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);
1820 /* Ensure we clear previously set non-block flag */
1821 if (!force_nonblock)
1822 req->rw.ki_flags &= ~IOCB_NOWAIT;
1826 if (req->flags & REQ_F_LINK)
1827 req->result = io_size;
1830 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1831 * we know to async punt it even if it was opened O_NONBLOCK
1833 if (force_nonblock && !io_file_supports_async(file)) {
1834 req->flags |= REQ_F_MUST_PUNT;
1838 iov_count = iov_iter_count(&iter);
1839 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1843 if (file->f_op->read_iter)
1844 ret2 = call_read_iter(file, kiocb, &iter);
1846 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1849 * In case of a short read, punt to async. This can happen
1850 * if we have data partially cached. Alternatively we can
1851 * return the short read, in which case the application will
1852 * need to issue another SQE and wait for it. That SQE will
1853 * need async punt anyway, so it's more efficient to do it
1856 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1857 (req->flags & REQ_F_ISREG) &&
1858 ret2 > 0 && ret2 < io_size)
1860 /* Catch -EAGAIN return for forced non-blocking submission */
1861 if (!force_nonblock || ret2 != -EAGAIN) {
1862 kiocb_done(kiocb, ret2, nxt, req->in_async);
1865 ret = io_setup_async_rw(req, io_size, iovec,
1866 inline_vecs, &iter);
1873 if (!io_wq_current_is_worker())
1878 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1879 struct iov_iter *iter, bool force_nonblock)
1883 ret = io_prep_rw(req, force_nonblock);
1887 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1890 return io_import_iovec(WRITE, req, iovec, iter);
1893 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1894 bool force_nonblock)
1896 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1897 struct kiocb *kiocb = &req->rw;
1898 struct iov_iter iter;
1901 ssize_t ret, io_size;
1904 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1908 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1913 /* Ensure we clear previously set non-block flag */
1914 if (!force_nonblock)
1915 req->rw.ki_flags &= ~IOCB_NOWAIT;
1917 file = kiocb->ki_filp;
1919 if (req->flags & REQ_F_LINK)
1920 req->result = io_size;
1923 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1924 * we know to async punt it even if it was opened O_NONBLOCK
1926 if (force_nonblock && !io_file_supports_async(req->file)) {
1927 req->flags |= REQ_F_MUST_PUNT;
1931 /* file path doesn't support NOWAIT for non-direct_IO */
1932 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
1933 (req->flags & REQ_F_ISREG))
1936 iov_count = iov_iter_count(&iter);
1937 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1942 * Open-code file_start_write here to grab freeze protection,
1943 * which will be released by another thread in
1944 * io_complete_rw(). Fool lockdep by telling it the lock got
1945 * released so that it doesn't complain about the held lock when
1946 * we return to userspace.
1948 if (req->flags & REQ_F_ISREG) {
1949 __sb_start_write(file_inode(file)->i_sb,
1950 SB_FREEZE_WRITE, true);
1951 __sb_writers_release(file_inode(file)->i_sb,
1954 kiocb->ki_flags |= IOCB_WRITE;
1956 if (file->f_op->write_iter)
1957 ret2 = call_write_iter(file, kiocb, &iter);
1959 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1960 if (!force_nonblock || ret2 != -EAGAIN) {
1961 kiocb_done(kiocb, ret2, nxt, req->in_async);
1964 ret = io_setup_async_rw(req, io_size, iovec,
1965 inline_vecs, &iter);
1972 if (!io_wq_current_is_worker())
1978 * IORING_OP_NOP just posts a completion event, nothing else.
1980 static int io_nop(struct io_kiocb *req)
1982 struct io_ring_ctx *ctx = req->ctx;
1984 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1987 io_cqring_add_event(req, 0);
1992 static int io_prep_fsync(struct io_kiocb *req)
1994 const struct io_uring_sqe *sqe = req->sqe;
1995 struct io_ring_ctx *ctx = req->ctx;
1997 if (req->flags & REQ_F_PREPPED)
2002 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2004 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2007 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2008 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2011 req->sync.off = READ_ONCE(sqe->off);
2012 req->sync.len = READ_ONCE(sqe->len);
2013 req->flags |= REQ_F_PREPPED;
2017 static bool io_req_cancelled(struct io_kiocb *req)
2019 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2020 req_set_fail_links(req);
2021 io_cqring_add_event(req, -ECANCELED);
2029 static void io_fsync_finish(struct io_wq_work **workptr)
2031 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2032 loff_t end = req->sync.off + req->sync.len;
2033 struct io_kiocb *nxt = NULL;
2036 if (io_req_cancelled(req))
2039 ret = vfs_fsync_range(req->rw.ki_filp, req->sync.off,
2040 end > 0 ? end : LLONG_MAX,
2041 req->sync.flags & IORING_FSYNC_DATASYNC);
2043 req_set_fail_links(req);
2044 io_cqring_add_event(req, ret);
2045 io_put_req_find_next(req, &nxt);
2047 *workptr = &nxt->work;
2050 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2051 bool force_nonblock)
2053 struct io_wq_work *work, *old_work;
2056 ret = io_prep_fsync(req);
2060 /* fsync always requires a blocking context */
2061 if (force_nonblock) {
2063 req->work.func = io_fsync_finish;
2067 work = old_work = &req->work;
2068 io_fsync_finish(&work);
2069 if (work && work != old_work)
2070 *nxt = container_of(work, struct io_kiocb, work);
2074 static int io_prep_sfr(struct io_kiocb *req)
2076 const struct io_uring_sqe *sqe = req->sqe;
2077 struct io_ring_ctx *ctx = req->ctx;
2079 if (req->flags & REQ_F_PREPPED)
2084 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2086 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2089 req->sync.off = READ_ONCE(sqe->off);
2090 req->sync.len = READ_ONCE(sqe->len);
2091 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2092 req->flags |= REQ_F_PREPPED;
2096 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2098 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2099 struct io_kiocb *nxt = NULL;
2102 if (io_req_cancelled(req))
2105 ret = sync_file_range(req->rw.ki_filp, req->sync.off, req->sync.len,
2108 req_set_fail_links(req);
2109 io_cqring_add_event(req, ret);
2110 io_put_req_find_next(req, &nxt);
2112 *workptr = &nxt->work;
2115 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2116 bool force_nonblock)
2118 struct io_wq_work *work, *old_work;
2121 ret = io_prep_sfr(req);
2125 /* sync_file_range always requires a blocking context */
2126 if (force_nonblock) {
2128 req->work.func = io_sync_file_range_finish;
2132 work = old_work = &req->work;
2133 io_sync_file_range_finish(&work);
2134 if (work && work != old_work)
2135 *nxt = container_of(work, struct io_kiocb, work);
2139 #if defined(CONFIG_NET)
2140 static void io_sendrecv_async(struct io_wq_work **workptr)
2142 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2143 struct iovec *iov = NULL;
2145 if (req->io->rw.iov != req->io->rw.fast_iov)
2146 iov = req->io->msg.iov;
2147 io_wq_submit_work(workptr);
2152 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2154 #if defined(CONFIG_NET)
2155 const struct io_uring_sqe *sqe = req->sqe;
2156 struct user_msghdr __user *msg;
2159 flags = READ_ONCE(sqe->msg_flags);
2160 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2161 io->msg.iov = io->msg.fast_iov;
2162 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2168 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2169 bool force_nonblock)
2171 #if defined(CONFIG_NET)
2172 const struct io_uring_sqe *sqe = req->sqe;
2173 struct io_async_msghdr *kmsg = NULL;
2174 struct socket *sock;
2177 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2180 sock = sock_from_file(req->file, &ret);
2182 struct io_async_ctx io;
2183 struct sockaddr_storage addr;
2186 flags = READ_ONCE(sqe->msg_flags);
2187 if (flags & MSG_DONTWAIT)
2188 req->flags |= REQ_F_NOWAIT;
2189 else if (force_nonblock)
2190 flags |= MSG_DONTWAIT;
2193 kmsg = &req->io->msg;
2194 kmsg->msg.msg_name = &addr;
2195 /* if iov is set, it's allocated already */
2197 kmsg->iov = kmsg->fast_iov;
2198 kmsg->msg.msg_iter.iov = kmsg->iov;
2201 kmsg->msg.msg_name = &addr;
2202 ret = io_sendmsg_prep(req, &io);
2207 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
2208 if (force_nonblock && ret == -EAGAIN) {
2211 if (io_alloc_async_ctx(req))
2213 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2214 req->work.func = io_sendrecv_async;
2217 if (ret == -ERESTARTSYS)
2222 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2224 io_cqring_add_event(req, ret);
2226 req_set_fail_links(req);
2227 io_put_req_find_next(req, nxt);
