_lock(&ctx->uring_lock);
1243 while (!list_empty(&ctx->poll_list)) {
1244 unsigned int nr_events = 0;
1246 io_iopoll_getevents(ctx, &nr_events, 1);
1249 * Ensure we allow local-to-the-cpu processing to take place,
1250 * in this case we need to ensure that we reap all events.
1254 mutex_unlock(&ctx->uring_lock);
1257 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1260 int iters = 0, ret = 0;
1266 * Don't enter poll loop if we already have events pending.
1267 * If we do, we can potentially be spinning for commands that
1268 * already triggered a CQE (eg in error).
1270 if (io_cqring_events(ctx, false))
1274 * If a submit got punted to a workqueue, we can have the
1275 * application entering polling for a command before it gets
1276 * issued. That app will hold the uring_lock for the duration
1277 * of the poll right here, so we need to take a breather every
1278 * now and then to ensure that the issue has a chance to add
1279 * the poll to the issued list. Otherwise we can spin here
1280 * forever, while the workqueue is stuck trying to acquire the
1283 if (!(++iters & 7)) {
1284 mutex_unlock(&ctx->uring_lock);
1285 mutex_lock(&ctx->uring_lock);
1288 if (*nr_events < min)
1289 tmin = min - *nr_events;
1291 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1295 } while (min && !*nr_events && !need_resched());
1300 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1306 * We disallow the app entering submit/complete with polling, but we
1307 * still need to lock the ring to prevent racing with polled issue
1308 * that got punted to a workqueue.
1310 mutex_lock(&ctx->uring_lock);
1311 ret = __io_iopoll_check(ctx, nr_events, min);
1312 mutex_unlock(&ctx->uring_lock);
1316 static void kiocb_end_write(struct io_kiocb *req)
1319 * Tell lockdep we inherited freeze protection from submission
1322 if (req->flags & REQ_F_ISREG) {
1323 struct inode *inode = file_inode(req->file);
1325 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1327 file_end_write(req->file);
1330 static inline void req_set_fail_links(struct io_kiocb *req)
1332 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1333 req->flags |= REQ_F_FAIL_LINK;
1336 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1338 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1340 if (kiocb->ki_flags & IOCB_WRITE)
1341 kiocb_end_write(req);
1343 if (res != req->result)
1344 req_set_fail_links(req);
1345 io_cqring_add_event(req, res);
1348 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1350 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1352 io_complete_rw_common(kiocb, res);
1356 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1358 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1359 struct io_kiocb *nxt = NULL;
1361 io_complete_rw_common(kiocb, res);
1362 io_put_req_find_next(req, &nxt);
1367 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1369 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1371 if (kiocb->ki_flags & IOCB_WRITE)
1372 kiocb_end_write(req);
1374 if (res != req->result)
1375 req_set_fail_links(req);
1378 req->flags |= REQ_F_IOPOLL_COMPLETED;
1382 * After the iocb has been issued, it's safe to be found on the poll list.
1383 * Adding the kiocb to the list AFTER submission ensures that we don't
1384 * find it from a io_iopoll_getevents() thread before the issuer is done
1385 * accessing the kiocb cookie.
1387 static void io_iopoll_req_issued(struct io_kiocb *req)
1389 struct io_ring_ctx *ctx = req->ctx;
1392 * Track whether we have multiple files in our lists. This will impact
1393 * how we do polling eventually, not spinning if we're on potentially
1394 * different devices.
1396 if (list_empty(&ctx->poll_list)) {
1397 ctx->poll_multi_file = false;
1398 } else if (!ctx->poll_multi_file) {
1399 struct io_kiocb *list_req;
1401 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1403 if (list_req->rw.ki_filp != req->rw.ki_filp)
1404 ctx->poll_multi_file = true;
1408 * For fast devices, IO may have already completed. If it has, add
1409 * it to the front so we find it first.
1411 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1412 list_add(&req->list, &ctx->poll_list);
1414 list_add_tail(&req->list, &ctx->poll_list);
1417 static void io_file_put(struct io_submit_state *state)
1420 int diff = state->has_refs - state->used_refs;
1423 fput_many(state->file, diff);
1429 * Get as many references to a file as we have IOs left in this submission,
1430 * assuming most submissions are for one file, or at least that each file
1431 * has more than one submission.
1433 static struct file *io_file_get(struct io_submit_state *state, int fd)
1439 if (state->fd == fd) {
1446 state->file = fget_many(fd, state->ios_left);
1451 state->has_refs = state->ios_left;
1452 state->used_refs = 1;
1458 * If we tracked the file through the SCM inflight mechanism, we could support
1459 * any file. For now, just ensure that anything potentially problematic is done
1462 static bool io_file_supports_async(struct file *file)
1464 umode_t mode = file_inode(file)->i_mode;
1466 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1468 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1474 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1476 const struct io_uring_sqe *sqe = req->sqe;
1477 struct io_ring_ctx *ctx = req->ctx;
1478 struct kiocb *kiocb = &req->rw;
1485 if (S_ISREG(file_inode(req->file)->i_mode))
1486 req->flags |= REQ_F_ISREG;
1488 kiocb->ki_pos = READ_ONCE(sqe->off);
1489 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1490 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1492 ioprio = READ_ONCE(sqe->ioprio);
1494 ret = ioprio_check_cap(ioprio);
1498 kiocb->ki_ioprio = ioprio;
1500 kiocb->ki_ioprio = get_current_ioprio();
1502 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1506 /* don't allow async punt if RWF_NOWAIT was requested */
1507 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1508 (req->file->f_flags & O_NONBLOCK))
1509 req->flags |= REQ_F_NOWAIT;
1512 kiocb->ki_flags |= IOCB_NOWAIT;
1514 if (ctx->flags & IORING_SETUP_IOPOLL) {
1515 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1516 !kiocb->ki_filp->f_op->iopoll)
1519 kiocb->ki_flags |= IOCB_HIPRI;
1520 kiocb->ki_complete = io_complete_rw_iopoll;
1523 if (kiocb->ki_flags & IOCB_HIPRI)
1525 kiocb->ki_complete = io_complete_rw;
1530 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1536 case -ERESTARTNOINTR:
1537 case -ERESTARTNOHAND:
1538 case -ERESTART_RESTARTBLOCK:
1540 * We can't just restart the syscall, since previously
1541 * submitted sqes may already be in progress. Just fail this
1547 kiocb->ki_complete(kiocb, ret, 0);
1551 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1554 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1555 *nxt = __io_complete_rw(kiocb, ret);
1557 io_rw_done(kiocb, ret);
1560 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1561 const struct io_uring_sqe *sqe,
1562 struct iov_iter *iter)
1564 size_t len = READ_ONCE(sqe->len);
1565 struct io_mapped_ubuf *imu;
1566 unsigned index, buf_index;
1570 /* attempt to use fixed buffers without having provided iovecs */
1571 if (unlikely(!ctx->user_bufs))
1574 buf_index = READ_ONCE(sqe->buf_index);
1575 if (unlikely(buf_index >= ctx->nr_user_bufs))
1578 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1579 imu = &ctx->user_bufs[index];
1580 buf_addr = READ_ONCE(sqe->addr);
1583 if (buf_addr + len < buf_addr)
1585 /* not inside the mapped region */
1586 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1590 * May not be a start of buffer, set size appropriately
1591 * and advance us to the beginning.
1593 offset = buf_addr - imu->ubuf;
1594 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1598 * Don't use iov_iter_advance() here, as it's really slow for
1599 * using the latter parts of a big fixed buffer - it iterates
1600 * over each segment manually. We can cheat a bit here, because
1603 * 1) it's a BVEC iter, we set it up
1604 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1605 * first and last bvec
1607 * So just find our index, and adjust the iterator afterwards.
1608 * If the offset is within the first bvec (or the whole first
1609 * bvec, just use iov_iter_advance(). This makes it easier
1610 * since we can just skip the first segment, which may not
1611 * be PAGE_SIZE aligned.
1613 const struct bio_vec *bvec = imu->bvec;
1615 if (offset <= bvec->bv_len) {
1616 iov_iter_advance(iter, offset);
1618 unsigned long seg_skip;
1620 /* skip first vec */
1621 offset -= bvec->bv_len;
1622 seg_skip = 1 + (offset >> PAGE_SHIFT);
1624 iter->bvec = bvec + seg_skip;
1625 iter->nr_segs -= seg_skip;
1626 iter->count -= bvec->bv_len + offset;
1627 iter->iov_offset = offset & ~PAGE_MASK;
1634 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1635 struct iovec **iovec, struct iov_iter *iter)
1637 const struct io_uring_sqe *sqe = req->sqe;
1638 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1639 size_t sqe_len = READ_ONCE(sqe->len);
1643 * We're reading ->opcode for the second time, but the first read
1644 * doesn't care whether it's _FIXED or not, so it doesn't matter
1645 * whether ->opcode changes concurrently. The first read does care
1646 * about whether it is a READ or a WRITE, so we don't trust this read
1647 * for that purpose and instead let the caller pass in the read/write
1650 opcode = req->opcode;
1651 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1653 return io_import_fixed(req->ctx, rw, sqe, iter);
1657 struct io_async_rw *iorw = &req->io->rw;
1660 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1661 if (iorw->iov == iorw->fast_iov)
1669 #ifdef CONFIG_COMPAT
1670 if (req->ctx->compat)
1671 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1675 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1679 * For files that don't have ->read_iter() and ->write_iter(), handle them
1680 * by looping over ->read() or ->write() manually.
1682 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1683 struct iov_iter *iter)
1688 * Don't support polled IO through this interface, and we can't
1689 * support non-blocking either. For the latter, this just causes
1690 * the kiocb to be handled from an async context.
