1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/kthread.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
74 #define CREATE_TRACE_POINTS
75 #include <trace/events/io_uring.h>
77 #include <uapi/linux/io_uring.h>
82 #define IORING_MAX_ENTRIES 32768
83 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
86 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
88 #define IORING_FILE_TABLE_SHIFT 9
89 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
90 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
91 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
94 u32 head ____cacheline_aligned_in_smp;
95 u32 tail ____cacheline_aligned_in_smp;
99 * This data is shared with the application through the mmap at offsets
100 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
102 * The offsets to the member fields are published through struct
103 * io_sqring_offsets when calling io_uring_setup.
107 * Head and tail offsets into the ring; the offsets need to be
108 * masked to get valid indices.
110 * The kernel controls head of the sq ring and the tail of the cq ring,
111 * and the application controls tail of the sq ring and the head of the
114 struct io_uring sq, cq;
116 * Bitmasks to apply to head and tail offsets (constant, equals
119 u32 sq_ring_mask, cq_ring_mask;
120 /* Ring sizes (constant, power of 2) */
121 u32 sq_ring_entries, cq_ring_entries;
123 * Number of invalid entries dropped by the kernel due to
124 * invalid index stored in array
126 * Written by the kernel, shouldn't be modified by the
127 * application (i.e. get number of "new events" by comparing to
130 * After a new SQ head value was read by the application this
131 * counter includes all submissions that were dropped reaching
132 * the new SQ head (and possibly more).
138 * Written by the kernel, shouldn't be modified by the
141 * The application needs a full memory barrier before checking
142 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
146 * Number of completion events lost because the queue was full;
147 * this should be avoided by the application by making sure
148 * there are not more requests pending than there is space in
149 * the completion queue.
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * As completion events come in out of order this counter is not
156 * ordered with any other data.
160 * Ring buffer of completion events.
162 * The kernel writes completion events fresh every time they are
163 * produced, so the application is allowed to modify pending
166 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
169 struct io_mapped_ubuf {
172 struct bio_vec *bvec;
173 unsigned int nr_bvecs;
176 struct fixed_file_table {
182 struct percpu_ref refs;
183 } ____cacheline_aligned_in_smp;
189 bool cq_overflow_flushed;
193 * Ring buffer of indices into array of io_uring_sqe, which is
194 * mmapped by the application using the IORING_OFF_SQES offset.
196 * This indirection could e.g. be used to assign fixed
197 * io_uring_sqe entries to operations and only submit them to
198 * the queue when needed.
200 * The kernel modifies neither the indices array nor the entries
204 unsigned cached_sq_head;
207 unsigned sq_thread_idle;
208 unsigned cached_sq_dropped;
209 atomic_t cached_cq_overflow;
210 struct io_uring_sqe *sq_sqes;
212 struct list_head defer_list;
213 struct list_head timeout_list;
214 struct list_head cq_overflow_list;
216 wait_queue_head_t inflight_wait;
217 } ____cacheline_aligned_in_smp;
219 struct io_rings *rings;
223 struct task_struct *sqo_thread; /* if using sq thread polling */
224 struct mm_struct *sqo_mm;
225 wait_queue_head_t sqo_wait;
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
232 struct fixed_file_table *file_table;
233 unsigned nr_user_files;
235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
239 struct user_struct *user;
241 const struct cred *creds;
243 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
244 struct completion *completions;
246 /* if all else fails... */
247 struct io_kiocb *fallback_req;
249 #if defined(CONFIG_UNIX)
250 struct socket *ring_sock;
254 unsigned cached_cq_tail;
257 atomic_t cq_timeouts;
258 struct wait_queue_head cq_wait;
259 struct fasync_struct *cq_fasync;
260 struct eventfd_ctx *cq_ev_fd;
261 } ____cacheline_aligned_in_smp;
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
269 spinlock_t completion_lock;
270 bool poll_multi_file;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list;
278 struct hlist_head *cancel_hash;
279 unsigned cancel_hash_bits;
281 spinlock_t inflight_lock;
282 struct list_head inflight_list;
283 } ____cacheline_aligned_in_smp;
287 * First field must be the file pointer in all the
288 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
290 struct io_poll_iocb {
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;
309 struct sockaddr __user *addr;
310 int __user *addr_len;
321 struct io_async_connect {
322 struct sockaddr_storage address;
325 struct io_async_msghdr {
326 struct iovec fast_iov[UIO_FASTIOV];
328 struct sockaddr __user *uaddr;
333 struct iovec fast_iov[UIO_FASTIOV];
339 struct io_async_ctx {
340 struct io_uring_sqe sqe;
342 struct io_async_rw rw;
343 struct io_async_msghdr msg;
344 struct io_async_connect connect;
345 struct io_timeout_data timeout;
350 * NOTE! Each of the iocb union members has the file pointer
351 * as the first entry in their struct definition. So you can
352 * access the file pointer through any of the sub-structs,
353 * or directly as just 'ki_filp' in this struct.
359 struct io_poll_iocb poll;
360 struct io_accept accept;
364 const struct io_uring_sqe *sqe;
365 struct io_async_ctx *io;
366 struct file *ring_file;
370 bool needs_fixed_file;
372 struct io_ring_ctx *ctx;
374 struct list_head list;
375 struct hlist_node hash_node;
377 struct list_head link_list;
380 #define REQ_F_NOWAIT 1 /* must not punt to workers */
381 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
382 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
383 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
384 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
385 #define REQ_F_IO_DRAINED 32 /* drain done */
386 #define REQ_F_LINK 64 /* linked sqes */
387 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
388 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
389 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
390 #define REQ_F_TIMEOUT 1024 /* timeout request */
391 #define REQ_F_ISREG 2048 /* regular file */
392 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
393 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
394 #define REQ_F_INFLIGHT 16384 /* on inflight list */
395 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
396 #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */
397 #define REQ_F_PREPPED 131072 /* request already opcode prepared */
402 struct list_head inflight_entry;
404 struct io_wq_work work;
407 #define IO_PLUG_THRESHOLD 2
408 #define IO_IOPOLL_BATCH 8
410 struct io_submit_state {
411 struct blk_plug plug;
414 * io_kiocb alloc cache
416 void *reqs[IO_IOPOLL_BATCH];
417 unsigned int free_reqs;
418 unsigned int cur_req;
421 * File reference cache
425 unsigned int has_refs;
426 unsigned int used_refs;
427 unsigned int ios_left;
430 static void io_wq_submit_work(struct io_wq_work **workptr);
431 static void io_cqring_fill_event(struct io_kiocb *req, long res);
432 static void __io_free_req(struct io_kiocb *req);
433 static void io_put_req(struct io_kiocb *req);
434 static void io_double_put_req(struct io_kiocb *req);
435 static void __io_double_put_req(struct io_kiocb *req);
436 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
437 static void io_queue_linked_timeout(struct io_kiocb *req);
439 static struct kmem_cache *req_cachep;
441 static const struct file_operations io_uring_fops;
443 struct sock *io_uring_get_socket(struct file *file)
445 #if defined(CONFIG_UNIX)
446 if (file->f_op == &io_uring_fops) {
447 struct io_ring_ctx *ctx = file->private_data;
449 return ctx->ring_sock->sk;
454 EXPORT_SYMBOL(io_uring_get_socket);
456 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
458 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
460 complete(&ctx->completions[0]);
463 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
465 struct io_ring_ctx *ctx;
468 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
472 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
473 if (!ctx->fallback_req)
476 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
477 if (!ctx->completions)
481 * Use 5 bits less than the max cq entries, that should give us around
482 * 32 entries per hash list if totally full and uniformly spread.
484 hash_bits = ilog2(p->cq_entries);
488 ctx->cancel_hash_bits = hash_bits;
489 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
491 if (!ctx->cancel_hash)
493 __hash_init(ctx->cancel_hash, 1U << hash_bits);
495 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
496 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
499 ctx->flags = p->flags;
500 init_waitqueue_head(&ctx->cq_wait);
501 INIT_LIST_HEAD(&ctx->cq_overflow_list);
502 init_completion(&ctx->completions[0]);
503 init_completion(&ctx->completions[1]);
504 mutex_init(&ctx->uring_lock);
505 init_waitqueue_head(&ctx->wait);
506 spin_lock_init(&ctx->completion_lock);
507 INIT_LIST_HEAD(&ctx->poll_list);
508 INIT_LIST_HEAD(&ctx->defer_list);
509 INIT_LIST_HEAD(&ctx->timeout_list);
510 init_waitqueue_head(&ctx->inflight_wait);
511 spin_lock_init(&ctx->inflight_lock);
512 INIT_LIST_HEAD(&ctx->inflight_list);
515 if (ctx->fallback_req)
516 kmem_cache_free(req_cachep, ctx->fallback_req);
517 kfree(ctx->completions);
518 kfree(ctx->cancel_hash);
523 static inline bool __req_need_defer(struct io_kiocb *req)
525 struct io_ring_ctx *ctx = req->ctx;
527 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
528 + atomic_read(&ctx->cached_cq_overflow);
531 static inline bool req_need_defer(struct io_kiocb *req)
533 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
534 return __req_need_defer(req);
539 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
541 struct io_kiocb *req;
543 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
544 if (req && !req_need_defer(req)) {
545 list_del_init(&req->list);
552 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
554 struct io_kiocb *req;
556 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
558 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
560 if (!__req_need_defer(req)) {
561 list_del_init(&req->list);
569 static void __io_commit_cqring(struct io_ring_ctx *ctx)
571 struct io_rings *rings = ctx->rings;
573 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
574 /* order cqe stores with ring update */
575 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
577 if (wq_has_sleeper(&ctx->cq_wait)) {
578 wake_up_interruptible(&ctx->cq_wait);
579 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
584 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
586 u8 opcode = READ_ONCE(sqe->opcode);
588 return !(opcode == IORING_OP_READ_FIXED ||
589 opcode == IORING_OP_WRITE_FIXED);
592 static inline bool io_prep_async_work(struct io_kiocb *req,
593 struct io_kiocb **link)
595 bool do_hashed = false;
598 switch (req->sqe->opcode) {
599 case IORING_OP_WRITEV:
600 case IORING_OP_WRITE_FIXED:
601 /* only regular files should be hashed for writes */
602 if (req->flags & REQ_F_ISREG)
605 case IORING_OP_READV:
606 case IORING_OP_READ_FIXED:
607 case IORING_OP_SENDMSG:
608 case IORING_OP_RECVMSG:
609 case IORING_OP_ACCEPT:
610 case IORING_OP_POLL_ADD:
611 case IORING_OP_CONNECT:
613 * We know REQ_F_ISREG is not set on some of these
614 * opcodes, but this enables us to keep the check in
617 if (!(req->flags & REQ_F_ISREG))
618 req->work.flags |= IO_WQ_WORK_UNBOUND;
621 if (io_sqe_needs_user(req->sqe))
622 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
625 *link = io_prep_linked_timeout(req);
629 static inline void io_queue_async_work(struct io_kiocb *req)
631 struct io_ring_ctx *ctx = req->ctx;
632 struct io_kiocb *link;
635 do_hashed = io_prep_async_work(req, &link);
637 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
640 io_wq_enqueue(ctx->io_wq, &req->work);
642 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
643 file_inode(req->file));
647 io_queue_linked_timeout(link);
650 static void io_kill_timeout(struct io_kiocb *req)
654 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
656 atomic_inc(&req->ctx->cq_timeouts);
657 list_del_init(&req->list);
658 io_cqring_fill_event(req, 0);
663 static void io_kill_timeouts(struct io_ring_ctx *ctx)
665 struct io_kiocb *req, *tmp;
667 spin_lock_irq(&ctx->completion_lock);
668 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
669 io_kill_timeout(req);
670 spin_unlock_irq(&ctx->completion_lock);
673 static void io_commit_cqring(struct io_ring_ctx *ctx)
675 struct io_kiocb *req;
677 while ((req = io_get_timeout_req(ctx)) != NULL)
678 io_kill_timeout(req);
680 __io_commit_cqring(ctx);
682 while ((req = io_get_deferred_req(ctx)) != NULL) {
683 req->flags |= REQ_F_IO_DRAINED;
684 io_queue_async_work(req);
688 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
690 struct io_rings *rings = ctx->rings;
693 tail = ctx->cached_cq_tail;
695 * writes to the cq entry need to come after reading head; the
696 * control dependency is enough as we're using WRITE_ONCE to
699 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
702 ctx->cached_cq_tail++;
703 return &rings->cqes[tail & ctx->cq_mask];
706 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
708 if (waitqueue_active(&ctx->wait))
710 if (waitqueue_active(&ctx->sqo_wait))
711 wake_up(&ctx->sqo_wait);
713 eventfd_signal(ctx->cq_ev_fd, 1);
716 /* Returns true if there are no backlogged entries after the flush */
717 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
719 struct io_rings *rings = ctx->rings;
720 struct io_uring_cqe *cqe;
721 struct io_kiocb *req;
726 if (list_empty_careful(&ctx->cq_overflow_list))
728 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
729 rings->cq_ring_entries))
733 spin_lock_irqsave(&ctx->completion_lock, flags);
735 /* if force is set, the ring is going away. always drop after that */
737 ctx->cq_overflow_flushed = true;
740 while (!list_empty(&ctx->cq_overflow_list)) {
741 cqe = io_get_cqring(ctx);
745 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
747 list_move(&req->list, &list);
749 WRITE_ONCE(cqe->user_data, req->user_data);
750 WRITE_ONCE(cqe->res, req->result);
751 WRITE_ONCE(cqe->flags, 0);
753 WRITE_ONCE(ctx->rings->cq_overflow,
754 atomic_inc_return(&ctx->cached_cq_overflow));
758 io_commit_cqring(ctx);
759 spin_unlock_irqrestore(&ctx->completion_lock, flags);
760 io_cqring_ev_posted(ctx);
762 while (!list_empty(&list)) {
763 req = list_first_entry(&list, struct io_kiocb, list);
764 list_del(&req->list);
771 static void io_cqring_fill_event(struct io_kiocb *req, long res)
773 struct io_ring_ctx *ctx = req->ctx;
774 struct io_uring_cqe *cqe;
776 trace_io_uring_complete(ctx, req->user_data, res);
779 * If we can't get a cq entry, userspace overflowed the
780 * submission (by quite a lot). Increment the overflow count in
783 cqe = io_get_cqring(ctx);
785 WRITE_ONCE(cqe->user_data, req->user_data);
786 WRITE_ONCE(cqe->res, res);
787 WRITE_ONCE(cqe->flags, 0);
788 } else if (ctx->cq_overflow_flushed) {
789 WRITE_ONCE(ctx->rings->cq_overflow,
790 atomic_inc_return(&ctx->cached_cq_overflow));
792 refcount_inc(&req->refs);
794 list_add_tail(&req->list, &ctx->cq_overflow_list);
798 static void io_cqring_add_event(struct io_kiocb *req, long res)
800 struct io_ring_ctx *ctx = req->ctx;
803 spin_lock_irqsave(&ctx->completion_lock, flags);
804 io_cqring_fill_event(req, res);
805 io_commit_cqring(ctx);
806 spin_unlock_irqrestore(&ctx->completion_lock, flags);
808 io_cqring_ev_posted(ctx);
811 static inline bool io_is_fallback_req(struct io_kiocb *req)
813 return req == (struct io_kiocb *)
814 ((unsigned long) req->ctx->fallback_req & ~1UL);
817 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
819 struct io_kiocb *req;
821 req = ctx->fallback_req;
822 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
828 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
829 struct io_submit_state *state)
831 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
832 struct io_kiocb *req;
834 if (!percpu_ref_tryget(&ctx->refs))
838 req = kmem_cache_alloc(req_cachep, gfp);
841 } else if (!state->free_reqs) {
845 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
846 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
849 * Bulk alloc is all-or-nothing. If we fail to get a batch,
850 * retry single alloc to be on the safe side.