2234 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2236 #if defined(CONFIG_NET)
2237 const struct io_uring_sqe *sqe = req->sqe;
2238 struct user_msghdr __user *msg;
2241 flags = READ_ONCE(sqe->msg_flags);
2242 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2243 io->msg.iov = io->msg.fast_iov;
2244 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2251 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2252 bool force_nonblock)
2254 #if defined(CONFIG_NET)
2255 const struct io_uring_sqe *sqe = req->sqe;
2256 struct io_async_msghdr *kmsg = NULL;
2257 struct socket *sock;
2260 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2263 sock = sock_from_file(req->file, &ret);
2265 struct user_msghdr __user *msg;
2266 struct io_async_ctx io;
2267 struct sockaddr_storage addr;
2270 flags = READ_ONCE(sqe->msg_flags);
2271 if (flags & MSG_DONTWAIT)
2272 req->flags |= REQ_F_NOWAIT;
2273 else if (force_nonblock)
2274 flags |= MSG_DONTWAIT;
2276 msg = (struct user_msghdr __user *) (unsigned long)
2277 READ_ONCE(sqe->addr);
2279 kmsg = &req->io->msg;
2280 kmsg->msg.msg_name = &addr;
2281 /* if iov is set, it's allocated already */
2283 kmsg->iov = kmsg->fast_iov;
2284 kmsg->msg.msg_iter.iov = kmsg->iov;
2287 kmsg->msg.msg_name = &addr;
2288 ret = io_recvmsg_prep(req, &io);
2293 ret = __sys_recvmsg_sock(sock, &kmsg->msg, msg, kmsg->uaddr, flags);
2294 if (force_nonblock && ret == -EAGAIN) {
2297 if (io_alloc_async_ctx(req))
2299 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2300 req->work.func = io_sendrecv_async;
2303 if (ret == -ERESTARTSYS)
2308 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2310 io_cqring_add_event(req, ret);
2312 req_set_fail_links(req);
2313 io_put_req_find_next(req, nxt);
2320 static int io_accept_prep(struct io_kiocb *req)
2322 #if defined(CONFIG_NET)
2323 const struct io_uring_sqe *sqe = req->sqe;
2324 struct io_accept *accept = &req->accept;
2326 if (req->flags & REQ_F_PREPPED)
2329 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2331 if (sqe->ioprio || sqe->len || sqe->buf_index)
2334 accept->addr = (struct sockaddr __user *)
2335 (unsigned long) READ_ONCE(sqe->addr);
2336 accept->addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2337 accept->flags = READ_ONCE(sqe->accept_flags);
2338 req->flags |= REQ_F_PREPPED;
2345 #if defined(CONFIG_NET)
2346 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2347 bool force_nonblock)
2349 struct io_accept *accept = &req->accept;
2350 unsigned file_flags;
2353 file_flags = force_nonblock ? O_NONBLOCK : 0;
2354 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
2355 accept->addr_len, accept->flags);
2356 if (ret == -EAGAIN && force_nonblock)
2358 if (ret == -ERESTARTSYS)
2361 req_set_fail_links(req);
2362 io_cqring_add_event(req, ret);
2363 io_put_req_find_next(req, nxt);
2367 static void io_accept_finish(struct io_wq_work **workptr)
2369 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2370 struct io_kiocb *nxt = NULL;
2372 if (io_req_cancelled(req))
2374 __io_accept(req, &nxt, false);
2376 *workptr = &nxt->work;
2380 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2381 bool force_nonblock)
2383 #if defined(CONFIG_NET)
2386 ret = io_accept_prep(req);
2390 ret = __io_accept(req, nxt, force_nonblock);
2391 if (ret == -EAGAIN && force_nonblock) {
2392 req->work.func = io_accept_finish;
2393 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2403 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2405 #if defined(CONFIG_NET)
2406 const struct io_uring_sqe *sqe = req->sqe;
2407 struct sockaddr __user *addr;
2410 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2411 addr_len = READ_ONCE(sqe->addr2);
2412 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2418 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
2419 bool force_nonblock)
2421 #if defined(CONFIG_NET)
2422 const struct io_uring_sqe *sqe = req->sqe;
2423 struct io_async_ctx __io, *io;
2424 unsigned file_flags;
2427 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2429 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2432 addr_len = READ_ONCE(sqe->addr2);
2433 file_flags = force_nonblock ? O_NONBLOCK : 0;
2438 ret = io_connect_prep(req, &__io);
2444 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2446 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2449 if (io_alloc_async_ctx(req)) {
2453 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
2456 if (ret == -ERESTARTSYS)
2460 req_set_fail_links(req);
2461 io_cqring_add_event(req, ret);
2462 io_put_req_find_next(req, nxt);
2469 static void io_poll_remove_one(struct io_kiocb *req)
2471 struct io_poll_iocb *poll = &req->poll;
2473 spin_lock(&poll->head->lock);
2474 WRITE_ONCE(poll->canceled, true);
2475 if (!list_empty(&poll->wait.entry)) {
2476 list_del_init(&poll->wait.entry);
2477 io_queue_async_work(req);
2479 spin_unlock(&poll->head->lock);
2480 hash_del(&req->hash_node);
2483 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2485 struct hlist_node *tmp;
2486 struct io_kiocb *req;
2489 spin_lock_irq(&ctx->completion_lock);
2490 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2491 struct hlist_head *list;
2493 list = &ctx->cancel_hash[i];
2494 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2495 io_poll_remove_one(req);
2497 spin_unlock_irq(&ctx->completion_lock);
2500 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2502 struct hlist_head *list;
2503 struct io_kiocb *req;
2505 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2506 hlist_for_each_entry(req, list, hash_node) {
2507 if (sqe_addr == req->user_data) {
2508 io_poll_remove_one(req);
2516 static int io_poll_remove_prep(struct io_kiocb *req)
2518 const struct io_uring_sqe *sqe = req->sqe;
2520 if (req->flags & REQ_F_PREPPED)
2522 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2524 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2528 req->poll.addr = READ_ONCE(sqe->addr);
2529 req->flags |= REQ_F_PREPPED;
2534 * Find a running poll command that matches one specified in sqe->addr,
2535 * and remove it if found.
2537 static int io_poll_remove(struct io_kiocb *req)
2539 struct io_ring_ctx *ctx = req->ctx;
2543 ret = io_poll_remove_prep(req);
2547 addr = req->poll.addr;
2548 spin_lock_irq(&ctx->completion_lock);
2549 ret = io_poll_cancel(ctx, addr);
2550 spin_unlock_irq(&ctx->completion_lock);
2552 io_cqring_add_event(req, ret);
2554 req_set_fail_links(req);
2559 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2561 struct io_ring_ctx *ctx = req->ctx;
2563 req->poll.done = true;
2565 io_cqring_fill_event(req, error);
2567 io_cqring_fill_event(req, mangle_poll(mask));
2568 io_commit_cqring(ctx);
2571 static void io_poll_complete_work(struct io_wq_work **workptr)
2573 struct io_wq_work *work = *workptr;
2574 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2575 struct io_poll_iocb *poll = &req->poll;
2576 struct poll_table_struct pt = { ._key = poll->events };
2577 struct io_ring_ctx *ctx = req->ctx;
2578 struct io_kiocb *nxt = NULL;
2582 if (work->flags & IO_WQ_WORK_CANCEL) {
2583 WRITE_ONCE(poll->canceled, true);
2585 } else if (READ_ONCE(poll->canceled)) {
2589 if (ret != -ECANCELED)
2590 mask = vfs_poll(poll->file, &pt) & poll->events;
2593 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2594 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2595 * synchronize with them. In the cancellation case the list_del_init
2596 * itself is not actually needed, but harmless so we keep it in to
2597 * avoid further branches in the fast path.
2599 spin_lock_irq(&ctx->completion_lock);
2600 if (!mask && ret != -ECANCELED) {
2601 add_wait_queue(poll->head, &poll->wait);
2602 spin_unlock_irq(&ctx->completion_lock);
2605 hash_del(&req->hash_node);
2606 io_poll_complete(req, mask, ret);
2607 spin_unlock_irq(&ctx->completion_lock);
2609 io_cqring_ev_posted(ctx);
2612 req_set_fail_links(req);
2613 io_put_req_find_next(req, &nxt);
2615 *workptr = &nxt->work;
2618 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2621 struct io_poll_iocb *poll = wait->private;
2622 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2623 struct io_ring_ctx *ctx = req->ctx;
2624 __poll_t mask = key_to_poll(key);
2625 unsigned long flags;
2627 /* for instances that support it check for an event match first: */
2628 if (mask && !(mask & poll->events))
2631 list_del_init(&poll->wait.entry);
2634 * Run completion inline if we can. We're using trylock here because
2635 * we are violating the completion_lock -> poll wq lock ordering.
2636 * If we have a link timeout we're going to need the completion_lock
2637 * for finalizing the request, mark us as having grabbed that already.