1692 if (kiocb->ki_flags & IOCB_HIPRI)
1694 if (kiocb->ki_flags & IOCB_NOWAIT)
1697 while (iov_iter_count(iter)) {
1701 if (!iov_iter_is_bvec(iter)) {
1702 iovec = iov_iter_iovec(iter);
1704 /* fixed buffers import bvec */
1705 iovec.iov_base = kmap(iter->bvec->bv_page)
1707 iovec.iov_len = min(iter->count,
1708 iter->bvec->bv_len - iter->iov_offset);
1712 nr = file->f_op->read(file, iovec.iov_base,
1713 iovec.iov_len, &kiocb->ki_pos);
1715 nr = file->f_op->write(file, iovec.iov_base,
1716 iovec.iov_len, &kiocb->ki_pos);
1719 if (iov_iter_is_bvec(iter))
1720 kunmap(iter->bvec->bv_page);
1728 if (nr != iovec.iov_len)
1730 iov_iter_advance(iter, nr);
1736 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
1737 struct iovec *iovec, struct iovec *fast_iov,
1738 struct iov_iter *iter)
1740 req->io->rw.nr_segs = iter->nr_segs;
1741 req->io->rw.size = io_size;
1742 req->io->rw.iov = iovec;
1743 if (!req->io->rw.iov) {
1744 req->io->rw.iov = req->io->rw.fast_iov;
1745 memcpy(req->io->rw.iov, fast_iov,
1746 sizeof(struct iovec) * iter->nr_segs);
1750 static int io_alloc_async_ctx(struct io_kiocb *req)
1752 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1754 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1755 req->sqe = &req->io->sqe;
1762 static void io_rw_async(struct io_wq_work **workptr)
1764 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1765 struct iovec *iov = NULL;
1767 if (req->io->rw.iov != req->io->rw.fast_iov)
1768 iov = req->io->rw.iov;
1769 io_wq_submit_work(workptr);
1773 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
1774 struct iovec *iovec, struct iovec *fast_iov,
1775 struct iov_iter *iter)
1777 if (!req->io && io_alloc_async_ctx(req))
1780 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
1781 req->work.func = io_rw_async;
1785 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1786 struct iov_iter *iter, bool force_nonblock)
1790 ret = io_prep_rw(req, force_nonblock);
1794 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1797 return io_import_iovec(READ, req, iovec, iter);
1800 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1801 bool force_nonblock)
1803 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1804 struct kiocb *kiocb = &req->rw;
1805 struct iov_iter iter;
1808 ssize_t io_size, ret;
1811 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1815 ret = io_import_iovec(READ, req, &iovec, &iter);
1822 if (req->flags & REQ_F_LINK)
1823 req->result = io_size;
1826 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1827 * we know to async punt it even if it was opened O_NONBLOCK
1829 if (force_nonblock && !io_file_supports_async(file)) {
1830 req->flags |= REQ_F_MUST_PUNT;
1834 iov_count = iov_iter_count(&iter);
1835 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1839 if (file->f_op->read_iter)
1840 ret2 = call_read_iter(file, kiocb, &iter);
1842 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1845 * In case of a short read, punt to async. This can happen
1846 * if we have data partially cached. Alternatively we can
1847 * return the short read, in which case the application will
1848 * need to issue another SQE and wait for it. That SQE will
1849 * need async punt anyway, so it's more efficient to do it
1852 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1853 (req->flags & REQ_F_ISREG) &&
1854 ret2 > 0 && ret2 < io_size)
1856 /* Catch -EAGAIN return for forced non-blocking submission */
1857 if (!force_nonblock || ret2 != -EAGAIN) {
1858 kiocb_done(kiocb, ret2, nxt, req->in_async);
1861 ret = io_setup_async_rw(req, io_size, iovec,
1862 inline_vecs, &iter);
1869 if (!io_wq_current_is_worker())
1874 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1875 struct iov_iter *iter, bool force_nonblock)
1879 ret = io_prep_rw(req, force_nonblock);
1883 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1886 return io_import_iovec(WRITE, req, iovec, iter);
1889 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1890 bool force_nonblock)
1892 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1893 struct kiocb *kiocb = &req->rw;
1894 struct iov_iter iter;
1897 ssize_t ret, io_size;
1900 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1904 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1909 file = kiocb->ki_filp;
1911 if (req->flags & REQ_F_LINK)
1912 req->result = io_size;
1915 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1916 * we know to async punt it even if it was opened O_NONBLOCK
1918 if (force_nonblock && !io_file_supports_async(req->file)) {
1919 req->flags |= REQ_F_MUST_PUNT;
1923 /* file path doesn't support NOWAIT for non-direct_IO */
1924 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
1925 (req->flags & REQ_F_ISREG))
1928 iov_count = iov_iter_count(&iter);
1929 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1934 * Open-code file_start_write here to grab freeze protection,
1935 * which will be released by another thread in
1936 * io_complete_rw(). Fool lockdep by telling it the lock got
1937 * released so that it doesn't complain about the held lock when
1938 * we return to userspace.
1940 if (req->flags & REQ_F_ISREG) {
1941 __sb_start_write(file_inode(file)->i_sb,
1942 SB_FREEZE_WRITE, true);
1943 __sb_writers_release(file_inode(file)->i_sb,
1946 kiocb->ki_flags |= IOCB_WRITE;
1948 if (file->f_op->write_iter)
1949 ret2 = call_write_iter(file, kiocb, &iter);
1951 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1952 if (!force_nonblock || ret2 != -EAGAIN) {
1953 kiocb_done(kiocb, ret2, nxt, req->in_async);
1956 ret = io_setup_async_rw(req, io_size, iovec,
1957 inline_vecs, &iter);
1964 if (!io_wq_current_is_worker())
1970 * IORING_OP_NOP just posts a completion event, nothing else.
1972 static int io_nop(struct io_kiocb *req)
1974 struct io_ring_ctx *ctx = req->ctx;
1976 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1979 io_cqring_add_event(req, 0);
1984 static int io_prep_fsync(struct io_kiocb *req)
1986 const struct io_uring_sqe *sqe = req->sqe;
1987 struct io_ring_ctx *ctx = req->ctx;
1989 if (req->flags & REQ_F_PREPPED)
1994 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1996 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1999 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2000 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2003 req->sync.off = READ_ONCE(sqe->off);
2004 req->sync.len = READ_ONCE(sqe->len);
2005 req->flags |= REQ_F_PREPPED;
2009 static bool io_req_cancelled(struct io_kiocb *req)
2011 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2012 req_set_fail_links(req);
2013 io_cqring_add_event(req, -ECANCELED);
2021 static void io_fsync_finish(struct io_wq_work **workptr)
2023 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2024 loff_t end = req->sync.off + req->sync.len;
2025 struct io_kiocb *nxt = NULL;
2028 if (io_req_cancelled(req))
2031 ret = vfs_fsync_range(req->rw.ki_filp, req->sync.off,
2032 end > 0 ? end : LLONG_MAX,
2033 req->sync.flags & IORING_FSYNC_DATASYNC);
2035 req_set_fail_links(req);
2036 io_cqring_add_event(req, ret);
2037 io_put_req_find_next(req, &nxt);
2039 *workptr = &nxt->work;
2042 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2043 bool force_nonblock)
2045 struct io_wq_work *work, *old_work;
2048 ret = io_prep_fsync(req);
2052 /* fsync always requires a blocking context */
2053 if (force_nonblock) {
2055 req->work.func = io_fsync_finish;
2059 work = old_work = &req->work;
2060 io_fsync_finish(&work);
2061 if (work && work != old_work)
2062 *nxt = container_of(work, struct io_kiocb, work);
2066 static int io_prep_sfr(struct io_kiocb *req)
2068 const struct io_uring_sqe *sqe = req->sqe;
2069 struct io_ring_ctx *ctx = req->ctx;
2071 if (req->flags & REQ_F_PREPPED)
2076 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2078 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2081 req->sync.off = READ_ONCE(sqe->off);
2082 req->sync.len = READ_ONCE(sqe->len);
2083 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2084 req->flags |= REQ_F_PREPPED;
2088 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2090 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2091 struct io_kiocb *nxt = NULL;
2094 if (io_req_cancelled(req))
2097 ret = sync_file_range(req->rw.ki_filp, req->sync.off, req->sync.len,
2100 req_set_fail_links(req);
2101 io_cqring_add_event(req, ret);
2102 io_put_req_find_next(req, &nxt);
2104 *workptr = &nxt->work;
2107 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2108 bool force_nonblock)
2110 struct io_wq_work *work, *old_work;
2113 ret = io_prep_sfr(req);
2117 /* sync_file_range always requires a blocking context */
2118 if (force_nonblock) {
2120 req->work.func = io_sync_file_range_finish;
2124 work = old_work = &req->work;
2125 io_sync_file_range_finish(&work);
2126 if (work && work != old_work)
2127 *nxt = container_of(work, struct io_kiocb, work);
2131 #if defined(CONFIG_NET)
2132 static void io_sendrecv_async(struct io_wq_work **workptr)
2134 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2135 struct iovec *iov = NULL;
2137 if (req->io->rw.iov != req->io->rw.fast_iov)
2138 iov = req->io->msg.iov;
2139 io_wq_submit_work(workptr);
2144 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2146 #if defined(CONFIG_NET)
2147 const struct io_uring_sqe *sqe = req->sqe;
2148 struct user_msghdr __user *msg;
2151 flags = READ_ONCE(sqe->msg_flags);
2152 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2153 io->msg.iov = io->msg.fast_iov;
2154 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2160 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2161 bool force_nonblock)
2163 #if defined(CONFIG_NET)
2164 const struct io_uring_sqe *sqe = req->sqe;
2165 struct io_async_msghdr *kmsg = NULL;
2166 struct socket *sock;
2169 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2172 sock = sock_from_file(req->file, &ret);
2174 struct io_async_ctx io;
2175 struct sockaddr_storage addr;
2178 flags = READ_ONCE(sqe->msg_flags);
2179 if (flags & MSG_DONTWAIT)
2180 req->flags |= REQ_F_NOWAIT;
2181 else if (force_nonblock)
2182 flags |= MSG_DONTWAIT;
2185 kmsg = &req->io->msg;
2186 kmsg->msg.msg_name = &addr;
2187 /* if iov is set, it's allocated already */
2189 kmsg->iov = kmsg->fast_iov;
2190 kmsg->msg.msg_iter.iov = kmsg->iov;
2193 kmsg->msg.msg_name = &addr;
2194 ret = io_sendmsg_prep(req, &io);
2199 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
2200 if (force_nonblock && ret == -EAGAIN) {
2203 if (io_alloc_async_ctx(req))
2205 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2206 req->work.func = io_sendrecv_async;
2209 if (ret == -ERESTARTSYS)
2214 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2216 io_cqring_add_event(req, ret);
2218 req_set_fail_links(req);
2219 io_put_req_find_next(req, nxt);
2226 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2228 #if defined(CONFIG_NET)
2229 const struct io_uring_sqe *sqe = req->sqe;
2230 struct user_msghdr __user *msg;
2233 flags = READ_ONCE(sqe->msg_flags);
2234 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2235 io->msg.iov = io->msg.fast_iov;
2236 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2243 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2244 bool force_nonblock)
2246 #if defined(CONFIG_NET)
2247 const struct io_uring_sqe *sqe = req->sqe;
2248 struct io_async_msghdr *kmsg = NULL;
2249 struct socket *sock;
2252 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2255 sock = sock_from_file(req->file, &ret);
