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 thatn 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 {
292 struct wait_queue_head *head;
296 struct wait_queue_entry *wait;
299 struct io_timeout_data {
300 struct io_kiocb *req;
301 struct hrtimer timer;
302 struct timespec64 ts;
303 enum hrtimer_mode mode;
307 struct io_async_connect {
308 struct sockaddr_storage address;
311 struct io_async_msghdr {
312 struct iovec fast_iov[UIO_FASTIOV];
314 struct sockaddr __user *uaddr;
319 struct iovec fast_iov[UIO_FASTIOV];
325 struct io_async_ctx {
326 struct io_uring_sqe sqe;
328 struct io_async_rw rw;
329 struct io_async_msghdr msg;
330 struct io_async_connect connect;
331 struct io_timeout_data timeout;
336 * NOTE! Each of the iocb union members has the file pointer
337 * as the first entry in their struct definition. So you can
338 * access the file pointer through any of the sub-structs,
339 * or directly as just 'ki_filp' in this struct.
345 struct io_poll_iocb poll;
348 const struct io_uring_sqe *sqe;
349 struct io_async_ctx *io;
350 struct file *ring_file;
354 bool needs_fixed_file;
356 struct io_ring_ctx *ctx;
358 struct list_head list;
359 struct hlist_node hash_node;
361 struct list_head link_list;
364 #define REQ_F_NOWAIT 1 /* must not punt to workers */
365 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
366 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
367 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
368 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
369 #define REQ_F_IO_DRAINED 32 /* drain done */
370 #define REQ_F_LINK 64 /* linked sqes */
371 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
372 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
373 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
374 #define REQ_F_TIMEOUT 1024 /* timeout request */
375 #define REQ_F_ISREG 2048 /* regular file */
376 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
377 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
378 #define REQ_F_INFLIGHT 16384 /* on inflight list */
379 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
384 struct list_head inflight_entry;
386 struct io_wq_work work;
389 #define IO_PLUG_THRESHOLD 2
390 #define IO_IOPOLL_BATCH 8
392 struct io_submit_state {
393 struct blk_plug plug;
396 * io_kiocb alloc cache
398 void *reqs[IO_IOPOLL_BATCH];
399 unsigned int free_reqs;
400 unsigned int cur_req;
403 * File reference cache
407 unsigned int has_refs;
408 unsigned int used_refs;
409 unsigned int ios_left;
412 static void io_wq_submit_work(struct io_wq_work **workptr);
413 static void io_cqring_fill_event(struct io_kiocb *req, long res);
414 static void __io_free_req(struct io_kiocb *req);
415 static void io_put_req(struct io_kiocb *req);
416 static void io_double_put_req(struct io_kiocb *req);
417 static void __io_double_put_req(struct io_kiocb *req);
418 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
419 static void io_queue_linked_timeout(struct io_kiocb *req);
421 static struct kmem_cache *req_cachep;
423 static const struct file_operations io_uring_fops;
425 struct sock *io_uring_get_socket(struct file *file)
427 #if defined(CONFIG_UNIX)
428 if (file->f_op == &io_uring_fops) {
429 struct io_ring_ctx *ctx = file->private_data;
431 return ctx->ring_sock->sk;
436 EXPORT_SYMBOL(io_uring_get_socket);
438 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
440 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
442 complete(&ctx->completions[0]);
445 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
447 struct io_ring_ctx *ctx;
450 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
454 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
455 if (!ctx->fallback_req)
458 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
459 if (!ctx->completions)
463 * Use 5 bits less than the max cq entries, that should give us around
464 * 32 entries per hash list if totally full and uniformly spread.
466 hash_bits = ilog2(p->cq_entries);
470 ctx->cancel_hash_bits = hash_bits;
471 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
473 if (!ctx->cancel_hash)
475 __hash_init(ctx->cancel_hash, 1U << hash_bits);
477 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
478 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
481 ctx->flags = p->flags;
482 init_waitqueue_head(&ctx->cq_wait);
483 INIT_LIST_HEAD(&ctx->cq_overflow_list);
484 init_completion(&ctx->completions[0]);
485 init_completion(&ctx->completions[1]);
486 mutex_init(&ctx->uring_lock);
487 init_waitqueue_head(&ctx->wait);
488 spin_lock_init(&ctx->completion_lock);
489 INIT_LIST_HEAD(&ctx->poll_list);
490 INIT_LIST_HEAD(&ctx->defer_list);
491 INIT_LIST_HEAD(&ctx->timeout_list);
492 init_waitqueue_head(&ctx->inflight_wait);
493 spin_lock_init(&ctx->inflight_lock);
494 INIT_LIST_HEAD(&ctx->inflight_list);
497 if (ctx->fallback_req)
498 kmem_cache_free(req_cachep, ctx->fallback_req);
499 kfree(ctx->completions);
500 kfree(ctx->cancel_hash);
505 static inline bool __req_need_defer(struct io_kiocb *req)
507 struct io_ring_ctx *ctx = req->ctx;
509 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
510 + atomic_read(&ctx->cached_cq_overflow);
513 static inline bool req_need_defer(struct io_kiocb *req)
515 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
516 return __req_need_defer(req);
521 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
523 struct io_kiocb *req;
525 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
526 if (req && !req_need_defer(req)) {
527 list_del_init(&req->list);
534 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
536 struct io_kiocb *req;
538 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
540 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
542 if (!__req_need_defer(req)) {
543 list_del_init(&req->list);
551 static void __io_commit_cqring(struct io_ring_ctx *ctx)
553 struct io_rings *rings = ctx->rings;
555 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
556 /* order cqe stores with ring update */
557 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
559 if (wq_has_sleeper(&ctx->cq_wait)) {
560 wake_up_interruptible(&ctx->cq_wait);
561 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
566 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
568 u8 opcode = READ_ONCE(sqe->opcode);
570 return !(opcode == IORING_OP_READ_FIXED ||
571 opcode == IORING_OP_WRITE_FIXED);
574 static inline bool io_prep_async_work(struct io_kiocb *req,
575 struct io_kiocb **link)
577 bool do_hashed = false;
580 switch (req->sqe->opcode) {
581 case IORING_OP_WRITEV:
582 case IORING_OP_WRITE_FIXED:
585 case IORING_OP_READV:
586 case IORING_OP_READ_FIXED:
587 case IORING_OP_SENDMSG:
588 case IORING_OP_RECVMSG:
589 case IORING_OP_ACCEPT:
590 case IORING_OP_POLL_ADD:
591 case IORING_OP_CONNECT:
593 * We know REQ_F_ISREG is not set on some of these
594 * opcodes, but this enables us to keep the check in
597 if (!(req->flags & REQ_F_ISREG))
598 req->work.flags |= IO_WQ_WORK_UNBOUND;
601 if (io_sqe_needs_user(req->sqe))
602 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
605 *link = io_prep_linked_timeout(req);
609 static inline void io_queue_async_work(struct io_kiocb *req)
611 struct io_ring_ctx *ctx = req->ctx;
612 struct io_kiocb *link;
615 do_hashed = io_prep_async_work(req, &link);
617 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
620 io_wq_enqueue(ctx->io_wq, &req->work);
622 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
623 file_inode(req->file));
627 io_queue_linked_timeout(link);
630 static void io_kill_timeout(struct io_kiocb *req)
634 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
636 atomic_inc(&req->ctx->cq_timeouts);
637 list_del_init(&req->list);
638 io_cqring_fill_event(req, 0);
643 static void io_kill_timeouts(struct io_ring_ctx *ctx)
645 struct io_kiocb *req, *tmp;
647 spin_lock_irq(&ctx->completion_lock);
648 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
649 io_kill_timeout(req);
650 spin_unlock_irq(&ctx->completion_lock);
653 static void io_commit_cqring(struct io_ring_ctx *ctx)
655 struct io_kiocb *req;
657 while ((req = io_get_timeout_req(ctx)) != NULL)
658 io_kill_timeout(req);
660 __io_commit_cqring(ctx);
662 while ((req = io_get_deferred_req(ctx)) != NULL) {
663 req->flags |= REQ_F_IO_DRAINED;
664 io_queue_async_work(req);
668 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
670 struct io_rings *rings = ctx->rings;
673 tail = ctx->cached_cq_tail;
675 * writes to the cq entry need to come after reading head; the
676 * control dependency is enough as we're using WRITE_ONCE to
679 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
682 ctx->cached_cq_tail++;
683 return &rings->cqes[tail & ctx->cq_mask];
686 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
688 if (waitqueue_active(&ctx->wait))
690 if (waitqueue_active(&ctx->sqo_wait))
691 wake_up(&ctx->sqo_wait);
693 eventfd_signal(ctx->cq_ev_fd, 1);
696 /* Returns true if there are no backlogged entries after the flush */
697 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
699 struct io_rings *rings = ctx->rings;
700 struct io_uring_cqe *cqe;
701 struct io_kiocb *req;
706 if (list_empty_careful(&ctx->cq_overflow_list))
708 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
709 rings->cq_ring_entries))
713 spin_lock_irqsave(&ctx->completion_lock, flags);
715 /* if force is set, the ring is going away. always drop after that */
717 ctx->cq_overflow_flushed = true;
720 while (!list_empty(&ctx->cq_overflow_list)) {
721 cqe = io_get_cqring(ctx);
725 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
727 list_move(&req->list, &list);
729 WRITE_ONCE(cqe->user_data, req->user_data);
730 WRITE_ONCE(cqe->res, req->result);
731 WRITE_ONCE(cqe->flags, 0);
733 WRITE_ONCE(ctx->rings->cq_overflow,
734 atomic_inc_return(&ctx->cached_cq_overflow));
738 io_commit_cqring(ctx);
739 spin_unlock_irqrestore(&ctx->completion_lock, flags);
740 io_cqring_ev_posted(ctx);
742 while (!list_empty(&list)) {
743 req = list_first_entry(&list, struct io_kiocb, list);
744 list_del(&req->list);
751 static void io_cqring_fill_event(struct io_kiocb *req, long res)
753 struct io_ring_ctx *ctx = req->ctx;
754 struct io_uring_cqe *cqe;
756 trace_io_uring_complete(ctx, req->user_data, res);
759 * If we can't get a cq entry, userspace overflowed the
760 * submission (by quite a lot). Increment the overflow count in
763 cqe = io_get_cqring(ctx);
765 WRITE_ONCE(cqe->user_data, req->user_data);
766 WRITE_ONCE(cqe->res, res);
767 WRITE_ONCE(cqe->flags, 0);
768 } else if (ctx->cq_overflow_flushed) {
769 WRITE_ONCE(ctx->rings->cq_overflow,
770 atomic_inc_return(&ctx->cached_cq_overflow));
772 refcount_inc(&req->refs);
774 list_add_tail(&req->list, &ctx->cq_overflow_list);
778 static void io_cqring_add_event(struct io_kiocb *req, long res)
780 struct io_ring_ctx *ctx = req->ctx;
783 spin_lock_irqsave(&ctx->completion_lock, flags);
784 io_cqring_fill_event(req, res);
785 io_commit_cqring(ctx);
786 spin_unlock_irqrestore(&ctx->completion_lock, flags);
788 io_cqring_ev_posted(ctx);
791 static inline bool io_is_fallback_req(struct io_kiocb *req)
793 return req == (struct io_kiocb *)
794 ((unsigned long) req->ctx->fallback_req & ~1UL);
797 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
799 struct io_kiocb *req;
801 req = ctx->fallback_req;
802 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
808 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
809 struct io_submit_state *state)
811 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
812 struct io_kiocb *req;
814 if (!percpu_ref_tryget(&ctx->refs))
818 req = kmem_cache_alloc(req_cachep, gfp);
821 } else if (!state->free_reqs) {
825 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
826 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
829 * Bulk alloc is all-or-nothing. If we fail to get a batch,
830 * retry single alloc to be on the safe side.