2639 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2640 hash_del(&req->hash_node);
2641 io_poll_complete(req, mask, 0);
2642 req->flags |= REQ_F_COMP_LOCKED;
2644 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2646 io_cqring_ev_posted(ctx);
2648 io_queue_async_work(req);
2654 struct io_poll_table {
2655 struct poll_table_struct pt;
2656 struct io_kiocb *req;
2660 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2661 struct poll_table_struct *p)
2663 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2665 if (unlikely(pt->req->poll.head)) {
2666 pt->error = -EINVAL;
2671 pt->req->poll.head = head;
2672 add_wait_queue(head, &pt->req->poll.wait);
2675 static void io_poll_req_insert(struct io_kiocb *req)
2677 struct io_ring_ctx *ctx = req->ctx;
2678 struct hlist_head *list;
2680 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2681 hlist_add_head(&req->hash_node, list);
2684 static int io_poll_add_prep(struct io_kiocb *req)
2686 const struct io_uring_sqe *sqe = req->sqe;
2687 struct io_poll_iocb *poll = &req->poll;
2690 if (req->flags & REQ_F_PREPPED)
2692 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2694 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2699 req->flags |= REQ_F_PREPPED;
2700 events = READ_ONCE(sqe->poll_events);
2701 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2705 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
2707 struct io_poll_iocb *poll = &req->poll;
2708 struct io_ring_ctx *ctx = req->ctx;
2709 struct io_poll_table ipt;
2710 bool cancel = false;
2714 ret = io_poll_add_prep(req);
2718 INIT_IO_WORK(&req->work, io_poll_complete_work);
2719 INIT_HLIST_NODE(&req->hash_node);
2723 poll->canceled = false;
2725 ipt.pt._qproc = io_poll_queue_proc;
2726 ipt.pt._key = poll->events;
2728 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2730 /* initialized the list so that we can do list_empty checks */
2731 INIT_LIST_HEAD(&poll->wait.entry);
2732 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
2733 poll->wait.private = poll;
2735 INIT_LIST_HEAD(&req->list);
2737 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2739 spin_lock_irq(&ctx->completion_lock);
2740 if (likely(poll->head)) {
2741 spin_lock(&poll->head->lock);
2742 if (unlikely(list_empty(&poll->wait.entry))) {
2748 if (mask || ipt.error)
2749 list_del_init(&poll->wait.entry);
2751 WRITE_ONCE(poll->canceled, true);
2752 else if (!poll->done) /* actually waiting for an event */
2753 io_poll_req_insert(req);
2754 spin_unlock(&poll->head->lock);
2756 if (mask) { /* no async, we'd stolen it */
2758 io_poll_complete(req, mask, 0);
2760 spin_unlock_irq(&ctx->completion_lock);
2763 io_cqring_ev_posted(ctx);
2764 io_put_req_find_next(req, nxt);
2769 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2771 struct io_timeout_data *data = container_of(timer,
2772 struct io_timeout_data, timer);
2773 struct io_kiocb *req = data->req;
2774 struct io_ring_ctx *ctx = req->ctx;
2775 unsigned long flags;
2777 atomic_inc(&ctx->cq_timeouts);
2779 spin_lock_irqsave(&ctx->completion_lock, flags);
2781 * We could be racing with timeout deletion. If the list is empty,
2782 * then timeout lookup already found it and will be handling it.
2784 if (!list_empty(&req->list)) {
2785 struct io_kiocb *prev;
2788 * Adjust the reqs sequence before the current one because it
2789 * will consume a slot in the cq_ring and the cq_tail
2790 * pointer will be increased, otherwise other timeout reqs may
2791 * return in advance without waiting for enough wait_nr.
2794 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2796 list_del_init(&req->list);
2799 io_cqring_fill_event(req, -ETIME);
2800 io_commit_cqring(ctx);
2801 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2803 io_cqring_ev_posted(ctx);
2804 req_set_fail_links(req);
2806 return HRTIMER_NORESTART;
2809 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2811 struct io_kiocb *req;
2814 list_for_each_entry(req, &ctx->timeout_list, list) {
2815 if (user_data == req->user_data) {
2816 list_del_init(&req->list);
2825 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2829 req_set_fail_links(req);
2830 io_cqring_fill_event(req, -ECANCELED);
2835 static int io_timeout_remove_prep(struct io_kiocb *req)
2837 const struct io_uring_sqe *sqe = req->sqe;
2839 if (req->flags & REQ_F_PREPPED)
2841 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2843 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2846 req->timeout.addr = READ_ONCE(sqe->addr);
2847 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
2848 if (req->timeout.flags)
2851 req->flags |= REQ_F_PREPPED;
2856 * Remove or update an existing timeout command
2858 static int io_timeout_remove(struct io_kiocb *req)
2860 struct io_ring_ctx *ctx = req->ctx;
2863 ret = io_timeout_remove_prep(req);
2867 spin_lock_irq(&ctx->completion_lock);
2868 ret = io_timeout_cancel(ctx, req->timeout.addr);
2870 io_cqring_fill_event(req, ret);
2871 io_commit_cqring(ctx);
2872 spin_unlock_irq(&ctx->completion_lock);
2873 io_cqring_ev_posted(ctx);
2875 req_set_fail_links(req);
2880 static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2881 bool is_timeout_link)
2883 const struct io_uring_sqe *sqe = req->sqe;
2884 struct io_timeout_data *data;
2887 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2889 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2891 if (sqe->off && is_timeout_link)
2893 flags = READ_ONCE(sqe->timeout_flags);
2894 if (flags & ~IORING_TIMEOUT_ABS)
2897 data = &io->timeout;
2899 req->flags |= REQ_F_TIMEOUT;
2901 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2904 if (flags & IORING_TIMEOUT_ABS)
2905 data->mode = HRTIMER_MODE_ABS;
2907 data->mode = HRTIMER_MODE_REL;
2909 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2913 static int io_timeout(struct io_kiocb *req)
2915 const struct io_uring_sqe *sqe = req->sqe;
2917 struct io_ring_ctx *ctx = req->ctx;
2918 struct io_timeout_data *data;
2919 struct list_head *entry;
2924 if (io_alloc_async_ctx(req))
2926 ret = io_timeout_prep(req, req->io, false);
2930 data = &req->io->timeout;
2933 * sqe->off holds how many events that need to occur for this
2934 * timeout event to be satisfied. If it isn't set, then this is
2935 * a pure timeout request, sequence isn't used.
2937 count = READ_ONCE(sqe->off);
2939 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2940 spin_lock_irq(&ctx->completion_lock);
2941 entry = ctx->timeout_list.prev;
2945 req->sequence = ctx->cached_sq_head + count - 1;
2946 data->seq_offset = count;
2949 * Insertion sort, ensuring the first entry in the list is always
2950 * the one we need first.
2952 spin_lock_irq(&ctx->completion_lock);
2953 list_for_each_prev(entry, &ctx->timeout_list) {
2954 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2955 unsigned nxt_sq_head;
2956 long long tmp, tmp_nxt;
2957 u32 nxt_offset = nxt->io->timeout.seq_offset;
2959 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2963 * Since cached_sq_head + count - 1 can overflow, use type long
2966 tmp = (long long)ctx->cached_sq_head + count - 1;
2967 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2968 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2971 * cached_sq_head may overflow, and it will never overflow twice
2972 * once there is some timeout req still be valid.
2974 if (ctx->cached_sq_head < nxt_sq_head)
2981 * Sequence of reqs after the insert one and itself should
2982 * be adjusted because each timeout req consumes a slot.
2987 req->sequence -= span;
2989 list_add(&req->list, entry);
2990 data->timer.function = io_timeout_fn;
2991 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2992 spin_unlock_irq(&ctx->completion_lock);
2996 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2998 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3000 return req->user_data == (unsigned long) data;
3003 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
3005 enum io_wq_cancel cancel_ret;
3008 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
3009 switch (cancel_ret) {
3010 case IO_WQ_CANCEL_OK:
3013 case IO_WQ_CANCEL_RUNNING:
3016 case IO_WQ_CANCEL_NOTFOUND:
3024 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
3025 struct io_kiocb *req, __u64 sqe_addr,
3026 struct io_kiocb **nxt, int success_ret)
3028 unsigned long flags;
3031 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
3032 if (ret != -ENOENT) {
3033 spin_lock_irqsave(&ctx->completion_lock, flags);
3037 spin_lock_irqsave(&ctx->completion_lock, flags);
3038 ret = io_timeout_cancel(ctx, sqe_addr);
3041 ret = io_poll_cancel(ctx, sqe_addr);
3045 io_cqring_fill_event(req, ret);
3046 io_commit_cqring(ctx);
3047 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3048 io_cqring_ev_posted(ctx);
3051 req_set_fail_links(req);
3052 io_put_req_find_next(req, nxt);
3055 static int io_async_cancel_prep(struct io_kiocb *req)
3057 const struct io_uring_sqe *sqe = req->sqe;
3059 if (req->flags & REQ_F_PREPPED)
3061 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3063 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
3067 req->flags |= REQ_F_PREPPED;
3068 req->cancel.addr = READ_ONCE(sqe->addr);
3072 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
3074 struct io_ring_ctx *ctx = req->ctx;
3077 ret = io_async_cancel_prep(req);
3081 io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0);
3085 static int io_req_defer_prep(struct io_kiocb *req)
3087 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3088 struct io_async_ctx *io = req->io;
3089 struct iov_iter iter;
3092 switch (req->opcode) {
3095 case IORING_OP_READV:
3096 case IORING_OP_READ_FIXED:
3097 /* ensure prep does right import */
3099 ret = io_read_prep(req, &iovec, &iter, true);
3103 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3106 case IORING_OP_WRITEV:
3107 case IORING_OP_WRITE_FIXED:
3108 /* ensure prep does right import */
3110 ret = io_write_prep(req, &iovec, &iter, true);
3114 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3117 case IORING_OP_POLL_ADD:
3118 ret = io_poll_add_prep(req);
3120 case IORING_OP_POLL_REMOVE:
3121 ret = io_poll_remove_prep(req);
3123 case IORING_OP_FSYNC:
3124 ret = io_prep_fsync(req);
3126 case IORING_OP_SYNC_FILE_RANGE:
3127 ret = io_prep_sfr(req);
3129 case IORING_OP_SENDMSG:
3130 ret = io_sendmsg_prep(req, io);
3132 case IORING_OP_RECVMSG:
3133 ret = io_recvmsg_prep(req, io);
3135 case IORING_OP_CONNECT:
3136 ret = io_connect_prep(req, io);
3138 case IORING_OP_TIMEOUT:
3139 ret = io_timeout_prep(req, io, false);
3141 case IORING_OP_TIMEOUT_REMOVE:
3142 ret = io_timeout_remove_prep(req);
3144 case IORING_OP_ASYNC_CANCEL:
3145 ret = io_async_cancel_prep(req);
3147 case IORING_OP_LINK_TIMEOUT:
3148 ret = io_timeout_prep(req, io, true);
3150 case IORING_OP_ACCEPT:
3151 ret = io_accept_prep(req);
3154 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
3163 static int io_req_defer(struct io_kiocb *req)
3165 struct io_ring_ctx *ctx = req->ctx;
3168 /* Still need defer if there is pending req in defer list. */
3169 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
3172 if (io_alloc_async_ctx(req))
3175 ret = io_req_defer_prep(req);
3179 spin_lock_irq(&ctx->completion_lock);
3180 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
3181 spin_unlock_irq(&ctx->completion_lock);
3185 trace_io_uring_defer(ctx, req, req->user_data);
3186 list_add_tail(&req->list, &ctx->defer_list);
3187 spin_unlock_irq(&ctx->completion_lock);
3188 return -EIOCBQUEUED;
3191 __attribute__((nonnull))
3192 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
3193 bool force_nonblock)
3195 struct io_ring_ctx *ctx = req->ctx;
3198 switch (req->opcode) {
3202 case IORING_OP_READV:
3203 if (unlikely(req->sqe->buf_index))
3205 ret = io_read(req, nxt, force_nonblock);
3207 case IORING_OP_WRITEV:
3208 if (unlikely(req->sqe->buf_index))
3210 ret = io_write(req, nxt, force_nonblock);
3212 case IORING_OP_READ_FIXED:
3213 ret = io_read(req, nxt, force_nonblock);
3215 case IORING_OP_WRITE_FIXED:
3216 ret = io_write(req, nxt, force_nonblock);
3218 case IORING_OP_FSYNC:
3219 ret = io_fsync(req, nxt, force_nonblock);
3221 case IORING_OP_POLL_ADD:
3222 ret = io_poll_add(req, nxt);
3224 case IORING_OP_POLL_REMOVE:
3225 ret = io_poll_remove(req);
3227 case IORING_OP_SYNC_FILE_RANGE:
3228 ret = io_sync_file_range(req, nxt, force_nonblock);
3230 case IORING_OP_SENDMSG:
3231 ret = io_sendmsg(req, nxt, force_nonblock);
3233 case IORING_OP_RECVMSG:
3234 ret = io_recvmsg(req, nxt, force_nonblock);
3236 case IORING_OP_TIMEOUT:
3237 ret = io_timeout(req);
3239 case IORING_OP_TIMEOUT_REMOVE:
3240 ret = io_timeout_remove(req);
3242 case IORING_OP_ACCEPT:
3243 ret = io_accept(req, nxt, force_nonblock);
3245 case IORING_OP_CONNECT:
3246 ret = io_connect(req, nxt, force_nonblock);
3248 case IORING_OP_ASYNC_CANCEL:
3249 ret = io_async_cancel(req, nxt);
3259 if (ctx->flags & IORING_SETUP_IOPOLL) {
3260 if (req->result == -EAGAIN)
3263 io_iopoll_req_issued(req);
3269 static void io_link_work_cb(struct io_wq_work **workptr)
3271 struct io_wq_work *work = *workptr;
3272 struct io_kiocb *link = work->data;
3274 io_queue_linked_timeout(link);
3275 work->func = io_wq_submit_work;
3278 static void io_wq_submit_work(struct io_wq_work **workptr)
3280 struct io_wq_work *work = *workptr;
3281 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3282 struct io_kiocb *nxt = NULL;
3285 if (work->flags & IO_WQ_WORK_CANCEL)
3289 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3290 req->in_async = true;
3292 ret = io_issue_sqe(req, &nxt, false);
3294 * We can get EAGAIN for polled IO even though we're
3295 * forcing a sync submission from here, since we can't
3296 * wait for request slots on the block side.
3304 /* drop submission reference */
3308 req_set_fail_links(req);
3309 io_cqring_add_event(req, ret);
3313 /* if a dependent link is ready, pass it back */
3315 struct io_kiocb *link;
3317 io_prep_async_work(nxt, &link);
3318 *workptr = &nxt->work;
3320 nxt->work.flags |= IO_WQ_WORK_CB;
3321 nxt->work.func = io_link_work_cb;
3322 nxt->work.data = link;
3327 static bool io_req_op_valid(int op)
3329 return op >= IORING_OP_NOP && op < IORING_OP_LAST;
3332 static int io_req_needs_file(struct io_kiocb *req)
3334 switch (req->opcode) {
3336 case IORING_OP_POLL_REMOVE:
3337 case IORING_OP_TIMEOUT:
3338 case IORING_OP_TIMEOUT_REMOVE:
3339 case IORING_OP_ASYNC_CANCEL:
3340 case IORING_OP_LINK_TIMEOUT:
3343 if (io_req_op_valid(req->opcode))
3349 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3352 struct fixed_file_table *table;
3354 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3355 return table->files[index & IORING_FILE_TABLE_MASK];
3358 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3360 struct io_ring_ctx *ctx = req->ctx;
3364 flags = READ_ONCE(req->sqe->flags);
3365 fd = READ_ONCE(req->sqe->fd);
3367 if (flags & IOSQE_IO_DRAIN)
3368 req->flags |= REQ_F_IO_DRAIN;
3370 ret = io_req_needs_file(req);
3374 if (flags & IOSQE_FIXED_FILE) {
3375 if (unlikely(!ctx->file_table ||
3376 (unsigned) fd >= ctx->nr_user_files))
3378 fd = array_index_nospec(fd, ctx->nr_user_files);
3379 req->file = io_file_from_index(ctx, fd);
3382 req->flags |= REQ_F_FIXED_FILE;
3384 if (req->needs_fixed_file)
3386 trace_io_uring_file_get(ctx, fd);
3387 req->file = io_file_get(state, fd);
3388 if (unlikely(!req->file))
3395 static int io_grab_files(struct io_kiocb *req)
3398 struct io_ring_ctx *ctx = req->ctx;
3401 spin_lock_irq(&ctx->inflight_lock);
3403 * We use the f_ops->flush() handler to ensure that we can flush
3404 * out work accessing these files if the fd is closed. Check if
3405 * the fd has changed since we started down this path, and disallow
3406 * this operation if it has.
3408 if (fcheck(req->ring_fd) == req->ring_file) {
3409 list_add(&req->inflight_entry, &ctx->inflight_list);
3410 req->flags |= REQ_F_INFLIGHT;
3411 req->work.files = current->files;
3414 spin_unlock_irq(&ctx->inflight_lock);
3420 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3422 struct io_timeout_data *data = container_of(timer,
3423 struct io_timeout_data, timer);
3424 struct io_kiocb *req = data->req;
3425 struct io_ring_ctx *ctx = req->ctx;
3426 struct io_kiocb *prev = NULL;
3427 unsigned long flags;
3429 spin_lock_irqsave(&ctx->completion_lock, flags);
3432 * We don't expect the list to be empty, that will only happen if we
3433 * race with the completion of the linked work.
3435 if (!list_empty(&req->link_list)) {
3436 prev = list_entry(req->link_list.prev, struct io_kiocb,
3438 if (refcount_inc_not_zero(&prev->refs)) {
3439 list_del_init(&req->link_list);
3440 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3445 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3448 req_set_fail_links(prev);
3449 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3453 io_cqring_add_event(req, -ETIME);
3456 return HRTIMER_NORESTART;
3459 static void io_queue_linked_timeout(struct io_kiocb *req)
3461 struct io_ring_ctx *ctx = req->ctx;
3464 * If the list is now empty, then our linked request finished before
3465 * we got a chance to setup the timer
3467 spin_lock_irq(&ctx->completion_lock);
3468 if (!list_empty(&req->link_list)) {
3469 struct io_timeout_data *data = &req->io->timeout;
3471 data->timer.function = io_link_timeout_fn;
3472 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3475 spin_unlock_irq(&ctx->completion_lock);
3477 /* drop submission reference */
3481 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3483 struct io_kiocb *nxt;
3485 if (!(req->flags & REQ_F_LINK))
3488 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3490 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
3493 req->flags |= REQ_F_LINK_TIMEOUT;
3497 static void __io_queue_sqe(struct io_kiocb *req)
3499 struct io_kiocb *linked_timeout;
3500 struct io_kiocb *nxt = NULL;
3504 linked_timeout = io_prep_linked_timeout(req);
3506 ret = io_issue_sqe(req, &nxt, true);
3509 * We async punt it if the file wasn't marked NOWAIT, or if the file
3510 * doesn't support non-blocking read/write attempts
3512 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3513 (req->flags & REQ_F_MUST_PUNT))) {
3514 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3515 ret = io_grab_files(req);
3521 * Queued up for async execution, worker will release
3522 * submit reference when the iocb is actually submitted.
3524 io_queue_async_work(req);
3529 /* drop submission reference */
3532 if (linked_timeout) {
3534 io_queue_linked_timeout(linked_timeout);
3536 io_put_req(linked_timeout);
3539 /* and drop final reference, if we failed */
3541 io_cqring_add_event(req, ret);
3542 req_set_fail_links(req);
3553 static void io_queue_sqe(struct io_kiocb *req)
3557 if (unlikely(req->ctx->drain_next)) {
3558 req->flags |= REQ_F_IO_DRAIN;
3559 req->ctx->drain_next = false;
3561 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3563 ret = io_req_defer(req);
3565 if (ret != -EIOCBQUEUED) {
3566 io_cqring_add_event(req, ret);
3567 req_set_fail_links(req);
3568 io_double_put_req(req);
3571 __io_queue_sqe(req);
3574 static inline void io_queue_link_head(struct io_kiocb *req)
3576 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3577 io_cqring_add_event(req, -ECANCELED);
3578 io_double_put_req(req);
3583 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3586 static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3587 struct io_kiocb **link)
3589 struct io_ring_ctx *ctx = req->ctx;
3592 /* enforce forwards compatibility on users */
3593 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3598 ret = io_req_set_file(state, req);
3599 if (unlikely(ret)) {
3601 io_cqring_add_event(req, ret);
3602 io_double_put_req(req);
3607 * If we already have a head request, queue this one for async
3608 * submittal once the head completes. If we don't have a head but
3609 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3610 * submitted sync once the chain is complete. If none of those
3611 * conditions are true (normal request), then just queue it.