2257 struct user_msghdr __user *msg;
2258 struct io_async_ctx io;
2259 struct sockaddr_storage addr;
2262 flags = READ_ONCE(sqe->msg_flags);
2263 if (flags & MSG_DONTWAIT)
2264 req->flags |= REQ_F_NOWAIT;
2265 else if (force_nonblock)
2266 flags |= MSG_DONTWAIT;
2268 msg = (struct user_msghdr __user *) (unsigned long)
2269 READ_ONCE(sqe->addr);
2271 kmsg = &req->io->msg;
2272 kmsg->msg.msg_name = &addr;
2273 /* if iov is set, it's allocated already */
2275 kmsg->iov = kmsg->fast_iov;
2276 kmsg->msg.msg_iter.iov = kmsg->iov;
2279 kmsg->msg.msg_name = &addr;
2280 ret = io_recvmsg_prep(req, &io);
2285 ret = __sys_recvmsg_sock(sock, &kmsg->msg, msg, kmsg->uaddr, flags);
2286 if (force_nonblock && ret == -EAGAIN) {
2289 if (io_alloc_async_ctx(req))
2291 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2292 req->work.func = io_sendrecv_async;
2295 if (ret == -ERESTARTSYS)
2300 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2302 io_cqring_add_event(req, ret);
2304 req_set_fail_links(req);
2305 io_put_req_find_next(req, nxt);
2312 static int io_accept_prep(struct io_kiocb *req)
2314 #if defined(CONFIG_NET)
2315 const struct io_uring_sqe *sqe = req->sqe;
2316 struct io_accept *accept = &req->accept;
2318 if (req->flags & REQ_F_PREPPED)
2321 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2323 if (sqe->ioprio || sqe->len || sqe->buf_index)
2326 accept->addr = (struct sockaddr __user *)
2327 (unsigned long) READ_ONCE(sqe->addr);
2328 accept->addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2329 accept->flags = READ_ONCE(sqe->accept_flags);
2330 req->flags |= REQ_F_PREPPED;
2337 #if defined(CONFIG_NET)
2338 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2339 bool force_nonblock)
2341 struct io_accept *accept = &req->accept;
2342 unsigned file_flags;
2345 file_flags = force_nonblock ? O_NONBLOCK : 0;
2346 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
2347 accept->addr_len, accept->flags);
2348 if (ret == -EAGAIN && force_nonblock)
2350 if (ret == -ERESTARTSYS)
2353 req_set_fail_links(req);
2354 io_cqring_add_event(req, ret);
2355 io_put_req_find_next(req, nxt);
2359 static void io_accept_finish(struct io_wq_work **workptr)
2361 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2362 struct io_kiocb *nxt = NULL;
2364 if (io_req_cancelled(req))
2366 __io_accept(req, &nxt, false);
2368 *workptr = &nxt->work;
2372 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2373 bool force_nonblock)
2375 #if defined(CONFIG_NET)
2378 ret = io_accept_prep(req);
2382 ret = __io_accept(req, nxt, force_nonblock);
2383 if (ret == -EAGAIN && force_nonblock) {
2384 req->work.func = io_accept_finish;
2385 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2395 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2397 #if defined(CONFIG_NET)
2398 const struct io_uring_sqe *sqe = req->sqe;
2399 struct sockaddr __user *addr;
2402 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2403 addr_len = READ_ONCE(sqe->addr2);
2404 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2410 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
2411 bool force_nonblock)
2413 #if defined(CONFIG_NET)
2414 const struct io_uring_sqe *sqe = req->sqe;
2415 struct io_async_ctx __io, *io;
2416 unsigned file_flags;
2419 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2421 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2424 addr_len = READ_ONCE(sqe->addr2);
2425 file_flags = force_nonblock ? O_NONBLOCK : 0;
2430 ret = io_connect_prep(req, &__io);
2436 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2438 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2441 if (io_alloc_async_ctx(req)) {
2445 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
2448 if (ret == -ERESTARTSYS)
2452 req_set_fail_links(req);
2453 io_cqring_add_event(req, ret);
2454 io_put_req_find_next(req, nxt);
2461 static void io_poll_remove_one(struct io_kiocb *req)
2463 struct io_poll_iocb *poll = &req->poll;
2465 spin_lock(&poll->head->lock);
2466 WRITE_ONCE(poll->canceled, true);
2467 if (!list_empty(&poll->wait.entry)) {
2468 list_del_init(&poll->wait.entry);
2469 io_queue_async_work(req);
2471 spin_unlock(&poll->head->lock);
2472 hash_del(&req->hash_node);
2475 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2477 struct hlist_node *tmp;
2478 struct io_kiocb *req;
2481 spin_lock_irq(&ctx->completion_lock);
2482 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2483 struct hlist_head *list;
2485 list = &ctx->cancel_hash[i];
2486 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2487 io_poll_remove_one(req);
2489 spin_unlock_irq(&ctx->completion_lock);
2492 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2494 struct hlist_head *list;
2495 struct io_kiocb *req;
2497 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2498 hlist_for_each_entry(req, list, hash_node) {
2499 if (sqe_addr == req->user_data) {
2500 io_poll_remove_one(req);
2508 static int io_poll_remove_prep(struct io_kiocb *req)
2510 const struct io_uring_sqe *sqe = req->sqe;
2512 if (req->flags & REQ_F_PREPPED)
2514 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2516 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2520 req->poll.addr = READ_ONCE(sqe->addr);
2521 req->flags |= REQ_F_PREPPED;
2526 * Find a running poll command that matches one specified in sqe->addr,
2527 * and remove it if found.
2529 static int io_poll_remove(struct io_kiocb *req)
2531 struct io_ring_ctx *ctx = req->ctx;
2535 ret = io_poll_remove_prep(req);
2539 addr = req->poll.addr;
2540 spin_lock_irq(&ctx->completion_lock);
2541 ret = io_poll_cancel(ctx, addr);
2542 spin_unlock_irq(&ctx->completion_lock);
2544 io_cqring_add_event(req, ret);
2546 req_set_fail_links(req);
2551 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2553 struct io_ring_ctx *ctx = req->ctx;
2555 req->poll.done = true;
2557 io_cqring_fill_event(req, error);
2559 io_cqring_fill_event(req, mangle_poll(mask));
2560 io_commit_cqring(ctx);
2563 static void io_poll_complete_work(struct io_wq_work **workptr)
2565 struct io_wq_work *work = *workptr;
2566 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2567 struct io_poll_iocb *poll = &req->poll;
2568 struct poll_table_struct pt = { ._key = poll->events };
2569 struct io_ring_ctx *ctx = req->ctx;
2570 struct io_kiocb *nxt = NULL;
2574 if (work->flags & IO_WQ_WORK_CANCEL) {
2575 WRITE_ONCE(poll->canceled, true);
2577 } else if (READ_ONCE(poll->canceled)) {
2581 if (ret != -ECANCELED)
2582 mask = vfs_poll(poll->file, &pt) & poll->events;
2585 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2586 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2587 * synchronize with them. In the cancellation case the list_del_init
2588 * itself is not actually needed, but harmless so we keep it in to
2589 * avoid further branches in the fast path.
2591 spin_lock_irq(&ctx->completion_lock);
2592 if (!mask && ret != -ECANCELED) {
2593 add_wait_queue(poll->head, &poll->wait);
2594 spin_unlock_irq(&ctx->completion_lock);
2597 hash_del(&req->hash_node);
2598 io_poll_complete(req, mask, ret);
2599 spin_unlock_irq(&ctx->completion_lock);
2601 io_cqring_ev_posted(ctx);
2604 req_set_fail_links(req);
2605 io_put_req_find_next(req, &nxt);
2607 *workptr = &nxt->work;
2610 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2613 struct io_poll_iocb *poll = wait->private;
2614 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2615 struct io_ring_ctx *ctx = req->ctx;
2616 __poll_t mask = key_to_poll(key);
2617 unsigned long flags;
2619 /* for instances that support it check for an event match first: */
2620 if (mask && !(mask & poll->events))
2623 list_del_init(&poll->wait.entry);
2626 * Run completion inline if we can. We're using trylock here because
2627 * we are violating the completion_lock -> poll wq lock ordering.
2628 * If we have a link timeout we're going to need the completion_lock
2629 * for finalizing the request, mark us as having grabbed that already.
2631 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2632 hash_del(&req->hash_node);
2633 io_poll_complete(req, mask, 0);
2634 req->flags |= REQ_F_COMP_LOCKED;
2636 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2638 io_cqring_ev_posted(ctx);
2640 io_queue_async_work(req);
2646 struct io_poll_table {
2647 struct poll_table_struct pt;
2648 struct io_kiocb *req;
2652 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2653 struct poll_table_struct *p)
2655 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2657 if (unlikely(pt->req->poll.head)) {
2658 pt->error = -EINVAL;
2663 pt->req->poll.head = head;
2664 add_wait_queue(head, &pt->req->poll.wait);
2667 static void io_poll_req_insert(struct io_kiocb *req)
2669 struct io_ring_ctx *ctx = req->ctx;
2670 struct hlist_head *list;
2672 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2673 hlist_add_head(&req->hash_node, list);
2676 static int io_poll_add_prep(struct io_kiocb *req)
2678 const struct io_uring_sqe *sqe = req->sqe;
2679 struct io_poll_iocb *poll = &req->poll;
2682 if (req->flags & REQ_F_PREPPED)
2684 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2686 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2691 req->flags |= REQ_F_PREPPED;
2692 events = READ_ONCE(sqe->poll_events);
2693 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2697 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
2699 struct io_poll_iocb *poll = &req->poll;
2700 struct io_ring_ctx *ctx = req->ctx;
2701 struct io_poll_table ipt;
2702 bool cancel = false;
2706 ret = io_poll_add_prep(req);
2710 INIT_IO_WORK(&req->work, io_poll_complete_work);
2711 INIT_HLIST_NODE(&req->hash_node);
2715 poll->canceled = false;
2717 ipt.pt._qproc = io_poll_queue_proc;
2718 ipt.pt._key = poll->events;
2720 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2722 /* initialized the list so that we can do list_empty checks */
2723 INIT_LIST_HEAD(&poll->wait.entry);
2724 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
2725 poll->wait.private = poll;
2727 INIT_LIST_HEAD(&req->list);
2729 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2731 spin_lock_irq(&ctx->completion_lock);
2732 if (likely(poll->head)) {
2733 spin_lock(&poll->head->lock);
2734 if (unlikely(list_empty(&poll->wait.entry))) {
2740 if (mask || ipt.error)
2741 list_del_init(&poll->wait.entry);
2743 WRITE_ONCE(poll->canceled, true);
2744 else if (!poll->done) /* actually waiting for an event */
2745 io_poll_req_insert(req);
2746 spin_unlock(&poll->head->lock);
2748 if (mask) { /* no async, we'd stolen it */
2750 io_poll_complete(req, mask, 0);
2752 spin_unlock_irq(&ctx->completion_lock);
2755 io_cqring_ev_posted(ctx);
2756 io_put_req_find_next(req, nxt);
2761 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2763 struct io_timeout_data *data = container_of(timer,
2764 struct io_timeout_data, timer);
2765 struct io_kiocb *req = data->req;
2766 struct io_ring_ctx *ctx = req->ctx;
2767 unsigned long flags;
2769 atomic_inc(&ctx->cq_timeouts);
2771 spin_lock_irqsave(&ctx->completion_lock, flags);
2773 * We could be racing with timeout deletion. If the list is empty,
2774 * then timeout lookup already found it and will be handling it.
2776 if (!list_empty(&req->list)) {
2777 struct io_kiocb *prev;
2780 * Adjust the reqs sequence before the current one because it
2781 * will consume a slot in the cq_ring and the cq_tail
2782 * pointer will be increased, otherwise other timeout reqs may
2783 * return in advance without waiting for enough wait_nr.