832 if (unlikely(ret <= 0)) {
833 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
838 state->free_reqs = ret - 1;
840 req = state->reqs[0];
842 req = state->reqs[state->cur_req];
849 req->ring_file = NULL;
853 /* one is dropped after submission, the other at completion */
854 refcount_set(&req->refs, 2);
856 INIT_IO_WORK(&req->work, io_wq_submit_work);
859 req = io_get_fallback_req(ctx);
862 percpu_ref_put(&ctx->refs);
866 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
869 kmem_cache_free_bulk(req_cachep, *nr, reqs);
870 percpu_ref_put_many(&ctx->refs, *nr);
875 static void __io_free_req(struct io_kiocb *req)
877 struct io_ring_ctx *ctx = req->ctx;
881 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
883 if (req->flags & REQ_F_INFLIGHT) {
886 spin_lock_irqsave(&ctx->inflight_lock, flags);
887 list_del(&req->inflight_entry);
888 if (waitqueue_active(&ctx->inflight_wait))
889 wake_up(&ctx->inflight_wait);
890 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
892 percpu_ref_put(&ctx->refs);
893 if (likely(!io_is_fallback_req(req)))
894 kmem_cache_free(req_cachep, req);
896 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
899 static bool io_link_cancel_timeout(struct io_kiocb *req)
901 struct io_ring_ctx *ctx = req->ctx;
904 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
906 io_cqring_fill_event(req, -ECANCELED);
907 io_commit_cqring(ctx);
908 req->flags &= ~REQ_F_LINK;
916 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
918 struct io_ring_ctx *ctx = req->ctx;
919 struct io_kiocb *nxt;
920 bool wake_ev = false;
922 /* Already got next link */
923 if (req->flags & REQ_F_LINK_NEXT)
927 * The list should never be empty when we are called here. But could
928 * potentially happen if the chain is messed up, check to be on the
931 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
933 list_del_init(&nxt->list);
935 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
936 (nxt->flags & REQ_F_TIMEOUT)) {
937 wake_ev |= io_link_cancel_timeout(nxt);
938 nxt = list_first_entry_or_null(&req->link_list,
939 struct io_kiocb, list);
940 req->flags &= ~REQ_F_LINK_TIMEOUT;
943 if (!list_empty(&req->link_list)) {
944 INIT_LIST_HEAD(&nxt->link_list);
945 list_splice(&req->link_list, &nxt->link_list);
946 nxt->flags |= REQ_F_LINK;
953 req->flags |= REQ_F_LINK_NEXT;
955 io_cqring_ev_posted(ctx);
959 * Called if REQ_F_LINK is set, and we fail the head request
961 static void io_fail_links(struct io_kiocb *req)
963 struct io_ring_ctx *ctx = req->ctx;
964 struct io_kiocb *link;
967 spin_lock_irqsave(&ctx->completion_lock, flags);
969 while (!list_empty(&req->link_list)) {
970 link = list_first_entry(&req->link_list, struct io_kiocb, list);
971 list_del_init(&link->list);
973 trace_io_uring_fail_link(req, link);
975 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
976 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
977 io_link_cancel_timeout(link);
979 io_cqring_fill_event(link, -ECANCELED);
980 __io_double_put_req(link);
982 req->flags &= ~REQ_F_LINK_TIMEOUT;
985 io_commit_cqring(ctx);
986 spin_unlock_irqrestore(&ctx->completion_lock, flags);
987 io_cqring_ev_posted(ctx);
990 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
992 if (likely(!(req->flags & REQ_F_LINK)))
996 * If LINK is set, we have dependent requests in this chain. If we
997 * didn't fail this request, queue the first one up, moving any other
998 * dependencies to the next request. In case of failure, fail the rest
1001 if (req->flags & REQ_F_FAIL_LINK) {
1003 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1004 REQ_F_LINK_TIMEOUT) {
1005 struct io_ring_ctx *ctx = req->ctx;
1006 unsigned long flags;
1009 * If this is a timeout link, we could be racing with the
1010 * timeout timer. Grab the completion lock for this case to
1011 * protect against that.
1013 spin_lock_irqsave(&ctx->completion_lock, flags);
1014 io_req_link_next(req, nxt);
1015 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1017 io_req_link_next(req, nxt);
1021 static void io_free_req(struct io_kiocb *req)
1023 struct io_kiocb *nxt = NULL;
1025 io_req_find_next(req, &nxt);
1029 io_queue_async_work(nxt);
1033 * Drop reference to request, return next in chain (if there is one) if this
1034 * was the last reference to this request.
1036 __attribute__((nonnull))
1037 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1039 io_req_find_next(req, nxtptr);
1041 if (refcount_dec_and_test(&req->refs))
1045 static void io_put_req(struct io_kiocb *req)
1047 if (refcount_dec_and_test(&req->refs))
1052 * Must only be used if we don't need to care about links, usually from
1053 * within the completion handling itself.
1055 static void __io_double_put_req(struct io_kiocb *req)
1057 /* drop both submit and complete references */
1058 if (refcount_sub_and_test(2, &req->refs))
1062 static void io_double_put_req(struct io_kiocb *req)
1064 /* drop both submit and complete references */
1065 if (refcount_sub_and_test(2, &req->refs))
1069 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1071 struct io_rings *rings = ctx->rings;
1074 * noflush == true is from the waitqueue handler, just ensure we wake
1075 * up the task, and the next invocation will flush the entries. We
1076 * cannot safely to it from here.
1078 if (noflush && !list_empty(&ctx->cq_overflow_list))
1081 io_cqring_overflow_flush(ctx, false);
1083 /* See comment at the top of this file */
1085 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1088 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1090 struct io_rings *rings = ctx->rings;
1092 /* make sure SQ entry isn't read before tail */
1093 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1097 * Find and free completed poll iocbs
1099 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1100 struct list_head *done)
1102 void *reqs[IO_IOPOLL_BATCH];
1103 struct io_kiocb *req;
1107 while (!list_empty(done)) {
1108 req = list_first_entry(done, struct io_kiocb, list);
1109 list_del(&req->list);
1111 io_cqring_fill_event(req, req->result);
1114 if (refcount_dec_and_test(&req->refs)) {
1115 /* If we're not using fixed files, we have to pair the
1116 * completion part with the file put. Use regular
1117 * completions for those, only batch free for fixed
1118 * file and non-linked commands.
1120 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1121 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1123 reqs[to_free++] = req;
1124 if (to_free == ARRAY_SIZE(reqs))
1125 io_free_req_many(ctx, reqs, &to_free);
1132 io_commit_cqring(ctx);
1133 io_free_req_many(ctx, reqs, &to_free);
1136 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1139 struct io_kiocb *req, *tmp;
1145 * Only spin for completions if we don't have multiple devices hanging
1146 * off our complete list, and we're under the requested amount.
1148 spin = !ctx->poll_multi_file && *nr_events < min;
1151 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1152 struct kiocb *kiocb = &req->rw;
1155 * Move completed entries to our local list. If we find a
1156 * request that requires polling, break out and complete
1157 * the done list first, if we have entries there.
1159 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1160 list_move_tail(&req->list, &done);
1163 if (!list_empty(&done))
1166 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1175 if (!list_empty(&done))
1176 io_iopoll_complete(ctx, nr_events, &done);
1182 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
1183 * non-spinning poll check - we'll still enter the driver poll loop, but only
1184 * as a non-spinning completion check.
1186 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1189 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1192 ret = io_do_iopoll(ctx, nr_events, min);
1195 if (!min || *nr_events >= min)
1203 * We can't just wait for polled events to come to us, we have to actively
1204 * find and complete them.
1206 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1208 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1211 mutex_lock(&ctx->uring_lock);
1212 while (!list_empty(&ctx->poll_list)) {
1213 unsigned int nr_events = 0;
1215 io_iopoll_getevents(ctx, &nr_events, 1);
1218 * Ensure we allow local-to-the-cpu processing to take place,
1219 * in this case we need to ensure that we reap all events.
1223 mutex_unlock(&ctx->uring_lock);
1226 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1229 int iters = 0, ret = 0;
1235 * Don't enter poll loop if we already have events pending.
1236 * If we do, we can potentially be spinning for commands that
1237 * already triggered a CQE (eg in error).
1239 if (io_cqring_events(ctx, false))
1243 * If a submit got punted to a workqueue, we can have the
1244 * application entering polling for a command before it gets
1245 * issued. That app will hold the uring_lock for the duration
1246 * of the poll right here, so we need to take a breather every
1247 * now and then to ensure that the issue has a chance to add
1248 * the poll to the issued list. Otherwise we can spin here
1249 * forever, while the workqueue is stuck trying to acquire the
1252 if (!(++iters & 7)) {
1253 mutex_unlock(&ctx->uring_lock);
1254 mutex_lock(&ctx->uring_lock);
1257 if (*nr_events < min)
1258 tmin = min - *nr_events;
1260 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1264 } while (min && !*nr_events && !need_resched());
1269 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1275 * We disallow the app entering submit/complete with polling, but we
1276 * still need to lock the ring to prevent racing with polled issue
1277 * that got punted to a workqueue.
1279 mutex_lock(&ctx->uring_lock);
1280 ret = __io_iopoll_check(ctx, nr_events, min);
1281 mutex_unlock(&ctx->uring_lock);
1285 static void kiocb_end_write(struct io_kiocb *req)
1288 * Tell lockdep we inherited freeze protection from submission
1291 if (req->flags & REQ_F_ISREG) {
1292 struct inode *inode = file_inode(req->file);
1294 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1296 file_end_write(req->file);
1299 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1301 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1303 if (kiocb->ki_flags & IOCB_WRITE)
1304 kiocb_end_write(req);
1306 if ((req->flags & REQ_F_LINK) && res != req->result)
1307 req->flags |= REQ_F_FAIL_LINK;
1308 io_cqring_add_event(req, res);
1311 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1313 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1315 io_complete_rw_common(kiocb, res);
1319 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1321 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1322 struct io_kiocb *nxt = NULL;
1324 io_complete_rw_common(kiocb, res);
1325 io_put_req_find_next(req, &nxt);
1330 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1332 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1334 if (kiocb->ki_flags & IOCB_WRITE)
1335 kiocb_end_write(req);
1337 if ((req->flags & REQ_F_LINK) && res != req->result)
1338 req->flags |= REQ_F_FAIL_LINK;
1341 req->flags |= REQ_F_IOPOLL_COMPLETED;
1345 * After the iocb has been issued, it's safe to be found on the poll list.
1346 * Adding the kiocb to the list AFTER submission ensures that we don't
1347 * find it from a io_iopoll_getevents() thread before the issuer is done
1348 * accessing the kiocb cookie.
1350 static void io_iopoll_req_issued(struct io_kiocb *req)
1352 struct io_ring_ctx *ctx = req->ctx;
1355 * Track whether we have multiple files in our lists. This will impact
1356 * how we do polling eventually, not spinning if we're on potentially
1357 * different devices.
1359 if (list_empty(&ctx->poll_list)) {
1360 ctx->poll_multi_file = false;
1361 } else if (!ctx->poll_multi_file) {
1362 struct io_kiocb *list_req;
1364 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1366 if (list_req->rw.ki_filp != req->rw.ki_filp)
1367 ctx->poll_multi_file = true;
1371 * For fast devices, IO may have already completed. If it has, add
1372 * it to the front so we find it first.
1374 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1375 list_add(&req->list, &ctx->poll_list);
1377 list_add_tail(&req->list, &ctx->poll_list);
1380 static void io_file_put(struct io_submit_state *state)
1383 int diff = state->has_refs - state->used_refs;
1386 fput_many(state->file, diff);
1392 * Get as many references to a file as we have IOs left in this submission,
1393 * assuming most submissions are for one file, or at least that each file
1394 * has more than one submission.
1396 static struct file *io_file_get(struct io_submit_state *state, int fd)
1402 if (state->fd == fd) {
1409 state->file = fget_many(fd, state->ios_left);
1414 state->has_refs = state->ios_left;
1415 state->used_refs = 1;
1421 * If we tracked the file through the SCM inflight mechanism, we could support
1422 * any file. For now, just ensure that anything potentially problematic is done
1425 static bool io_file_supports_async(struct file *file)
1427 umode_t mode = file_inode(file)->i_mode;
1429 if (S_ISBLK(mode) || S_ISCHR(mode))
1431 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1437 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1439 const struct io_uring_sqe *sqe = req->sqe;
1440 struct io_ring_ctx *ctx = req->ctx;
1441 struct kiocb *kiocb = &req->rw;
1448 if (S_ISREG(file_inode(req->file)->i_mode))
1449 req->flags |= REQ_F_ISREG;
1451 kiocb->ki_pos = READ_ONCE(sqe->off);
1452 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1453 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1455 ioprio = READ_ONCE(sqe->ioprio);
1457 ret = ioprio_check_cap(ioprio);
1461 kiocb->ki_ioprio = ioprio;
1463 kiocb->ki_ioprio = get_current_ioprio();
1465 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1469 /* don't allow async punt if RWF_NOWAIT was requested */
1470 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1471 (req->file->f_flags & O_NONBLOCK))
1472 req->flags |= REQ_F_NOWAIT;
1475 kiocb->ki_flags |= IOCB_NOWAIT;
1477 if (ctx->flags & IORING_SETUP_IOPOLL) {
1478 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1479 !kiocb->ki_filp->f_op->iopoll)
1482 kiocb->ki_flags |= IOCB_HIPRI;
1483 kiocb->ki_complete = io_complete_rw_iopoll;
1486 if (kiocb->ki_flags & IOCB_HIPRI)
1488 kiocb->ki_complete = io_complete_rw;
1493 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1499 case -ERESTARTNOINTR:
1500 case -ERESTARTNOHAND:
1501 case -ERESTART_RESTARTBLOCK:
1503 * We can't just restart the syscall, since previously
1504 * submitted sqes may already be in progress. Just fail this
1510 kiocb->ki_complete(kiocb, ret, 0);
1514 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1517 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1518 *nxt = __io_complete_rw(kiocb, ret);
1520 io_rw_done(kiocb, ret);
1523 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1524 const struct io_uring_sqe *sqe,
1525 struct iov_iter *iter)
1527 size_t len = READ_ONCE(sqe->len);
1528 struct io_mapped_ubuf *imu;
1529 unsigned index, buf_index;
1533 /* attempt to use fixed buffers without having provided iovecs */
1534 if (unlikely(!ctx->user_bufs))
1537 buf_index = READ_ONCE(sqe->buf_index);
1538 if (unlikely(buf_index >= ctx->nr_user_bufs))
1541 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1542 imu = &ctx->user_bufs[index];
1543 buf_addr = READ_ONCE(sqe->addr);
1546 if (buf_addr + len < buf_addr)
1548 /* not inside the mapped region */
1549 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1553 * May not be a start of buffer, set size appropriately
1554 * and advance us to the beginning.