3614 struct io_kiocb *prev = *link;
3616 if (req->sqe->flags & IOSQE_IO_DRAIN)
3617 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3619 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3620 req->flags |= REQ_F_HARDLINK;
3622 if (io_alloc_async_ctx(req)) {
3627 ret = io_req_defer_prep(req);
3629 /* fail even hard links since we don't submit */
3630 prev->flags |= REQ_F_FAIL_LINK;
3633 trace_io_uring_link(ctx, req, prev);
3634 list_add_tail(&req->link_list, &prev->link_list);
3635 } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
3636 req->flags |= REQ_F_LINK;
3637 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3638 req->flags |= REQ_F_HARDLINK;
3640 INIT_LIST_HEAD(&req->link_list);
3650 * Batched submission is done, ensure local IO is flushed out.
3652 static void io_submit_state_end(struct io_submit_state *state)
3654 blk_finish_plug(&state->plug);
3656 if (state->free_reqs)
3657 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3658 &state->reqs[state->cur_req]);
3662 * Start submission side cache.
3664 static void io_submit_state_start(struct io_submit_state *state,
3665 unsigned int max_ios)
3667 blk_start_plug(&state->plug);
3668 state->free_reqs = 0;
3670 state->ios_left = max_ios;
3673 static void io_commit_sqring(struct io_ring_ctx *ctx)
3675 struct io_rings *rings = ctx->rings;
3677 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3679 * Ensure any loads from the SQEs are done at this point,
3680 * since once we write the new head, the application could
3681 * write new data to them.
3683 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3688 * Fetch an sqe, if one is available. Note that req->sqe will point to memory
3689 * that is mapped by userspace. This means that care needs to be taken to
3690 * ensure that reads are stable, as we cannot rely on userspace always
3691 * being a good citizen. If members of the sqe are validated and then later
3692 * used, it's important that those reads are done through READ_ONCE() to
3693 * prevent a re-load down the line.
3695 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3697 struct io_rings *rings = ctx->rings;
3698 u32 *sq_array = ctx->sq_array;
3702 * The cached sq head (or cq tail) serves two purposes:
3704 * 1) allows us to batch the cost of updating the user visible
3706 * 2) allows the kernel side to track the head on its own, even
3707 * though the application is the one updating it.
3709 head = ctx->cached_sq_head;
3710 /* make sure SQ entry isn't read before tail */
3711 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3714 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3715 if (likely(head < ctx->sq_entries)) {
3717 * All io need record the previous position, if LINK vs DARIN,
3718 * it can be used to mark the position of the first IO in the
3721 req->sequence = ctx->cached_sq_head;
3722 req->sqe = &ctx->sq_sqes[head];
3723 req->opcode = READ_ONCE(req->sqe->opcode);
3724 req->user_data = READ_ONCE(req->sqe->user_data);
3725 ctx->cached_sq_head++;
3729 /* drop invalid entries */
3730 ctx->cached_sq_head++;
3731 ctx->cached_sq_dropped++;
3732 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3736 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3737 struct file *ring_file, int ring_fd,
3738 struct mm_struct **mm, bool async)
3740 struct io_submit_state state, *statep = NULL;
3741 struct io_kiocb *link = NULL;
3742 int i, submitted = 0;
3743 bool mm_fault = false;
3745 /* if we have a backlog and couldn't flush it all, return BUSY */
3746 if (!list_empty(&ctx->cq_overflow_list) &&
3747 !io_cqring_overflow_flush(ctx, false))
3750 if (nr > IO_PLUG_THRESHOLD) {
3751 io_submit_state_start(&state, nr);
3755 for (i = 0; i < nr; i++) {
3756 struct io_kiocb *req;
3757 unsigned int sqe_flags;
3759 req = io_get_req(ctx, statep);
3760 if (unlikely(!req)) {
3762 submitted = -EAGAIN;
3765 if (!io_get_sqring(ctx, req)) {
3770 if (io_req_needs_user(req) && !*mm) {
3771 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3773 use_mm(ctx->sqo_mm);
3779 sqe_flags = req->sqe->flags;
3781 req->ring_file = ring_file;
3782 req->ring_fd = ring_fd;
3783 req->has_user = *mm != NULL;
3784 req->in_async = async;
3785 req->needs_fixed_file = async;
3786 trace_io_uring_submit_sqe(ctx, req->user_data, true, async);
3787 if (!io_submit_sqe(req, statep, &link))
3790 * If previous wasn't linked and we have a linked command,
3791 * that's the end of the chain. Submit the previous link.
3793 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) && link) {
3794 io_queue_link_head(link);
3800 io_queue_link_head(link);
3802 io_submit_state_end(&state);
3804 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3805 io_commit_sqring(ctx);
3810 static int io_sq_thread(void *data)
3812 struct io_ring_ctx *ctx = data;
3813 struct mm_struct *cur_mm = NULL;
3814 const struct cred *old_cred;
3815 mm_segment_t old_fs;
3818 unsigned long timeout;
3821 complete(&ctx->completions[1]);
3825 old_cred = override_creds(ctx->creds);
3827 ret = timeout = inflight = 0;
3828 while (!kthread_should_park()) {
3829 unsigned int to_submit;
3832 unsigned nr_events = 0;
3834 if (ctx->flags & IORING_SETUP_IOPOLL) {
3836 * inflight is the count of the maximum possible
3837 * entries we submitted, but it can be smaller
3838 * if we dropped some of them. If we don't have
3839 * poll entries available, then we know that we
3840 * have nothing left to poll for. Reset the
3841 * inflight count to zero in that case.
3843 mutex_lock(&ctx->uring_lock);
3844 if (!list_empty(&ctx->poll_list))
3845 __io_iopoll_check(ctx, &nr_events, 0);
3848 mutex_unlock(&ctx->uring_lock);
3851 * Normal IO, just pretend everything completed.
3852 * We don't have to poll completions for that.
3854 nr_events = inflight;
3857 inflight -= nr_events;
3859 timeout = jiffies + ctx->sq_thread_idle;
3862 to_submit = io_sqring_entries(ctx);
3865 * If submit got -EBUSY, flag us as needing the application
3866 * to enter the kernel to reap and flush events.
3868 if (!to_submit || ret == -EBUSY) {
3870 * We're polling. If we're within the defined idle
3871 * period, then let us spin without work before going
3872 * to sleep. The exception is if we got EBUSY doing
3873 * more IO, we should wait for the application to
3874 * reap events and wake us up.
3877 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3883 * Drop cur_mm before scheduling, we can't hold it for
3884 * long periods (or over schedule()). Do this before
3885 * adding ourselves to the waitqueue, as the unuse/drop
3894 prepare_to_wait(&ctx->sqo_wait, &wait,
3895 TASK_INTERRUPTIBLE);
3897 /* Tell userspace we may need a wakeup call */
3898 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3899 /* make sure to read SQ tail after writing flags */
3902 to_submit = io_sqring_entries(ctx);
3903 if (!to_submit || ret == -EBUSY) {
3904 if (kthread_should_park()) {
3905 finish_wait(&ctx->sqo_wait, &wait);
3908 if (signal_pending(current))
3909 flush_signals(current);
3911 finish_wait(&ctx->sqo_wait, &wait);
3913 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3916 finish_wait(&ctx->sqo_wait, &wait);
3918 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3921 to_submit = min(to_submit, ctx->sq_entries);
3922 mutex_lock(&ctx->uring_lock);
3923 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3924 mutex_unlock(&ctx->uring_lock);
3934 revert_creds(old_cred);
3941 struct io_wait_queue {
3942 struct wait_queue_entry wq;
3943 struct io_ring_ctx *ctx;
3945 unsigned nr_timeouts;
3948 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3950 struct io_ring_ctx *ctx = iowq->ctx;
3953 * Wake up if we have enough events, or if a timeout occurred since we
3954 * started waiting. For timeouts, we always want to return to userspace,
3955 * regardless of event count.
3957 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3958 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3961 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3962 int wake_flags, void *key)
3964 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3967 /* use noflush == true, as we can't safely rely on locking context */
3968 if (!io_should_wake(iowq, true))
3971 return autoremove_wake_function(curr, mode, wake_flags, key);
3975 * Wait until events become available, if we don't already have some. The
3976 * application must reap them itself, as they reside on the shared cq ring.
3978 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3979 const sigset_t __user *sig, size_t sigsz)
3981 struct io_wait_queue iowq = {
3984 .func = io_wake_function,
3985 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3988 .to_wait = min_events,
3990 struct io_rings *rings = ctx->rings;
3993 if (io_cqring_events(ctx, false) >= min_events)
3997 #ifdef CONFIG_COMPAT
3998 if (in_compat_syscall())
3999 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
4003 ret = set_user_sigmask(sig, sigsz);
4009 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
4010 trace_io_uring_cqring_wait(ctx, min_events);
4012 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
4013 TASK_INTERRUPTIBLE);
4014 if (io_should_wake(&iowq, false))
4017 if (signal_pending(current)) {
4022 finish_wait(&ctx->wait, &iowq.wq);
4024 restore_saved_sigmask_unless(ret == -EINTR);
4026 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
4029 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
4031 #if defined(CONFIG_UNIX)
4032 if (ctx->ring_sock) {
4033 struct sock *sock = ctx->ring_sock->sk;
4034 struct sk_buff *skb;
4036 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
4042 for (i = 0; i < ctx->nr_user_files; i++) {
4045 file = io_file_from_index(ctx, i);
4052 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
4054 unsigned nr_tables, i;
4056 if (!ctx->file_table)
4059 __io_sqe_files_unregister(ctx);
4060 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
4061 for (i = 0; i < nr_tables; i++)
4062 kfree(ctx->file_table[i].files);
4063 kfree(ctx->file_table);
4064 ctx->file_table = NULL;
4065 ctx->nr_user_files = 0;
4069 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
4071 if (ctx->sqo_thread) {
4072 wait_for_completion(&ctx->completions[1]);
4074 * The park is a bit of a work-around, without it we get
4075 * warning spews on shutdown with SQPOLL set and affinity
4076 * set to a single CPU.