852 if (unlikely(ret <= 0)) {
853 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
858 state->free_reqs = ret - 1;
860 req = state->reqs[0];
862 req = state->reqs[state->cur_req];
869 req->ring_file = NULL;
873 /* one is dropped after submission, the other at completion */
874 refcount_set(&req->refs, 2);
876 INIT_IO_WORK(&req->work, io_wq_submit_work);
879 req = io_get_fallback_req(ctx);
882 percpu_ref_put(&ctx->refs);
886 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
889 kmem_cache_free_bulk(req_cachep, *nr, reqs);
890 percpu_ref_put_many(&ctx->refs, *nr);
895 static void __io_free_req(struct io_kiocb *req)
897 struct io_ring_ctx *ctx = req->ctx;
901 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
903 if (req->flags & REQ_F_INFLIGHT) {
906 spin_lock_irqsave(&ctx->inflight_lock, flags);
907 list_del(&req->inflight_entry);
908 if (waitqueue_active(&ctx->inflight_wait))
909 wake_up(&ctx->inflight_wait);
910 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
912 percpu_ref_put(&ctx->refs);
913 if (likely(!io_is_fallback_req(req)))
914 kmem_cache_free(req_cachep, req);
916 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
919 static bool io_link_cancel_timeout(struct io_kiocb *req)
921 struct io_ring_ctx *ctx = req->ctx;
924 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
926 io_cqring_fill_event(req, -ECANCELED);
927 io_commit_cqring(ctx);
928 req->flags &= ~REQ_F_LINK;
936 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
938 struct io_ring_ctx *ctx = req->ctx;
939 bool wake_ev = false;
941 /* Already got next link */
942 if (req->flags & REQ_F_LINK_NEXT)
946 * The list should never be empty when we are called here. But could
947 * potentially happen if the chain is messed up, check to be on the
950 while (!list_empty(&req->link_list)) {
951 struct io_kiocb *nxt = list_first_entry(&req->link_list,
952 struct io_kiocb, link_list);
954 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
955 (nxt->flags & REQ_F_TIMEOUT))) {
956 list_del_init(&nxt->link_list);
957 wake_ev |= io_link_cancel_timeout(nxt);
958 req->flags &= ~REQ_F_LINK_TIMEOUT;
962 list_del_init(&req->link_list);
963 if (!list_empty(&nxt->link_list))
964 nxt->flags |= REQ_F_LINK;
969 req->flags |= REQ_F_LINK_NEXT;
971 io_cqring_ev_posted(ctx);
975 * Called if REQ_F_LINK is set, and we fail the head request
977 static void io_fail_links(struct io_kiocb *req)
979 struct io_ring_ctx *ctx = req->ctx;
982 spin_lock_irqsave(&ctx->completion_lock, flags);
984 while (!list_empty(&req->link_list)) {
985 struct io_kiocb *link = list_first_entry(&req->link_list,
986 struct io_kiocb, link_list);
988 list_del_init(&link->link_list);
989 trace_io_uring_fail_link(req, link);
991 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
992 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
993 io_link_cancel_timeout(link);
995 io_cqring_fill_event(link, -ECANCELED);
996 __io_double_put_req(link);
998 req->flags &= ~REQ_F_LINK_TIMEOUT;
1001 io_commit_cqring(ctx);
1002 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1003 io_cqring_ev_posted(ctx);
1006 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1008 if (likely(!(req->flags & REQ_F_LINK)))
1012 * If LINK is set, we have dependent requests in this chain. If we
1013 * didn't fail this request, queue the first one up, moving any other
1014 * dependencies to the next request. In case of failure, fail the rest
1017 if (req->flags & REQ_F_FAIL_LINK) {
1019 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1020 REQ_F_LINK_TIMEOUT) {
1021 struct io_ring_ctx *ctx = req->ctx;
1022 unsigned long flags;
1025 * If this is a timeout link, we could be racing with the
1026 * timeout timer. Grab the completion lock for this case to
1027 * protect against that.
1029 spin_lock_irqsave(&ctx->completion_lock, flags);
1030 io_req_link_next(req, nxt);
1031 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1033 io_req_link_next(req, nxt);
1037 static void io_free_req(struct io_kiocb *req)
1039 struct io_kiocb *nxt = NULL;
1041 io_req_find_next(req, &nxt);
1045 io_queue_async_work(nxt);
1049 * Drop reference to request, return next in chain (if there is one) if this
1050 * was the last reference to this request.
1052 __attribute__((nonnull))
1053 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1055 io_req_find_next(req, nxtptr);
1057 if (refcount_dec_and_test(&req->refs))
1061 static void io_put_req(struct io_kiocb *req)
1063 if (refcount_dec_and_test(&req->refs))
1068 * Must only be used if we don't need to care about links, usually from
1069 * within the completion handling itself.
1071 static void __io_double_put_req(struct io_kiocb *req)
1073 /* drop both submit and complete references */
1074 if (refcount_sub_and_test(2, &req->refs))
1078 static void io_double_put_req(struct io_kiocb *req)
1080 /* drop both submit and complete references */
1081 if (refcount_sub_and_test(2, &req->refs))
1085 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1087 struct io_rings *rings = ctx->rings;
1090 * noflush == true is from the waitqueue handler, just ensure we wake
1091 * up the task, and the next invocation will flush the entries. We
1092 * cannot safely to it from here.
1094 if (noflush && !list_empty(&ctx->cq_overflow_list))
1097 io_cqring_overflow_flush(ctx, false);
1099 /* See comment at the top of this file */
1101 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1104 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1106 struct io_rings *rings = ctx->rings;
1108 /* make sure SQ entry isn't read before tail */
1109 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1113 * Find and free completed poll iocbs
1115 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1116 struct list_head *done)
1118 void *reqs[IO_IOPOLL_BATCH];
1119 struct io_kiocb *req;
1123 while (!list_empty(done)) {
1124 req = list_first_entry(done, struct io_kiocb, list);
1125 list_del(&req->list);
1127 io_cqring_fill_event(req, req->result);
1130 if (refcount_dec_and_test(&req->refs)) {
1131 /* If we're not using fixed files, we have to pair the
1132 * completion part with the file put. Use regular
1133 * completions for those, only batch free for fixed
1134 * file and non-linked commands.
1136 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1137 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1139 reqs[to_free++] = req;
1140 if (to_free == ARRAY_SIZE(reqs))
1141 io_free_req_many(ctx, reqs, &to_free);
1148 io_commit_cqring(ctx);
1149 io_free_req_many(ctx, reqs, &to_free);
1152 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1155 struct io_kiocb *req, *tmp;
1161 * Only spin for completions if we don't have multiple devices hanging
1162 * off our complete list, and we're under the requested amount.
1164 spin = !ctx->poll_multi_file && *nr_events < min;
1167 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1168 struct kiocb *kiocb = &req->rw;
1171 * Move completed entries to our local list. If we find a
1172 * request that requires polling, break out and complete
1173 * the done list first, if we have entries there.
1175 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1176 list_move_tail(&req->list, &done);
1179 if (!list_empty(&done))
1182 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1191 if (!list_empty(&done))
1192 io_iopoll_complete(ctx, nr_events, &done);
1198 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1199 * non-spinning poll check - we'll still enter the driver poll loop, but only
1200 * as a non-spinning completion check.
1202 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1205 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1208 ret = io_do_iopoll(ctx, nr_events, min);
1211 if (!min || *nr_events >= min)
1219 * We can't just wait for polled events to come to us, we have to actively
1220 * find and complete them.
1222 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1224 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1227 mutex_lock(&ctx->uring_lock);
1228 while (!list_empty(&ctx->poll_list)) {
1229 unsigned int nr_events = 0;
1231 io_iopoll_getevents(ctx, &nr_events, 1);
1234 * Ensure we allow local-to-the-cpu processing to take place,
1235 * in this case we need to ensure that we reap all events.
1239 mutex_unlock(&ctx->uring_lock);
1242 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1245 int iters = 0, ret = 0;
1251 * Don't enter poll loop if we already have events pending.
1252 * If we do, we can potentially be spinning for commands that
1253 * already triggered a CQE (eg in error).
1255 if (io_cqring_events(ctx, false))
1259 * If a submit got punted to a workqueue, we can have the
1260 * application entering polling for a command before it gets
1261 * issued. That app will hold the uring_lock for the duration
1262 * of the poll right here, so we need to take a breather every
1263 * now and then to ensure that the issue has a chance to add
1264 * the poll to the issued list. Otherwise we can spin here
1265 * forever, while the workqueue is stuck trying to acquire the
1268 if (!(++iters & 7)) {
1269 mutex_unlock(&ctx->uring_lock);
1270 mutex_lock(&ctx->uring_lock);
1273 if (*nr_events < min)
1274 tmin = min - *nr_events;
1276 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1280 } while (min && !*nr_events && !need_resched());
1285 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1291 * We disallow the app entering submit/complete with polling, but we
1292 * still need to lock the ring to prevent racing with polled issue
1293 * that got punted to a workqueue.
1295 mutex_lock(&ctx->uring_lock);
1296 ret = __io_iopoll_check(ctx, nr_events, min);
1297 mutex_unlock(&ctx->uring_lock);
1301 static void kiocb_end_write(struct io_kiocb *req)
1304 * Tell lockdep we inherited freeze protection from submission
1307 if (req->flags & REQ_F_ISREG) {
1308 struct inode *inode = file_inode(req->file);
1310 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1312 file_end_write(req->file);
1315 static inline void req_set_fail_links(struct io_kiocb *req)
1317 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1318 req->flags |= REQ_F_FAIL_LINK;
1321 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1323 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1325 if (kiocb->ki_flags & IOCB_WRITE)
1326 kiocb_end_write(req);
1328 if (res != req->result)
1329 req_set_fail_links(req);
1330 io_cqring_add_event(req, res);
1333 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1335 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1337 io_complete_rw_common(kiocb, res);
1341 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1343 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1344 struct io_kiocb *nxt = NULL;
1346 io_complete_rw_common(kiocb, res);
1347 io_put_req_find_next(req, &nxt);
1352 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1354 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1356 if (kiocb->ki_flags & IOCB_WRITE)
1357 kiocb_end_write(req);
1359 if (res != req->result)
1360 req_set_fail_links(req);
1363 req->flags |= REQ_F_IOPOLL_COMPLETED;
1367 * After the iocb has been issued, it's safe to be found on the poll list.
1368 * Adding the kiocb to the list AFTER submission ensures that we don't
1369 * find it from a io_iopoll_getevents() thread before the issuer is done
1370 * accessing the kiocb cookie.
1372 static void io_iopoll_req_issued(struct io_kiocb *req)
1374 struct io_ring_ctx *ctx = req->ctx;
1377 * Track whether we have multiple files in our lists. This will impact
1378 * how we do polling eventually, not spinning if we're on potentially
1379 * different devices.
1381 if (list_empty(&ctx->poll_list)) {
1382 ctx->poll_multi_file = false;
1383 } else if (!ctx->poll_multi_file) {
1384 struct io_kiocb *list_req;
1386 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1388 if (list_req->rw.ki_filp != req->rw.ki_filp)
1389 ctx->poll_multi_file = true;
1393 * For fast devices, IO may have already completed. If it has, add
1394 * it to the front so we find it first.
1396 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1397 list_add(&req->list, &ctx->poll_list);
1399 list_add_tail(&req->list, &ctx->poll_list);
1402 static void io_file_put(struct io_submit_state *state)
1405 int diff = state->has_refs - state->used_refs;
1408 fput_many(state->file, diff);
1414 * Get as many references to a file as we have IOs left in this submission,
1415 * assuming most submissions are for one file, or at least that each file
1416 * has more than one submission.