1556 offset = buf_addr - imu->ubuf;
1557 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1561 * Don't use iov_iter_advance() here, as it's really slow for
1562 * using the latter parts of a big fixed buffer - it iterates
1563 * over each segment manually. We can cheat a bit here, because
1566 * 1) it's a BVEC iter, we set it up
1567 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1568 * first and last bvec
1570 * So just find our index, and adjust the iterator afterwards.
1571 * If the offset is within the first bvec (or the whole first
1572 * bvec, just use iov_iter_advance(). This makes it easier
1573 * since we can just skip the first segment, which may not
1574 * be PAGE_SIZE aligned.
1576 const struct bio_vec *bvec = imu->bvec;
1578 if (offset <= bvec->bv_len) {
1579 iov_iter_advance(iter, offset);
1581 unsigned long seg_skip;
1583 /* skip first vec */
1584 offset -= bvec->bv_len;
1585 seg_skip = 1 + (offset >> PAGE_SHIFT);
1587 iter->bvec = bvec + seg_skip;
1588 iter->nr_segs -= seg_skip;
1589 iter->count -= bvec->bv_len + offset;
1590 iter->iov_offset = offset & ~PAGE_MASK;
1597 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1598 struct iovec **iovec, struct iov_iter *iter)
1600 const struct io_uring_sqe *sqe = req->sqe;
1601 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1602 size_t sqe_len = READ_ONCE(sqe->len);
1606 * We're reading ->opcode for the second time, but the first read
1607 * doesn't care whether it's _FIXED or not, so it doesn't matter
1608 * whether ->opcode changes concurrently. The first read does care
1609 * about whether it is a READ or a WRITE, so we don't trust this read
1610 * for that purpose and instead let the caller pass in the read/write
1613 opcode = READ_ONCE(sqe->opcode);
1614 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1616 return io_import_fixed(req->ctx, rw, sqe, iter);
1620 struct io_async_rw *iorw = &req->io->rw;
1623 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1624 if (iorw->iov == iorw->fast_iov)
1632 #ifdef CONFIG_COMPAT
1633 if (req->ctx->compat)
1634 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1638 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1642 * For files that don't have ->read_iter() and ->write_iter(), handle them
1643 * by looping over ->read() or ->write() manually.
1645 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1646 struct iov_iter *iter)
1651 * Don't support polled IO through this interface, and we can't
1652 * support non-blocking either. For the latter, this just causes
1653 * the kiocb to be handled from an async context.
1655 if (kiocb->ki_flags & IOCB_HIPRI)
1657 if (kiocb->ki_flags & IOCB_NOWAIT)
1660 while (iov_iter_count(iter)) {
1664 if (!iov_iter_is_bvec(iter)) {
1665 iovec = iov_iter_iovec(iter);
1667 /* fixed buffers import bvec */
1668 iovec.iov_base = kmap(iter->bvec->bv_page)
1670 iovec.iov_len = min(iter->count,
1671 iter->bvec->bv_len - iter->iov_offset);
1675 nr = file->f_op->read(file, iovec.iov_base,
1676 iovec.iov_len, &kiocb->ki_pos);
1678 nr = file->f_op->write(file, iovec.iov_base,
1679 iovec.iov_len, &kiocb->ki_pos);
1682 if (iov_iter_is_bvec(iter))
1683 kunmap(iter->bvec->bv_page);
1691 if (nr != iovec.iov_len)
1693 iov_iter_advance(iter, nr);
1699 static void io_req_map_io(struct io_kiocb *req, ssize_t io_size,
1700 struct iovec *iovec, struct iovec *fast_iov,
1701 struct iov_iter *iter)
1703 req->io->rw.nr_segs = iter->nr_segs;
1704 req->io->rw.size = io_size;
1705 req->io->rw.iov = iovec;
1706 if (!req->io->rw.iov) {
1707 req->io->rw.iov = req->io->rw.fast_iov;
1708 memcpy(req->io->rw.iov, fast_iov,
1709 sizeof(struct iovec) * iter->nr_segs);
1713 static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size,
1714 struct iovec *iovec, struct iovec *fast_iov,
1715 struct iov_iter *iter)
1717 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1719 io_req_map_io(req, io_size, iovec, fast_iov, iter);
1720 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1721 req->sqe = &req->io->sqe;
1728 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1729 struct iov_iter *iter, bool force_nonblock)
1733 ret = io_prep_rw(req, force_nonblock);
1737 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1740 return io_import_iovec(READ, req, iovec, iter);
1743 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1744 bool force_nonblock)
1746 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1747 struct kiocb *kiocb = &req->rw;
1748 struct iov_iter iter;
1751 ssize_t io_size, ret;
1754 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1758 ret = io_import_iovec(READ, req, &iovec, &iter);
1765 if (req->flags & REQ_F_LINK)
1766 req->result = io_size;
1769 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1770 * we know to async punt it even if it was opened O_NONBLOCK
1772 if (force_nonblock && !io_file_supports_async(file)) {
1773 req->flags |= REQ_F_MUST_PUNT;
1777 iov_count = iov_iter_count(&iter);
1778 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1782 if (file->f_op->read_iter)
1783 ret2 = call_read_iter(file, kiocb, &iter);
1785 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1788 * In case of a short read, punt to async. This can happen
1789 * if we have data partially cached. Alternatively we can
1790 * return the short read, in which case the application will
1791 * need to issue another SQE and wait for it. That SQE will
1792 * need async punt anyway, so it's more efficient to do it
1795 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1796 (req->flags & REQ_F_ISREG) &&
1797 ret2 > 0 && ret2 < io_size)
1799 /* Catch -EAGAIN return for forced non-blocking submission */
1800 if (!force_nonblock || ret2 != -EAGAIN) {
1801 kiocb_done(kiocb, ret2, nxt, req->in_async);
1804 ret = io_setup_async_io(req, io_size, iovec,
1805 inline_vecs, &iter);
1816 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1817 struct iov_iter *iter, bool force_nonblock)
1821 ret = io_prep_rw(req, force_nonblock);
1825 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1828 return io_import_iovec(WRITE, req, iovec, iter);
1831 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1832 bool force_nonblock)
1834 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1835 struct kiocb *kiocb = &req->rw;
1836 struct iov_iter iter;
1839 ssize_t ret, io_size;
1842 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1846 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1851 file = kiocb->ki_filp;
1853 if (req->flags & REQ_F_LINK)
1854 req->result = io_size;
1857 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1858 * we know to async punt it even if it was opened O_NONBLOCK
1860 if (force_nonblock && !io_file_supports_async(req->file)) {
1861 req->flags |= REQ_F_MUST_PUNT;
1865 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
1868 iov_count = iov_iter_count(&iter);
1869 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1874 * Open-code file_start_write here to grab freeze protection,
1875 * which will be released by another thread in
1876 * io_complete_rw(). Fool lockdep by telling it the lock got
1877 * released so that it doesn't complain about the held lock when
1878 * we return to userspace.
1880 if (req->flags & REQ_F_ISREG) {
1881 __sb_start_write(file_inode(file)->i_sb,
1882 SB_FREEZE_WRITE, true);
1883 __sb_writers_release(file_inode(file)->i_sb,
1886 kiocb->ki_flags |= IOCB_WRITE;
1888 if (file->f_op->write_iter)
1889 ret2 = call_write_iter(file, kiocb, &iter);
1891 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1892 if (!force_nonblock || ret2 != -EAGAIN) {
1893 kiocb_done(kiocb, ret2, nxt, req->in_async);
1896 ret = io_setup_async_io(req, io_size, iovec,
1897 inline_vecs, &iter);
1909 * IORING_OP_NOP just posts a completion event, nothing else.
1911 static int io_nop(struct io_kiocb *req)
1913 struct io_ring_ctx *ctx = req->ctx;
1915 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1918 io_cqring_add_event(req, 0);
1923 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1925 struct io_ring_ctx *ctx = req->ctx;
1930 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1932 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1938 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1939 struct io_kiocb **nxt, bool force_nonblock)
1941 loff_t sqe_off = READ_ONCE(sqe->off);
1942 loff_t sqe_len = READ_ONCE(sqe->len);
1943 loff_t end = sqe_off + sqe_len;
1944 unsigned fsync_flags;
1947 fsync_flags = READ_ONCE(sqe->fsync_flags);
1948 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1951 ret = io_prep_fsync(req, sqe);
1955 /* fsync always requires a blocking context */
1959 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1960 end > 0 ? end : LLONG_MAX,
1961 fsync_flags & IORING_FSYNC_DATASYNC);
1963 if (ret < 0 && (req->flags & REQ_F_LINK))
1964 req->flags |= REQ_F_FAIL_LINK;
1965 io_cqring_add_event(req, ret);
1966 io_put_req_find_next(req, nxt);
1970 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1972 struct io_ring_ctx *ctx = req->ctx;
1978 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1980 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1986 static int io_sync_file_range(struct io_kiocb *req,
1987 const struct io_uring_sqe *sqe,
1988 struct io_kiocb **nxt,
1989 bool force_nonblock)
1996 ret = io_prep_sfr(req, sqe);
2000 /* sync_file_range always requires a blocking context */
2004 sqe_off = READ_ONCE(sqe->off);
2005 sqe_len = READ_ONCE(sqe->len);
2006 flags = READ_ONCE(sqe->sync_range_flags);
2008 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
2010 if (ret < 0 && (req->flags & REQ_F_LINK))
2011 req->flags |= REQ_F_FAIL_LINK;
2012 io_cqring_add_event(req, ret);
2013 io_put_req_find_next(req, nxt);
2017 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2019 #if defined(CONFIG_NET)
2020 const struct io_uring_sqe *sqe = req->sqe;
2021 struct user_msghdr __user *msg;
2024 flags = READ_ONCE(sqe->msg_flags);
2025 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2026 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2032 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2033 struct io_kiocb **nxt, bool force_nonblock)
2035 #if defined(CONFIG_NET)
2036 struct socket *sock;
2039 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2042 sock = sock_from_file(req->file, &ret);
2044 struct io_async_ctx io, *copy;
2045 struct sockaddr_storage addr;
2046 struct msghdr *kmsg;
2049 flags = READ_ONCE(sqe->msg_flags);
2050 if (flags & MSG_DONTWAIT)
2051 req->flags |= REQ_F_NOWAIT;
2052 else if (force_nonblock)
2053 flags |= MSG_DONTWAIT;
2056 kmsg = &req->io->msg.msg;
2057 kmsg->msg_name = &addr;
2060 kmsg->msg_name = &addr;
2061 io.msg.iov = io.msg.fast_iov;
2062 ret = io_sendmsg_prep(req, &io);
2067 ret = __sys_sendmsg_sock(sock, kmsg, flags);
2068 if (force_nonblock && ret == -EAGAIN) {
2069 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2074 memcpy(©->msg, &io.msg, sizeof(copy->msg));
2076 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2077 req->sqe = &req->io->sqe;
2080 if (ret == -ERESTARTSYS)
2085 io_cqring_add_event(req, ret);
2086 if (ret < 0 && (req->flags & REQ_F_LINK))
2087 req->flags |= REQ_F_FAIL_LINK;
2088 io_put_req_find_next(req, nxt);
2095 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2097 #if defined(CONFIG_NET)
2098 const struct io_uring_sqe *sqe = req->sqe;
2099 struct user_msghdr __user *msg;
2102 flags = READ_ONCE(sqe->msg_flags);
2103 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2104 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2111 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2112 struct io_kiocb **nxt, bool force_nonblock)
2114 #if defined(CONFIG_NET)
2115 struct socket *sock;
2118 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2121 sock = sock_from_file(req->file, &ret);
2123 struct user_msghdr __user *msg;
2124 struct io_async_ctx io, *copy;
2125 struct sockaddr_storage addr;
2126 struct msghdr *kmsg;
2129 flags = READ_ONCE(sqe->msg_flags);
2130 if (flags & MSG_DONTWAIT)
2131 req->flags |= REQ_F_NOWAIT;
2132 else if (force_nonblock)
2133 flags |= MSG_DONTWAIT;
2135 msg = (struct user_msghdr __user *) (unsigned long)
2136 READ_ONCE(sqe->addr);
2138 kmsg = &req->io->msg.msg;
2139 kmsg->msg_name = &addr;
2142 kmsg->msg_name = &addr;
2143 io.msg.iov = io.msg.fast_iov;
2144 ret = io_recvmsg_prep(req, &io);
2149 ret = __sys_recvmsg_sock(sock, kmsg, msg, io.msg.uaddr, flags);
2150 if (force_nonblock && ret == -EAGAIN) {