4078 kthread_park(ctx->sqo_thread);
4079 kthread_stop(ctx->sqo_thread);
4080 ctx->sqo_thread = NULL;
4084 static void io_finish_async(struct io_ring_ctx *ctx)
4086 io_sq_thread_stop(ctx);
4089 io_wq_destroy(ctx->io_wq);
4094 #if defined(CONFIG_UNIX)
4095 static void io_destruct_skb(struct sk_buff *skb)
4097 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
4100 io_wq_flush(ctx->io_wq);
4102 unix_destruct_scm(skb);
4106 * Ensure the UNIX gc is aware of our file set, so we are certain that
4107 * the io_uring can be safely unregistered on process exit, even if we have
4108 * loops in the file referencing.
4110 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
4112 struct sock *sk = ctx->ring_sock->sk;
4113 struct scm_fp_list *fpl;
4114 struct sk_buff *skb;
4117 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
4118 unsigned long inflight = ctx->user->unix_inflight + nr;
4120 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
4124 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
4128 skb = alloc_skb(0, GFP_KERNEL);
4137 fpl->user = get_uid(ctx->user);
4138 for (i = 0; i < nr; i++) {
4139 struct file *file = io_file_from_index(ctx, i + offset);
4143 fpl->fp[nr_files] = get_file(file);
4144 unix_inflight(fpl->user, fpl->fp[nr_files]);
4149 fpl->max = SCM_MAX_FD;
4150 fpl->count = nr_files;
4151 UNIXCB(skb).fp = fpl;
4152 skb->destructor = io_destruct_skb;
4153 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
4154 skb_queue_head(&sk->sk_receive_queue, skb);
4156 for (i = 0; i < nr_files; i++)
4167 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
4168 * causes regular reference counting to break down. We rely on the UNIX
4169 * garbage collection to take care of this problem for us.
4171 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4173 unsigned left, total;
4177 left = ctx->nr_user_files;
4179 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
4181 ret = __io_sqe_files_scm(ctx, this_files, total);
4185 total += this_files;
4191 while (total < ctx->nr_user_files) {
4192 struct file *file = io_file_from_index(ctx, total);
4202 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4208 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
4213 for (i = 0; i < nr_tables; i++) {
4214 struct fixed_file_table *table = &ctx->file_table[i];
4215 unsigned this_files;
4217 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
4218 table->files = kcalloc(this_files, sizeof(struct file *),
4222 nr_files -= this_files;
4228 for (i = 0; i < nr_tables; i++) {
4229 struct fixed_file_table *table = &ctx->file_table[i];
4230 kfree(table->files);
4235 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
4238 __s32 __user *fds = (__s32 __user *) arg;
4243 if (ctx->file_table)
4247 if (nr_args > IORING_MAX_FIXED_FILES)
4250 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
4251 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
4253 if (!ctx->file_table)
4256 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
4257 kfree(ctx->file_table);
4258 ctx->file_table = NULL;
4262 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
4263 struct fixed_file_table *table;
4267 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4269 /* allow sparse sets */
4275 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4276 index = i & IORING_FILE_TABLE_MASK;
4277 table->files[index] = fget(fd);
4280 if (!table->files[index])
4283 * Don't allow io_uring instances to be registered. If UNIX
4284 * isn't enabled, then this causes a reference cycle and this
4285 * instance can never get freed. If UNIX is enabled we'll
4286 * handle it just fine, but there's still no point in allowing
4287 * a ring fd as it doesn't support regular read/write anyway.
4289 if (table->files[index]->f_op == &io_uring_fops) {
4290 fput(table->files[index]);
4297 for (i = 0; i < ctx->nr_user_files; i++) {
4300 file = io_file_from_index(ctx, i);
4304 for (i = 0; i < nr_tables; i++)
4305 kfree(ctx->file_table[i].files);
4307 kfree(ctx->file_table);
4308 ctx->file_table = NULL;
4309 ctx->nr_user_files = 0;
4313 ret = io_sqe_files_scm(ctx);
4315 io_sqe_files_unregister(ctx);
4320 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4322 #if defined(CONFIG_UNIX)
4323 struct file *file = io_file_from_index(ctx, index);
4324 struct sock *sock = ctx->ring_sock->sk;
4325 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4326 struct sk_buff *skb;
4329 __skb_queue_head_init(&list);
4332 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4333 * remove this entry and rearrange the file array.
4335 skb = skb_dequeue(head);
4337 struct scm_fp_list *fp;
4339 fp = UNIXCB(skb).fp;
4340 for (i = 0; i < fp->count; i++) {
4343 if (fp->fp[i] != file)
4346 unix_notinflight(fp->user, fp->fp[i]);
4347 left = fp->count - 1 - i;
4349 memmove(&fp->fp[i], &fp->fp[i + 1],
4350 left * sizeof(struct file *));
4357 __skb_queue_tail(&list, skb);
4367 __skb_queue_tail(&list, skb);
4369 skb = skb_dequeue(head);
4372 if (skb_peek(&list)) {
4373 spin_lock_irq(&head->lock);
4374 while ((skb = __skb_dequeue(&list)) != NULL)
4375 __skb_queue_tail(head, skb);
4376 spin_unlock_irq(&head->lock);
4379 fput(io_file_from_index(ctx, index));
4383 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4386 #if defined(CONFIG_UNIX)
4387 struct sock *sock = ctx->ring_sock->sk;
4388 struct sk_buff_head *head = &sock->sk_receive_queue;
4389 struct sk_buff *skb;
4392 * See if we can merge this file into an existing skb SCM_RIGHTS
4393 * file set. If there's no room, fall back to allocating a new skb
4394 * and filling it in.
4396 spin_lock_irq(&head->lock);
4397 skb = skb_peek(head);
4399 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4401 if (fpl->count < SCM_MAX_FD) {
4402 __skb_unlink(skb, head);
4403 spin_unlock_irq(&head->lock);
4404 fpl->fp[fpl->count] = get_file(file);
4405 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4407 spin_lock_irq(&head->lock);
4408 __skb_queue_head(head, skb);
4413 spin_unlock_irq(&head->lock);
4420 return __io_sqe_files_scm(ctx, 1, index);
4426 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4429 struct io_uring_files_update up;
4434 if (!ctx->file_table)
4438 if (copy_from_user(&up, arg, sizeof(up)))
4440 if (check_add_overflow(up.offset, nr_args, &done))
4442 if (done > ctx->nr_user_files)
4446 fds = (__s32 __user *) up.fds;
4448 struct fixed_file_table *table;
4452 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4456 i = array_index_nospec(up.offset, ctx->nr_user_files);
4457 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4458 index = i & IORING_FILE_TABLE_MASK;
4459 if (table->files[index]) {
4460 io_sqe_file_unregister(ctx, i);
4461 table->files[index] = NULL;
4472 * Don't allow io_uring instances to be registered. If
4473 * UNIX isn't enabled, then this causes a reference
4474 * cycle and this instance can never get freed. If UNIX
4475 * is enabled we'll handle it just fine, but there's
4476 * still no point in allowing a ring fd as it doesn't
4477 * support regular read/write anyway.