1418 static struct file *io_file_get(struct io_submit_state *state, int fd)
1424 if (state->fd == fd) {
1431 state->file = fget_many(fd, state->ios_left);
1436 state->has_refs = state->ios_left;
1437 state->used_refs = 1;
1443 * If we tracked the file through the SCM inflight mechanism, we could support
1444 * any file. For now, just ensure that anything potentially problematic is done
1447 static bool io_file_supports_async(struct file *file)
1449 umode_t mode = file_inode(file)->i_mode;
1451 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1453 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1459 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1461 const struct io_uring_sqe *sqe = req->sqe;
1462 struct io_ring_ctx *ctx = req->ctx;
1463 struct kiocb *kiocb = &req->rw;
1470 if (S_ISREG(file_inode(req->file)->i_mode))
1471 req->flags |= REQ_F_ISREG;
1473 kiocb->ki_pos = READ_ONCE(sqe->off);
1474 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1475 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1477 ioprio = READ_ONCE(sqe->ioprio);
1479 ret = ioprio_check_cap(ioprio);
1483 kiocb->ki_ioprio = ioprio;
1485 kiocb->ki_ioprio = get_current_ioprio();
1487 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1491 /* don't allow async punt if RWF_NOWAIT was requested */
1492 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1493 (req->file->f_flags & O_NONBLOCK))
1494 req->flags |= REQ_F_NOWAIT;
1497 kiocb->ki_flags |= IOCB_NOWAIT;
1499 if (ctx->flags & IORING_SETUP_IOPOLL) {
1500 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1501 !kiocb->ki_filp->f_op->iopoll)
1504 kiocb->ki_flags |= IOCB_HIPRI;
1505 kiocb->ki_complete = io_complete_rw_iopoll;
1508 if (kiocb->ki_flags & IOCB_HIPRI)
1510 kiocb->ki_complete = io_complete_rw;
1515 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1521 case -ERESTARTNOINTR:
1522 case -ERESTARTNOHAND:
1523 case -ERESTART_RESTARTBLOCK:
1525 * We can't just restart the syscall, since previously
1526 * submitted sqes may already be in progress. Just fail this
1532 kiocb->ki_complete(kiocb, ret, 0);
1536 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1539 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1540 *nxt = __io_complete_rw(kiocb, ret);
1542 io_rw_done(kiocb, ret);
1545 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1546 const struct io_uring_sqe *sqe,
1547 struct iov_iter *iter)
1549 size_t len = READ_ONCE(sqe->len);
1550 struct io_mapped_ubuf *imu;
1551 unsigned index, buf_index;
1555 /* attempt to use fixed buffers without having provided iovecs */
1556 if (unlikely(!ctx->user_bufs))
1559 buf_index = READ_ONCE(sqe->buf_index);
1560 if (unlikely(buf_index >= ctx->nr_user_bufs))
1563 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1564 imu = &ctx->user_bufs[index];
1565 buf_addr = READ_ONCE(sqe->addr);
1568 if (buf_addr + len < buf_addr)
1570 /* not inside the mapped region */
1571 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1575 * May not be a start of buffer, set size appropriately
1576 * and advance us to the beginning.
1578 offset = buf_addr - imu->ubuf;
1579 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1583 * Don't use iov_iter_advance() here, as it's really slow for
1584 * using the latter parts of a big fixed buffer - it iterates
1585 * over each segment manually. We can cheat a bit here, because
1588 * 1) it's a BVEC iter, we set it up
1589 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1590 * first and last bvec
1592 * So just find our index, and adjust the iterator afterwards.
1593 * If the offset is within the first bvec (or the whole first
1594 * bvec, just use iov_iter_advance(). This makes it easier
1595 * since we can just skip the first segment, which may not
1596 * be PAGE_SIZE aligned.
1598 const struct bio_vec *bvec = imu->bvec;
1600 if (offset <= bvec->bv_len) {
1601 iov_iter_advance(iter, offset);
1603 unsigned long seg_skip;
1605 /* skip first vec */
1606 offset -= bvec->bv_len;
1607 seg_skip = 1 + (offset >> PAGE_SHIFT);
1609 iter->bvec = bvec + seg_skip;
1610 iter->nr_segs -= seg_skip;
1611 iter->count -= bvec->bv_len + offset;
1612 iter->iov_offset = offset & ~PAGE_MASK;
1619 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1620 struct iovec **iovec, struct iov_iter *iter)
1622 const struct io_uring_sqe *sqe = req->sqe;
1623 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1624 size_t sqe_len = READ_ONCE(sqe->len);
1628 * We're reading ->opcode for the second time, but the first read
1629 * doesn't care whether it's _FIXED or not, so it doesn't matter
1630 * whether ->opcode changes concurrently. The first read does care
1631 * about whether it is a READ or a WRITE, so we don't trust this read
1632 * for that purpose and instead let the caller pass in the read/write
1635 opcode = READ_ONCE(sqe->opcode);
1636 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1638 return io_import_fixed(req->ctx, rw, sqe, iter);
1642 struct io_async_rw *iorw = &req->io->rw;
1645 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1646 if (iorw->iov == iorw->fast_iov)
1654 #ifdef CONFIG_COMPAT
1655 if (req->ctx->compat)
1656 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1660 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1664 * For files that don't have ->read_iter() and ->write_iter(), handle them
1665 * by looping over ->read() or ->write() manually.
1667 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1668 struct iov_iter *iter)
1673 * Don't support polled IO through this interface, and we can't
1674 * support non-blocking either. For the latter, this just causes
1675 * the kiocb to be handled from an async context.
1677 if (kiocb->ki_flags & IOCB_HIPRI)
1679 if (kiocb->ki_flags & IOCB_NOWAIT)
1682 while (iov_iter_count(iter)) {
1686 if (!iov_iter_is_bvec(iter)) {
1687 iovec = iov_iter_iovec(iter);
1689 /* fixed buffers import bvec */
1690 iovec.iov_base = kmap(iter->bvec->bv_page)
1692 iovec.iov_len = min(iter->count,
1693 iter->bvec->bv_len - iter->iov_offset);
1697 nr = file->f_op->read(file, iovec.iov_base,
1698 iovec.iov_len, &kiocb->ki_pos);
1700 nr = file->f_op->write(file, iovec.iov_base,
1701 iovec.iov_len, &kiocb->ki_pos);
1704 if (iov_iter_is_bvec(iter))
1705 kunmap(iter->bvec->bv_page);
1713 if (nr != iovec.iov_len)
1715 iov_iter_advance(iter, nr);
1721 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
1722 struct iovec *iovec, struct iovec *fast_iov,
1723 struct iov_iter *iter)
1725 req->io->rw.nr_segs = iter->nr_segs;
1726 req->io->rw.size = io_size;
1727 req->io->rw.iov = iovec;
1728 if (!req->io->rw.iov) {
1729 req->io->rw.iov = req->io->rw.fast_iov;
1730 memcpy(req->io->rw.iov, fast_iov,
1731 sizeof(struct iovec) * iter->nr_segs);
1735 static int io_alloc_async_ctx(struct io_kiocb *req)
1737 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1739 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1740 req->sqe = &req->io->sqe;
1747 static void io_rw_async(struct io_wq_work **workptr)
1749 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1750 struct iovec *iov = NULL;
1752 if (req->io->rw.iov != req->io->rw.fast_iov)
1753 iov = req->io->rw.iov;
1754 io_wq_submit_work(workptr);
1758 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
1759 struct iovec *iovec, struct iovec *fast_iov,
1760 struct iov_iter *iter)
1762 if (!req->io && io_alloc_async_ctx(req))
1765 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
1766 req->work.func = io_rw_async;
1770 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1771 struct iov_iter *iter, bool force_nonblock)
1775 ret = io_prep_rw(req, force_nonblock);
1779 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1782 return io_import_iovec(READ, req, iovec, iter);
1785 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1786 bool force_nonblock)
1788 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1789 struct kiocb *kiocb = &req->rw;
1790 struct iov_iter iter;
1793 ssize_t io_size, ret;
1796 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1800 ret = io_import_iovec(READ, req, &iovec, &iter);
1807 if (req->flags & REQ_F_LINK)
1808 req->result = io_size;
1811 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1812 * we know to async punt it even if it was opened O_NONBLOCK
1814 if (force_nonblock && !io_file_supports_async(file)) {
1815 req->flags |= REQ_F_MUST_PUNT;
1819 iov_count = iov_iter_count(&iter);
1820 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1824 if (file->f_op->read_iter)
1825 ret2 = call_read_iter(file, kiocb, &iter);
1827 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1830 * In case of a short read, punt to async. This can happen
1831 * if we have data partially cached. Alternatively we can
1832 * return the short read, in which case the application will
1833 * need to issue another SQE and wait for it. That SQE will
1834 * need async punt anyway, so it's more efficient to do it
1837 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1838 (req->flags & REQ_F_ISREG) &&
1839 ret2 > 0 && ret2 < io_size)
1841 /* Catch -EAGAIN return for forced non-blocking submission */
1842 if (!force_nonblock || ret2 != -EAGAIN) {
1843 kiocb_done(kiocb, ret2, nxt, req->in_async);
1846 ret = io_setup_async_rw(req, io_size, iovec,
1847 inline_vecs, &iter);
1854 if (!io_wq_current_is_worker())
1859 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1860 struct iov_iter *iter, bool force_nonblock)
1864 ret = io_prep_rw(req, force_nonblock);
1868 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1871 return io_import_iovec(WRITE, req, iovec, iter);
1874 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1875 bool force_nonblock)
1877 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1878 struct kiocb *kiocb = &req->rw;
1879 struct iov_iter iter;
1882 ssize_t ret, io_size;
1885 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1889 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1894 file = kiocb->ki_filp;
1896 if (req->flags & REQ_F_LINK)
1897 req->result = io_size;
1900 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1901 * we know to async punt it even if it was opened O_NONBLOCK
1903 if (force_nonblock && !io_file_supports_async(req->file)) {
1904 req->flags |= REQ_F_MUST_PUNT;
1908 /* file path doesn't support NOWAIT for non-direct_IO */
1909 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
1910 (req->flags & REQ_F_ISREG))
1913 iov_count = iov_iter_count(&iter);
1914 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1919 * Open-code file_start_write here to grab freeze protection,
1920 * which will be released by another thread in
1921 * io_complete_rw(). Fool lockdep by telling it the lock got
1922 * released so that it doesn't complain about the held lock when
1923 * we return to userspace.
1925 if (req->flags & REQ_F_ISREG) {
1926 __sb_start_write(file_inode(file)->i_sb,
1927 SB_FREEZE_WRITE, true);
1928 __sb_writers_release(file_inode(file)->i_sb,
1931 kiocb->ki_flags |= IOCB_WRITE;
1933 if (file->f_op->write_iter)
1934 ret2 = call_write_iter(file, kiocb, &iter);
1936 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1937 if (!force_nonblock || ret2 != -EAGAIN) {
1938 kiocb_done(kiocb, ret2, nxt, req->in_async);
1941 ret = io_setup_async_rw(req, io_size, iovec,
1942 inline_vecs, &iter);
1949 if (!io_wq_current_is_worker())
1955 * IORING_OP_NOP just posts a completion event, nothing else.