2151 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2156 memcpy(copy, &io, sizeof(*copy));
2158 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2159 req->sqe = &req->io->sqe;
2162 if (ret == -ERESTARTSYS)
2167 io_cqring_add_event(req, ret);
2168 if (ret < 0 && (req->flags & REQ_F_LINK))
2169 req->flags |= REQ_F_FAIL_LINK;
2170 io_put_req_find_next(req, nxt);
2177 static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2178 struct io_kiocb **nxt, bool force_nonblock)
2180 #if defined(CONFIG_NET)
2181 struct sockaddr __user *addr;
2182 int __user *addr_len;
2183 unsigned file_flags;
2186 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2188 if (sqe->ioprio || sqe->len || sqe->buf_index)
2191 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2192 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2193 flags = READ_ONCE(sqe->accept_flags);
2194 file_flags = force_nonblock ? O_NONBLOCK : 0;
2196 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
2197 if (ret == -EAGAIN && force_nonblock) {
2198 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2201 if (ret == -ERESTARTSYS)
2203 if (ret < 0 && (req->flags & REQ_F_LINK))
2204 req->flags |= REQ_F_FAIL_LINK;
2205 io_cqring_add_event(req, ret);
2206 io_put_req_find_next(req, nxt);
2213 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2215 #if defined(CONFIG_NET)
2216 const struct io_uring_sqe *sqe = req->sqe;
2217 struct sockaddr __user *addr;
2220 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2221 addr_len = READ_ONCE(sqe->addr2);
2222 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2228 static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2229 struct io_kiocb **nxt, bool force_nonblock)
2231 #if defined(CONFIG_NET)
2232 struct io_async_ctx __io, *io;
2233 unsigned file_flags;
2236 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2238 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2241 addr_len = READ_ONCE(sqe->addr2);
2242 file_flags = force_nonblock ? O_NONBLOCK : 0;
2247 ret = io_connect_prep(req, &__io);
2253 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2255 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2256 io = kmalloc(sizeof(*io), GFP_KERNEL);
2261 memcpy(&io->connect, &__io.connect, sizeof(io->connect));
2263 memcpy(&io->sqe, req->sqe, sizeof(*req->sqe));
2264 req->sqe = &io->sqe;
2267 if (ret == -ERESTARTSYS)
2270 if (ret < 0 && (req->flags & REQ_F_LINK))
2271 req->flags |= REQ_F_FAIL_LINK;
2272 io_cqring_add_event(req, ret);
2273 io_put_req_find_next(req, nxt);
2280 static void io_poll_remove_one(struct io_kiocb *req)
2282 struct io_poll_iocb *poll = &req->poll;
2284 spin_lock(&poll->head->lock);
2285 WRITE_ONCE(poll->canceled, true);
2286 if (!list_empty(&poll->wait->entry)) {
2287 list_del_init(&poll->wait->entry);
2288 io_queue_async_work(req);
2290 spin_unlock(&poll->head->lock);
2291 hash_del(&req->hash_node);
2294 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2296 struct hlist_node *tmp;
2297 struct io_kiocb *req;
2300 spin_lock_irq(&ctx->completion_lock);
2301 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2302 struct hlist_head *list;
2304 list = &ctx->cancel_hash[i];
2305 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2306 io_poll_remove_one(req);
2308 spin_unlock_irq(&ctx->completion_lock);
2311 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2313 struct hlist_head *list;
2314 struct io_kiocb *req;
2316 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2317 hlist_for_each_entry(req, list, hash_node) {
2318 if (sqe_addr == req->user_data) {
2319 io_poll_remove_one(req);
2328 * Find a running poll command that matches one specified in sqe->addr,
2329 * and remove it if found.
2331 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2333 struct io_ring_ctx *ctx = req->ctx;
2336 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2338 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2342 spin_lock_irq(&ctx->completion_lock);
2343 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
2344 spin_unlock_irq(&ctx->completion_lock);
2346 io_cqring_add_event(req, ret);
2347 if (ret < 0 && (req->flags & REQ_F_LINK))
2348 req->flags |= REQ_F_FAIL_LINK;
2353 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2355 struct io_ring_ctx *ctx = req->ctx;
2357 req->poll.done = true;
2358 kfree(req->poll.wait);
2360 io_cqring_fill_event(req, error);
2362 io_cqring_fill_event(req, mangle_poll(mask));
2363 io_commit_cqring(ctx);
2366 static void io_poll_complete_work(struct io_wq_work **workptr)
2368 struct io_wq_work *work = *workptr;
2369 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2370 struct io_poll_iocb *poll = &req->poll;
2371 struct poll_table_struct pt = { ._key = poll->events };
2372 struct io_ring_ctx *ctx = req->ctx;
2373 struct io_kiocb *nxt = NULL;
2377 if (work->flags & IO_WQ_WORK_CANCEL) {
2378 WRITE_ONCE(poll->canceled, true);
2380 } else if (READ_ONCE(poll->canceled)) {
2384 if (ret != -ECANCELED)
2385 mask = vfs_poll(poll->file, &pt) & poll->events;
2388 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2389 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2390 * synchronize with them. In the cancellation case the list_del_init
2391 * itself is not actually needed, but harmless so we keep it in to
2392 * avoid further branches in the fast path.
2394 spin_lock_irq(&ctx->completion_lock);
2395 if (!mask && ret != -ECANCELED) {
2396 add_wait_queue(poll->head, poll->wait);
2397 spin_unlock_irq(&ctx->completion_lock);
2400 hash_del(&req->hash_node);
2401 io_poll_complete(req, mask, ret);
2402 spin_unlock_irq(&ctx->completion_lock);
2404 io_cqring_ev_posted(ctx);
2406 if (ret < 0 && req->flags & REQ_F_LINK)
2407 req->flags |= REQ_F_FAIL_LINK;
2408 io_put_req_find_next(req, &nxt);
2410 *workptr = &nxt->work;
2413 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2416 struct io_poll_iocb *poll = wait->private;
2417 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2418 struct io_ring_ctx *ctx = req->ctx;
2419 __poll_t mask = key_to_poll(key);
2420 unsigned long flags;
2422 /* for instances that support it check for an event match first: */
2423 if (mask && !(mask & poll->events))
2426 list_del_init(&poll->wait->entry);
2429 * Run completion inline if we can. We're using trylock here because
2430 * we are violating the completion_lock -> poll wq lock ordering.
2431 * If we have a link timeout we're going to need the completion_lock
2432 * for finalizing the request, mark us as having grabbed that already.
2434 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2435 hash_del(&req->hash_node);
2436 io_poll_complete(req, mask, 0);
2437 req->flags |= REQ_F_COMP_LOCKED;
2439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2441 io_cqring_ev_posted(ctx);
2443 io_queue_async_work(req);
2449 struct io_poll_table {
2450 struct poll_table_struct pt;
2451 struct io_kiocb *req;
2455 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2456 struct poll_table_struct *p)
2458 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2460 if (unlikely(pt->req->poll.head)) {
2461 pt->error = -EINVAL;
2466 pt->req->poll.head = head;
2467 add_wait_queue(head, pt->req->poll.wait);
2470 static void io_poll_req_insert(struct io_kiocb *req)
2472 struct io_ring_ctx *ctx = req->ctx;
2473 struct hlist_head *list;
2475 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2476 hlist_add_head(&req->hash_node, list);
2479 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2480 struct io_kiocb **nxt)
2482 struct io_poll_iocb *poll = &req->poll;
2483 struct io_ring_ctx *ctx = req->ctx;
2484 struct io_poll_table ipt;
2485 bool cancel = false;
2489 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2491 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2496 poll->wait = kmalloc(sizeof(*poll->wait), GFP_KERNEL);
2501 INIT_IO_WORK(&req->work, io_poll_complete_work);
2502 events = READ_ONCE(sqe->poll_events);
2503 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2504 INIT_HLIST_NODE(&req->hash_node);
2508 poll->canceled = false;
2510 ipt.pt._qproc = io_poll_queue_proc;
2511 ipt.pt._key = poll->events;
2513 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2515 /* initialized the list so that we can do list_empty checks */
2516 INIT_LIST_HEAD(&poll->wait->entry);
2517 init_waitqueue_func_entry(poll->wait, io_poll_wake);
2518 poll->wait->private = poll;
2520 INIT_LIST_HEAD(&req->list);
2522 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2524 spin_lock_irq(&ctx->completion_lock);
2525 if (likely(poll->head)) {
2526 spin_lock(&poll->head->lock);
2527 if (unlikely(list_empty(&poll->wait->entry))) {
2533 if (mask || ipt.error)
2534 list_del_init(&poll->wait->entry);
2536 WRITE_ONCE(poll->canceled, true);
2537 else if (!poll->done) /* actually waiting for an event */
2538 io_poll_req_insert(req);
2539 spin_unlock(&poll->head->lock);
2541 if (mask) { /* no async, we'd stolen it */
2543 io_poll_complete(req, mask, 0);
2545 spin_unlock_irq(&ctx->completion_lock);
2548 io_cqring_ev_posted(ctx);
2549 io_put_req_find_next(req, nxt);
2554 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2556 struct io_timeout_data *data = container_of(timer,
2557 struct io_timeout_data, timer);
2558 struct io_kiocb *req = data->req;
2559 struct io_ring_ctx *ctx = req->ctx;
2560 unsigned long flags;
2562 atomic_inc(&ctx->cq_timeouts);
2564 spin_lock_irqsave(&ctx->completion_lock, flags);
2566 * We could be racing with timeout deletion. If the list is empty,
2567 * then timeout lookup already found it and will be handling it.
2569 if (!list_empty(&req->list)) {
2570 struct io_kiocb *prev;
2573 * Adjust the reqs sequence before the current one because it
2574 * will consume a slot in the cq_ring and the the cq_tail
2575 * pointer will be increased, otherwise other timeout reqs may
2576 * return in advance without waiting for enough wait_nr.
2579 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2581 list_del_init(&req->list);
2584 io_cqring_fill_event(req, -ETIME);
2585 io_commit_cqring(ctx);
2586 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2588 io_cqring_ev_posted(ctx);
2589 if (req->flags & REQ_F_LINK)
2590 req->flags |= REQ_F_FAIL_LINK;
2592 return HRTIMER_NORESTART;
2595 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2597 struct io_kiocb *req;
2600 list_for_each_entry(req, &ctx->timeout_list, list) {
2601 if (user_data == req->user_data) {
2602 list_del_init(&req->list);
2611 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2615 if (req->flags & REQ_F_LINK)
2616 req->flags |= REQ_F_FAIL_LINK;
2617 io_cqring_fill_event(req, -ECANCELED);
2623 * Remove or update an existing timeout command
2625 static int io_timeout_remove(struct io_kiocb *req,
2626 const struct io_uring_sqe *sqe)
2628 struct io_ring_ctx *ctx = req->ctx;
2632 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2634 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2636 flags = READ_ONCE(sqe->timeout_flags);
2640 spin_lock_irq(&ctx->completion_lock);
2641 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
2643 io_cqring_fill_event(req, ret);
2644 io_commit_cqring(ctx);
2645 spin_unlock_irq(&ctx->completion_lock);
2646 io_cqring_ev_posted(ctx);
2647 if (ret < 0 && req->flags & REQ_F_LINK)
2648 req->flags |= REQ_F_FAIL_LINK;
2653 static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2654 bool is_timeout_link)
2656 const struct io_uring_sqe *sqe = req->sqe;
2657 struct io_timeout_data *data;
2660 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2662 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2664 if (sqe->off && is_timeout_link)
2666 flags = READ_ONCE(sqe->timeout_flags);
2667 if (flags & ~IORING_TIMEOUT_ABS)
2670 data = &io->timeout;
2672 req->flags |= REQ_F_TIMEOUT;
2674 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2677 if (flags & IORING_TIMEOUT_ABS)
2678 data->mode = HRTIMER_MODE_ABS;
2680 data->mode = HRTIMER_MODE_REL;
2682 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2687 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2690 struct io_ring_ctx *ctx = req->ctx;
2691 struct io_timeout_data *data;
2692 struct io_async_ctx *io;
2693 struct list_head *entry;
2700 io = kmalloc(sizeof(*io), GFP_KERNEL);
2703 ret = io_timeout_prep(req, io, false);
2709 data = &req->io->timeout;
2712 * sqe->off holds how many events that need to occur for this
2713 * timeout event to be satisfied. If it isn't set, then this is
2714 * a pure timeout request, sequence isn't used.