4479 if (file->f_op == &io_uring_fops) {
4484 table->files[index] = file;
4485 err = io_sqe_file_register(ctx, file, i);
4494 return done ? done : err;
4497 static void io_put_work(struct io_wq_work *work)
4499 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4504 static void io_get_work(struct io_wq_work *work)
4506 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4508 refcount_inc(&req->refs);
4511 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4512 struct io_uring_params *p)
4514 struct io_wq_data data;
4515 unsigned concurrency;
4518 init_waitqueue_head(&ctx->sqo_wait);
4519 mmgrab(current->mm);
4520 ctx->sqo_mm = current->mm;
4522 if (ctx->flags & IORING_SETUP_SQPOLL) {
4524 if (!capable(CAP_SYS_ADMIN))
4527 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4528 if (!ctx->sq_thread_idle)
4529 ctx->sq_thread_idle = HZ;
4531 if (p->flags & IORING_SETUP_SQ_AFF) {
4532 int cpu = p->sq_thread_cpu;
4535 if (cpu >= nr_cpu_ids)
4537 if (!cpu_online(cpu))
4540 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4544 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4547 if (IS_ERR(ctx->sqo_thread)) {
4548 ret = PTR_ERR(ctx->sqo_thread);
4549 ctx->sqo_thread = NULL;
4552 wake_up_process(ctx->sqo_thread);
4553 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4554 /* Can't have SQ_AFF without SQPOLL */
4559 data.mm = ctx->sqo_mm;
4560 data.user = ctx->user;
4561 data.creds = ctx->creds;
4562 data.get_work = io_get_work;
4563 data.put_work = io_put_work;
4565 /* Do QD, or 4 * CPUS, whatever is smallest */
4566 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4567 ctx->io_wq = io_wq_create(concurrency, &data);
4568 if (IS_ERR(ctx->io_wq)) {
4569 ret = PTR_ERR(ctx->io_wq);
4576 io_finish_async(ctx);
4577 mmdrop(ctx->sqo_mm);
4582 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4584 atomic_long_sub(nr_pages, &user->locked_vm);
4587 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4589 unsigned long page_limit, cur_pages, new_pages;
4591 /* Don't allow more pages than we can safely lock */
4592 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4595 cur_pages = atomic_long_read(&user->locked_vm);
4596 new_pages = cur_pages + nr_pages;
4597 if (new_pages > page_limit)
4599 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4600 new_pages) != cur_pages);
4605 static void io_mem_free(void *ptr)
4612 page = virt_to_head_page(ptr);
4613 if (put_page_testzero(page))
4614 free_compound_page(page);
4617 static void *io_mem_alloc(size_t size)
4619 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4622 return (void *) __get_free_pages(gfp_flags, get_order(size));
4625 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4628 struct io_rings *rings;
4629 size_t off, sq_array_size;
4631 off = struct_size(rings, cqes, cq_entries);
4632 if (off == SIZE_MAX)
4636 off = ALIGN(off, SMP_CACHE_BYTES);
4641 sq_array_size = array_size(sizeof(u32), sq_entries);
4642 if (sq_array_size == SIZE_MAX)
4645 if (check_add_overflow(off, sq_array_size, &off))
4654 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4658 pages = (size_t)1 << get_order(
4659 rings_size(sq_entries, cq_entries, NULL));
4660 pages += (size_t)1 << get_order(
4661 array_size(sizeof(struct io_uring_sqe), sq_entries));
4666 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4670 if (!ctx->user_bufs)
4673 for (i = 0; i < ctx->nr_user_bufs; i++) {
4674 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4676 for (j = 0; j < imu->nr_bvecs; j++)
4677 put_user_page(imu->bvec[j].bv_page);
4679 if (ctx->account_mem)
4680 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4685 kfree(ctx->user_bufs);
4686 ctx->user_bufs = NULL;
4687 ctx->nr_user_bufs = 0;
4691 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4692 void __user *arg, unsigned index)
4694 struct iovec __user *src;
4696 #ifdef CONFIG_COMPAT
4698 struct compat_iovec __user *ciovs;
4699 struct compat_iovec ciov;
4701 ciovs = (struct compat_iovec __user *) arg;
4702 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4705 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4706 dst->iov_len = ciov.iov_len;
4710 src = (struct iovec __user *) arg;
4711 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4716 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4719 struct vm_area_struct **vmas = NULL;
4720 struct page **pages = NULL;
4721 int i, j, got_pages = 0;
4726 if (!nr_args || nr_args > UIO_MAXIOV)
4729 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4731 if (!ctx->user_bufs)
4734 for (i = 0; i < nr_args; i++) {
4735 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4736 unsigned long off, start, end, ubuf;
4741 ret = io_copy_iov(ctx, &iov, arg, i);
4746 * Don't impose further limits on the size and buffer
4747 * constraints here, we'll -EINVAL later when IO is
4748 * submitted if they are wrong.
4751 if (!iov.iov_base || !iov.iov_len)
4754 /* arbitrary limit, but we need something */
4755 if (iov.iov_len > SZ_1G)
4758 ubuf = (unsigned long) iov.iov_base;
4759 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4760 start = ubuf >> PAGE_SHIFT;
4761 nr_pages = end - start;
4763 if (ctx->account_mem) {
4764 ret = io_account_mem(ctx->user, nr_pages);
4770 if (!pages || nr_pages > got_pages) {
4773 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4775 vmas = kvmalloc_array(nr_pages,
4776 sizeof(struct vm_area_struct *),
4778 if (!pages || !vmas) {
4780 if (ctx->account_mem)
4781 io_unaccount_mem(ctx->user, nr_pages);
4784 got_pages = nr_pages;
4787 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4791 if (ctx->account_mem)
4792 io_unaccount_mem(ctx->user, nr_pages);
4797 down_read(¤t->mm->mmap_sem);
4798 pret = get_user_pages(ubuf, nr_pages,
4799 FOLL_WRITE | FOLL_LONGTERM,
4801 if (pret == nr_pages) {
4802 /* don't support file backed memory */
4803 for (j = 0; j < nr_pages; j++) {
4804 struct vm_area_struct *vma = vmas[j];
4807 !is_file_hugepages(vma->vm_file)) {
4813 ret = pret < 0 ? pret : -EFAULT;
4815 up_read(¤t->mm->mmap_sem);
4818 * if we did partial map, or found file backed vmas,
4819 * release any pages we did get
4822 put_user_pages(pages, pret);
4823 if (ctx->account_mem)
4824 io_unaccount_mem(ctx->user, nr_pages);
4829 off = ubuf & ~PAGE_MASK;
4831 for (j = 0; j < nr_pages; j++) {
4834 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4835 imu->bvec[j].bv_page = pages[j];
4836 imu->bvec[j].bv_len = vec_len;
4837 imu->bvec[j].bv_offset = off;
4841 /* store original address for later verification */
4843 imu->len = iov.iov_len;
4844 imu->nr_bvecs = nr_pages;
4846 ctx->nr_user_bufs++;
4854 io_sqe_buffer_unregister(ctx);
4858 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4860 __s32 __user *fds = arg;
4866 if (copy_from_user(&fd, fds, sizeof(*fds)))
4869 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4870 if (IS_ERR(ctx->cq_ev_fd)) {
4871 int ret = PTR_ERR(ctx->cq_ev_fd);
4872 ctx->cq_ev_fd = NULL;
4879 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4881 if (ctx->cq_ev_fd) {
4882 eventfd_ctx_put(ctx->cq_ev_fd);
4883 ctx->cq_ev_fd = NULL;
4890 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4892 io_finish_async(ctx);
4894 mmdrop(ctx->sqo_mm);
4896 io_iopoll_reap_events(ctx);
4897 io_sqe_buffer_unregister(ctx);
4898 io_sqe_files_unregister(ctx);
4899 io_eventfd_unregister(ctx);
4901 #if defined(CONFIG_UNIX)
4902 if (ctx->ring_sock) {
4903 ctx->ring_sock->file = NULL; /* so that iput() is called */
4904 sock_release(ctx->ring_sock);
4908 io_mem_free(ctx->rings);
4909 io_mem_free(ctx->sq_sqes);
4911 percpu_ref_exit(&ctx->refs);
4912 if (ctx->account_mem)
4913 io_unaccount_mem(ctx->user,
4914 ring_pages(ctx->sq_entries, ctx->cq_entries));
4915 free_uid(ctx->user);
4916 put_cred(ctx->creds);
4917 kfree(ctx->completions);
4918 kfree(ctx->cancel_hash);
4919 kmem_cache_free(req_cachep, ctx->fallback_req);
4923 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4925 struct io_ring_ctx *ctx = file->private_data;
4928 poll_wait(file, &ctx->cq_wait, wait);
4930 * synchronizes with barrier from wq_has_sleeper call in
4934 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4935 ctx->rings->sq_ring_entries)
4936 mask |= EPOLLOUT | EPOLLWRNORM;
4937 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4938 mask |= EPOLLIN | EPOLLRDNORM;
4943 static int io_uring_fasync(int fd, struct file *file, int on)
4945 struct io_ring_ctx *ctx = file->private_data;
4947 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4950 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4952 mutex_lock(&ctx->uring_lock);
4953 percpu_ref_kill(&ctx->refs);
4954 mutex_unlock(&ctx->uring_lock);
4956 io_kill_timeouts(ctx);
4957 io_poll_remove_all(ctx);
4960 io_wq_cancel_all(ctx->io_wq);
4962 io_iopoll_reap_events(ctx);
4963 /* if we failed setting up the ctx, we might not have any rings */
4965 io_cqring_overflow_flush(ctx, true);
4966 wait_for_completion(&ctx->completions[0]);
4967 io_ring_ctx_free(ctx);
4970 static int io_uring_release(struct inode *inode, struct file *file)
4972 struct io_ring_ctx *ctx = file->private_data;
4974 file->private_data = NULL;
4975 io_ring_ctx_wait_and_kill(ctx);
4979 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4980 struct files_struct *files)
4982 struct io_kiocb *req;
4985 while (!list_empty_careful(&ctx->inflight_list)) {
4986 struct io_kiocb *cancel_req = NULL;
4988 spin_lock_irq(&ctx->inflight_lock);
4989 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4990 if (req->work.files != files)
4992 /* req is being completed, ignore */
4993 if (!refcount_inc_not_zero(&req->refs))
4999 prepare_to_wait(&ctx->inflight_wait, &wait,
5000 TASK_UNINTERRUPTIBLE);
5001 spin_unlock_irq(&ctx->inflight_lock);
5003 /* We need to keep going until we don't find a matching req */
5007 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
5008 io_put_req(cancel_req);
5011 finish_wait(&ctx->inflight_wait, &wait);
5014 static int io_uring_flush(struct file *file, void *data)
5016 struct io_ring_ctx *ctx = file->private_data;
5018 io_uring_cancel_files(ctx, data);
5019 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
5020 io_cqring_overflow_flush(ctx, true);
5021 io_wq_cancel_all(ctx->io_wq);
5026 static void *io_uring_validate_mmap_request(struct file *file,
5027 loff_t pgoff, size_t sz)
5029 struct io_ring_ctx *ctx = file->private_data;
5030 loff_t offset = pgoff << PAGE_SHIFT;
5035 case IORING_OFF_SQ_RING:
5036 case IORING_OFF_CQ_RING:
5039 case IORING_OFF_SQES:
5043 return ERR_PTR(-EINVAL);
5046 page = virt_to_head_page(ptr);
5047 if (sz > page_size(page))
5048 return ERR_PTR(-EINVAL);
5055 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5057 size_t sz = vma->vm_end - vma->vm_start;
5061 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
5063 return PTR_ERR(ptr);
5065 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
5066 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
5069 #else /* !CONFIG_MMU */
5071 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
5073 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
5076 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
5078 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
5081 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
5082 unsigned long addr, unsigned long len,
5083 unsigned long pgoff, unsigned long flags)
5087 ptr = io_uring_validate_mmap_request(file, pgoff, len);
5089 return PTR_ERR(ptr);
5091 return (unsigned long) ptr;
5094 #endif /* !CONFIG_MMU */
5096 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
5097 u32, min_complete, u32, flags, const sigset_t __user *, sig,
5100 struct io_ring_ctx *ctx;
5105 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
5113 if (f.file->f_op != &io_uring_fops)
5117 ctx = f.file->private_data;
5118 if (!percpu_ref_tryget(&ctx->refs))
5122 * For SQ polling, the thread will do all submissions and completions.