1957 static int io_nop(struct io_kiocb *req)
1959 struct io_ring_ctx *ctx = req->ctx;
1961 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1964 io_cqring_add_event(req, 0);
1969 static int io_prep_fsync(struct io_kiocb *req)
1971 const struct io_uring_sqe *sqe = req->sqe;
1972 struct io_ring_ctx *ctx = req->ctx;
1974 if (req->flags & REQ_F_PREPPED)
1979 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1981 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1984 req->sync.flags = READ_ONCE(sqe->fsync_flags);
1985 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
1988 req->sync.off = READ_ONCE(sqe->off);
1989 req->sync.len = READ_ONCE(sqe->len);
1990 req->flags |= REQ_F_PREPPED;
1994 static bool io_req_cancelled(struct io_kiocb *req)
1996 if (req->work.flags & IO_WQ_WORK_CANCEL) {
1997 req_set_fail_links(req);
1998 io_cqring_add_event(req, -ECANCELED);
2006 static void io_fsync_finish(struct io_wq_work **workptr)
2008 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2009 loff_t end = req->sync.off + req->sync.len;
2010 struct io_kiocb *nxt = NULL;
2013 if (io_req_cancelled(req))
2016 ret = vfs_fsync_range(req->rw.ki_filp, req->sync.off,
2017 end > 0 ? end : LLONG_MAX,
2018 req->sync.flags & IORING_FSYNC_DATASYNC);
2020 req_set_fail_links(req);
2021 io_cqring_add_event(req, ret);
2022 io_put_req_find_next(req, &nxt);
2024 *workptr = &nxt->work;
2027 static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt,
2028 bool force_nonblock)
2030 struct io_wq_work *work, *old_work;
2033 ret = io_prep_fsync(req);
2037 /* fsync always requires a blocking context */
2038 if (force_nonblock) {
2040 req->work.func = io_fsync_finish;
2044 work = old_work = &req->work;
2045 io_fsync_finish(&work);
2046 if (work && work != old_work)
2047 *nxt = container_of(work, struct io_kiocb, work);
2051 static int io_prep_sfr(struct io_kiocb *req)
2053 const struct io_uring_sqe *sqe = req->sqe;
2054 struct io_ring_ctx *ctx = req->ctx;
2056 if (req->flags & REQ_F_PREPPED)
2061 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2063 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2066 req->sync.off = READ_ONCE(sqe->off);
2067 req->sync.len = READ_ONCE(sqe->len);
2068 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
2069 req->flags |= REQ_F_PREPPED;
2073 static void io_sync_file_range_finish(struct io_wq_work **workptr)
2075 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2076 struct io_kiocb *nxt = NULL;
2079 if (io_req_cancelled(req))
2082 ret = sync_file_range(req->rw.ki_filp, req->sync.off, req->sync.len,
2085 req_set_fail_links(req);
2086 io_cqring_add_event(req, ret);
2087 io_put_req_find_next(req, &nxt);
2089 *workptr = &nxt->work;
2092 static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt,
2093 bool force_nonblock)
2095 struct io_wq_work *work, *old_work;
2098 ret = io_prep_sfr(req);
2102 /* sync_file_range always requires a blocking context */
2103 if (force_nonblock) {
2105 req->work.func = io_sync_file_range_finish;
2109 work = old_work = &req->work;
2110 io_sync_file_range_finish(&work);
2111 if (work && work != old_work)
2112 *nxt = container_of(work, struct io_kiocb, work);
2116 #if defined(CONFIG_NET)
2117 static void io_sendrecv_async(struct io_wq_work **workptr)
2119 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2120 struct iovec *iov = NULL;
2122 if (req->io->rw.iov != req->io->rw.fast_iov)
2123 iov = req->io->msg.iov;
2124 io_wq_submit_work(workptr);
2129 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2131 #if defined(CONFIG_NET)
2132 const struct io_uring_sqe *sqe = req->sqe;
2133 struct user_msghdr __user *msg;
2136 flags = READ_ONCE(sqe->msg_flags);
2137 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2138 io->msg.iov = io->msg.fast_iov;
2139 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2145 static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2146 bool force_nonblock)
2148 #if defined(CONFIG_NET)
2149 const struct io_uring_sqe *sqe = req->sqe;
2150 struct io_async_msghdr *kmsg = NULL;
2151 struct socket *sock;
2154 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2157 sock = sock_from_file(req->file, &ret);
2159 struct io_async_ctx io;
2160 struct sockaddr_storage addr;
2163 flags = READ_ONCE(sqe->msg_flags);
2164 if (flags & MSG_DONTWAIT)
2165 req->flags |= REQ_F_NOWAIT;
2166 else if (force_nonblock)
2167 flags |= MSG_DONTWAIT;
2170 kmsg = &req->io->msg;
2171 kmsg->msg.msg_name = &addr;
2172 /* if iov is set, it's allocated already */
2174 kmsg->iov = kmsg->fast_iov;
2175 kmsg->msg.msg_iter.iov = kmsg->iov;
2178 kmsg->msg.msg_name = &addr;
2179 ret = io_sendmsg_prep(req, &io);
2184 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
2185 if (force_nonblock && ret == -EAGAIN) {
2188 if (io_alloc_async_ctx(req))
2190 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2191 req->work.func = io_sendrecv_async;
2194 if (ret == -ERESTARTSYS)
2199 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2201 io_cqring_add_event(req, ret);
2203 req_set_fail_links(req);
2204 io_put_req_find_next(req, nxt);
2211 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2213 #if defined(CONFIG_NET)
2214 const struct io_uring_sqe *sqe = req->sqe;
2215 struct user_msghdr __user *msg;
2218 flags = READ_ONCE(sqe->msg_flags);
2219 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2220 io->msg.iov = io->msg.fast_iov;
2221 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2228 static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt,
2229 bool force_nonblock)
2231 #if defined(CONFIG_NET)
2232 const struct io_uring_sqe *sqe = req->sqe;
2233 struct io_async_msghdr *kmsg = NULL;
2234 struct socket *sock;
2237 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2240 sock = sock_from_file(req->file, &ret);
2242 struct user_msghdr __user *msg;
2243 struct io_async_ctx io;
2244 struct sockaddr_storage addr;
2247 flags = READ_ONCE(sqe->msg_flags);
2248 if (flags & MSG_DONTWAIT)
2249 req->flags |= REQ_F_NOWAIT;
2250 else if (force_nonblock)
2251 flags |= MSG_DONTWAIT;
2253 msg = (struct user_msghdr __user *) (unsigned long)
2254 READ_ONCE(sqe->addr);
2256 kmsg = &req->io->msg;
2257 kmsg->msg.msg_name = &addr;
2258 /* if iov is set, it's allocated already */
2260 kmsg->iov = kmsg->fast_iov;
2261 kmsg->msg.msg_iter.iov = kmsg->iov;
2264 kmsg->msg.msg_name = &addr;
2265 ret = io_recvmsg_prep(req, &io);
2270 ret = __sys_recvmsg_sock(sock, &kmsg->msg, msg, kmsg->uaddr, flags);
2271 if (force_nonblock && ret == -EAGAIN) {
2274 if (io_alloc_async_ctx(req))
2276 memcpy(&req->io->msg, &io.msg, sizeof(io.msg));
2277 req->work.func = io_sendrecv_async;
2280 if (ret == -ERESTARTSYS)
2285 if (!io_wq_current_is_worker() && kmsg && kmsg->iov != kmsg->fast_iov)
2287 io_cqring_add_event(req, ret);
2289 req_set_fail_links(req);
2290 io_put_req_find_next(req, nxt);
2297 static int io_accept_prep(struct io_kiocb *req)
2299 #if defined(CONFIG_NET)
2300 const struct io_uring_sqe *sqe = req->sqe;
2301 struct io_accept *accept = &req->accept;
2303 if (req->flags & REQ_F_PREPPED)
2306 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2308 if (sqe->ioprio || sqe->len || sqe->buf_index)
2311 accept->addr = (struct sockaddr __user *)
2312 (unsigned long) READ_ONCE(sqe->addr);
2313 accept->addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2314 accept->flags = READ_ONCE(sqe->accept_flags);
2315 req->flags |= REQ_F_PREPPED;
2322 #if defined(CONFIG_NET)
2323 static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2324 bool force_nonblock)
2326 struct io_accept *accept = &req->accept;
2327 unsigned file_flags;
2330 file_flags = force_nonblock ? O_NONBLOCK : 0;
2331 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
2332 accept->addr_len, accept->flags);
2333 if (ret == -EAGAIN && force_nonblock)
2335 if (ret == -ERESTARTSYS)
2338 req_set_fail_links(req);
2339 io_cqring_add_event(req, ret);
2340 io_put_req_find_next(req, nxt);
2344 static void io_accept_finish(struct io_wq_work **workptr)
2346 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2347 struct io_kiocb *nxt = NULL;
2349 if (io_req_cancelled(req))
2351 __io_accept(req, &nxt, false);
2353 *workptr = &nxt->work;
2357 static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt,
2358 bool force_nonblock)
2360 #if defined(CONFIG_NET)
2363 ret = io_accept_prep(req);
2367 ret = __io_accept(req, nxt, force_nonblock);
2368 if (ret == -EAGAIN && force_nonblock) {
2369 req->work.func = io_accept_finish;
2370 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2380 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2382 #if defined(CONFIG_NET)
2383 const struct io_uring_sqe *sqe = req->sqe;
2384 struct sockaddr __user *addr;
2387 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2388 addr_len = READ_ONCE(sqe->addr2);
2389 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2395 static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt,
2396 bool force_nonblock)
2398 #if defined(CONFIG_NET)
2399 const struct io_uring_sqe *sqe = req->sqe;
2400 struct io_async_ctx __io, *io;
2401 unsigned file_flags;
2404 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2406 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2409 addr_len = READ_ONCE(sqe->addr2);
2410 file_flags = force_nonblock ? O_NONBLOCK : 0;
2415 ret = io_connect_prep(req, &__io);
2421 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2423 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2426 if (io_alloc_async_ctx(req)) {
2430 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
2433 if (ret == -ERESTARTSYS)
2437 req_set_fail_links(req);
2438 io_cqring_add_event(req, ret);
2439 io_put_req_find_next(req, nxt);
2446 static void io_poll_remove_one(struct io_kiocb *req)
2448 struct io_poll_iocb *poll = &req->poll;
2450 spin_lock(&poll->head->lock);
2451 WRITE_ONCE(poll->canceled, true);
2452 if (!list_empty(&poll->wait.entry)) {
2453 list_del_init(&poll->wait.entry);
2454 io_queue_async_work(req);
2456 spin_unlock(&poll->head->lock);
2457 hash_del(&req->hash_node);
2460 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2462 struct hlist_node *tmp;
2463 struct io_kiocb *req;
2466 spin_lock_irq(&ctx->completion_lock);
2467 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2468 struct hlist_head *list;
2470 list = &ctx->cancel_hash[i];
2471 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2472 io_poll_remove_one(req);
2474 spin_unlock_irq(&ctx->completion_lock);
2477 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2479 struct hlist_head *list;
2480 struct io_kiocb *req;
2482 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2483 hlist_for_each_entry(req, list, hash_node) {
2484 if (sqe_addr == req->user_data) {
2485 io_poll_remove_one(req);
2494 * Find a running poll command that matches one specified in sqe->addr,
2495 * and remove it if found.
2497 static int io_poll_remove(struct io_kiocb *req)
2499 const struct io_uring_sqe *sqe = req->sqe;
2500 struct io_ring_ctx *ctx = req->ctx;
2503 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2505 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2509 spin_lock_irq(&ctx->completion_lock);
2510 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
2511 spin_unlock_irq(&ctx->completion_lock);
2513 io_cqring_add_event(req, ret);
2515 req_set_fail_links(req);
2520 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2522 struct io_ring_ctx *ctx = req->ctx;
2524 req->poll.done = true;
2526 io_cqring_fill_event(req, error);
2528 io_cqring_fill_event(req, mangle_poll(mask));
2529 io_commit_cqring(ctx);
2532 static void io_poll_complete_work(struct io_wq_work **workptr)
2534 struct io_wq_work *work = *workptr;
2535 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2536 struct io_poll_iocb *poll = &req->poll;
2537 struct poll_table_struct pt = { ._key = poll->events };
2538 struct io_ring_ctx *ctx = req->ctx;
2539 struct io_kiocb *nxt = NULL;
2543 if (work->flags & IO_WQ_WORK_CANCEL) {
2544 WRITE_ONCE(poll->canceled, true);
2546 } else if (READ_ONCE(poll->canceled)) {
2550 if (ret != -ECANCELED)
2551 mask = vfs_poll(poll->file, &pt) & poll->events;
2554 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2555 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2556 * synchronize with them. In the cancellation case the list_del_init
2557 * itself is not actually needed, but harmless so we keep it in to
2558 * avoid further branches in the fast path.
2560 spin_lock_irq(&ctx->completion_lock);
2561 if (!mask && ret != -ECANCELED) {
2562 add_wait_queue(poll->head, &poll->wait);
2563 spin_unlock_irq(&ctx->completion_lock);
2566 hash_del(&req->hash_node);
2567 io_poll_complete(req, mask, ret);
2568 spin_unlock_irq(&ctx->completion_lock);
2570 io_cqring_ev_posted(ctx);
2573 req_set_fail_links(req);
2574 io_put_req_find_next(req, &nxt);
2576 *workptr = &nxt->work;
2579 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2582 struct io_poll_iocb *poll = wait->private;
2583 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2584 struct io_ring_ctx *ctx = req->ctx;
2585 __poll_t mask = key_to_poll(key);
2586 unsigned long flags;
2588 /* for instances that support it check for an event match first: */
2589 if (mask && !(mask & poll->events))
2592 list_del_init(&poll->wait.entry);
2595 * Run completion inline if we can. We're using trylock here because
2596 * we are violating the completion_lock -> poll wq lock ordering.
2597 * If we have a link timeout we're going to need the completion_lock
2598 * for finalizing the request, mark us as having grabbed that already.
2600 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2601 hash_del(&req->hash_node);
2602 io_poll_complete(req, mask, 0);
2603 req->flags |= REQ_F_COMP_LOCKED;
2605 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2607 io_cqring_ev_posted(ctx);
2609 io_queue_async_work(req);
2615 struct io_poll_table {
2616 struct poll_table_struct pt;
2617 struct io_kiocb *req;
2621 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2622 struct poll_table_struct *p)
2624 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2626 if (unlikely(pt->req->poll.head)) {
2627 pt->error = -EINVAL;
2632 pt->req->poll.head = head;
2633 add_wait_queue(head, &pt->req->poll.wait);
2636 static void io_poll_req_insert(struct io_kiocb *req)
2638 struct io_ring_ctx *ctx = req->ctx;
2639 struct hlist_head *list;
2641 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2642 hlist_add_head(&req->hash_node, list);
2645 static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt)
2647 const struct io_uring_sqe *sqe = req->sqe;
2648 struct io_poll_iocb *poll = &req->poll;
2649 struct io_ring_ctx *ctx = req->ctx;
2650 struct io_poll_table ipt;
2651 bool cancel = false;
2655 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2657 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2662 INIT_IO_WORK(&req->work, io_poll_complete_work);
2663 events = READ_ONCE(sqe->poll_events);
2664 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2665 INIT_HLIST_NODE(&req->hash_node);
2669 poll->canceled = false;
2671 ipt.pt._qproc = io_poll_queue_proc;
2672 ipt.pt._key = poll->events;
2674 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2676 /* initialized the list so that we can do list_empty checks */
2677 INIT_LIST_HEAD(&poll->wait.entry);
2678 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
2679 poll->wait.private = poll;
2681 INIT_LIST_HEAD(&req->list);
2683 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2685 spin_lock_irq(&ctx->completion_lock);
2686 if (likely(poll->head)) {
2687 spin_lock(&poll->head->lock);
2688 if (unlikely(list_empty(&poll->wait.entry))) {
2694 if (mask || ipt.error)
2695 list_del_init(&poll->wait.entry);
2697 WRITE_ONCE(poll->canceled, true);
2698 else if (!poll->done) /* actually waiting for an event */
2699 io_poll_req_insert(req);
2700 spin_unlock(&poll->head->lock);
2702 if (mask) { /* no async, we'd stolen it */
2704 io_poll_complete(req, mask, 0);
2706 spin_unlock_irq(&ctx->completion_lock);
2709 io_cqring_ev_posted(ctx);
2710 io_put_req_find_next(req, nxt);
2715 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2717 struct io_timeout_data *data = container_of(timer,
2718 struct io_timeout_data, timer);
2719 struct io_kiocb *req = data->req;
2720 struct io_ring_ctx *ctx = req->ctx;
2721 unsigned long flags;
2723 atomic_inc(&ctx->cq_timeouts);
2725 spin_lock_irqsave(&ctx->completion_lock, flags);
2727 * We could be racing with timeout deletion. If the list is empty,
2728 * then timeout lookup already found it and will be handling it.