2716 count = READ_ONCE(sqe->off);
2718 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2719 spin_lock_irq(&ctx->completion_lock);
2720 entry = ctx->timeout_list.prev;
2724 req->sequence = ctx->cached_sq_head + count - 1;
2725 data->seq_offset = count;
2728 * Insertion sort, ensuring the first entry in the list is always
2729 * the one we need first.
2731 spin_lock_irq(&ctx->completion_lock);
2732 list_for_each_prev(entry, &ctx->timeout_list) {
2733 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2734 unsigned nxt_sq_head;
2735 long long tmp, tmp_nxt;
2736 u32 nxt_offset = nxt->io->timeout.seq_offset;
2738 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2742 * Since cached_sq_head + count - 1 can overflow, use type long
2745 tmp = (long long)ctx->cached_sq_head + count - 1;
2746 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2747 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2750 * cached_sq_head may overflow, and it will never overflow twice
2751 * once there is some timeout req still be valid.
2753 if (ctx->cached_sq_head < nxt_sq_head)
2760 * Sequence of reqs after the insert one and itself should
2761 * be adjusted because each timeout req consumes a slot.
2766 req->sequence -= span;
2768 list_add(&req->list, entry);
2769 data->timer.function = io_timeout_fn;
2770 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2771 spin_unlock_irq(&ctx->completion_lock);
2775 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2779 return req->user_data == (unsigned long) data;
2782 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2784 enum io_wq_cancel cancel_ret;
2787 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2788 switch (cancel_ret) {
2789 case IO_WQ_CANCEL_OK:
2792 case IO_WQ_CANCEL_RUNNING:
2795 case IO_WQ_CANCEL_NOTFOUND:
2803 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2804 struct io_kiocb *req, __u64 sqe_addr,
2805 struct io_kiocb **nxt, int success_ret)
2807 unsigned long flags;
2810 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2811 if (ret != -ENOENT) {
2812 spin_lock_irqsave(&ctx->completion_lock, flags);
2816 spin_lock_irqsave(&ctx->completion_lock, flags);
2817 ret = io_timeout_cancel(ctx, sqe_addr);
2820 ret = io_poll_cancel(ctx, sqe_addr);
2824 io_cqring_fill_event(req, ret);
2825 io_commit_cqring(ctx);
2826 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2827 io_cqring_ev_posted(ctx);
2829 if (ret < 0 && (req->flags & REQ_F_LINK))
2830 req->flags |= REQ_F_FAIL_LINK;
2831 io_put_req_find_next(req, nxt);
2834 static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2835 struct io_kiocb **nxt)
2837 struct io_ring_ctx *ctx = req->ctx;
2839 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2841 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2845 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
2849 static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io)
2851 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2852 struct iov_iter iter;
2855 memcpy(&io->sqe, req->sqe, sizeof(io->sqe));
2856 req->sqe = &io->sqe;
2858 switch (io->sqe.opcode) {
2859 case IORING_OP_READV:
2860 case IORING_OP_READ_FIXED:
2861 ret = io_read_prep(req, &iovec, &iter, true);
2863 case IORING_OP_WRITEV:
2864 case IORING_OP_WRITE_FIXED:
2865 ret = io_write_prep(req, &iovec, &iter, true);
2867 case IORING_OP_SENDMSG:
2868 ret = io_sendmsg_prep(req, io);
2870 case IORING_OP_RECVMSG:
2871 ret = io_recvmsg_prep(req, io);
2873 case IORING_OP_CONNECT:
2874 ret = io_connect_prep(req, io);
2876 case IORING_OP_TIMEOUT:
2877 return io_timeout_prep(req, io, false);
2878 case IORING_OP_LINK_TIMEOUT:
2879 return io_timeout_prep(req, io, true);
2889 io_req_map_io(req, ret, iovec, inline_vecs, &iter);
2893 static int io_req_defer(struct io_kiocb *req)
2895 struct io_ring_ctx *ctx = req->ctx;
2896 struct io_async_ctx *io;
2899 /* Still need defer if there is pending req in defer list. */
2900 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
2903 io = kmalloc(sizeof(*io), GFP_KERNEL);
2907 ret = io_req_defer_prep(req, io);
2913 spin_lock_irq(&ctx->completion_lock);
2914 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
2915 spin_unlock_irq(&ctx->completion_lock);
2919 trace_io_uring_defer(ctx, req, req->user_data);
2920 list_add_tail(&req->list, &ctx->defer_list);
2921 spin_unlock_irq(&ctx->completion_lock);
2922 return -EIOCBQUEUED;
2925 __attribute__((nonnull))
2926 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
2927 bool force_nonblock)
2930 struct io_ring_ctx *ctx = req->ctx;
2932 opcode = READ_ONCE(req->sqe->opcode);
2937 case IORING_OP_READV:
2938 if (unlikely(req->sqe->buf_index))
2940 ret = io_read(req, nxt, force_nonblock);
2942 case IORING_OP_WRITEV:
2943 if (unlikely(req->sqe->buf_index))
2945 ret = io_write(req, nxt, force_nonblock);
2947 case IORING_OP_READ_FIXED:
2948 ret = io_read(req, nxt, force_nonblock);
2950 case IORING_OP_WRITE_FIXED:
2951 ret = io_write(req, nxt, force_nonblock);
2953 case IORING_OP_FSYNC:
2954 ret = io_fsync(req, req->sqe, nxt, force_nonblock);
2956 case IORING_OP_POLL_ADD:
2957 ret = io_poll_add(req, req->sqe, nxt);
2959 case IORING_OP_POLL_REMOVE:
2960 ret = io_poll_remove(req, req->sqe);
2962 case IORING_OP_SYNC_FILE_RANGE:
2963 ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock);
2965 case IORING_OP_SENDMSG:
2966 ret = io_sendmsg(req, req->sqe, nxt, force_nonblock);
2968 case IORING_OP_RECVMSG:
2969 ret = io_recvmsg(req, req->sqe, nxt, force_nonblock);
2971 case IORING_OP_TIMEOUT:
2972 ret = io_timeout(req, req->sqe);
2974 case IORING_OP_TIMEOUT_REMOVE:
2975 ret = io_timeout_remove(req, req->sqe);
2977 case IORING_OP_ACCEPT:
2978 ret = io_accept(req, req->sqe, nxt, force_nonblock);
2980 case IORING_OP_CONNECT:
2981 ret = io_connect(req, req->sqe, nxt, force_nonblock);
2983 case IORING_OP_ASYNC_CANCEL:
2984 ret = io_async_cancel(req, req->sqe, nxt);
2994 if (ctx->flags & IORING_SETUP_IOPOLL) {
2995 if (req->result == -EAGAIN)
2998 /* workqueue context doesn't hold uring_lock, grab it now */
3000 mutex_lock(&ctx->uring_lock);
3001 io_iopoll_req_issued(req);
3003 mutex_unlock(&ctx->uring_lock);
3009 static void io_link_work_cb(struct io_wq_work **workptr)
3011 struct io_wq_work *work = *workptr;
3012 struct io_kiocb *link = work->data;
3014 io_queue_linked_timeout(link);
3015 work->func = io_wq_submit_work;
3018 static void io_wq_submit_work(struct io_wq_work **workptr)
3020 struct io_wq_work *work = *workptr;
3021 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3022 struct io_kiocb *nxt = NULL;
3025 /* Ensure we clear previously set non-block flag */
3026 req->rw.ki_flags &= ~IOCB_NOWAIT;
3028 if (work->flags & IO_WQ_WORK_CANCEL)
3032 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3033 req->in_async = true;
3035 ret = io_issue_sqe(req, &nxt, false);
3037 * We can get EAGAIN for polled IO even though we're
3038 * forcing a sync submission from here, since we can't
3039 * wait for request slots on the block side.
3047 /* drop submission reference */
3051 if (req->flags & REQ_F_LINK)
3052 req->flags |= REQ_F_FAIL_LINK;
3053 io_cqring_add_event(req, ret);
3057 /* if a dependent link is ready, pass it back */
3059 struct io_kiocb *link;
3061 io_prep_async_work(nxt, &link);
3062 *workptr = &nxt->work;
3064 nxt->work.flags |= IO_WQ_WORK_CB;
3065 nxt->work.func = io_link_work_cb;
3066 nxt->work.data = link;
3071 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
3073 int op = READ_ONCE(sqe->opcode);
3077 case IORING_OP_POLL_REMOVE:
3078 case IORING_OP_TIMEOUT:
3079 case IORING_OP_TIMEOUT_REMOVE:
3080 case IORING_OP_ASYNC_CANCEL:
3081 case IORING_OP_LINK_TIMEOUT:
3088 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3091 struct fixed_file_table *table;
3093 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3094 return table->files[index & IORING_FILE_TABLE_MASK];
3097 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3099 struct io_ring_ctx *ctx = req->ctx;
3103 flags = READ_ONCE(req->sqe->flags);
3104 fd = READ_ONCE(req->sqe->fd);
3106 if (flags & IOSQE_IO_DRAIN)
3107 req->flags |= REQ_F_IO_DRAIN;
3109 if (!io_op_needs_file(req->sqe))
3112 if (flags & IOSQE_FIXED_FILE) {
3113 if (unlikely(!ctx->file_table ||
3114 (unsigned) fd >= ctx->nr_user_files))
3116 fd = array_index_nospec(fd, ctx->nr_user_files);
3117 req->file = io_file_from_index(ctx, fd);
3120 req->flags |= REQ_F_FIXED_FILE;
3122 if (req->needs_fixed_file)
3124 trace_io_uring_file_get(ctx, fd);
3125 req->file = io_file_get(state, fd);
3126 if (unlikely(!req->file))
3133 static int io_grab_files(struct io_kiocb *req)
3136 struct io_ring_ctx *ctx = req->ctx;
3139 spin_lock_irq(&ctx->inflight_lock);
3141 * We use the f_ops->flush() handler to ensure that we can flush
3142 * out work accessing these files if the fd is closed. Check if
3143 * the fd has changed since we started down this path, and disallow
3144 * this operation if it has.
3146 if (fcheck(req->ring_fd) == req->ring_file) {
3147 list_add(&req->inflight_entry, &ctx->inflight_list);
3148 req->flags |= REQ_F_INFLIGHT;
3149 req->work.files = current->files;
3152 spin_unlock_irq(&ctx->inflight_lock);
3158 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3160 struct io_timeout_data *data = container_of(timer,
3161 struct io_timeout_data, timer);
3162 struct io_kiocb *req = data->req;
3163 struct io_ring_ctx *ctx = req->ctx;
3164 struct io_kiocb *prev = NULL;
3165 unsigned long flags;
3167 spin_lock_irqsave(&ctx->completion_lock, flags);
3170 * We don't expect the list to be empty, that will only happen if we
3171 * race with the completion of the linked work.
3173 if (!list_empty(&req->list)) {
3174 prev = list_entry(req->list.prev, struct io_kiocb, link_list);
3175 if (refcount_inc_not_zero(&prev->refs)) {
3176 list_del_init(&req->list);
3177 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3182 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3185 if (prev->flags & REQ_F_LINK)
3186 prev->flags |= REQ_F_FAIL_LINK;
3187 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3191 io_cqring_add_event(req, -ETIME);
3194 return HRTIMER_NORESTART;
3197 static void io_queue_linked_timeout(struct io_kiocb *req)
3199 struct io_ring_ctx *ctx = req->ctx;
3202 * If the list is now empty, then our linked request finished before
3203 * we got a chance to setup the timer
3205 spin_lock_irq(&ctx->completion_lock);
3206 if (!list_empty(&req->list)) {
3207 struct io_timeout_data *data = &req->io->timeout;
3209 data->timer.function = io_link_timeout_fn;
3210 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3213 spin_unlock_irq(&ctx->completion_lock);
3215 /* drop submission reference */
3219 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3221 struct io_kiocb *nxt;
3223 if (!(req->flags & REQ_F_LINK))
3226 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
3227 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
3230 req->flags |= REQ_F_LINK_TIMEOUT;
3234 static void __io_queue_sqe(struct io_kiocb *req)
3236 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
3237 struct io_kiocb *nxt = NULL;
3240 ret = io_issue_sqe(req, &nxt, true);
3242 io_queue_async_work(nxt);
3245 * We async punt it if the file wasn't marked NOWAIT, or if the file
3246 * doesn't support non-blocking read/write attempts
3248 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3249 (req->flags & REQ_F_MUST_PUNT))) {
3250 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3251 ret = io_grab_files(req);
3257 * Queued up for async execution, worker will release
3258 * submit reference when the iocb is actually submitted.