5123 * Just return the requested submit count, and wake the thread if
5127 if (ctx->flags & IORING_SETUP_SQPOLL) {
5128 if (!list_empty_careful(&ctx->cq_overflow_list))
5129 io_cqring_overflow_flush(ctx, false);
5130 if (flags & IORING_ENTER_SQ_WAKEUP)
5131 wake_up(&ctx->sqo_wait);
5132 submitted = to_submit;
5133 } else if (to_submit) {
5134 struct mm_struct *cur_mm;
5136 to_submit = min(to_submit, ctx->sq_entries);
5137 mutex_lock(&ctx->uring_lock);
5138 /* already have mm, so io_submit_sqes() won't try to grab it */
5139 cur_mm = ctx->sqo_mm;
5140 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
5142 mutex_unlock(&ctx->uring_lock);
5144 if (submitted != to_submit)
5147 if (flags & IORING_ENTER_GETEVENTS) {
5148 unsigned nr_events = 0;
5150 min_complete = min(min_complete, ctx->cq_entries);
5152 if (ctx->flags & IORING_SETUP_IOPOLL) {
5153 ret = io_iopoll_check(ctx, &nr_events, min_complete);
5155 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
5160 percpu_ref_put(&ctx->refs);
5163 return submitted ? submitted : ret;
5166 static const struct file_operations io_uring_fops = {
5167 .release = io_uring_release,
5168 .flush = io_uring_flush,
5169 .mmap = io_uring_mmap,
5171 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
5172 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
5174 .poll = io_uring_poll,
5175 .fasync = io_uring_fasync,
5178 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
5179 struct io_uring_params *p)
5181 struct io_rings *rings;
5182 size_t size, sq_array_offset;
5184 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
5185 if (size == SIZE_MAX)
5188 rings = io_mem_alloc(size);
5193 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
5194 rings->sq_ring_mask = p->sq_entries - 1;
5195 rings->cq_ring_mask = p->cq_entries - 1;
5196 rings->sq_ring_entries = p->sq_entries;
5197 rings->cq_ring_entries = p->cq_entries;
5198 ctx->sq_mask = rings->sq_ring_mask;
5199 ctx->cq_mask = rings->cq_ring_mask;
5200 ctx->sq_entries = rings->sq_ring_entries;
5201 ctx->cq_entries = rings->cq_ring_entries;
5203 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
5204 if (size == SIZE_MAX) {
5205 io_mem_free(ctx->rings);
5210 ctx->sq_sqes = io_mem_alloc(size);
5211 if (!ctx->sq_sqes) {
5212 io_mem_free(ctx->rings);
5221 * Allocate an anonymous fd, this is what constitutes the application
5222 * visible backing of an io_uring instance. The application mmaps this
5223 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
5224 * we have to tie this fd to a socket for file garbage collection purposes.
5226 static int io_uring_get_fd(struct io_ring_ctx *ctx)
5231 #if defined(CONFIG_UNIX)
5232 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
5238 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
5242 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
5243 O_RDWR | O_CLOEXEC);
5246 ret = PTR_ERR(file);
5250 #if defined(CONFIG_UNIX)
5251 ctx->ring_sock->file = file;
5252 ctx->ring_sock->sk->sk_user_data = ctx;
5254 fd_install(ret, file);
5257 #if defined(CONFIG_UNIX)
5258 sock_release(ctx->ring_sock);
5259 ctx->ring_sock = NULL;
5264 static int io_uring_create(unsigned entries, struct io_uring_params *p)
5266 struct user_struct *user = NULL;
5267 struct io_ring_ctx *ctx;
5271 if (!entries || entries > IORING_MAX_ENTRIES)
5275 * Use twice as many entries for the CQ ring. It's possible for the
5276 * application to drive a higher depth than the size of the SQ ring,
5277 * since the sqes are only used at submission time. This allows for
5278 * some flexibility in overcommitting a bit. If the application has
5279 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5280 * of CQ ring entries manually.
5282 p->sq_entries = roundup_pow_of_two(entries);
5283 if (p->flags & IORING_SETUP_CQSIZE) {
5285 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5286 * to a power-of-two, if it isn't already. We do NOT impose
5287 * any cq vs sq ring sizing.
5289 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5291 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5293 p->cq_entries = 2 * p->sq_entries;
5296 user = get_uid(current_user());
5297 account_mem = !capable(CAP_IPC_LOCK);
5300 ret = io_account_mem(user,
5301 ring_pages(p->sq_entries, p->cq_entries));
5308 ctx = io_ring_ctx_alloc(p);
5311 io_unaccount_mem(user, ring_pages(p->sq_entries,
5316 ctx->compat = in_compat_syscall();
5317 ctx->account_mem = account_mem;
5319 ctx->creds = get_current_cred();
5321 ret = io_allocate_scq_urings(ctx, p);
5325 ret = io_sq_offload_start(ctx, p);
5329 memset(&p->sq_off, 0, sizeof(p->sq_off));
5330 p->sq_off.head = offsetof(struct io_rings, sq.head);
5331 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5332 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5333 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5334 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5335 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5336 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5338 memset(&p->cq_off, 0, sizeof(p->cq_off));
5339 p->cq_off.head = offsetof(struct io_rings, cq.head);
5340 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5341 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5342 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5343 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5344 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5347 * Install ring fd as the very last thing, so we don't risk someone
5348 * having closed it before we finish setup
5350 ret = io_uring_get_fd(ctx);
5354 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5355 IORING_FEAT_SUBMIT_STABLE;
5356 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5359 io_ring_ctx_wait_and_kill(ctx);
5364 * Sets up an aio uring context, and returns the fd. Applications asks for a
5365 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5366 * params structure passed in.
5368 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5370 struct io_uring_params p;
5374 if (copy_from_user(&p, params, sizeof(p)))
5376 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5381 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5382 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5385 ret = io_uring_create(entries, &p);
5389 if (copy_to_user(params, &p, sizeof(p)))
5395 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5396 struct io_uring_params __user *, params)
5398 return io_uring_setup(entries, params);
5401 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5402 void __user *arg, unsigned nr_args)
5403 __releases(ctx->uring_lock)
5404 __acquires(ctx->uring_lock)
5409 * We're inside the ring mutex, if the ref is already dying, then
5410 * someone else killed the ctx or is already going through
5411 * io_uring_register().
5413 if (percpu_ref_is_dying(&ctx->refs))
5416 percpu_ref_kill(&ctx->refs);
5419 * Drop uring mutex before waiting for references to exit. If another
5420 * thread is currently inside io_uring_enter() it might need to grab
5421 * the uring_lock to make progress. If we hold it here across the drain
5422 * wait, then we can deadlock. It's safe to drop the mutex here, since
5423 * no new references will come in after we've killed the percpu ref.
5425 mutex_unlock(&ctx->uring_lock);
5426 wait_for_completion(&ctx->completions[0]);
5427 mutex_lock(&ctx->uring_lock);
5430 case IORING_REGISTER_BUFFERS:
5431 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5433 case IORING_UNREGISTER_BUFFERS:
5437 ret = io_sqe_buffer_unregister(ctx);
5439 case IORING_REGISTER_FILES:
5440 ret = io_sqe_files_register(ctx, arg, nr_args);
5442 case IORING_UNREGISTER_FILES:
5446 ret = io_sqe_files_unregister(ctx);
5448 case IORING_REGISTER_FILES_UPDATE:
5449 ret = io_sqe_files_update(ctx, arg, nr_args);
5451 case IORING_REGISTER_EVENTFD:
5455 ret = io_eventfd_register(ctx, arg);
5457 case IORING_UNREGISTER_EVENTFD:
5461 ret = io_eventfd_unregister(ctx);
5468 /* bring the ctx back to life */
5469 reinit_completion(&ctx->completions[0]);
5470 percpu_ref_reinit(&ctx->refs);
5474 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5475 void __user *, arg, unsigned int, nr_args)
5477 struct io_ring_ctx *ctx;
5486 if (f.file->f_op != &io_uring_fops)
5489 ctx = f.file->private_data;
5491 mutex_lock(&ctx->uring_lock);
5492 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5493 mutex_unlock(&ctx->uring_lock);
5494 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5495 ctx->cq_ev_fd != NULL, ret);
5501 static int __init io_uring_init(void)
5503 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5506 __initcall(io_uring_init);