2730 if (!list_empty(&req->list)) {
2731 struct io_kiocb *prev;
2734 * Adjust the reqs sequence before the current one because it
2735 * will consume a slot in the cq_ring and the cq_tail
2736 * pointer will be increased, otherwise other timeout reqs may
2737 * return in advance without waiting for enough wait_nr.
2740 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2742 list_del_init(&req->list);
2745 io_cqring_fill_event(req, -ETIME);
2746 io_commit_cqring(ctx);
2747 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2749 io_cqring_ev_posted(ctx);
2750 req_set_fail_links(req);
2752 return HRTIMER_NORESTART;
2755 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2757 struct io_kiocb *req;
2760 list_for_each_entry(req, &ctx->timeout_list, list) {
2761 if (user_data == req->user_data) {
2762 list_del_init(&req->list);
2771 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2775 req_set_fail_links(req);
2776 io_cqring_fill_event(req, -ECANCELED);
2782 * Remove or update an existing timeout command
2784 static int io_timeout_remove(struct io_kiocb *req)
2786 const struct io_uring_sqe *sqe = req->sqe;
2787 struct io_ring_ctx *ctx = req->ctx;
2791 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2793 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2795 flags = READ_ONCE(sqe->timeout_flags);
2799 spin_lock_irq(&ctx->completion_lock);
2800 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
2802 io_cqring_fill_event(req, ret);
2803 io_commit_cqring(ctx);
2804 spin_unlock_irq(&ctx->completion_lock);
2805 io_cqring_ev_posted(ctx);
2807 req_set_fail_links(req);
2812 static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2813 bool is_timeout_link)
2815 const struct io_uring_sqe *sqe = req->sqe;
2816 struct io_timeout_data *data;
2819 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2821 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2823 if (sqe->off && is_timeout_link)
2825 flags = READ_ONCE(sqe->timeout_flags);
2826 if (flags & ~IORING_TIMEOUT_ABS)
2829 data = &io->timeout;
2831 req->flags |= REQ_F_TIMEOUT;
2833 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2836 if (flags & IORING_TIMEOUT_ABS)
2837 data->mode = HRTIMER_MODE_ABS;
2839 data->mode = HRTIMER_MODE_REL;
2841 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2845 static int io_timeout(struct io_kiocb *req)
2847 const struct io_uring_sqe *sqe = req->sqe;
2849 struct io_ring_ctx *ctx = req->ctx;
2850 struct io_timeout_data *data;
2851 struct list_head *entry;
2856 if (io_alloc_async_ctx(req))
2858 ret = io_timeout_prep(req, req->io, false);
2862 data = &req->io->timeout;
2865 * sqe->off holds how many events that need to occur for this
2866 * timeout event to be satisfied. If it isn't set, then this is
2867 * a pure timeout request, sequence isn't used.
2869 count = READ_ONCE(sqe->off);
2871 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2872 spin_lock_irq(&ctx->completion_lock);
2873 entry = ctx->timeout_list.prev;
2877 req->sequence = ctx->cached_sq_head + count - 1;
2878 data->seq_offset = count;
2881 * Insertion sort, ensuring the first entry in the list is always
2882 * the one we need first.
2884 spin_lock_irq(&ctx->completion_lock);
2885 list_for_each_prev(entry, &ctx->timeout_list) {
2886 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2887 unsigned nxt_sq_head;
2888 long long tmp, tmp_nxt;
2889 u32 nxt_offset = nxt->io->timeout.seq_offset;
2891 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2895 * Since cached_sq_head + count - 1 can overflow, use type long
2898 tmp = (long long)ctx->cached_sq_head + count - 1;
2899 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2900 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2903 * cached_sq_head may overflow, and it will never overflow twice
2904 * once there is some timeout req still be valid.
2906 if (ctx->cached_sq_head < nxt_sq_head)
2913 * Sequence of reqs after the insert one and itself should
2914 * be adjusted because each timeout req consumes a slot.
2919 req->sequence -= span;
2921 list_add(&req->list, entry);
2922 data->timer.function = io_timeout_fn;
2923 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2924 spin_unlock_irq(&ctx->completion_lock);
2928 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2930 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2932 return req->user_data == (unsigned long) data;
2935 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2937 enum io_wq_cancel cancel_ret;
2940 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2941 switch (cancel_ret) {
2942 case IO_WQ_CANCEL_OK:
2945 case IO_WQ_CANCEL_RUNNING:
2948 case IO_WQ_CANCEL_NOTFOUND:
2956 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2957 struct io_kiocb *req, __u64 sqe_addr,
2958 struct io_kiocb **nxt, int success_ret)
2960 unsigned long flags;
2963 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2964 if (ret != -ENOENT) {
2965 spin_lock_irqsave(&ctx->completion_lock, flags);
2969 spin_lock_irqsave(&ctx->completion_lock, flags);
2970 ret = io_timeout_cancel(ctx, sqe_addr);
2973 ret = io_poll_cancel(ctx, sqe_addr);
2977 io_cqring_fill_event(req, ret);
2978 io_commit_cqring(ctx);
2979 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2980 io_cqring_ev_posted(ctx);
2983 req_set_fail_links(req);
2984 io_put_req_find_next(req, nxt);
2987 static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt)
2989 const struct io_uring_sqe *sqe = req->sqe;
2990 struct io_ring_ctx *ctx = req->ctx;
2992 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2994 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2998 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
3002 static int io_req_defer_prep(struct io_kiocb *req)
3004 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
3005 struct io_async_ctx *io = req->io;
3006 struct iov_iter iter;
3009 switch (io->sqe.opcode) {
3010 case IORING_OP_READV:
3011 case IORING_OP_READ_FIXED:
3012 /* ensure prep does right import */
3014 ret = io_read_prep(req, &iovec, &iter, true);
3018 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3021 case IORING_OP_WRITEV:
3022 case IORING_OP_WRITE_FIXED:
3023 /* ensure prep does right import */
3025 ret = io_write_prep(req, &iovec, &iter, true);
3029 io_req_map_rw(req, ret, iovec, inline_vecs, &iter);
3032 case IORING_OP_FSYNC:
3033 ret = io_prep_fsync(req);
3035 case IORING_OP_SYNC_FILE_RANGE:
3036 ret = io_prep_sfr(req);
3038 case IORING_OP_SENDMSG:
3039 ret = io_sendmsg_prep(req, io);
3041 case IORING_OP_RECVMSG:
3042 ret = io_recvmsg_prep(req, io);
3044 case IORING_OP_CONNECT:
3045 ret = io_connect_prep(req, io);
3047 case IORING_OP_TIMEOUT:
3048 ret = io_timeout_prep(req, io, false);
3050 case IORING_OP_LINK_TIMEOUT:
3051 ret = io_timeout_prep(req, io, true);
3053 case IORING_OP_ACCEPT:
3054 ret = io_accept_prep(req);
3064 static int io_req_defer(struct io_kiocb *req)
3066 struct io_ring_ctx *ctx = req->ctx;
3069 /* Still need defer if there is pending req in defer list. */
3070 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
3073 if (io_alloc_async_ctx(req))
3076 ret = io_req_defer_prep(req);
3080 spin_lock_irq(&ctx->completion_lock);
3081 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
3082 spin_unlock_irq(&ctx->completion_lock);
3086 trace_io_uring_defer(ctx, req, req->user_data);
3087 list_add_tail(&req->list, &ctx->defer_list);
3088 spin_unlock_irq(&ctx->completion_lock);
3089 return -EIOCBQUEUED;
3092 __attribute__((nonnull))
3093 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
3094 bool force_nonblock)
3097 struct io_ring_ctx *ctx = req->ctx;
3099 opcode = READ_ONCE(req->sqe->opcode);
3104 case IORING_OP_READV:
3105 if (unlikely(req->sqe->buf_index))
3107 ret = io_read(req, nxt, force_nonblock);
3109 case IORING_OP_WRITEV:
3110 if (unlikely(req->sqe->buf_index))
3112 ret = io_write(req, nxt, force_nonblock);
3114 case IORING_OP_READ_FIXED:
3115 ret = io_read(req, nxt, force_nonblock);
3117 case IORING_OP_WRITE_FIXED:
3118 ret = io_write(req, nxt, force_nonblock);
3120 case IORING_OP_FSYNC:
3121 ret = io_fsync(req, nxt, force_nonblock);
3123 case IORING_OP_POLL_ADD:
3124 ret = io_poll_add(req, nxt);
3126 case IORING_OP_POLL_REMOVE:
3127 ret = io_poll_remove(req);
3129 case IORING_OP_SYNC_FILE_RANGE:
3130 ret = io_sync_file_range(req, nxt, force_nonblock);
3132 case IORING_OP_SENDMSG:
3133 ret = io_sendmsg(req, nxt, force_nonblock);
3135 case IORING_OP_RECVMSG:
3136 ret = io_recvmsg(req, nxt, force_nonblock);
3138 case IORING_OP_TIMEOUT:
3139 ret = io_timeout(req);
3141 case IORING_OP_TIMEOUT_REMOVE:
3142 ret = io_timeout_remove(req);
3144 case IORING_OP_ACCEPT:
3145 ret = io_accept(req, nxt, force_nonblock);
3147 case IORING_OP_CONNECT:
3148 ret = io_connect(req, nxt, force_nonblock);
3150 case IORING_OP_ASYNC_CANCEL:
3151 ret = io_async_cancel(req, nxt);
3161 if (ctx->flags & IORING_SETUP_IOPOLL) {
3162 if (req->result == -EAGAIN)
3165 io_iopoll_req_issued(req);
3171 static void io_link_work_cb(struct io_wq_work **workptr)
3173 struct io_wq_work *work = *workptr;
3174 struct io_kiocb *link = work->data;
3176 io_queue_linked_timeout(link);
3177 work->func = io_wq_submit_work;
3180 static void io_wq_submit_work(struct io_wq_work **workptr)
3182 struct io_wq_work *work = *workptr;
3183 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3184 struct io_kiocb *nxt = NULL;
3187 /* Ensure we clear previously set non-block flag */
3188 req->rw.ki_flags &= ~IOCB_NOWAIT;
3190 if (work->flags & IO_WQ_WORK_CANCEL)
3194 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3195 req->in_async = true;
3197 ret = io_issue_sqe(req, &nxt, false);
3199 * We can get EAGAIN for polled IO even though we're
3200 * forcing a sync submission from here, since we can't
3201 * wait for request slots on the block side.
3209 /* drop submission reference */
3213 req_set_fail_links(req);
3214 io_cqring_add_event(req, ret);
3218 /* if a dependent link is ready, pass it back */
3220 struct io_kiocb *link;
3222 io_prep_async_work(nxt, &link);
3223 *workptr = &nxt->work;
3225 nxt->work.flags |= IO_WQ_WORK_CB;
3226 nxt->work.func = io_link_work_cb;
3227 nxt->work.data = link;
3232 static bool io_req_op_valid(int op)
3234 return op >= IORING_OP_NOP && op < IORING_OP_LAST;
3237 static int io_op_needs_file(const struct io_uring_sqe *sqe)
3239 int op = READ_ONCE(sqe->opcode);
3243 case IORING_OP_POLL_REMOVE:
3244 case IORING_OP_TIMEOUT:
3245 case IORING_OP_TIMEOUT_REMOVE:
3246 case IORING_OP_ASYNC_CANCEL:
3247 case IORING_OP_LINK_TIMEOUT:
3250 if (io_req_op_valid(op))
3256 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3259 struct fixed_file_table *table;
3261 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3262 return table->files[index & IORING_FILE_TABLE_MASK];
3265 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3267 struct io_ring_ctx *ctx = req->ctx;
3271 flags = READ_ONCE(req->sqe->flags);
3272 fd = READ_ONCE(req->sqe->fd);
3274 if (flags & IOSQE_IO_DRAIN)
3275 req->flags |= REQ_F_IO_DRAIN;
3277 ret = io_op_needs_file(req->sqe);
3281 if (flags & IOSQE_FIXED_FILE) {
3282 if (unlikely(!ctx->file_table ||
3283 (unsigned) fd >= ctx->nr_user_files))
3285 fd = array_index_nospec(fd, ctx->nr_user_files);
3286 req->file = io_file_from_index(ctx, fd);
3289 req->flags |= REQ_F_FIXED_FILE;
3291 if (req->needs_fixed_file)
3293 trace_io_uring_file_get(ctx, fd);
3294 req->file = io_file_get(state, fd);
3295 if (unlikely(!req->file))
3302 static int io_grab_files(struct io_kiocb *req)
3305 struct io_ring_ctx *ctx = req->ctx;
3308 spin_lock_irq(&ctx->inflight_lock);
3310 * We use the f_ops->flush() handler to ensure that we can flush
3311 * out work accessing these files if the fd is closed. Check if
3312 * the fd has changed since we started down this path, and disallow
3313 * this operation if it has.