3260 io_queue_async_work(req);
3265 /* drop submission reference */
3268 if (linked_timeout) {
3270 io_queue_linked_timeout(linked_timeout);
3272 io_put_req(linked_timeout);
3275 /* and drop final reference, if we failed */
3277 io_cqring_add_event(req, ret);
3278 if (req->flags & REQ_F_LINK)
3279 req->flags |= REQ_F_FAIL_LINK;
3284 static void io_queue_sqe(struct io_kiocb *req)
3288 if (unlikely(req->ctx->drain_next)) {
3289 req->flags |= REQ_F_IO_DRAIN;
3290 req->ctx->drain_next = false;
3292 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3294 ret = io_req_defer(req);
3296 if (ret != -EIOCBQUEUED) {
3297 io_cqring_add_event(req, ret);
3298 if (req->flags & REQ_F_LINK)
3299 req->flags |= REQ_F_FAIL_LINK;
3300 io_double_put_req(req);
3303 __io_queue_sqe(req);
3306 static inline void io_queue_link_head(struct io_kiocb *req)
3308 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3309 io_cqring_add_event(req, -ECANCELED);
3310 io_double_put_req(req);
3316 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
3318 static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3319 struct io_kiocb **link)
3321 struct io_ring_ctx *ctx = req->ctx;
3324 req->user_data = req->sqe->user_data;
3326 /* enforce forwards compatibility on users */
3327 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3332 ret = io_req_set_file(state, req);
3333 if (unlikely(ret)) {
3335 io_cqring_add_event(req, ret);
3336 io_double_put_req(req);
3341 * If we already have a head request, queue this one for async
3342 * submittal once the head completes. If we don't have a head but
3343 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3344 * submitted sync once the chain is complete. If none of those
3345 * conditions are true (normal request), then just queue it.
3348 struct io_kiocb *prev = *link;
3349 struct io_async_ctx *io;
3351 if (req->sqe->flags & IOSQE_IO_DRAIN)
3352 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3354 io = kmalloc(sizeof(*io), GFP_KERNEL);
3360 ret = io_req_defer_prep(req, io);
3363 prev->flags |= REQ_F_FAIL_LINK;
3366 trace_io_uring_link(ctx, req, prev);
3367 list_add_tail(&req->list, &prev->link_list);
3368 } else if (req->sqe->flags & IOSQE_IO_LINK) {
3369 req->flags |= REQ_F_LINK;
3371 INIT_LIST_HEAD(&req->link_list);
3381 * Batched submission is done, ensure local IO is flushed out.
3383 static void io_submit_state_end(struct io_submit_state *state)
3385 blk_finish_plug(&state->plug);
3387 if (state->free_reqs)
3388 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3389 &state->reqs[state->cur_req]);
3393 * Start submission side cache.
3395 static void io_submit_state_start(struct io_submit_state *state,
3396 unsigned int max_ios)
3398 blk_start_plug(&state->plug);
3399 state->free_reqs = 0;
3401 state->ios_left = max_ios;
3404 static void io_commit_sqring(struct io_ring_ctx *ctx)
3406 struct io_rings *rings = ctx->rings;
3408 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3410 * Ensure any loads from the SQEs are done at this point,
3411 * since once we write the new head, the application could
3412 * write new data to them.
3414 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3419 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
3420 * that is mapped by userspace. This means that care needs to be taken to
3421 * ensure that reads are stable, as we cannot rely on userspace always
3422 * being a good citizen. If members of the sqe are validated and then later
3423 * used, it's important that those reads are done through READ_ONCE() to
3424 * prevent a re-load down the line.
3426 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3428 struct io_rings *rings = ctx->rings;
3429 u32 *sq_array = ctx->sq_array;
3433 * The cached sq head (or cq tail) serves two purposes:
3435 * 1) allows us to batch the cost of updating the user visible
3437 * 2) allows the kernel side to track the head on its own, even
3438 * though the application is the one updating it.
3440 head = ctx->cached_sq_head;
3441 /* make sure SQ entry isn't read before tail */
3442 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3445 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3446 if (likely(head < ctx->sq_entries)) {
3448 * All io need record the previous position, if LINK vs DARIN,
3449 * it can be used to mark the position of the first IO in the
3452 req->sequence = ctx->cached_sq_head;
3453 req->sqe = &ctx->sq_sqes[head];
3454 ctx->cached_sq_head++;
3458 /* drop invalid entries */
3459 ctx->cached_sq_head++;
3460 ctx->cached_sq_dropped++;
3461 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3465 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3466 struct file *ring_file, int ring_fd,
3467 struct mm_struct **mm, bool async)
3469 struct io_submit_state state, *statep = NULL;
3470 struct io_kiocb *link = NULL;
3471 int i, submitted = 0;
3472 bool mm_fault = false;
3474 /* if we have a backlog and couldn't flush it all, return BUSY */
3475 if (!list_empty(&ctx->cq_overflow_list) &&
3476 !io_cqring_overflow_flush(ctx, false))
3479 if (nr > IO_PLUG_THRESHOLD) {
3480 io_submit_state_start(&state, nr);
3484 for (i = 0; i < nr; i++) {
3485 struct io_kiocb *req;
3486 unsigned int sqe_flags;
3488 req = io_get_req(ctx, statep);
3489 if (unlikely(!req)) {
3491 submitted = -EAGAIN;
3494 if (!io_get_sqring(ctx, req)) {
3499 if (io_sqe_needs_user(req->sqe) && !*mm) {
3500 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3502 use_mm(ctx->sqo_mm);
3508 sqe_flags = req->sqe->flags;
3510 req->ring_file = ring_file;
3511 req->ring_fd = ring_fd;
3512 req->has_user = *mm != NULL;
3513 req->in_async = async;
3514 req->needs_fixed_file = async;
3515 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
3517 if (!io_submit_sqe(req, statep, &link))
3520 * If previous wasn't linked and we have a linked command,
3521 * that's the end of the chain. Submit the previous link.
3523 if (!(sqe_flags & IOSQE_IO_LINK) && link) {
3524 io_queue_link_head(link);
3530 io_queue_link_head(link);
3532 io_submit_state_end(&state);
3534 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3535 io_commit_sqring(ctx);
3540 static int io_sq_thread(void *data)
3542 struct io_ring_ctx *ctx = data;
3543 struct mm_struct *cur_mm = NULL;
3544 const struct cred *old_cred;
3545 mm_segment_t old_fs;
3548 unsigned long timeout;
3551 complete(&ctx->completions[1]);
3555 old_cred = override_creds(ctx->creds);
3557 ret = timeout = inflight = 0;
3558 while (!kthread_should_park()) {
3559 unsigned int to_submit;
3562 unsigned nr_events = 0;
3564 if (ctx->flags & IORING_SETUP_IOPOLL) {
3566 * inflight is the count of the maximum possible
3567 * entries we submitted, but it can be smaller
3568 * if we dropped some of them. If we don't have
3569 * poll entries available, then we know that we
3570 * have nothing left to poll for. Reset the
3571 * inflight count to zero in that case.
3573 mutex_lock(&ctx->uring_lock);
3574 if (!list_empty(&ctx->poll_list))
3575 __io_iopoll_check(ctx, &nr_events, 0);
3578 mutex_unlock(&ctx->uring_lock);
3581 * Normal IO, just pretend everything completed.
3582 * We don't have to poll completions for that.
3584 nr_events = inflight;
3587 inflight -= nr_events;
3589 timeout = jiffies + ctx->sq_thread_idle;
3592 to_submit = io_sqring_entries(ctx);
3595 * If submit got -EBUSY, flag us as needing the application
3596 * to enter the kernel to reap and flush events.
3598 if (!to_submit || ret == -EBUSY) {
3600 * We're polling. If we're within the defined idle
3601 * period, then let us spin without work before going
3602 * to sleep. The exception is if we got EBUSY doing
3603 * more IO, we should wait for the application to
3604 * reap events and wake us up.
3607 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3613 * Drop cur_mm before scheduling, we can't hold it for
3614 * long periods (or over schedule()). Do this before
3615 * adding ourselves to the waitqueue, as the unuse/drop
3624 prepare_to_wait(&ctx->sqo_wait, &wait,
3625 TASK_INTERRUPTIBLE);
3627 /* Tell userspace we may need a wakeup call */
3628 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3629 /* make sure to read SQ tail after writing flags */
3632 to_submit = io_sqring_entries(ctx);
3633 if (!to_submit || ret == -EBUSY) {
3634 if (kthread_should_park()) {
3635 finish_wait(&ctx->sqo_wait, &wait);
3638 if (signal_pending(current))
3639 flush_signals(current);
3641 finish_wait(&ctx->sqo_wait, &wait);
3643 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3646 finish_wait(&ctx->sqo_wait, &wait);
3648 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3651 to_submit = min(to_submit, ctx->sq_entries);
3652 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3662 revert_creds(old_cred);
3669 struct io_wait_queue {
3670 struct wait_queue_entry wq;
3671 struct io_ring_ctx *ctx;
3673 unsigned nr_timeouts;
3676 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3678 struct io_ring_ctx *ctx = iowq->ctx;
3681 * Wake up if we have enough events, or if a timeout occured since we
3682 * started waiting. For timeouts, we always want to return to userspace,
3683 * regardless of event count.
3685 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3686 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3689 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3690 int wake_flags, void *key)
3692 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3695 /* use noflush == true, as we can't safely rely on locking context */
3696 if (!io_should_wake(iowq, true))
3699 return autoremove_wake_function(curr, mode, wake_flags, key);
3703 * Wait until events become available, if we don't already have some. The
3704 * application must reap them itself, as they reside on the shared cq ring.
3706 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3707 const sigset_t __user *sig, size_t sigsz)
3709 struct io_wait_queue iowq = {
3712 .func = io_wake_function,
3713 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3716 .to_wait = min_events,
3718 struct io_rings *rings = ctx->rings;
3721 if (io_cqring_events(ctx, false) >= min_events)
3725 #ifdef CONFIG_COMPAT
3726 if (in_compat_syscall())
3727 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3731 ret = set_user_sigmask(sig, sigsz);
3737 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
3738 trace_io_uring_cqring_wait(ctx, min_events);
3740 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3741 TASK_INTERRUPTIBLE);
3742 if (io_should_wake(&iowq, false))
3745 if (signal_pending(current)) {
3750 finish_wait(&ctx->wait, &iowq.wq);
3752 restore_saved_sigmask_unless(ret == -EINTR);
3754 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
3757 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3759 #if defined(CONFIG_UNIX)
3760 if (ctx->ring_sock) {
3761 struct sock *sock = ctx->ring_sock->sk;
3762 struct sk_buff *skb;
3764 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3770 for (i = 0; i < ctx->nr_user_files; i++) {
3773 file = io_file_from_index(ctx, i);
3780 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3782 unsigned nr_tables, i;
3784 if (!ctx->file_table)
3787 __io_sqe_files_unregister(ctx);
3788 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3789 for (i = 0; i < nr_tables; i++)
3790 kfree(ctx->file_table[i].files);
3791 kfree(ctx->file_table);
3792 ctx->file_table = NULL;
3793 ctx->nr_user_files = 0;
3797 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3799 if (ctx->sqo_thread) {
3800 wait_for_completion(&ctx->completions[1]);
3802 * The park is a bit of a work-around, without it we get
3803 * warning spews on shutdown with SQPOLL set and affinity
3804 * set to a single CPU.
3806 kthread_park(ctx->sqo_thread);
3807 kthread_stop(ctx->sqo_thread);
3808 ctx->sqo_thread = NULL;
3812 static void io_finish_async(struct io_ring_ctx *ctx)
3814 io_sq_thread_stop(ctx);
3817 io_wq_destroy(ctx->io_wq);
3822 #if defined(CONFIG_UNIX)
3823 static void io_destruct_skb(struct sk_buff *skb)
3825 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3828 io_wq_flush(ctx->io_wq);
3830 unix_destruct_scm(skb);
3834 * Ensure the UNIX gc is aware of our file set, so we are certain that
3835 * the io_uring can be safely unregistered on process exit, even if we have
3836 * loops in the file referencing.
3838 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3840 struct sock *sk = ctx->ring_sock->sk;
3841 struct scm_fp_list *fpl;
3842 struct sk_buff *skb;
3845 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3846 unsigned long inflight = ctx->user->unix_inflight + nr;
3848 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3852 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3856 skb = alloc_skb(0, GFP_KERNEL);
3865 fpl->user = get_uid(ctx->user);
3866 for (i = 0; i < nr; i++) {
3867 struct file *file = io_file_from_index(ctx, i + offset);
3871 fpl->fp[nr_files] = get_file(file);
3872 unix_inflight(fpl->user, fpl->fp[nr_files]);
3877 fpl->max = SCM_MAX_FD;
3878 fpl->count = nr_files;
3879 UNIXCB(skb).fp = fpl;
3880 skb->destructor = io_destruct_skb;
3881 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3882 skb_queue_head(&sk->sk_receive_queue, skb);
3884 for (i = 0; i < nr_files; i++)
3895 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3896 * causes regular reference counting to break down. We rely on the UNIX
3897 * garbage collection to take care of this problem for us.