2786 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2788 list_del_init(&req->list);
2791 io_cqring_fill_event(req, -ETIME);
2792 io_commit_cqring(ctx);
2793 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2795 io_cqring_ev_posted(ctx);
2796 req_set_fail_links(req);
2798 return HRTIMER_NORESTART;
2801 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2803 struct io_kiocb *req;
2806 list_for_each_entry(req, &ctx->timeout_list, list) {
2807 if (user_data == req->user_data) {
2808 list_del_init(&req->list);
2817 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2821 req_set_fail_links(req);
2822 io_cqring_fill_event(req, -ECANCELED);
2827 static int io_timeout_remove_prep(struct io_kiocb *req)
2829 const struct io_uring_sqe *sqe = req->sqe;
2831 if (req->flags & REQ_F_PREPPED)
2833 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2835 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2838 req->timeout.addr = READ_ONCE(sqe->addr);
2839 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
2840 if (req->timeout.flags)
2843 req->flags |= REQ_F_PREPPED;
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 ret = io_timeout_remove_prep(req);
2859 spin_lock_irq(&ctx->completion_lock);
2860 ret = io_timeout_cancel(ctx, req->timeout.addr);
2862 io_cqring_fill_event(req, ret);
2863 io_commit_cqring(ctx);
2864 spin_unlock_irq(&ctx->completion_lock);
2865 io_cqring_ev_posted(ctx);
2867 req_set_fail_links(req);
2872 static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2873 bool is_timeout_link)
2875 const struct io_uring_sqe *sqe = req->sqe;
2876 struct io_timeout_data *data;
2879 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2881 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2883 if (sqe->off && is_timeout_link)
2885 flags = READ_ONCE(sqe->timeout_flags);
2886 if (flags & ~IORING_TIMEOUT_ABS)
2889 data = &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)
2907 const struct io_uring_sqe *sqe = req->sqe;
2909 struct io_ring_ctx *ctx = req->ctx;
2910 struct io_timeout_data *data;
2911 struct list_head *entry;
2916 if (io_alloc_async_ctx(req))
2918 ret = io_timeout_prep(req, req->io, false);
2922 data = &req->io->timeout;
2925 * sqe->off holds how many events that need to occur for this
2926 * timeout event to be satisfied. If it isn't set, then this is
2927 * a pure timeout request, sequence isn't used.
2929 count = READ_ONCE(sqe->off);
2931 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2932 spin_lock_irq(&ctx->completion_lock);
2933 entry = ctx->timeout_list.prev;
2937 req->sequence = ctx->cached_sq_head + count - 1;
2938 data->seq_offset = count;
2941 * Insertion sort, ensuring the first entry in the list is always
2942 * the one we need first.
2944 spin_lock_irq(&ctx->completion_lock);
2945 list_for_each_prev(entry, &ctx->timeout_list) {
2946 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2947 unsigned nxt_sq_head;
2948 long long tmp, tmp_nxt;
2949 u32 nxt_offset = nxt->io->timeout.seq_offset;
2951 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2955 * Since cached_sq_head + count - 1 can overflow, use type long
2958 tmp = (long long)ctx->cached_sq_head + count - 1;
2959 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2960 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2963 * cached_sq_head may overflow, and it will never overflow twice
2964 * once there is some timeout req still be valid.
2966 if (ctx->cached_sq_head < nxt_sq_head)
2973 * Sequence of reqs after the insert one and itself should
2974 * be adjusted because each timeout req consumes a slot.
2979 req->sequence -= span;
2981 list_add(&req->list, entry);
2982 data->timer.function = io_timeout_fn;
2983 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2984 spin_unlock_irq(&ctx->completion_lock);
2988 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2990 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2992 return req->user_data == (unsigned long) data;
2995 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2997 enum io_wq_cancel cancel_ret;
3000 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3001 switch (cancel_ret) {
3002 case IO_WQ_CANCEL_OK:
3005 case IO_WQ_CANCEL_RUNNING:
3008 case IO_WQ_CANCEL_NOTFOUND:
3016 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3017 struct io_kiocb *req, __u64 sqe_addr,
3018 struct io_kiocb **nxt, int success_ret)
3020 unsigned long flags;
3023 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3024 if (ret != -ENOENT) {
3025 spin_lock_irqsave(&ctx->completion_lock, flags);
3029 spin_lock_irqsave(&ctx->completion_lock, flags);
3030 ret = io_timeout_cancel(ctx, sqe_addr);
3033 ret = io_poll_cancel(ctx, sqe_addr);
3037 io_cqring_fill_event(req, ret);
3038 io_commit_cqring(ctx);
3039 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3040 io_cqring_ev_posted(ctx);
3043 req_set_fail_links(req);
3044 io_put_req_find_next(req, nxt);
3047 static int io_async_cancel_prep(struct io_kiocb *req)
3049 const struct io_uring_sqe *sqe = req->sqe;
3051 if (req->flags & REQ_F_PREPPED)
3053 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3055 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
3059 req->flags |= REQ_F_PREPPED;
3060 req->cancel.addr = READ_ONCE(sqe->addr);
3064 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
3066 struct io_ring_ctx *ctx = req->ctx;
3069 ret = io_async_cancel_prep(req);
3073 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
3077 static int io_req_defer_prep(struct io_kiocb *req)
3079 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3080 struct io_async_ctx *io = req->io;
3081 struct iov_iter iter;
3084 switch (req->opcode) {
3085 case IORING_OP_READV:
3086 case IORING_OP_READ_FIXED:
3087 /* ensure prep does right import */
3089 ret = io_read_prep(req, &iovec, &iter, true);
3093 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3096 case IORING_OP_WRITEV:
3097 case IORING_OP_WRITE_FIXED:
3098 /* ensure prep does right import */
3100 ret = io_write_prep(req, &iovec, &iter, true);
3104 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3107 case IORING_OP_POLL_ADD:
3108 ret = io_poll_add_prep(req);
3110 case IORING_OP_POLL_REMOVE:
3111 ret = io_poll_remove_prep(req);
3113 case IORING_OP_FSYNC:
3114 ret = io_prep_fsync(req);
3116 case IORING_OP_SYNC_FILE_RANGE:
3117 ret = io_prep_sfr(req);
3119 case IORING_OP_SENDMSG:
3120 ret = io_sendmsg_prep(req, io);
3122 case IORING_OP_RECVMSG:
3123 ret = io_recvmsg_prep(req, io);
3125 case IORING_OP_CONNECT:
3126 ret = io_connect_prep(req, io);
3128 case IORING_OP_TIMEOUT:
3129 ret = io_timeout_prep(req, io, false);
3131 case IORING_OP_TIMEOUT_REMOVE:
3132 ret = io_timeout_remove_prep(req);
3134 case IORING_OP_ASYNC_CANCEL:
3135 ret = io_async_cancel_prep(req);
3137 case IORING_OP_LINK_TIMEOUT:
3138 ret = io_timeout_prep(req, io, true);
3140 case IORING_OP_ACCEPT:
3141 ret = io_accept_prep(req);
3151 static int io_req_defer(struct io_kiocb *req)
3153 struct io_ring_ctx *ctx = req->ctx;
3156 /* Still need defer if there is pending req in defer list. */
3157 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
3160 if (io_alloc_async_ctx(req))
3163 ret = io_req_defer_prep(req);
3167 spin_lock_irq(&ctx->completion_lock);
3168 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
3169 spin_unlock_irq(&ctx->completion_lock);
3173 trace_io_uring_defer(ctx, req, req->user_data);
3174 list_add_tail(&req->list, &ctx->defer_list);
3175 spin_unlock_irq(&ctx->completion_lock);
3176 return -EIOCBQUEUED;
3179 __attribute__((nonnull))
3180 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
3181 bool force_nonblock)
3183 struct io_ring_ctx *ctx = req->ctx;
3186 switch (req->opcode) {
3190 case IORING_OP_READV:
3191 if (unlikely(req->sqe->buf_index))
3193 ret = io_read(req, nxt, force_nonblock);
3195 case IORING_OP_WRITEV:
3196 if (unlikely(req->sqe->buf_index))
3198 ret = io_write(req, nxt, force_nonblock);
3200 case IORING_OP_READ_FIXED:
3201 ret = io_read(req, nxt, force_nonblock);
3203 case IORING_OP_WRITE_FIXED:
3204 ret = io_write(req, nxt, force_nonblock);
3206 case IORING_OP_FSYNC:
3207 ret = io_fsync(req, nxt, force_nonblock);
3209 case IORING_OP_POLL_ADD:
3210 ret = io_poll_add(req, nxt);
3212 case IORING_OP_POLL_REMOVE:
3213 ret = io_poll_remove(req);
3215 case IORING_OP_SYNC_FILE_RANGE:
3216 ret = io_sync_file_range(req, nxt, force_nonblock);
3218 case IORING_OP_SENDMSG:
3219 ret = io_sendmsg(req, nxt, force_nonblock);
3221 case IORING_OP_RECVMSG:
3222 ret = io_recvmsg(req, nxt, force_nonblock);
3224 case IORING_OP_TIMEOUT:
3225 ret = io_timeout(req);
3227 case IORING_OP_TIMEOUT_REMOVE:
3228 ret = io_timeout_remove(req);
3230 case IORING_OP_ACCEPT:
3231 ret = io_accept(req, nxt, force_nonblock);
3233 case IORING_OP_CONNECT:
3234 ret = io_connect(req, nxt, force_nonblock);
3236 case IORING_OP_ASYNC_CANCEL:
3237 ret = io_async_cancel(req, nxt);
3247 if (ctx->flags & IORING_SETUP_IOPOLL) {
3248 if (req->result == -EAGAIN)
3251 io_iopoll_req_issued(req);
3257 static void io_link_work_cb(struct io_wq_work **workptr)
3259 struct io_wq_work *work = *workptr;
3260 struct io_kiocb *link = work->data;
3262 io_queue_linked_timeout(link);
3263 work->func = io_wq_submit_work;
3266 static void io_wq_submit_work(struct io_wq_work **workptr)
3268 struct io_wq_work *work = *workptr;
3269 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3270 struct io_kiocb *nxt = NULL;
3273 /* Ensure we clear previously set non-block flag */
3274 req->rw.ki_flags &= ~IOCB_NOWAIT;
3276 if (work->flags & IO_WQ_WORK_CANCEL)
3280 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3281 req->in_async = true;
3283 ret = io_issue_sqe(req, &nxt, false);
3285 * We can get EAGAIN for polled IO even though we're
3286 * forcing a sync submission from here, since we can't
3287 * wait for request slots on the block side.