3315 if (fcheck(req->ring_fd) == req->ring_file) {
3316 list_add(&req->inflight_entry, &ctx->inflight_list);
3317 req->flags |= REQ_F_INFLIGHT;
3318 req->work.files = current->files;
3321 spin_unlock_irq(&ctx->inflight_lock);
3327 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3329 struct io_timeout_data *data = container_of(timer,
3330 struct io_timeout_data, timer);
3331 struct io_kiocb *req = data->req;
3332 struct io_ring_ctx *ctx = req->ctx;
3333 struct io_kiocb *prev = NULL;
3334 unsigned long flags;
3336 spin_lock_irqsave(&ctx->completion_lock, flags);
3339 * We don't expect the list to be empty, that will only happen if we
3340 * race with the completion of the linked work.
3342 if (!list_empty(&req->link_list)) {
3343 prev = list_entry(req->link_list.prev, struct io_kiocb,
3345 if (refcount_inc_not_zero(&prev->refs)) {
3346 list_del_init(&req->link_list);
3347 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3352 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3355 req_set_fail_links(prev);
3356 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3360 io_cqring_add_event(req, -ETIME);
3363 return HRTIMER_NORESTART;
3366 static void io_queue_linked_timeout(struct io_kiocb *req)
3368 struct io_ring_ctx *ctx = req->ctx;
3371 * If the list is now empty, then our linked request finished before
3372 * we got a chance to setup the timer
3374 spin_lock_irq(&ctx->completion_lock);
3375 if (!list_empty(&req->link_list)) {
3376 struct io_timeout_data *data = &req->io->timeout;
3378 data->timer.function = io_link_timeout_fn;
3379 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3382 spin_unlock_irq(&ctx->completion_lock);
3384 /* drop submission reference */
3388 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3390 struct io_kiocb *nxt;
3392 if (!(req->flags & REQ_F_LINK))
3395 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3397 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
3400 req->flags |= REQ_F_LINK_TIMEOUT;
3404 static void __io_queue_sqe(struct io_kiocb *req)
3406 struct io_kiocb *linked_timeout;
3407 struct io_kiocb *nxt = NULL;
3411 linked_timeout = io_prep_linked_timeout(req);
3413 ret = io_issue_sqe(req, &nxt, true);
3416 * We async punt it if the file wasn't marked NOWAIT, or if the file
3417 * doesn't support non-blocking read/write attempts
3419 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3420 (req->flags & REQ_F_MUST_PUNT))) {
3421 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3422 ret = io_grab_files(req);
3428 * Queued up for async execution, worker will release
3429 * submit reference when the iocb is actually submitted.
3431 io_queue_async_work(req);
3436 /* drop submission reference */
3439 if (linked_timeout) {
3441 io_queue_linked_timeout(linked_timeout);
3443 io_put_req(linked_timeout);
3446 /* and drop final reference, if we failed */
3448 io_cqring_add_event(req, ret);
3449 req_set_fail_links(req);
3460 static void io_queue_sqe(struct io_kiocb *req)
3464 if (unlikely(req->ctx->drain_next)) {
3465 req->flags |= REQ_F_IO_DRAIN;
3466 req->ctx->drain_next = false;
3468 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3470 ret = io_req_defer(req);
3472 if (ret != -EIOCBQUEUED) {
3473 io_cqring_add_event(req, ret);
3474 req_set_fail_links(req);
3475 io_double_put_req(req);
3478 __io_queue_sqe(req);
3481 static inline void io_queue_link_head(struct io_kiocb *req)
3483 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3484 io_cqring_add_event(req, -ECANCELED);
3485 io_double_put_req(req);
3490 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3493 static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3494 struct io_kiocb **link)
3496 struct io_ring_ctx *ctx = req->ctx;
3499 req->user_data = req->sqe->user_data;
3501 /* enforce forwards compatibility on users */
3502 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3507 ret = io_req_set_file(state, req);
3508 if (unlikely(ret)) {
3510 io_cqring_add_event(req, ret);
3511 io_double_put_req(req);
3516 * If we already have a head request, queue this one for async
3517 * submittal once the head completes. If we don't have a head but
3518 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3519 * submitted sync once the chain is complete. If none of those
3520 * conditions are true (normal request), then just queue it.
3523 struct io_kiocb *prev = *link;
3525 if (req->sqe->flags & IOSQE_IO_DRAIN)
3526 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3528 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3529 req->flags |= REQ_F_HARDLINK;
3531 if (io_alloc_async_ctx(req)) {
3536 ret = io_req_defer_prep(req);
3538 /* fail even hard links since we don't submit */
3539 prev->flags |= REQ_F_FAIL_LINK;
3542 trace_io_uring_link(ctx, req, prev);
3543 list_add_tail(&req->link_list, &prev->link_list);
3544 } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
3545 req->flags |= REQ_F_LINK;
3546 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3547 req->flags |= REQ_F_HARDLINK;
3549 INIT_LIST_HEAD(&req->link_list);
3559 * Batched submission is done, ensure local IO is flushed out.
3561 static void io_submit_state_end(struct io_submit_state *state)
3563 blk_finish_plug(&state->plug);
3565 if (state->free_reqs)
3566 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3567 &state->reqs[state->cur_req]);
3571 * Start submission side cache.
3573 static void io_submit_state_start(struct io_submit_state *state,
3574 unsigned int max_ios)
3576 blk_start_plug(&state->plug);
3577 state->free_reqs = 0;
3579 state->ios_left = max_ios;
3582 static void io_commit_sqring(struct io_ring_ctx *ctx)
3584 struct io_rings *rings = ctx->rings;
3586 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3588 * Ensure any loads from the SQEs are done at this point,
3589 * since once we write the new head, the application could
3590 * write new data to them.
3592 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3597 * Fetch an sqe, if one is available. Note that req->sqe will point to memory
3598 * that is mapped by userspace. This means that care needs to be taken to
3599 * ensure that reads are stable, as we cannot rely on userspace always
3600 * being a good citizen. If members of the sqe are validated and then later
3601 * used, it's important that those reads are done through READ_ONCE() to
3602 * prevent a re-load down the line.
3604 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3606 struct io_rings *rings = ctx->rings;
3607 u32 *sq_array = ctx->sq_array;
3611 * The cached sq head (or cq tail) serves two purposes:
3613 * 1) allows us to batch the cost of updating the user visible
3615 * 2) allows the kernel side to track the head on its own, even
3616 * though the application is the one updating it.
3618 head = ctx->cached_sq_head;
3619 /* make sure SQ entry isn't read before tail */
3620 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3623 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3624 if (likely(head < ctx->sq_entries)) {
3626 * All io need record the previous position, if LINK vs DARIN,
3627 * it can be used to mark the position of the first IO in the
3630 req->sequence = ctx->cached_sq_head;
3631 req->sqe = &ctx->sq_sqes[head];
3632 ctx->cached_sq_head++;
3636 /* drop invalid entries */
3637 ctx->cached_sq_head++;
3638 ctx->cached_sq_dropped++;
3639 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3643 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3644 struct file *ring_file, int ring_fd,
3645 struct mm_struct **mm, bool async)
3647 struct io_submit_state state, *statep = NULL;
3648 struct io_kiocb *link = NULL;
3649 int i, submitted = 0;
3650 bool mm_fault = false;
3652 /* if we have a backlog and couldn't flush it all, return BUSY */
3653 if (!list_empty(&ctx->cq_overflow_list) &&
3654 !io_cqring_overflow_flush(ctx, false))
3657 if (nr > IO_PLUG_THRESHOLD) {
3658 io_submit_state_start(&state, nr);
3662 for (i = 0; i < nr; i++) {
3663 struct io_kiocb *req;
3664 unsigned int sqe_flags;
3666 req = io_get_req(ctx, statep);
3667 if (unlikely(!req)) {
3669 submitted = -EAGAIN;
3672 if (!io_get_sqring(ctx, req)) {
3677 if (io_sqe_needs_user(req->sqe) && !*mm) {
3678 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3680 use_mm(ctx->sqo_mm);
3686 sqe_flags = req->sqe->flags;
3688 req->ring_file = ring_file;
3689 req->ring_fd = ring_fd;
3690 req->has_user = *mm != NULL;
3691 req->in_async = async;
3692 req->needs_fixed_file = async;
3693 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
3695 if (!io_submit_sqe(req, statep, &link))
3698 * If previous wasn't linked and we have a linked command,
3699 * that's the end of the chain. Submit the previous link.
3701 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) && link) {
3702 io_queue_link_head(link);
3708 io_queue_link_head(link);
3710 io_submit_state_end(&state);
3712 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3713 io_commit_sqring(ctx);
3718 static int io_sq_thread(void *data)
3720 struct io_ring_ctx *ctx = data;
3721 struct mm_struct *cur_mm = NULL;
3722 const struct cred *old_cred;
3723 mm_segment_t old_fs;
3726 unsigned long timeout;
3729 complete(&ctx->completions[1]);
3733 old_cred = override_creds(ctx->creds);
3735 ret = timeout = inflight = 0;
3736 while (!kthread_should_park()) {
3737 unsigned int to_submit;
3740 unsigned nr_events = 0;
3742 if (ctx->flags & IORING_SETUP_IOPOLL) {
3744 * inflight is the count of the maximum possible
3745 * entries we submitted, but it can be smaller
3746 * if we dropped some of them. If we don't have
3747 * poll entries available, then we know that we
3748 * have nothing left to poll for. Reset the
3749 * inflight count to zero in that case.
3751 mutex_lock(&ctx->uring_lock);
3752 if (!list_empty(&ctx->poll_list))
3753 __io_iopoll_check(ctx, &nr_events, 0);
3756 mutex_unlock(&ctx->uring_lock);
3759 * Normal IO, just pretend everything completed.
3760 * We don't have to poll completions for that.
3762 nr_events = inflight;
3765 inflight -= nr_events;
3767 timeout = jiffies + ctx->sq_thread_idle;
3770 to_submit = io_sqring_entries(ctx);
3773 * If submit got -EBUSY, flag us as needing the application
3774 * to enter the kernel to reap and flush events.
3776 if (!to_submit || ret == -EBUSY) {
3778 * We're polling. If we're within the defined idle
3779 * period, then let us spin without work before going
3780 * to sleep. The exception is if we got EBUSY doing
3781 * more IO, we should wait for the application to
3782 * reap events and wake us up.
3785 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3791 * Drop cur_mm before scheduling, we can't hold it for
3792 * long periods (or over schedule()). Do this before
3793 * adding ourselves to the waitqueue, as the unuse/drop
3802 prepare_to_wait(&ctx->sqo_wait, &wait,
3803 TASK_INTERRUPTIBLE);
3805 /* Tell userspace we may need a wakeup call */
3806 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3807 /* make sure to read SQ tail after writing flags */
3810 to_submit = io_sqring_entries(ctx);
3811 if (!to_submit || ret == -EBUSY) {
3812 if (kthread_should_park()) {
3813 finish_wait(&ctx->sqo_wait, &wait);
3816 if (signal_pending(current))
3817 flush_signals(current);
3819 finish_wait(&ctx->sqo_wait, &wait);
3821 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3824 finish_wait(&ctx->sqo_wait, &wait);
3826 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3829 to_submit = min(to_submit, ctx->sq_entries);
3830 mutex_lock(&ctx->uring_lock);
3831 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3832 mutex_unlock(&ctx->uring_lock);
3842 revert_creds(old_cred);
3849 struct io_wait_queue {
3850 struct wait_queue_entry wq;
3851 struct io_ring_ctx *ctx;
3853 unsigned nr_timeouts;
3856 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3858 struct io_ring_ctx *ctx = iowq->ctx;
3861 * Wake up if we have enough events, or if a timeout occurred since we
3862 * started waiting. For timeouts, we always want to return to userspace,
3863 * regardless of event count.
3865 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3866 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3869 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3870 int wake_flags, void *key)
3872 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3875 /* use noflush == true, as we can't safely rely on locking context */
3876 if (!io_should_wake(iowq, true))
3879 return autoremove_wake_function(curr, mode, wake_flags, key);
3883 * Wait until events become available, if we don't already have some. The
3884 * application must reap them itself, as they reside on the shared cq ring.
3886 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3887 const sigset_t __user *sig, size_t sigsz)
3889 struct io_wait_queue iowq = {
3892 .func = io_wake_function,
3893 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3896 .to_wait = min_events,
3898 struct io_rings *rings = ctx->rings;
3901 if (io_cqring_events(ctx, false) >= min_events)
3905 #ifdef CONFIG_COMPAT
3906 if (in_compat_syscall())
3907 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3911 ret = set_user_sigmask(sig, sigsz);
3917 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
3918 trace_io_uring_cqring_wait(ctx, min_events);
3920 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3921 TASK_INTERRUPTIBLE);
3922 if (io_should_wake(&iowq, false))
3925 if (signal_pending(current)) {
3930 finish_wait(&ctx->wait, &iowq.wq);
3932 restore_saved_sigmask_unless(ret == -EINTR);
3934 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
3937 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3939 #if defined(CONFIG_UNIX)
3940 if (ctx->ring_sock) {
3941 struct sock *sock = ctx->ring_sock->sk;
3942 struct sk_buff *skb;
3944 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3950 for (i = 0; i < ctx->nr_user_files; i++) {
3953 file = io_file_from_index(ctx, i);
3960 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3962 unsigned nr_tables, i;
3964 if (!ctx->file_table)
3967 __io_sqe_files_unregister(ctx);
3968 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3969 for (i = 0; i < nr_tables; i++)
3970 kfree(ctx->file_table[i].files);
3971 kfree(ctx->file_table);
3972 ctx->file_table = NULL;
3973 ctx->nr_user_files = 0;
3977 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3979 if (ctx->sqo_thread) {
3980 wait_for_completion(&ctx->completions[1]);
3982 * The park is a bit of a work-around, without it we get
3983 * warning spews on shutdown with SQPOLL set and affinity
3984 * set to a single CPU.