3899 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3901 unsigned left, total;
3905 left = ctx->nr_user_files;
3907 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
3909 ret = __io_sqe_files_scm(ctx, this_files, total);
3913 total += this_files;
3919 while (total < ctx->nr_user_files) {
3920 struct file *file = io_file_from_index(ctx, total);
3930 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3936 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3941 for (i = 0; i < nr_tables; i++) {
3942 struct fixed_file_table *table = &ctx->file_table[i];
3943 unsigned this_files;
3945 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3946 table->files = kcalloc(this_files, sizeof(struct file *),
3950 nr_files -= this_files;
3956 for (i = 0; i < nr_tables; i++) {
3957 struct fixed_file_table *table = &ctx->file_table[i];
3958 kfree(table->files);
3963 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3966 __s32 __user *fds = (__s32 __user *) arg;
3971 if (ctx->file_table)
3975 if (nr_args > IORING_MAX_FIXED_FILES)
3978 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
3979 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
3981 if (!ctx->file_table)
3984 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
3985 kfree(ctx->file_table);
3986 ctx->file_table = NULL;
3990 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
3991 struct fixed_file_table *table;
3995 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3997 /* allow sparse sets */
4003 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4004 index = i & IORING_FILE_TABLE_MASK;
4005 table->files[index] = fget(fd);
4008 if (!table->files[index])
4011 * Don't allow io_uring instances to be registered. If UNIX
4012 * isn't enabled, then this causes a reference cycle and this
4013 * instance can never get freed. If UNIX is enabled we'll
4014 * handle it just fine, but there's still no point in allowing
4015 * a ring fd as it doesn't support regular read/write anyway.
4017 if (table->files[index]->f_op == &io_uring_fops) {
4018 fput(table->files[index]);
4025 for (i = 0; i < ctx->nr_user_files; i++) {
4028 file = io_file_from_index(ctx, i);
4032 for (i = 0; i < nr_tables; i++)
4033 kfree(ctx->file_table[i].files);
4035 kfree(ctx->file_table);
4036 ctx->file_table = NULL;
4037 ctx->nr_user_files = 0;
4041 ret = io_sqe_files_scm(ctx);
4043 io_sqe_files_unregister(ctx);
4048 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4050 #if defined(CONFIG_UNIX)
4051 struct file *file = io_file_from_index(ctx, index);
4052 struct sock *sock = ctx->ring_sock->sk;
4053 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4054 struct sk_buff *skb;
4057 __skb_queue_head_init(&list);
4060 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4061 * remove this entry and rearrange the file array.
4063 skb = skb_dequeue(head);
4065 struct scm_fp_list *fp;
4067 fp = UNIXCB(skb).fp;
4068 for (i = 0; i < fp->count; i++) {
4071 if (fp->fp[i] != file)
4074 unix_notinflight(fp->user, fp->fp[i]);
4075 left = fp->count - 1 - i;
4077 memmove(&fp->fp[i], &fp->fp[i + 1],
4078 left * sizeof(struct file *));
4085 __skb_queue_tail(&list, skb);
4095 __skb_queue_tail(&list, skb);
4097 skb = skb_dequeue(head);
4100 if (skb_peek(&list)) {
4101 spin_lock_irq(&head->lock);
4102 while ((skb = __skb_dequeue(&list)) != NULL)
4103 __skb_queue_tail(head, skb);
4104 spin_unlock_irq(&head->lock);
4107 fput(io_file_from_index(ctx, index));
4111 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4114 #if defined(CONFIG_UNIX)
4115 struct sock *sock = ctx->ring_sock->sk;
4116 struct sk_buff_head *head = &sock->sk_receive_queue;
4117 struct sk_buff *skb;
4120 * See if we can merge this file into an existing skb SCM_RIGHTS
4121 * file set. If there's no room, fall back to allocating a new skb
4122 * and filling it in.
4124 spin_lock_irq(&head->lock);
4125 skb = skb_peek(head);
4127 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4129 if (fpl->count < SCM_MAX_FD) {
4130 __skb_unlink(skb, head);
4131 spin_unlock_irq(&head->lock);
4132 fpl->fp[fpl->count] = get_file(file);
4133 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4135 spin_lock_irq(&head->lock);
4136 __skb_queue_head(head, skb);
4141 spin_unlock_irq(&head->lock);
4148 return __io_sqe_files_scm(ctx, 1, index);
4154 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4157 struct io_uring_files_update up;
4162 if (!ctx->file_table)
4166 if (copy_from_user(&up, arg, sizeof(up)))
4168 if (check_add_overflow(up.offset, nr_args, &done))
4170 if (done > ctx->nr_user_files)
4174 fds = (__s32 __user *) up.fds;
4176 struct fixed_file_table *table;
4180 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4184 i = array_index_nospec(up.offset, ctx->nr_user_files);
4185 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4186 index = i & IORING_FILE_TABLE_MASK;
4187 if (table->files[index]) {
4188 io_sqe_file_unregister(ctx, i);
4189 table->files[index] = NULL;
4200 * Don't allow io_uring instances to be registered. If
4201 * UNIX isn't enabled, then this causes a reference
4202 * cycle and this instance can never get freed. If UNIX
4203 * is enabled we'll handle it just fine, but there's
4204 * still no point in allowing a ring fd as it doesn't
4205 * support regular read/write anyway.
4207 if (file->f_op == &io_uring_fops) {
4212 table->files[index] = file;
4213 err = io_sqe_file_register(ctx, file, i);
4222 return done ? done : err;
4225 static void io_put_work(struct io_wq_work *work)
4227 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4232 static void io_get_work(struct io_wq_work *work)
4234 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4236 refcount_inc(&req->refs);
4239 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4240 struct io_uring_params *p)
4242 struct io_wq_data data;
4243 unsigned concurrency;
4246 init_waitqueue_head(&ctx->sqo_wait);
4247 mmgrab(current->mm);
4248 ctx->sqo_mm = current->mm;
4250 if (ctx->flags & IORING_SETUP_SQPOLL) {
4252 if (!capable(CAP_SYS_ADMIN))
4255 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4256 if (!ctx->sq_thread_idle)
4257 ctx->sq_thread_idle = HZ;
4259 if (p->flags & IORING_SETUP_SQ_AFF) {
4260 int cpu = p->sq_thread_cpu;
4263 if (cpu >= nr_cpu_ids)
4265 if (!cpu_online(cpu))
4268 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4272 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4275 if (IS_ERR(ctx->sqo_thread)) {
4276 ret = PTR_ERR(ctx->sqo_thread);
4277 ctx->sqo_thread = NULL;
4280 wake_up_process(ctx->sqo_thread);
4281 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4282 /* Can't have SQ_AFF without SQPOLL */
4287 data.mm = ctx->sqo_mm;
4288 data.user = ctx->user;
4289 data.creds = ctx->creds;
4290 data.get_work = io_get_work;
4291 data.put_work = io_put_work;
4293 /* Do QD, or 4 * CPUS, whatever is smallest */
4294 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4295 ctx->io_wq = io_wq_create(concurrency, &data);
4296 if (IS_ERR(ctx->io_wq)) {
4297 ret = PTR_ERR(ctx->io_wq);
4304 io_finish_async(ctx);
4305 mmdrop(ctx->sqo_mm);
4310 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4312 atomic_long_sub(nr_pages, &user->locked_vm);
4315 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4317 unsigned long page_limit, cur_pages, new_pages;
4319 /* Don't allow more pages than we can safely lock */
4320 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4323 cur_pages = atomic_long_read(&user->locked_vm);
4324 new_pages = cur_pages + nr_pages;
4325 if (new_pages > page_limit)
4327 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4328 new_pages) != cur_pages);
4333 static void io_mem_free(void *ptr)
4340 page = virt_to_head_page(ptr);
4341 if (put_page_testzero(page))
4342 free_compound_page(page);
4345 static void *io_mem_alloc(size_t size)
4347 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4350 return (void *) __get_free_pages(gfp_flags, get_order(size));
4353 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4356 struct io_rings *rings;
4357 size_t off, sq_array_size;
4359 off = struct_size(rings, cqes, cq_entries);
4360 if (off == SIZE_MAX)
4364 off = ALIGN(off, SMP_CACHE_BYTES);
4369 sq_array_size = array_size(sizeof(u32), sq_entries);
4370 if (sq_array_size == SIZE_MAX)
4373 if (check_add_overflow(off, sq_array_size, &off))
4382 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4386 pages = (size_t)1 << get_order(
4387 rings_size(sq_entries, cq_entries, NULL));
4388 pages += (size_t)1 << get_order(
4389 array_size(sizeof(struct io_uring_sqe), sq_entries));
4394 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4398 if (!ctx->user_bufs)
4401 for (i = 0; i < ctx->nr_user_bufs; i++) {
4402 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4404 for (j = 0; j < imu->nr_bvecs; j++)
4405 put_user_page(imu->bvec[j].bv_page);
4407 if (ctx->account_mem)
4408 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4413 kfree(ctx->user_bufs);
4414 ctx->user_bufs = NULL;
4415 ctx->nr_user_bufs = 0;
4419 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4420 void __user *arg, unsigned index)
4422 struct iovec __user *src;
4424 #ifdef CONFIG_COMPAT
4426 struct compat_iovec __user *ciovs;
4427 struct compat_iovec ciov;
4429 ciovs = (struct compat_iovec __user *) arg;
4430 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4433 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4434 dst->iov_len = ciov.iov_len;
4438 src = (struct iovec __user *) arg;
4439 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4444 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4447 struct vm_area_struct **vmas = NULL;
4448 struct page **pages = NULL;
4449 int i, j, got_pages = 0;
4454 if (!nr_args || nr_args > UIO_MAXIOV)
4457 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4459 if (!ctx->user_bufs)
4462 for (i = 0; i < nr_args; i++) {
4463 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4464 unsigned long off, start, end, ubuf;
4469 ret = io_copy_iov(ctx, &iov, arg, i);
4474 * Don't impose further limits on the size and buffer
4475 * constraints here, we'll -EINVAL later when IO is
4476 * submitted if they are wrong.