3295 /* drop submission reference */
3299 req_set_fail_links(req);
3300 io_cqring_add_event(req, ret);
3304 /* if a dependent link is ready, pass it back */
3306 struct io_kiocb *link;
3308 io_prep_async_work(nxt, &link);
3309 *workptr = &nxt->work;
3311 nxt->work.flags |= IO_WQ_WORK_CB;
3312 nxt->work.func = io_link_work_cb;
3313 nxt->work.data = link;
3318 static bool io_req_op_valid(int op)
3320 return op >= IORING_OP_NOP && op < IORING_OP_LAST;
3323 static int io_req_needs_file(struct io_kiocb *req)
3325 switch (req->opcode) {
3327 case IORING_OP_POLL_REMOVE:
3328 case IORING_OP_TIMEOUT:
3329 case IORING_OP_TIMEOUT_REMOVE:
3330 case IORING_OP_ASYNC_CANCEL:
3331 case IORING_OP_LINK_TIMEOUT:
3334 if (io_req_op_valid(req->opcode))
3340 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3343 struct fixed_file_table *table;
3345 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3346 return table->files[index & IORING_FILE_TABLE_MASK];
3349 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3351 struct io_ring_ctx *ctx = req->ctx;
3355 flags = READ_ONCE(req->sqe->flags);
3356 fd = READ_ONCE(req->sqe->fd);
3358 if (flags & IOSQE_IO_DRAIN)
3359 req->flags |= REQ_F_IO_DRAIN;
3361 ret = io_req_needs_file(req);
3365 if (flags & IOSQE_FIXED_FILE) {
3366 if (unlikely(!ctx->file_table ||
3367 (unsigned) fd >= ctx->nr_user_files))
3369 fd = array_index_nospec(fd, ctx->nr_user_files);
3370 req->file = io_file_from_index(ctx, fd);
3373 req->flags |= REQ_F_FIXED_FILE;
3375 if (req->needs_fixed_file)
3377 trace_io_uring_file_get(ctx, fd);
3378 req->file = io_file_get(state, fd);
3379 if (unlikely(!req->file))
3386 static int io_grab_files(struct io_kiocb *req)
3389 struct io_ring_ctx *ctx = req->ctx;
3392 spin_lock_irq(&ctx->inflight_lock);
3394 * We use the f_ops->flush() handler to ensure that we can flush
3395 * out work accessing these files if the fd is closed. Check if
3396 * the fd has changed since we started down this path, and disallow
3397 * this operation if it has.
3399 if (fcheck(req->ring_fd) == req->ring_file) {
3400 list_add(&req->inflight_entry, &ctx->inflight_list);
3401 req->flags |= REQ_F_INFLIGHT;
3402 req->work.files = current->files;
3405 spin_unlock_irq(&ctx->inflight_lock);
3411 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3413 struct io_timeout_data *data = container_of(timer,
3414 struct io_timeout_data, timer);
3415 struct io_kiocb *req = data->req;
3416 struct io_ring_ctx *ctx = req->ctx;
3417 struct io_kiocb *prev = NULL;
3418 unsigned long flags;
3420 spin_lock_irqsave(&ctx->completion_lock, flags);
3423 * We don't expect the list to be empty, that will only happen if we
3424 * race with the completion of the linked work.
3426 if (!list_empty(&req->link_list)) {
3427 prev = list_entry(req->link_list.prev, struct io_kiocb,
3429 if (refcount_inc_not_zero(&prev->refs)) {
3430 list_del_init(&req->link_list);
3431 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3436 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3439 req_set_fail_links(prev);
3440 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3444 io_cqring_add_event(req, -ETIME);
3447 return HRTIMER_NORESTART;
3450 static void io_queue_linked_timeout(struct io_kiocb *req)
3452 struct io_ring_ctx *ctx = req->ctx;
3455 * If the list is now empty, then our linked request finished before
3456 * we got a chance to setup the timer
3458 spin_lock_irq(&ctx->completion_lock);
3459 if (!list_empty(&req->link_list)) {
3460 struct io_timeout_data *data = &req->io->timeout;
3462 data->timer.function = io_link_timeout_fn;
3463 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3466 spin_unlock_irq(&ctx->completion_lock);
3468 /* drop submission reference */
3472 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3474 struct io_kiocb *nxt;
3476 if (!(req->flags & REQ_F_LINK))
3479 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3481 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
3484 req->flags |= REQ_F_LINK_TIMEOUT;
3488 static void __io_queue_sqe(struct io_kiocb *req)
3490 struct io_kiocb *linked_timeout;
3491 struct io_kiocb *nxt = NULL;
3495 linked_timeout = io_prep_linked_timeout(req);
3497 ret = io_issue_sqe(req, &nxt, true);
3500 * We async punt it if the file wasn't marked NOWAIT, or if the file
3501 * doesn't support non-blocking read/write attempts
3503 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3504 (req->flags & REQ_F_MUST_PUNT))) {
3505 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3506 ret = io_grab_files(req);
3512 * Queued up for async execution, worker will release
3513 * submit reference when the iocb is actually submitted.
3515 io_queue_async_work(req);
3520 /* drop submission reference */
3523 if (linked_timeout) {
3525 io_queue_linked_timeout(linked_timeout);
3527 io_put_req(linked_timeout);
3530 /* and drop final reference, if we failed */
3532 io_cqring_add_event(req, ret);
3533 req_set_fail_links(req);
3544 static void io_queue_sqe(struct io_kiocb *req)
3548 if (unlikely(req->ctx->drain_next)) {
3549 req->flags |= REQ_F_IO_DRAIN;
3550 req->ctx->drain_next = false;
3552 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3554 ret = io_req_defer(req);
3556 if (ret != -EIOCBQUEUED) {
3557 io_cqring_add_event(req, ret);
3558 req_set_fail_links(req);
3559 io_double_put_req(req);
3562 __io_queue_sqe(req);
3565 static inline void io_queue_link_head(struct io_kiocb *req)
3567 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3568 io_cqring_add_event(req, -ECANCELED);
3569 io_double_put_req(req);
3574 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3577 static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3578 struct io_kiocb **link)
3580 struct io_ring_ctx *ctx = req->ctx;
3583 /* enforce forwards compatibility on users */
3584 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3589 ret = io_req_set_file(state, req);
3590 if (unlikely(ret)) {
3592 io_cqring_add_event(req, ret);
3593 io_double_put_req(req);
3598 * If we already have a head request, queue this one for async
3599 * submittal once the head completes. If we don't have a head but
3600 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3601 * submitted sync once the chain is complete. If none of those
3602 * conditions are true (normal request), then just queue it.
3605 struct io_kiocb *prev = *link;
3607 if (req->sqe->flags & IOSQE_IO_DRAIN)
3608 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3610 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3611 req->flags |= REQ_F_HARDLINK;
3613 if (io_alloc_async_ctx(req)) {
3618 ret = io_req_defer_prep(req);
3620 /* fail even hard links since we don't submit */
3621 prev->flags |= REQ_F_FAIL_LINK;
3624 trace_io_uring_link(ctx, req, prev);
3625 list_add_tail(&req->link_list, &prev->link_list);
3626 } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
3627 req->flags |= REQ_F_LINK;
3628 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3629 req->flags |= REQ_F_HARDLINK;
3631 INIT_LIST_HEAD(&req->link_list);
3641 * Batched submission is done, ensure local IO is flushed out.
3643 static void io_submit_state_end(struct io_submit_state *state)
3645 blk_finish_plug(&state->plug);
3647 if (state->free_reqs)
3648 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3649 &state->reqs[state->cur_req]);
3653 * Start submission side cache.
3655 static void io_submit_state_start(struct io_submit_state *state,
3656 unsigned int max_ios)
3658 blk_start_plug(&state->plug);
3659 state->free_reqs = 0;
3661 state->ios_left = max_ios;
3664 static void io_commit_sqring(struct io_ring_ctx *ctx)
3666 struct io_rings *rings = ctx->rings;
3668 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3670 * Ensure any loads from the SQEs are done at this point,
3671 * since once we write the new head, the application could
3672 * write new data to them.
3674 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3679 * Fetch an sqe, if one is available. Note that req->sqe will point to memory
3680 * that is mapped by userspace. This means that care needs to be taken to
3681 * ensure that reads are stable, as we cannot rely on userspace always
3682 * being a good citizen. If members of the sqe are validated and then later
3683 * used, it's important that those reads are done through READ_ONCE() to
3684 * prevent a re-load down the line.
3686 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3688 struct io_rings *rings = ctx->rings;
3689 u32 *sq_array = ctx->sq_array;
3693 * The cached sq head (or cq tail) serves two purposes:
3695 * 1) allows us to batch the cost of updating the user visible
3697 * 2) allows the kernel side to track the head on its own, even
3698 * though the application is the one updating it.
3700 head = ctx->cached_sq_head;
3701 /* make sure SQ entry isn't read before tail */
3702 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3705 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3706 if (likely(head < ctx->sq_entries)) {
3708 * All io need record the previous position, if LINK vs DARIN,
3709 * it can be used to mark the position of the first IO in the
3712 req->sequence = ctx->cached_sq_head;
3713 req->sqe = &ctx->sq_sqes[head];
3714 req->opcode = READ_ONCE(req->sqe->opcode);
3715 req->user_data = READ_ONCE(req->sqe->user_data);
3716 ctx->cached_sq_head++;
3720 /* drop invalid entries */
3721 ctx->cached_sq_head++;
3722 ctx->cached_sq_dropped++;
3723 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3727 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3728 struct file *ring_file, int ring_fd,
3729 struct mm_struct **mm, bool async)
3731 struct io_submit_state state, *statep = NULL;
3732 struct io_kiocb *link = NULL;
3733 int i, submitted = 0;
3734 bool mm_fault = false;
3736 /* if we have a backlog and couldn't flush it all, return BUSY */
3737 if (!list_empty(&ctx->cq_overflow_list) &&
3738 !io_cqring_overflow_flush(ctx, false))
3741 if (nr > IO_PLUG_THRESHOLD) {
3742 io_submit_state_start(&state, nr);
3746 for (i = 0; i < nr; i++) {
3747 struct io_kiocb *req;
3748 unsigned int sqe_flags;
3750 req = io_get_req(ctx, statep);
3751 if (unlikely(!req)) {
3753 submitted = -EAGAIN;
3756 if (!io_get_sqring(ctx, req)) {
3761 if (io_req_needs_user(req) && !*mm) {
3762 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3764 use_mm(ctx->sqo_mm);
3770 sqe_flags = req->sqe->flags;
3772 req->ring_file = ring_file;
3773 req->ring_fd = ring_fd;
3774 req->has_user = *mm != NULL;
3775 req->in_async = async;
3776 req->needs_fixed_file = async;
3777 trace_io_uring_submit_sqe(ctx, req->user_data, true, async);
3778 if (!io_submit_sqe(req, statep, &link))
3781 * If previous wasn't linked and we have a linked command,
3782 * that's the end of the chain. Submit the previous link.