3986 kthread_park(ctx->sqo_thread);
3987 kthread_stop(ctx->sqo_thread);
3988 ctx->sqo_thread = NULL;
3992 static void io_finish_async(struct io_ring_ctx *ctx)
3994 io_sq_thread_stop(ctx);
3997 io_wq_destroy(ctx->io_wq);
4002 #if defined(CONFIG_UNIX)
4003 static void io_destruct_skb(struct sk_buff *skb)
4005 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
4008 io_wq_flush(ctx->io_wq);
4010 unix_destruct_scm(skb);
4014 * Ensure the UNIX gc is aware of our file set, so we are certain that
4015 * the io_uring can be safely unregistered on process exit, even if we have
4016 * loops in the file referencing.
4018 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
4020 struct sock *sk = ctx->ring_sock->sk;
4021 struct scm_fp_list *fpl;
4022 struct sk_buff *skb;
4025 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
4026 unsigned long inflight = ctx->user->unix_inflight + nr;
4028 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
4032 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
4036 skb = alloc_skb(0, GFP_KERNEL);
4045 fpl->user = get_uid(ctx->user);
4046 for (i = 0; i < nr; i++) {
4047 struct file *file = io_file_from_index(ctx, i + offset);
4051 fpl->fp[nr_files] = get_file(file);
4052 unix_inflight(fpl->user, fpl->fp[nr_files]);
4057 fpl->max = SCM_MAX_FD;
4058 fpl->count = nr_files;
4059 UNIXCB(skb).fp = fpl;
4060 skb->destructor = io_destruct_skb;
4061 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
4062 skb_queue_head(&sk->sk_receive_queue, skb);
4064 for (i = 0; i < nr_files; i++)
4075 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
4076 * causes regular reference counting to break down. We rely on the UNIX
4077 * garbage collection to take care of this problem for us.
4079 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4081 unsigned left, total;
4085 left = ctx->nr_user_files;
4087 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
4089 ret = __io_sqe_files_scm(ctx, this_files, total);
4093 total += this_files;
4099 while (total < ctx->nr_user_files) {
4100 struct file *file = io_file_from_index(ctx, total);
4110 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
4116 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
4121 for (i = 0; i < nr_tables; i++) {
4122 struct fixed_file_table *table = &ctx->file_table[i];
4123 unsigned this_files;
4125 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
4126 table->files = kcalloc(this_files, sizeof(struct file *),
4130 nr_files -= this_files;
4136 for (i = 0; i < nr_tables; i++) {
4137 struct fixed_file_table *table = &ctx->file_table[i];
4138 kfree(table->files);
4143 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
4146 __s32 __user *fds = (__s32 __user *) arg;
4151 if (ctx->file_table)
4155 if (nr_args > IORING_MAX_FIXED_FILES)
4158 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
4159 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
4161 if (!ctx->file_table)
4164 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
4165 kfree(ctx->file_table);
4166 ctx->file_table = NULL;
4170 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
4171 struct fixed_file_table *table;
4175 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4177 /* allow sparse sets */
4183 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4184 index = i & IORING_FILE_TABLE_MASK;
4185 table->files[index] = fget(fd);
4188 if (!table->files[index])
4191 * Don't allow io_uring instances to be registered. If UNIX
4192 * isn't enabled, then this causes a reference cycle and this
4193 * instance can never get freed. If UNIX is enabled we'll
4194 * handle it just fine, but there's still no point in allowing
4195 * a ring fd as it doesn't support regular read/write anyway.
4197 if (table->files[index]->f_op == &io_uring_fops) {
4198 fput(table->files[index]);
4205 for (i = 0; i < ctx->nr_user_files; i++) {
4208 file = io_file_from_index(ctx, i);
4212 for (i = 0; i < nr_tables; i++)
4213 kfree(ctx->file_table[i].files);
4215 kfree(ctx->file_table);
4216 ctx->file_table = NULL;
4217 ctx->nr_user_files = 0;
4221 ret = io_sqe_files_scm(ctx);
4223 io_sqe_files_unregister(ctx);
4228 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4230 #if defined(CONFIG_UNIX)
4231 struct file *file = io_file_from_index(ctx, index);
4232 struct sock *sock = ctx->ring_sock->sk;
4233 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4234 struct sk_buff *skb;
4237 __skb_queue_head_init(&list);
4240 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4241 * remove this entry and rearrange the file array.
4243 skb = skb_dequeue(head);
4245 struct scm_fp_list *fp;
4247 fp = UNIXCB(skb).fp;
4248 for (i = 0; i < fp->count; i++) {
4251 if (fp->fp[i] != file)
4254 unix_notinflight(fp->user, fp->fp[i]);
4255 left = fp->count - 1 - i;
4257 memmove(&fp->fp[i], &fp->fp[i + 1],
4258 left * sizeof(struct file *));
4265 __skb_queue_tail(&list, skb);
4275 __skb_queue_tail(&list, skb);
4277 skb = skb_dequeue(head);
4280 if (skb_peek(&list)) {
4281 spin_lock_irq(&head->lock);
4282 while ((skb = __skb_dequeue(&list)) != NULL)
4283 __skb_queue_tail(head, skb);
4284 spin_unlock_irq(&head->lock);
4287 fput(io_file_from_index(ctx, index));
4291 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4294 #if defined(CONFIG_UNIX)
4295 struct sock *sock = ctx->ring_sock->sk;
4296 struct sk_buff_head *head = &sock->sk_receive_queue;
4297 struct sk_buff *skb;
4300 * See if we can merge this file into an existing skb SCM_RIGHTS
4301 * file set. If there's no room, fall back to allocating a new skb
4302 * and filling it in.
4304 spin_lock_irq(&head->lock);
4305 skb = skb_peek(head);
4307 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4309 if (fpl->count < SCM_MAX_FD) {
4310 __skb_unlink(skb, head);
4311 spin_unlock_irq(&head->lock);
4312 fpl->fp[fpl->count] = get_file(file);
4313 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4315 spin_lock_irq(&head->lock);
4316 __skb_queue_head(head, skb);
4321 spin_unlock_irq(&head->lock);
4328 return __io_sqe_files_scm(ctx, 1, index);
4334 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4337 struct io_uring_files_update up;
4342 if (!ctx->file_table)
4346 if (copy_from_user(&up, arg, sizeof(up)))
4348 if (check_add_overflow(up.offset, nr_args, &done))
4350 if (done > ctx->nr_user_files)
4354 fds = (__s32 __user *) up.fds;
4356 struct fixed_file_table *table;
4360 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4364 i = array_index_nospec(up.offset, ctx->nr_user_files);
4365 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4366 index = i & IORING_FILE_TABLE_MASK;
4367 if (table->files[index]) {
4368 io_sqe_file_unregister(ctx, i);
4369 table->files[index] = NULL;
4380 * Don't allow io_uring instances to be registered. If
4381 * UNIX isn't enabled, then this causes a reference
4382 * cycle and this instance can never get freed. If UNIX
4383 * is enabled we'll handle it just fine, but there's
4384 * still no point in allowing a ring fd as it doesn't
4385 * support regular read/write anyway.
4387 if (file->f_op == &io_uring_fops) {
4392 table->files[index] = file;
4393 err = io_sqe_file_register(ctx, file, i);
4402 return done ? done : err;
4405 static void io_put_work(struct io_wq_work *work)
4407 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4412 static void io_get_work(struct io_wq_work *work)
4414 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4416 refcount_inc(&req->refs);
4419 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4420 struct io_uring_params *p)
4422 struct io_wq_data data;
4423 unsigned concurrency;
4426 init_waitqueue_head(&ctx->sqo_wait);
4427 mmgrab(current->mm);
4428 ctx->sqo_mm = current->mm;
4430 if (ctx->flags & IORING_SETUP_SQPOLL) {
4432 if (!capable(CAP_SYS_ADMIN))
4435 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4436 if (!ctx->sq_thread_idle)
4437 ctx->sq_thread_idle = HZ;
4439 if (p->flags & IORING_SETUP_SQ_AFF) {
4440 int cpu = p->sq_thread_cpu;
4443 if (cpu >= nr_cpu_ids)
4445 if (!cpu_online(cpu))
4448 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4452 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4455 if (IS_ERR(ctx->sqo_thread)) {
4456 ret = PTR_ERR(ctx->sqo_thread);
4457 ctx->sqo_thread = NULL;
4460 wake_up_process(ctx->sqo_thread);
4461 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4462 /* Can't have SQ_AFF without SQPOLL */
4467 data.mm = ctx->sqo_mm;
4468 data.user = ctx->user;
4469 data.creds = ctx->creds;
4470 data.get_work = io_get_work;
4471 data.put_work = io_put_work;
4473 /* Do QD, or 4 * CPUS, whatever is smallest */
4474 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4475 ctx->io_wq = io_wq_create(concurrency, &data);
4476 if (IS_ERR(ctx->io_wq)) {
4477 ret = PTR_ERR(ctx->io_wq);
4484 io_finish_async(ctx);
4485 mmdrop(ctx->sqo_mm);
4490 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4492 atomic_long_sub(nr_pages, &user->locked_vm);
4495 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4497 unsigned long page_limit, cur_pages, new_pages;
4499 /* Don't allow more pages than we can safely lock */
4500 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4503 cur_pages = atomic_long_read(&user->locked_vm);
4504 new_pages = cur_pages + nr_pages;
4505 if (new_pages > page_limit)
4507 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4508 new_pages) != cur_pages);
4513 static void io_mem_free(void *ptr)
4520 page = virt_to_head_page(ptr);
4521 if (put_page_testzero(page))
4522 free_compound_page(page);
4525 static void *io_mem_alloc(size_t size)
4527 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4530 return (void *) __get_free_pages(gfp_flags, get_order(size));
4533 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4536 struct io_rings *rings;
4537 size_t off, sq_array_size;
4539 off = struct_size(rings, cqes, cq_entries);
4540 if (off == SIZE_MAX)
4544 off = ALIGN(off, SMP_CACHE_BYTES);
4549 sq_array_size = array_size(sizeof(u32), sq_entries);
4550 if (sq_array_size == SIZE_MAX)
4553 if (check_add_overflow(off, sq_array_size, &off))
4562 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4566 pages = (size_t)1 << get_order(
4567 rings_size(sq_entries, cq_entries, NULL));
4568 pages += (size_t)1 << get_order(
4569 array_size(sizeof(struct io_uring_sqe), sq_entries));
4574 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4578 if (!ctx->user_bufs)
4581 for (i = 0; i < ctx->nr_user_bufs; i++) {
4582 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4584 for (j = 0; j < imu->nr_bvecs; j++)
4585 put_user_page(imu->bvec[j].bv_page);
4587 if (ctx->account_mem)
4588 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4593 kfree(ctx->user_bufs);
4594 ctx->user_bufs = NULL;
4595 ctx->nr_user_bufs = 0;
4599 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4600 void __user *arg, unsigned index)
4602 struct iovec __user *src;
4604 #ifdef CONFIG_COMPAT
4606 struct compat_iovec __user *ciovs;
4607 struct compat_iovec ciov;
4609 ciovs = (struct compat_iovec __user *) arg;
4610 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4613 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4614 dst->iov_len = ciov.iov_len;
4618 src = (struct iovec __user *) arg;
4619 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4624 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4627 struct vm_area_struct **vmas = NULL;
4628 struct page **pages = NULL;
4629 int i, j, got_pages = 0;
4634 if (!nr_args || nr_args > UIO_MAXIOV)
4637 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4639 if (!ctx->user_bufs)
4642 for (i = 0; i < nr_args; i++) {
4643 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4644 unsigned long off, start, end, ubuf;
4649 ret = io_copy_iov(ctx, &iov, arg, i);
4654 * Don't impose further limits on the size and buffer
4655 * constraints here, we'll -EINVAL later when IO is
4656 * submitted if they are wrong.