4479 if (!iov.iov_base || !iov.iov_len)
4482 /* arbitrary limit, but we need something */
4483 if (iov.iov_len > SZ_1G)
4486 ubuf = (unsigned long) iov.iov_base;
4487 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4488 start = ubuf >> PAGE_SHIFT;
4489 nr_pages = end - start;
4491 if (ctx->account_mem) {
4492 ret = io_account_mem(ctx->user, nr_pages);
4498 if (!pages || nr_pages > got_pages) {
4501 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4503 vmas = kvmalloc_array(nr_pages,
4504 sizeof(struct vm_area_struct *),
4506 if (!pages || !vmas) {
4508 if (ctx->account_mem)
4509 io_unaccount_mem(ctx->user, nr_pages);
4512 got_pages = nr_pages;
4515 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4519 if (ctx->account_mem)
4520 io_unaccount_mem(ctx->user, nr_pages);
4525 down_read(¤t->mm->mmap_sem);
4526 pret = get_user_pages(ubuf, nr_pages,
4527 FOLL_WRITE | FOLL_LONGTERM,
4529 if (pret == nr_pages) {
4530 /* don't support file backed memory */
4531 for (j = 0; j < nr_pages; j++) {
4532 struct vm_area_struct *vma = vmas[j];
4535 !is_file_hugepages(vma->vm_file)) {
4541 ret = pret < 0 ? pret : -EFAULT;
4543 up_read(¤t->mm->mmap_sem);
4546 * if we did partial map, or found file backed vmas,
4547 * release any pages we did get
4550 put_user_pages(pages, pret);
4551 if (ctx->account_mem)
4552 io_unaccount_mem(ctx->user, nr_pages);
4557 off = ubuf & ~PAGE_MASK;
4559 for (j = 0; j < nr_pages; j++) {
4562 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4563 imu->bvec[j].bv_page = pages[j];
4564 imu->bvec[j].bv_len = vec_len;
4565 imu->bvec[j].bv_offset = off;
4569 /* store original address for later verification */
4571 imu->len = iov.iov_len;
4572 imu->nr_bvecs = nr_pages;
4574 ctx->nr_user_bufs++;
4582 io_sqe_buffer_unregister(ctx);
4586 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4588 __s32 __user *fds = arg;
4594 if (copy_from_user(&fd, fds, sizeof(*fds)))
4597 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4598 if (IS_ERR(ctx->cq_ev_fd)) {
4599 int ret = PTR_ERR(ctx->cq_ev_fd);
4600 ctx->cq_ev_fd = NULL;
4607 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4609 if (ctx->cq_ev_fd) {
4610 eventfd_ctx_put(ctx->cq_ev_fd);
4611 ctx->cq_ev_fd = NULL;
4618 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4620 io_finish_async(ctx);
4622 mmdrop(ctx->sqo_mm);
4624 io_iopoll_reap_events(ctx);
4625 io_sqe_buffer_unregister(ctx);
4626 io_sqe_files_unregister(ctx);
4627 io_eventfd_unregister(ctx);
4629 #if defined(CONFIG_UNIX)
4630 if (ctx->ring_sock) {
4631 ctx->ring_sock->file = NULL; /* so that iput() is called */
4632 sock_release(ctx->ring_sock);
4636 io_mem_free(ctx->rings);
4637 io_mem_free(ctx->sq_sqes);
4639 percpu_ref_exit(&ctx->refs);
4640 if (ctx->account_mem)
4641 io_unaccount_mem(ctx->user,
4642 ring_pages(ctx->sq_entries, ctx->cq_entries));
4643 free_uid(ctx->user);
4644 put_cred(ctx->creds);
4645 kfree(ctx->completions);
4646 kfree(ctx->cancel_hash);
4647 kmem_cache_free(req_cachep, ctx->fallback_req);
4651 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4653 struct io_ring_ctx *ctx = file->private_data;
4656 poll_wait(file, &ctx->cq_wait, wait);
4658 * synchronizes with barrier from wq_has_sleeper call in
4662 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4663 ctx->rings->sq_ring_entries)
4664 mask |= EPOLLOUT | EPOLLWRNORM;
4665 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4666 mask |= EPOLLIN | EPOLLRDNORM;
4671 static int io_uring_fasync(int fd, struct file *file, int on)
4673 struct io_ring_ctx *ctx = file->private_data;
4675 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4678 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4680 mutex_lock(&ctx->uring_lock);
4681 percpu_ref_kill(&ctx->refs);
4682 mutex_unlock(&ctx->uring_lock);
4684 io_kill_timeouts(ctx);
4685 io_poll_remove_all(ctx);
4688 io_wq_cancel_all(ctx->io_wq);
4690 io_iopoll_reap_events(ctx);
4691 /* if we failed setting up the ctx, we might not have any rings */
4693 io_cqring_overflow_flush(ctx, true);
4694 wait_for_completion(&ctx->completions[0]);
4695 io_ring_ctx_free(ctx);
4698 static int io_uring_release(struct inode *inode, struct file *file)
4700 struct io_ring_ctx *ctx = file->private_data;
4702 file->private_data = NULL;
4703 io_ring_ctx_wait_and_kill(ctx);
4707 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4708 struct files_struct *files)
4710 struct io_kiocb *req;
4713 while (!list_empty_careful(&ctx->inflight_list)) {
4714 struct io_kiocb *cancel_req = NULL;
4716 spin_lock_irq(&ctx->inflight_lock);
4717 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4718 if (req->work.files != files)
4720 /* req is being completed, ignore */
4721 if (!refcount_inc_not_zero(&req->refs))
4727 prepare_to_wait(&ctx->inflight_wait, &wait,
4728 TASK_UNINTERRUPTIBLE);
4729 spin_unlock_irq(&ctx->inflight_lock);
4731 /* We need to keep going until we don't find a matching req */
4735 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4736 io_put_req(cancel_req);
4739 finish_wait(&ctx->inflight_wait, &wait);
4742 static int io_uring_flush(struct file *file, void *data)
4744 struct io_ring_ctx *ctx = file->private_data;
4746 io_uring_cancel_files(ctx, data);
4747 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4748 io_cqring_overflow_flush(ctx, true);
4749 io_wq_cancel_all(ctx->io_wq);
4754 static void *io_uring_validate_mmap_request(struct file *file,
4755 loff_t pgoff, size_t sz)
4757 struct io_ring_ctx *ctx = file->private_data;
4758 loff_t offset = pgoff << PAGE_SHIFT;
4763 case IORING_OFF_SQ_RING:
4764 case IORING_OFF_CQ_RING:
4767 case IORING_OFF_SQES:
4771 return ERR_PTR(-EINVAL);
4774 page = virt_to_head_page(ptr);
4775 if (sz > page_size(page))
4776 return ERR_PTR(-EINVAL);
4783 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4785 size_t sz = vma->vm_end - vma->vm_start;
4789 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4791 return PTR_ERR(ptr);
4793 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4794 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4797 #else /* !CONFIG_MMU */
4799 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4801 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4804 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4806 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4809 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4810 unsigned long addr, unsigned long len,
4811 unsigned long pgoff, unsigned long flags)
4815 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4817 return PTR_ERR(ptr);
4819 return (unsigned long) ptr;
4822 #endif /* !CONFIG_MMU */
4824 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
4825 u32, min_complete, u32, flags, const sigset_t __user *, sig,
4828 struct io_ring_ctx *ctx;
4833 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
4841 if (f.file->f_op != &io_uring_fops)
4845 ctx = f.file->private_data;
4846 if (!percpu_ref_tryget(&ctx->refs))
4850 * For SQ polling, the thread will do all submissions and completions.
4851 * Just return the requested submit count, and wake the thread if
4855 if (ctx->flags & IORING_SETUP_SQPOLL) {
4856 if (!list_empty_careful(&ctx->cq_overflow_list))
4857 io_cqring_overflow_flush(ctx, false);
4858 if (flags & IORING_ENTER_SQ_WAKEUP)
4859 wake_up(&ctx->sqo_wait);
4860 submitted = to_submit;
4861 } else if (to_submit) {
4862 struct mm_struct *cur_mm;
4864 to_submit = min(to_submit, ctx->sq_entries);
4865 mutex_lock(&ctx->uring_lock);
4866 /* already have mm, so io_submit_sqes() won't try to grab it */
4867 cur_mm = ctx->sqo_mm;
4868 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4870 mutex_unlock(&ctx->uring_lock);
4872 if (flags & IORING_ENTER_GETEVENTS) {
4873 unsigned nr_events = 0;
4875 min_complete = min(min_complete, ctx->cq_entries);
4877 if (ctx->flags & IORING_SETUP_IOPOLL) {
4878 ret = io_iopoll_check(ctx, &nr_events, min_complete);
4880 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4884 percpu_ref_put(&ctx->refs);
4887 return submitted ? submitted : ret;
4890 static const struct file_operations io_uring_fops = {
4891 .release = io_uring_release,
4892 .flush = io_uring_flush,
4893 .mmap = io_uring_mmap,
4895 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
4896 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
4898 .poll = io_uring_poll,
4899 .fasync = io_uring_fasync,
4902 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4903 struct io_uring_params *p)
4905 struct io_rings *rings;
4906 size_t size, sq_array_offset;
4908 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4909 if (size == SIZE_MAX)
4912 rings = io_mem_alloc(size);
4917 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4918 rings->sq_ring_mask = p->sq_entries - 1;
4919 rings->cq_ring_mask = p->cq_entries - 1;
4920 rings->sq_ring_entries = p->sq_entries;
4921 rings->cq_ring_entries = p->cq_entries;
4922 ctx->sq_mask = rings->sq_ring_mask;
4923 ctx->cq_mask = rings->cq_ring_mask;
4924 ctx->sq_entries = rings->sq_ring_entries;
4925 ctx->cq_entries = rings->cq_ring_entries;
4927 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
4928 if (size == SIZE_MAX) {
4929 io_mem_free(ctx->rings);
4934 ctx->sq_sqes = io_mem_alloc(size);
4935 if (!ctx->sq_sqes) {
4936 io_mem_free(ctx->rings);
4945 * Allocate an anonymous fd, this is what constitutes the application
4946 * visible backing of an io_uring instance. The application mmaps this
4947 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4948 * we have to tie this fd to a socket for file garbage collection purposes.
4950 static int io_uring_get_fd(struct io_ring_ctx *ctx)
4955 #if defined(CONFIG_UNIX)
4956 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4962 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4966 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4967 O_RDWR | O_CLOEXEC);
4970 ret = PTR_ERR(file);
4974 #if defined(CONFIG_UNIX)
4975 ctx->ring_sock->file = file;
4976 ctx->ring_sock->sk->sk_user_data = ctx;
4978 fd_install(ret, file);
4981 #if defined(CONFIG_UNIX)
4982 sock_release(ctx->ring_sock);
4983 ctx->ring_sock = NULL;
4988 static int io_uring_create(unsigned entries, struct io_uring_params *p)
4990 struct user_struct *user = NULL;
4991 struct io_ring_ctx *ctx;
4995 if (!entries || entries > IORING_MAX_ENTRIES)
4999 * Use twice as many entries for the CQ ring. It's possible for the
5000 * application to drive a higher depth than the size of the SQ ring,
5001 * since the sqes are only used at submission time. This allows for
5002 * some flexibility in overcommitting a bit. If the application has
5003 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5004 * of CQ ring entries manually.
5006 p->sq_entries = roundup_pow_of_two(entries);
5007 if (p->flags & IORING_SETUP_CQSIZE) {
5009 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5010 * to a power-of-two, if it isn't already. We do NOT impose
5011 * any cq vs sq ring sizing.
5013 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5015 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5017 p->cq_entries = 2 * p->sq_entries;
5020 user = get_uid(current_user());
5021 account_mem = !capable(CAP_IPC_LOCK);
5024 ret = io_account_mem(user,
5025 ring_pages(p->sq_entries, p->cq_entries));
5032 ctx = io_ring_ctx_alloc(p);
5035 io_unaccount_mem(user, ring_pages(p->sq_entries,
5040 ctx->compat = in_compat_syscall();
5041 ctx->account_mem = account_mem;
5043 ctx->creds = get_current_cred();
5045 ret = io_allocate_scq_urings(ctx, p);
5049 ret = io_sq_offload_start(ctx, p);
5053 memset(&p->sq_off, 0, sizeof(p->sq_off));
5054 p->sq_off.head = offsetof(struct io_rings, sq.head);
5055 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5056 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5057 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5058 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5059 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5060 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5062 memset(&p->cq_off, 0, sizeof(p->cq_off));
5063 p->cq_off.head = offsetof(struct io_rings, cq.head);
5064 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5065 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5066 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5067 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5068 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5071 * Install ring fd as the very last thing, so we don't risk someone
5072 * having closed it before we finish setup
5074 ret = io_uring_get_fd(ctx);
5078 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5079 IORING_FEAT_SUBMIT_STABLE;
5080 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5083 io_ring_ctx_wait_and_kill(ctx);
5088 * Sets up an aio uring context, and returns the fd. Applications asks for a
5089 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5090 * params structure passed in.
5092 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5094 struct io_uring_params p;
5098 if (copy_from_user(&p, params, sizeof(p)))
5100 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5105 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5106 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5109 ret = io_uring_create(entries, &p);
5113 if (copy_to_user(params, &p, sizeof(p)))
5119 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5120 struct io_uring_params __user *, params)
5122 return io_uring_setup(entries, params);
5125 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5126 void __user *arg, unsigned nr_args)
5127 __releases(ctx->uring_lock)
5128 __acquires(ctx->uring_lock)
5133 * We're inside the ring mutex, if the ref is already dying, then
5134 * someone else killed the ctx or is already going through
5135 * io_uring_register().
5137 if (percpu_ref_is_dying(&ctx->refs))
5140 percpu_ref_kill(&ctx->refs);
5143 * Drop uring mutex before waiting for references to exit. If another
5144 * thread is currently inside io_uring_enter() it might need to grab
5145 * the uring_lock to make progress. If we hold it here across the drain
5146 * wait, then we can deadlock. It's safe to drop the mutex here, since
5147 * no new references will come in after we've killed the percpu ref.
5149 mutex_unlock(&ctx->uring_lock);
5150 wait_for_completion(&ctx->completions[0]);
5151 mutex_lock(&ctx->uring_lock);
5154 case IORING_REGISTER_BUFFERS:
5155 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5157 case IORING_UNREGISTER_BUFFERS:
5161 ret = io_sqe_buffer_unregister(ctx);
5163 case IORING_REGISTER_FILES:
5164 ret = io_sqe_files_register(ctx, arg, nr_args);
5166 case IORING_UNREGISTER_FILES:
5170 ret = io_sqe_files_unregister(ctx);
5172 case IORING_REGISTER_FILES_UPDATE:
5173 ret = io_sqe_files_update(ctx, arg, nr_args);
5175 case IORING_REGISTER_EVENTFD:
5179 ret = io_eventfd_register(ctx, arg);
5181 case IORING_UNREGISTER_EVENTFD:
5185 ret = io_eventfd_unregister(ctx);
5192 /* bring the ctx back to life */
5193 reinit_completion(&ctx->completions[0]);
5194 percpu_ref_reinit(&ctx->refs);
5198 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5199 void __user *, arg, unsigned int, nr_args)
5201 struct io_ring_ctx *ctx;
5210 if (f.file->f_op != &io_uring_fops)
5213 ctx = f.file->private_data;
5215 mutex_lock(&ctx->uring_lock);
5216 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5217 mutex_unlock(&ctx->uring_lock);
5218 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5219 ctx->cq_ev_fd != NULL, ret);
5225 static int __init io_uring_init(void)
5227 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5230 __initcall(io_uring_init);