3784 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) && link) {
3785 io_queue_link_head(link);
3791 io_queue_link_head(link);
3793 io_submit_state_end(&state);
3795 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3796 io_commit_sqring(ctx);
3801 static int io_sq_thread(void *data)
3803 struct io_ring_ctx *ctx = data;
3804 struct mm_struct *cur_mm = NULL;
3805 const struct cred *old_cred;
3806 mm_segment_t old_fs;
3809 unsigned long timeout;
3812 complete(&ctx->completions[1]);
3816 old_cred = override_creds(ctx->creds);
3818 ret = timeout = inflight = 0;
3819 while (!kthread_should_park()) {
3820 unsigned int to_submit;
3823 unsigned nr_events = 0;
3825 if (ctx->flags & IORING_SETUP_IOPOLL) {
3827 * inflight is the count of the maximum possible
3828 * entries we submitted, but it can be smaller
3829 * if we dropped some of them. If we don't have
3830 * poll entries available, then we know that we
3831 * have nothing left to poll for. Reset the
3832 * inflight count to zero in that case.
3834 mutex_lock(&ctx->uring_lock);
3835 if (!list_empty(&ctx->poll_list))
3836 __io_iopoll_check(ctx, &nr_events, 0);
3839 mutex_unlock(&ctx->uring_lock);
3842 * Normal IO, just pretend everything completed.
3843 * We don't have to poll completions for that.
3845 nr_events = inflight;
3848 inflight -= nr_events;
3850 timeout = jiffies + ctx->sq_thread_idle;
3853 to_submit = io_sqring_entries(ctx);
3856 * If submit got -EBUSY, flag us as needing the application
3857 * to enter the kernel to reap and flush events.
3859 if (!to_submit || ret == -EBUSY) {
3861 * We're polling. If we're within the defined idle
3862 * period, then let us spin without work before going
3863 * to sleep. The exception is if we got EBUSY doing
3864 * more IO, we should wait for the application to
3865 * reap events and wake us up.
3868 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3874 * Drop cur_mm before scheduling, we can't hold it for
3875 * long periods (or over schedule()). Do this before
3876 * adding ourselves to the waitqueue, as the unuse/drop
3885 prepare_to_wait(&ctx->sqo_wait, &wait,
3886 TASK_INTERRUPTIBLE);
3888 /* Tell userspace we may need a wakeup call */
3889 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3890 /* make sure to read SQ tail after writing flags */
3893 to_submit = io_sqring_entries(ctx);
3894 if (!to_submit || ret == -EBUSY) {
3895 if (kthread_should_park()) {
3896 finish_wait(&ctx->sqo_wait, &wait);
3899 if (signal_pending(current))
3900 flush_signals(current);
3902 finish_wait(&ctx->sqo_wait, &wait);
3904 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3907 finish_wait(&ctx->sqo_wait, &wait);
3909 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3912 to_submit = min(to_submit, ctx->sq_entries);
3913 mutex_lock(&ctx->uring_lock);
3914 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3915 mutex_unlock(&ctx->uring_lock);
3925 revert_creds(old_cred);
3932 struct io_wait_queue {
3933 struct wait_queue_entry wq;
3934 struct io_ring_ctx *ctx;
3936 unsigned nr_timeouts;
3939 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3941 struct io_ring_ctx *ctx = iowq->ctx;
3944 * Wake up if we have enough events, or if a timeout occurred since we
3945 * started waiting. For timeouts, we always want to return to userspace,
3946 * regardless of event count.
3948 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3949 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3952 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3953 int wake_flags, void *key)
3955 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3958 /* use noflush == true, as we can't safely rely on locking context */
3959 if (!io_should_wake(iowq, true))
3962 return autoremove_wake_function(curr, mode, wake_flags, key);
3966 * Wait until events become available, if we don't already have some. The
3967 * application must reap them itself, as they reside on the shared cq ring.
3969 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3970 const sigset_t __user *sig, size_t sigsz)
3972 struct io_wait_queue iowq = {
3975 .func = io_wake_function,
3976 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3979 .to_wait = min_events,
3981 struct io_rings *rings = ctx->rings;
3984 if (io_cqring_events(ctx, false) >= min_events)
3988 #ifdef CONFIG_COMPAT
3989 if (in_compat_syscall())
3990 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3994 ret = set_user_sigmask(sig, sigsz);
4000 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
4001 trace_io_uring_cqring_wait(ctx, min_events);
4003 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
4004 TASK_INTERRUPTIBLE);
4005 if (io_should_wake(&iowq, false))
4008 if (signal_pending(current)) {
4013 finish_wait(&ctx->wait, &iowq.wq);
4015 restore_saved_sigmask_unless(ret == -EINTR);
4017 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
4020 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
4022 #if defined(CONFIG_UNIX)
4023 if (ctx->ring_sock) {
4024 struct sock *sock = ctx->ring_sock->sk;
4025 struct sk_buff *skb;
4027 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
4033 for (i = 0; i < ctx->nr_user_files; i++) {
4036 file = io_file_from_index(ctx, i);
4043 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
4045 unsigned nr_tables, i;
4047 if (!ctx->file_table)
4050 __io_sqe_files_unregister(ctx);
4051 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
4052 for (i = 0; i < nr_tables; i++)
4053 kfree(ctx->file_table[i].files);
4054 kfree(ctx->file_table);
4055 ctx->file_table = NULL;
4056 ctx->nr_user_files = 0;
4060 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
4062 if (ctx->sqo_thread) {
4063 wait_for_completion(&ctx->completions[1]);
4065 * The park is a bit of a work-around, without it we get
4066 * warning spews on shutdown with SQPOLL set and affinity
4067 * set to a single CPU.
4069 kthread_park(ctx->sqo_thread);
4070 kthread_stop(ctx->sqo_thread);
4071 ctx->sqo_thread = NULL;
4075 static void io_finish_async(struct io_ring_ctx *ctx)
4077 io_sq_thread_stop(ctx);
4080 io_wq_destroy(ctx->io_wq);
4085 #if defined(CONFIG_UNIX)
4086 static void io_destruct_skb(struct sk_buff *skb)
4088 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
4091 io_wq_flush(ctx->io_wq);
4093 unix_destruct_scm(skb);
4097 * Ensure the UNIX gc is aware of our file set, so we are certain that
4098 * the io_uring can be safely unregistered on process exit, even if we have
4099 * loops in the file referencing.
4101 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
4103 struct sock *sk = ctx->ring_sock->sk;
4104 struct scm_fp_list *fpl;
4105 struct sk_buff *skb;
4108 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
4109 unsigned long inflight = ctx->user->unix_inflight + nr;
4111 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
4115 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
4119 skb = alloc_skb(0, GFP_KERNEL);
4128 fpl->user = get_uid(ctx->user);
4129 for (i = 0; i < nr; i++) {
4130 struct file *file = io_file_from_index(ctx, i + offset);
4134 fpl->fp[nr_files] = get_file(file);
4135 unix_inflight(fpl->user, fpl->fp[nr_files]);
4140 fpl->max = SCM_MAX_FD;
4141 fpl->count = nr_files;
4142 UNIXCB(skb).fp = fpl;
4143 skb->destructor = io_destruct_skb;
4144 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
4145 skb_queue_head(&sk->sk_receive_queue, skb);
4147 for (i = 0; i < nr_files; i++)
4158 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
4159 * causes regular reference counting to break down. We rely on the UNIX
4160 * garbage collection to take care of this problem for us.
4162 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4164 unsigned left, total;
4168 left = ctx->nr_user_files;
4170 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
4172 ret = __io_sqe_files_scm(ctx, this_files, total);
4176 total += this_files;
4182 while (total < ctx->nr_user_files) {
4183 struct file *file = io_file_from_index(ctx, total);
4193 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4199 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
4204 for (i = 0; i < nr_tables; i++) {
4205 struct fixed_file_table *table = &ctx->file_table[i];
4206 unsigned this_files;
4208 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
4209 table->files = kcalloc(this_files, sizeof(struct file *),
4213 nr_files -= this_files;
4219 for (i = 0; i < nr_tables; i++) {
4220 struct fixed_file_table *table = &ctx->file_table[i];
4221 kfree(table->files);
4226 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
4229 __s32 __user *fds = (__s32 __user *) arg;
4234 if (ctx->file_table)
4238 if (nr_args > IORING_MAX_FIXED_FILES)
4241 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
4242 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
4244 if (!ctx->file_table)
4247 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
4248 kfree(ctx->file_table);
4249 ctx->file_table = NULL;
4253 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
4254 struct fixed_file_table *table;
4258 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4260 /* allow sparse sets */
4266 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4267 index = i & IORING_FILE_TABLE_MASK;
4268 table->files[index] = fget(fd);
4271 if (!table->files[index])
4274 * Don't allow io_uring instances to be registered. If UNIX
4275 * isn't enabled, then this causes a reference cycle and this
4276 * instance can never get freed. If UNIX is enabled we'll
4277 * handle it just fine, but there's still no point in allowing
4278 * a ring fd as it doesn't support regular read/write anyway.
4280 if (table->files[index]->f_op == &io_uring_fops) {
4281 fput(table->files[index]);
4288 for (i = 0; i < ctx->nr_user_files; i++) {
4291 file = io_file_from_index(ctx, i);
4295 for (i = 0; i < nr_tables; i++)
4296 kfree(ctx->file_table[i].files);
4298 kfree(ctx->file_table);
4299 ctx->file_table = NULL;
4300 ctx->nr_user_files = 0;
4304 ret = io_sqe_files_scm(ctx);
4306 io_sqe_files_unregister(ctx);
4311 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4313 #if defined(CONFIG_UNIX)
4314 struct file *file = io_file_from_index(ctx, index);
4315 struct sock *sock = ctx->ring_sock->sk;
4316 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4317 struct sk_buff *skb;
4320 __skb_queue_head_init(&list);
4323 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4324 * remove this entry and rearrange the file array.
4326 skb = skb_dequeue(head);
4328 struct scm_fp_list *fp;
4330 fp = UNIXCB(skb).fp;
4331 for (i = 0; i < fp->count; i++) {
4334 if (fp->fp[i] != file)
4337 unix_notinflight(fp->user, fp->fp[i]);
4338 left = fp->count - 1 - i;
4340 memmove(&fp->fp[i], &fp->fp[i + 1],
4341 left * sizeof(struct file *));
4348 __skb_queue_tail(&list, skb);
4358 __skb_queue_tail(&list, skb);
4360 skb = skb_dequeue(head);
4363 if (skb_peek(&list)) {
4364 spin_lock_irq(&head->lock);
4365 while ((skb = __skb_dequeue(&list)) != NULL)
4366 __skb_queue_tail(head, skb);
4367 spin_unlock_irq(&head->lock);
4370 fput(io_file_from_index(ctx, index));
4374 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4377 #if defined(CONFIG_UNIX)
4378 struct sock *sock = ctx->ring_sock->sk;
4379 struct sk_buff_head *head = &sock->sk_receive_queue;
4380 struct sk_buff *skb;
4383 * See if we can merge this file into an existing skb SCM_RIGHTS
4384 * file set. If there's no room, fall back to allocating a new skb
4385 * and filling it in.