4659 if (!iov.iov_base || !iov.iov_len)
4662 /* arbitrary limit, but we need something */
4663 if (iov.iov_len > SZ_1G)
4666 ubuf = (unsigned long) iov.iov_base;
4667 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4668 start = ubuf >> PAGE_SHIFT;
4669 nr_pages = end - start;
4671 if (ctx->account_mem) {
4672 ret = io_account_mem(ctx->user, nr_pages);
4678 if (!pages || nr_pages > got_pages) {
4681 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4683 vmas = kvmalloc_array(nr_pages,
4684 sizeof(struct vm_area_struct *),
4686 if (!pages || !vmas) {
4688 if (ctx->account_mem)
4689 io_unaccount_mem(ctx->user, nr_pages);
4692 got_pages = nr_pages;
4695 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4699 if (ctx->account_mem)
4700 io_unaccount_mem(ctx->user, nr_pages);
4705 down_read(¤t->mm->mmap_sem);
4706 pret = get_user_pages(ubuf, nr_pages,
4707 FOLL_WRITE | FOLL_LONGTERM,
4709 if (pret == nr_pages) {
4710 /* don't support file backed memory */
4711 for (j = 0; j < nr_pages; j++) {
4712 struct vm_area_struct *vma = vmas[j];
4715 !is_file_hugepages(vma->vm_file)) {
4721 ret = pret < 0 ? pret : -EFAULT;
4723 up_read(¤t->mm->mmap_sem);
4726 * if we did partial map, or found file backed vmas,
4727 * release any pages we did get
4730 put_user_pages(pages, pret);
4731 if (ctx->account_mem)
4732 io_unaccount_mem(ctx->user, nr_pages);
4737 off = ubuf & ~PAGE_MASK;
4739 for (j = 0; j < nr_pages; j++) {
4742 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4743 imu->bvec[j].bv_page = pages[j];
4744 imu->bvec[j].bv_len = vec_len;
4745 imu->bvec[j].bv_offset = off;
4749 /* store original address for later verification */
4751 imu->len = iov.iov_len;
4752 imu->nr_bvecs = nr_pages;
4754 ctx->nr_user_bufs++;
4762 io_sqe_buffer_unregister(ctx);
4766 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4768 __s32 __user *fds = arg;
4774 if (copy_from_user(&fd, fds, sizeof(*fds)))
4777 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4778 if (IS_ERR(ctx->cq_ev_fd)) {
4779 int ret = PTR_ERR(ctx->cq_ev_fd);
4780 ctx->cq_ev_fd = NULL;
4787 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4789 if (ctx->cq_ev_fd) {
4790 eventfd_ctx_put(ctx->cq_ev_fd);
4791 ctx->cq_ev_fd = NULL;
4798 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4800 io_finish_async(ctx);
4802 mmdrop(ctx->sqo_mm);
4804 io_iopoll_reap_events(ctx);
4805 io_sqe_buffer_unregister(ctx);
4806 io_sqe_files_unregister(ctx);
4807 io_eventfd_unregister(ctx);
4809 #if defined(CONFIG_UNIX)
4810 if (ctx->ring_sock) {
4811 ctx->ring_sock->file = NULL; /* so that iput() is called */
4812 sock_release(ctx->ring_sock);
4816 io_mem_free(ctx->rings);
4817 io_mem_free(ctx->sq_sqes);
4819 percpu_ref_exit(&ctx->refs);
4820 if (ctx->account_mem)
4821 io_unaccount_mem(ctx->user,
4822 ring_pages(ctx->sq_entries, ctx->cq_entries));
4823 free_uid(ctx->user);
4824 put_cred(ctx->creds);
4825 kfree(ctx->completions);
4826 kfree(ctx->cancel_hash);
4827 kmem_cache_free(req_cachep, ctx->fallback_req);
4831 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4833 struct io_ring_ctx *ctx = file->private_data;
4836 poll_wait(file, &ctx->cq_wait, wait);
4838 * synchronizes with barrier from wq_has_sleeper call in
4842 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4843 ctx->rings->sq_ring_entries)
4844 mask |= EPOLLOUT | EPOLLWRNORM;
4845 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4846 mask |= EPOLLIN | EPOLLRDNORM;
4851 static int io_uring_fasync(int fd, struct file *file, int on)
4853 struct io_ring_ctx *ctx = file->private_data;
4855 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4858 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4860 mutex_lock(&ctx->uring_lock);
4861 percpu_ref_kill(&ctx->refs);
4862 mutex_unlock(&ctx->uring_lock);
4864 io_kill_timeouts(ctx);
4865 io_poll_remove_all(ctx);
4868 io_wq_cancel_all(ctx->io_wq);
4870 io_iopoll_reap_events(ctx);
4871 /* if we failed setting up the ctx, we might not have any rings */
4873 io_cqring_overflow_flush(ctx, true);
4874 wait_for_completion(&ctx->completions[0]);
4875 io_ring_ctx_free(ctx);
4878 static int io_uring_release(struct inode *inode, struct file *file)
4880 struct io_ring_ctx *ctx = file->private_data;
4882 file->private_data = NULL;
4883 io_ring_ctx_wait_and_kill(ctx);
4887 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4888 struct files_struct *files)
4890 struct io_kiocb *req;
4893 while (!list_empty_careful(&ctx->inflight_list)) {
4894 struct io_kiocb *cancel_req = NULL;
4896 spin_lock_irq(&ctx->inflight_lock);
4897 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4898 if (req->work.files != files)
4900 /* req is being completed, ignore */
4901 if (!refcount_inc_not_zero(&req->refs))
4907 prepare_to_wait(&ctx->inflight_wait, &wait,
4908 TASK_UNINTERRUPTIBLE);
4909 spin_unlock_irq(&ctx->inflight_lock);
4911 /* We need to keep going until we don't find a matching req */
4915 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4916 io_put_req(cancel_req);
4919 finish_wait(&ctx->inflight_wait, &wait);
4922 static int io_uring_flush(struct file *file, void *data)
4924 struct io_ring_ctx *ctx = file->private_data;
4926 io_uring_cancel_files(ctx, data);
4927 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4928 io_cqring_overflow_flush(ctx, true);
4929 io_wq_cancel_all(ctx->io_wq);
4934 static void *io_uring_validate_mmap_request(struct file *file,
4935 loff_t pgoff, size_t sz)
4937 struct io_ring_ctx *ctx = file->private_data;
4938 loff_t offset = pgoff << PAGE_SHIFT;
4943 case IORING_OFF_SQ_RING:
4944 case IORING_OFF_CQ_RING:
4947 case IORING_OFF_SQES:
4951 return ERR_PTR(-EINVAL);
4954 page = virt_to_head_page(ptr);
4955 if (sz > page_size(page))
4956 return ERR_PTR(-EINVAL);
4963 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4965 size_t sz = vma->vm_end - vma->vm_start;
4969 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4971 return PTR_ERR(ptr);
4973 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4974 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4977 #else /* !CONFIG_MMU */
4979 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4981 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4984 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4986 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4989 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4990 unsigned long addr, unsigned long len,
4991 unsigned long pgoff, unsigned long flags)
4995 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4997 return PTR_ERR(ptr);
4999 return (unsigned long) ptr;
5002 #endif /* !CONFIG_MMU */
5004 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
5005 u32, min_complete, u32, flags, const sigset_t __user *, sig,
5008 struct io_ring_ctx *ctx;
5013 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
5021 if (f.file->f_op != &io_uring_fops)
5025 ctx = f.file->private_data;
5026 if (!percpu_ref_tryget(&ctx->refs))
5030 * For SQ polling, the thread will do all submissions and completions.
5031 * Just return the requested submit count, and wake the thread if
5035 if (ctx->flags & IORING_SETUP_SQPOLL) {
5036 if (!list_empty_careful(&ctx->cq_overflow_list))
5037 io_cqring_overflow_flush(ctx, false);
5038 if (flags & IORING_ENTER_SQ_WAKEUP)
5039 wake_up(&ctx->sqo_wait);
5040 submitted = to_submit;
5041 } else if (to_submit) {
5042 struct mm_struct *cur_mm;
5044 to_submit = min(to_submit, ctx->sq_entries);
5045 mutex_lock(&ctx->uring_lock);
5046 /* already have mm, so io_submit_sqes() won't try to grab it */
5047 cur_mm = ctx->sqo_mm;
5048 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
5050 mutex_unlock(&ctx->uring_lock);
5052 if (flags & IORING_ENTER_GETEVENTS) {
5053 unsigned nr_events = 0;
5055 min_complete = min(min_complete, ctx->cq_entries);
5057 if (ctx->flags & IORING_SETUP_IOPOLL) {
5058 ret = io_iopoll_check(ctx, &nr_events, min_complete);
5060 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
5064 percpu_ref_put(&ctx->refs);
5067 return submitted ? submitted : ret;
5070 static const struct file_operations io_uring_fops = {
5071 .release = io_uring_release,
5072 .flush = io_uring_flush,
5073 .mmap = io_uring_mmap,
5075 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
5076 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
5078 .poll = io_uring_poll,
5079 .fasync = io_uring_fasync,
5082 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
5083 struct io_uring_params *p)
5085 struct io_rings *rings;
5086 size_t size, sq_array_offset;
5088 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
5089 if (size == SIZE_MAX)
5092 rings = io_mem_alloc(size);
5097 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
5098 rings->sq_ring_mask = p->sq_entries - 1;
5099 rings->cq_ring_mask = p->cq_entries - 1;
5100 rings->sq_ring_entries = p->sq_entries;
5101 rings->cq_ring_entries = p->cq_entries;
5102 ctx->sq_mask = rings->sq_ring_mask;
5103 ctx->cq_mask = rings->cq_ring_mask;
5104 ctx->sq_entries = rings->sq_ring_entries;
5105 ctx->cq_entries = rings->cq_ring_entries;
5107 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
5108 if (size == SIZE_MAX) {
5109 io_mem_free(ctx->rings);
5114 ctx->sq_sqes = io_mem_alloc(size);
5115 if (!ctx->sq_sqes) {
5116 io_mem_free(ctx->rings);
5125 * Allocate an anonymous fd, this is what constitutes the application
5126 * visible backing of an io_uring instance. The application mmaps this
5127 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
5128 * we have to tie this fd to a socket for file garbage collection purposes.
5130 static int io_uring_get_fd(struct io_ring_ctx *ctx)
5135 #if defined(CONFIG_UNIX)
5136 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
5142 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
5146 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
5147 O_RDWR | O_CLOEXEC);
5150 ret = PTR_ERR(file);
5154 #if defined(CONFIG_UNIX)
5155 ctx->ring_sock->file = file;
5156 ctx->ring_sock->sk->sk_user_data = ctx;
5158 fd_install(ret, file);
5161 #if defined(CONFIG_UNIX)
5162 sock_release(ctx->ring_sock);
5163 ctx->ring_sock = NULL;
5168 static int io_uring_create(unsigned entries, struct io_uring_params *p)
5170 struct user_struct *user = NULL;
5171 struct io_ring_ctx *ctx;
5175 if (!entries || entries > IORING_MAX_ENTRIES)
5179 * Use twice as many entries for the CQ ring. It's possible for the
5180 * application to drive a higher depth than the size of the SQ ring,
5181 * since the sqes are only used at submission time. This allows for
5182 * some flexibility in overcommitting a bit. If the application has
5183 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5184 * of CQ ring entries manually.
5186 p->sq_entries = roundup_pow_of_two(entries);
5187 if (p->flags & IORING_SETUP_CQSIZE) {
5189 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5190 * to a power-of-two, if it isn't already. We do NOT impose
5191 * any cq vs sq ring sizing.
5193 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5195 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5197 p->cq_entries = 2 * p->sq_entries;
5200 user = get_uid(current_user());
5201 account_mem = !capable(CAP_IPC_LOCK);
5204 ret = io_account_mem(user,
5205 ring_pages(p->sq_entries, p->cq_entries));
5212 ctx = io_ring_ctx_alloc(p);
5215 io_unaccount_mem(user, ring_pages(p->sq_entries,
5220 ctx->compat = in_compat_syscall();
5221 ctx->account_mem = account_mem;
5223 ctx->creds = get_current_cred();
5225 ret = io_allocate_scq_urings(ctx, p);
5229 ret = io_sq_offload_start(ctx, p);
5233 memset(&p->sq_off, 0, sizeof(p->sq_off));
5234 p->sq_off.head = offsetof(struct io_rings, sq.head);
5235 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5236 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5237 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5238 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5239 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5240 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5242 memset(&p->cq_off, 0, sizeof(p->cq_off));
5243 p->cq_off.head = offsetof(struct io_rings, cq.head);
5244 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5245 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5246 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5247 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5248 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5251 * Install ring fd as the very last thing, so we don't risk someone
5252 * having closed it before we finish setup
5254 ret = io_uring_get_fd(ctx);
5258 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5259 IORING_FEAT_SUBMIT_STABLE;
5260 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5263 io_ring_ctx_wait_and_kill(ctx);
5268 * Sets up an aio uring context, and returns the fd. Applications asks for a
5269 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5270 * params structure passed in.
5272 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5274 struct io_uring_params p;
5278 if (copy_from_user(&p, params, sizeof(p)))
5280 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5285 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5286 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5289 ret = io_uring_create(entries, &p);
5293 if (copy_to_user(params, &p, sizeof(p)))
5299 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5300 struct io_uring_params __user *, params)
5302 return io_uring_setup(entries, params);
5305 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5306 void __user *arg, unsigned nr_args)
5307 __releases(ctx->uring_lock)
5308 __acquires(ctx->uring_lock)
5313 * We're inside the ring mutex, if the ref is already dying, then
5314 * someone else killed the ctx or is already going through
5315 * io_uring_register().
5317 if (percpu_ref_is_dying(&ctx->refs))
5320 percpu_ref_kill(&ctx->refs);
5323 * Drop uring mutex before waiting for references to exit. If another
5324 * thread is currently inside io_uring_enter() it might need to grab
5325 * the uring_lock to make progress. If we hold it here across the drain
5326 * wait, then we can deadlock. It's safe to drop the mutex here, since
5327 * no new references will come in after we've killed the percpu ref.
5329 mutex_unlock(&ctx->uring_lock);
5330 wait_for_completion(&ctx->completions[0]);
5331 mutex_lock(&ctx->uring_lock);
5334 case IORING_REGISTER_BUFFERS:
5335 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5337 case IORING_UNREGISTER_BUFFERS:
5341 ret = io_sqe_buffer_unregister(ctx);
5343 case IORING_REGISTER_FILES:
5344 ret = io_sqe_files_register(ctx, arg, nr_args);
5346 case IORING_UNREGISTER_FILES:
5350 ret = io_sqe_files_unregister(ctx);
5352 case IORING_REGISTER_FILES_UPDATE:
5353 ret = io_sqe_files_update(ctx, arg, nr_args);
5355 case IORING_REGISTER_EVENTFD:
5359 ret = io_eventfd_register(ctx, arg);
5361 case IORING_UNREGISTER_EVENTFD:
5365 ret = io_eventfd_unregister(ctx);
5372 /* bring the ctx back to life */
5373 reinit_completion(&ctx->completions[0]);
5374 percpu_ref_reinit(&ctx->refs);
5378 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5379 void __user *, arg, unsigned int, nr_args)
5381 struct io_ring_ctx *ctx;
5390 if (f.file->f_op != &io_uring_fops)
5393 ctx = f.file->private_data;
5395 mutex_lock(&ctx->uring_lock);
5396 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5397 mutex_unlock(&ctx->uring_lock);
5398 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5399 ctx->cq_ev_fd != NULL, ret);
5405 static int __init io_uring_init(void)
5407 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5410 __initcall(io_uring_init);