4387 spin_lock_irq(&head->lock);
4388 skb = skb_peek(head);
4390 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4392 if (fpl->count < SCM_MAX_FD) {
4393 __skb_unlink(skb, head);
4394 spin_unlock_irq(&head->lock);
4395 fpl->fp[fpl->count] = get_file(file);
4396 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4398 spin_lock_irq(&head->lock);
4399 __skb_queue_head(head, skb);
4404 spin_unlock_irq(&head->lock);
4411 return __io_sqe_files_scm(ctx, 1, index);
4417 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4420 struct io_uring_files_update up;
4425 if (!ctx->file_table)
4429 if (copy_from_user(&up, arg, sizeof(up)))
4431 if (check_add_overflow(up.offset, nr_args, &done))
4433 if (done > ctx->nr_user_files)
4437 fds = (__s32 __user *) up.fds;
4439 struct fixed_file_table *table;
4443 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4447 i = array_index_nospec(up.offset, ctx->nr_user_files);
4448 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4449 index = i & IORING_FILE_TABLE_MASK;
4450 if (table->files[index]) {
4451 io_sqe_file_unregister(ctx, i);
4452 table->files[index] = NULL;
4463 * Don't allow io_uring instances to be registered. If
4464 * UNIX isn't enabled, then this causes a reference
4465 * cycle and this instance can never get freed. If UNIX
4466 * is enabled we'll handle it just fine, but there's
4467 * still no point in allowing a ring fd as it doesn't
4468 * support regular read/write anyway.
4470 if (file->f_op == &io_uring_fops) {
4475 table->files[index] = file;
4476 err = io_sqe_file_register(ctx, file, i);
4485 return done ? done : err;
4488 static void io_put_work(struct io_wq_work *work)
4490 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4495 static void io_get_work(struct io_wq_work *work)
4497 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4499 refcount_inc(&req->refs);
4502 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4503 struct io_uring_params *p)
4505 struct io_wq_data data;
4506 unsigned concurrency;
4509 init_waitqueue_head(&ctx->sqo_wait);
4510 mmgrab(current->mm);
4511 ctx->sqo_mm = current->mm;
4513 if (ctx->flags & IORING_SETUP_SQPOLL) {
4515 if (!capable(CAP_SYS_ADMIN))
4518 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4519 if (!ctx->sq_thread_idle)
4520 ctx->sq_thread_idle = HZ;
4522 if (p->flags & IORING_SETUP_SQ_AFF) {
4523 int cpu = p->sq_thread_cpu;
4526 if (cpu >= nr_cpu_ids)
4528 if (!cpu_online(cpu))
4531 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4535 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4538 if (IS_ERR(ctx->sqo_thread)) {
4539 ret = PTR_ERR(ctx->sqo_thread);
4540 ctx->sqo_thread = NULL;
4543 wake_up_process(ctx->sqo_thread);
4544 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4545 /* Can't have SQ_AFF without SQPOLL */
4550 data.mm = ctx->sqo_mm;
4551 data.user = ctx->user;
4552 data.creds = ctx->creds;
4553 data.get_work = io_get_work;
4554 data.put_work = io_put_work;
4556 /* Do QD, or 4 * CPUS, whatever is smallest */
4557 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4558 ctx->io_wq = io_wq_create(concurrency, &data);
4559 if (IS_ERR(ctx->io_wq)) {
4560 ret = PTR_ERR(ctx->io_wq);
4567 io_finish_async(ctx);
4568 mmdrop(ctx->sqo_mm);
4573 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4575 atomic_long_sub(nr_pages, &user->locked_vm);
4578 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4580 unsigned long page_limit, cur_pages, new_pages;
4582 /* Don't allow more pages than we can safely lock */
4583 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4586 cur_pages = atomic_long_read(&user->locked_vm);
4587 new_pages = cur_pages + nr_pages;
4588 if (new_pages > page_limit)
4590 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4591 new_pages) != cur_pages);
4596 static void io_mem_free(void *ptr)
4603 page = virt_to_head_page(ptr);
4604 if (put_page_testzero(page))
4605 free_compound_page(page);
4608 static void *io_mem_alloc(size_t size)
4610 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4613 return (void *) __get_free_pages(gfp_flags, get_order(size));
4616 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4619 struct io_rings *rings;
4620 size_t off, sq_array_size;
4622 off = struct_size(rings, cqes, cq_entries);
4623 if (off == SIZE_MAX)
4627 off = ALIGN(off, SMP_CACHE_BYTES);
4632 sq_array_size = array_size(sizeof(u32), sq_entries);
4633 if (sq_array_size == SIZE_MAX)
4636 if (check_add_overflow(off, sq_array_size, &off))
4645 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4649 pages = (size_t)1 << get_order(
4650 rings_size(sq_entries, cq_entries, NULL));
4651 pages += (size_t)1 << get_order(
4652 array_size(sizeof(struct io_uring_sqe), sq_entries));
4657 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4661 if (!ctx->user_bufs)
4664 for (i = 0; i < ctx->nr_user_bufs; i++) {
4665 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4667 for (j = 0; j < imu->nr_bvecs; j++)
4668 put_user_page(imu->bvec[j].bv_page);
4670 if (ctx->account_mem)
4671 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4676 kfree(ctx->user_bufs);
4677 ctx->user_bufs = NULL;
4678 ctx->nr_user_bufs = 0;
4682 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4683 void __user *arg, unsigned index)
4685 struct iovec __user *src;
4687 #ifdef CONFIG_COMPAT
4689 struct compat_iovec __user *ciovs;
4690 struct compat_iovec ciov;
4692 ciovs = (struct compat_iovec __user *) arg;
4693 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4696 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4697 dst->iov_len = ciov.iov_len;
4701 src = (struct iovec __user *) arg;
4702 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4707 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4710 struct vm_area_struct **vmas = NULL;
4711 struct page **pages = NULL;
4712 int i, j, got_pages = 0;
4717 if (!nr_args || nr_args > UIO_MAXIOV)
4720 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4722 if (!ctx->user_bufs)
4725 for (i = 0; i < nr_args; i++) {
4726 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4727 unsigned long off, start, end, ubuf;
4732 ret = io_copy_iov(ctx, &iov, arg, i);
4737 * Don't impose further limits on the size and buffer
4738 * constraints here, we'll -EINVAL later when IO is
4739 * submitted if they are wrong.
4742 if (!iov.iov_base || !iov.iov_len)
4745 /* arbitrary limit, but we need something */
4746 if (iov.iov_len > SZ_1G)
4749 ubuf = (unsigned long) iov.iov_base;
4750 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4751 start = ubuf >> PAGE_SHIFT;
4752 nr_pages = end - start;
4754 if (ctx->account_mem) {
4755 ret = io_account_mem(ctx->user, nr_pages);
4761 if (!pages || nr_pages > got_pages) {
4764 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4766 vmas = kvmalloc_array(nr_pages,
4767 sizeof(struct vm_area_struct *),
4769 if (!pages || !vmas) {
4771 if (ctx->account_mem)
4772 io_unaccount_mem(ctx->user, nr_pages);
4775 got_pages = nr_pages;
4778 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4782 if (ctx->account_mem)
4783 io_unaccount_mem(ctx->user, nr_pages);
4788 down_read(¤t->mm->mmap_sem);
4789 pret = get_user_pages(ubuf, nr_pages,
4790 FOLL_WRITE | FOLL_LONGTERM,
4792 if (pret == nr_pages) {
4793 /* don't support file backed memory */
4794 for (j = 0; j < nr_pages; j++) {
4795 struct vm_area_struct *vma = vmas[j];
4798 !is_file_hugepages(vma->vm_file)) {
4804 ret = pret < 0 ? pret : -EFAULT;
4806 up_read(¤t->mm->mmap_sem);
4809 * if we did partial map, or found file backed vmas,
4810 * release any pages we did get
4813 put_user_pages(pages, pret);
4814 if (ctx->account_mem)
4815 io_unaccount_mem(ctx->user, nr_pages);
4820 off = ubuf & ~PAGE_MASK;
4822 for (j = 0; j < nr_pages; j++) {
4825 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4826 imu->bvec[j].bv_page = pages[j];
4827 imu->bvec[j].bv_len = vec_len;
4828 imu->bvec[j].bv_offset = off;
4832 /* store original address for later verification */
4834 imu->len = iov.iov_len;
4835 imu->nr_bvecs = nr_pages;
4837 ctx->nr_user_bufs++;
4845 io_sqe_buffer_unregister(ctx);
4849 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4851 __s32 __user *fds = arg;
4857 if (copy_from_user(&fd, fds, sizeof(*fds)))
4860 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4861 if (IS_ERR(ctx->cq_ev_fd)) {
4862 int ret = PTR_ERR(ctx->cq_ev_fd);
4863 ctx->cq_ev_fd = NULL;
4870 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4872 if (ctx->cq_ev_fd) {
4873 eventfd_ctx_put(ctx->cq_ev_fd);
4874 ctx->cq_ev_fd = NULL;
4881 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4883 io_finish_async(ctx);
4885 mmdrop(ctx->sqo_mm);
4887 io_iopoll_reap_events(ctx);
4888 io_sqe_buffer_unregister(ctx);
4889 io_sqe_files_unregister(ctx);
4890 io_eventfd_unregister(ctx);
4892 #if defined(CONFIG_UNIX)
4893 if (ctx->ring_sock) {
4894 ctx->ring_sock->file = NULL; /* so that iput() is called */
4895 sock_release(ctx->ring_sock);
4899 io_mem_free(ctx->rings);
4900 io_mem_free(ctx->sq_sqes);
4902 percpu_ref_exit(&ctx->refs);
4903 if (ctx->account_mem)
4904 io_unaccount_mem(ctx->user,
4905 ring_pages(ctx->sq_entries, ctx->cq_entries));
4906 free_uid(ctx->user);
4907 put_cred(ctx->creds);
4908 kfree(ctx->completions);
4909 kfree(ctx->cancel_hash);
4910 kmem_cache_free(req_cachep, ctx->fallback_req);
4914 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4916 struct io_ring_ctx *ctx = file->private_data;
4919 poll_wait(file, &ctx->cq_wait, wait);
4921 * synchronizes with barrier from wq_has_sleeper call in
4925 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4926 ctx->rings->sq_ring_entries)
4927 mask |= EPOLLOUT | EPOLLWRNORM;
4928 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4929 mask |= EPOLLIN | EPOLLRDNORM;
4934 static int io_uring_fasync(int fd, struct file *file, int on)
4936 struct io_ring_ctx *ctx = file->private_data;
4938 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4941 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4943 mutex_lock(&ctx->uring_lock);
4944 percpu_ref_kill(&ctx->refs);
4945 mutex_unlock(&ctx->uring_lock);
4947 io_kill_timeouts(ctx);
4948 io_poll_remove_all(ctx);
4951 io_wq_cancel_all(ctx->io_wq);
4953 io_iopoll_reap_events(ctx);
4954 /* if we failed setting up the ctx, we might not have any rings */
4956 io_cqring_overflow_flush(ctx, true);
4957 wait_for_completion(&ctx->completions[0]);
4958 io_ring_ctx_free(ctx);
4961 static int io_uring_release(struct inode *inode, struct file *file)
4963 struct io_ring_ctx *ctx = file->private_data;
4965 file->private_data = NULL;
4966 io_ring_ctx_wait_and_kill(ctx);
4970 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4971 struct files_struct *files)
4973 struct io_kiocb *req;
4976 while (!list_empty_careful(&ctx->inflight_list)) {
4977 struct io_kiocb *cancel_req = NULL;
4979 spin_lock_irq(&ctx->inflight_lock);
4980 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4981 if (req->work.files != files)
4983 /* req is being completed, ignore */
4984 if (!refcount_inc_not_zero(&req->refs))
4990 prepare_to_wait(&ctx->inflight_wait, &wait,
4991 TASK_UNINTERRUPTIBLE);
4992 spin_unlock_irq(&ctx->inflight_lock);
4994 /* We need to keep going until we don't find a matching req */
4998 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4999 io_put_req(cancel_req);
5002 finish_wait(&ctx->inflight_wait, &wait);
5005 static int io_uring_flush(struct file *file, void *data)
5007 struct io_ring_ctx *ctx = file->private_data;
5009 io_uring_cancel_files(ctx, data);
5010 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
5011 io_cqring_overflow_flush(ctx, true);
5012 io_wq_cancel_all(ctx->io_wq);
5017 static void *io_uring_validate_mmap_request(struct file *file,
5018 loff_t pgoff, size_t sz)
5020 struct io_ring_ctx *ctx = file->private_data;
5021 loff_t offset = pgoff << PAGE_SHIFT;
5026 case IORING_OFF_SQ_RING:
5027 case IORING_OFF_CQ_RING:
5030 case IORING_OFF_SQES:
5034 return ERR_PTR(-EINVAL);
5037 page = virt_to_head_page(ptr);
5038 if (sz > page_size(page))
5039 return ERR_PTR(-EINVAL);
5046 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5048 size_t sz = vma->vm_end - vma->vm_start;
5052 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
5054 return PTR_ERR(ptr);
5056 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
5057 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
5060 #else /* !CONFIG_MMU */
5062 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5064 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
5067 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
5069 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
5072 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
5073 unsigned long addr, unsigned long len,
5074 unsigned long pgoff, unsigned long flags)
5078 ptr = io_uring_validate_mmap_request(file, pgoff, len);
5080 return PTR_ERR(ptr);
5082 return (unsigned long) ptr;
5085 #endif /* !CONFIG_MMU */
5087 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
5088 u32, min_complete, u32, flags, const sigset_t __user *, sig,
5091 struct io_ring_ctx *ctx;
5096 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
5104 if (f.file->f_op != &io_uring_fops)
5108 ctx = f.file->private_data;
5109 if (!percpu_ref_tryget(&ctx->refs))
5113 * For SQ polling, the thread will do all submissions and completions.
5114 * Just return the requested submit count, and wake the thread if
5118 if (ctx->flags & IORING_SETUP_SQPOLL) {
5119 if (!list_empty_careful(&ctx->cq_overflow_list))
5120 io_cqring_overflow_flush(ctx, false);
5121 if (flags & IORING_ENTER_SQ_WAKEUP)
5122 wake_up(&ctx->sqo_wait);
5123 submitted = to_submit;
5124 } else if (to_submit) {
5125 struct mm_struct *cur_mm;
5127 to_submit = min(to_submit, ctx->sq_entries);
5128 mutex_lock(&ctx->uring_lock);
5129 /* already have mm, so io_submit_sqes() won't try to grab it */
5130 cur_mm = ctx->sqo_mm;
5131 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
5133 mutex_unlock(&ctx->uring_lock);
5135 if (flags & IORING_ENTER_GETEVENTS) {
5136 unsigned nr_events = 0;
5138 min_complete = min(min_complete, ctx->cq_entries);
5140 if (ctx->flags & IORING_SETUP_IOPOLL) {
5141 ret = io_iopoll_check(ctx, &nr_events, min_complete);
5143 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
5147 percpu_ref_put(&ctx->refs);
5150 return submitted ? submitted : ret;
5153 static const struct file_operations io_uring_fops = {
5154 .release = io_uring_release,
5155 .flush = io_uring_flush,
5156 .mmap = io_uring_mmap,
5158 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
5159 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
5161 .poll = io_uring_poll,
5162 .fasync = io_uring_fasync,
5165 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
5166 struct io_uring_params *p)
5168 struct io_rings *rings;
5169 size_t size, sq_array_offset;
5171 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
5172 if (size == SIZE_MAX)
5175 rings = io_mem_alloc(size);
5180 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
5181 rings->sq_ring_mask = p->sq_entries - 1;
5182 rings->cq_ring_mask = p->cq_entries - 1;
5183 rings->sq_ring_entries = p->sq_entries;
5184 rings->cq_ring_entries = p->cq_entries;
5185 ctx->sq_mask = rings->sq_ring_mask;
5186 ctx->cq_mask = rings->cq_ring_mask;
5187 ctx->sq_entries = rings->sq_ring_entries;
5188 ctx->cq_entries = rings->cq_ring_entries;
5190 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
5191 if (size == SIZE_MAX) {
5192 io_mem_free(ctx->rings);
5197 ctx->sq_sqes = io_mem_alloc(size);
5198 if (!ctx->sq_sqes) {
5199 io_mem_free(ctx->rings);
5208 * Allocate an anonymous fd, this is what constitutes the application
5209 * visible backing of an io_uring instance. The application mmaps this
5210 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
5211 * we have to tie this fd to a socket for file garbage collection purposes.
5213 static int io_uring_get_fd(struct io_ring_ctx *ctx)
5218 #if defined(CONFIG_UNIX)
5219 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
5225 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
5229 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
5230 O_RDWR | O_CLOEXEC);
5233 ret = PTR_ERR(file);
5237 #if defined(CONFIG_UNIX)
5238 ctx->ring_sock->file = file;
5239 ctx->ring_sock->sk->sk_user_data = ctx;
5241 fd_install(ret, file);
5244 #if defined(CONFIG_UNIX)
5245 sock_release(ctx->ring_sock);
5246 ctx->ring_sock = NULL;
5251 static int io_uring_create(unsigned entries, struct io_uring_params *p)
5253 struct user_struct *user = NULL;
5254 struct io_ring_ctx *ctx;
5258 if (!entries || entries > IORING_MAX_ENTRIES)
5262 * Use twice as many entries for the CQ ring. It's possible for the
5263 * application to drive a higher depth than the size of the SQ ring,
5264 * since the sqes are only used at submission time. This allows for
5265 * some flexibility in overcommitting a bit. If the application has
5266 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5267 * of CQ ring entries manually.
5269 p->sq_entries = roundup_pow_of_two(entries);
5270 if (p->flags & IORING_SETUP_CQSIZE) {
5272 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5273 * to a power-of-two, if it isn't already. We do NOT impose
5274 * any cq vs sq ring sizing.
5276 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5278 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5280 p->cq_entries = 2 * p->sq_entries;
5283 user = get_uid(current_user());
5284 account_mem = !capable(CAP_IPC_LOCK);
5287 ret = io_account_mem(user,
5288 ring_pages(p->sq_entries, p->cq_entries));
5295 ctx = io_ring_ctx_alloc(p);
5298 io_unaccount_mem(user, ring_pages(p->sq_entries,
5303 ctx->compat = in_compat_syscall();
5304 ctx->account_mem = account_mem;
5306 ctx->creds = get_current_cred();
5308 ret = io_allocate_scq_urings(ctx, p);
5312 ret = io_sq_offload_start(ctx, p);
5316 memset(&p->sq_off, 0, sizeof(p->sq_off));
5317 p->sq_off.head = offsetof(struct io_rings, sq.head);
5318 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5319 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5320 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5321 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5322 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5323 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5325 memset(&p->cq_off, 0, sizeof(p->cq_off));
5326 p->cq_off.head = offsetof(struct io_rings, cq.head);
5327 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5328 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5329 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5330 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5331 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5334 * Install ring fd as the very last thing, so we don't risk someone
5335 * having closed it before we finish setup
5337 ret = io_uring_get_fd(ctx);
5341 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5342 IORING_FEAT_SUBMIT_STABLE;
5343 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5346 io_ring_ctx_wait_and_kill(ctx);
5351 * Sets up an aio uring context, and returns the fd. Applications asks for a
5352 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5353 * params structure passed in.
5355 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5357 struct io_uring_params p;
5361 if (copy_from_user(&p, params, sizeof(p)))
5363 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5368 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5369 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5372 ret = io_uring_create(entries, &p);
5376 if (copy_to_user(params, &p, sizeof(p)))
5382 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5383 struct io_uring_params __user *, params)
5385 return io_uring_setup(entries, params);
5388 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5389 void __user *arg, unsigned nr_args)
5390 __releases(ctx->uring_lock)
5391 __acquires(ctx->uring_lock)
5396 * We're inside the ring mutex, if the ref is already dying, then
5397 * someone else killed the ctx or is already going through
5398 * io_uring_register().
5400 if (percpu_ref_is_dying(&ctx->refs))
5403 percpu_ref_kill(&ctx->refs);
5406 * Drop uring mutex before waiting for references to exit. If another
5407 * thread is currently inside io_uring_enter() it might need to grab
5408 * the uring_lock to make progress. If we hold it here across the drain
5409 * wait, then we can deadlock. It's safe to drop the mutex here, since
5410 * no new references will come in after we've killed the percpu ref.
5412 mutex_unlock(&ctx->uring_lock);
5413 wait_for_completion(&ctx->completions[0]);
5414 mutex_lock(&ctx->uring_lock);
5417 case IORING_REGISTER_BUFFERS:
5418 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5420 case IORING_UNREGISTER_BUFFERS:
5424 ret = io_sqe_buffer_unregister(ctx);
5426 case IORING_REGISTER_FILES:
5427 ret = io_sqe_files_register(ctx, arg, nr_args);
5429 case IORING_UNREGISTER_FILES:
5433 ret = io_sqe_files_unregister(ctx);
5435 case IORING_REGISTER_FILES_UPDATE:
5436 ret = io_sqe_files_update(ctx, arg, nr_args);
5438 case IORING_REGISTER_EVENTFD:
5442 ret = io_eventfd_register(ctx, arg);
5444 case IORING_UNREGISTER_EVENTFD:
5448 ret = io_eventfd_unregister(ctx);
5455 /* bring the ctx back to life */
5456 reinit_completion(&ctx->completions[0]);
5457 percpu_ref_reinit(&ctx->refs);
5461 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5462 void __user *, arg, unsigned int, nr_args)
5464 struct io_ring_ctx *ctx;
5473 if (f.file->f_op != &io_uring_fops)
5476 ctx = f.file->private_data;
5478 mutex_lock(&ctx->uring_lock);
5479 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5480 mutex_unlock(&ctx->uring_lock);
5481 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5482 ctx->cq_ev_fd != NULL, ret);
5488 static int __init io_uring_init(void)
5490 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5493 __initcall(io_uring_init);