1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/kthread.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
74 #define CREATE_TRACE_POINTS
75 #include <trace/events/io_uring.h>
77 #include <uapi/linux/io_uring.h>
82 #define IORING_MAX_ENTRIES 32768
83 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
86 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
88 #define IORING_FILE_TABLE_SHIFT 9
89 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
90 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
91 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
94 u32 head ____cacheline_aligned_in_smp;
95 u32 tail ____cacheline_aligned_in_smp;
99 * This data is shared with the application through the mmap at offsets
100 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
102 * The offsets to the member fields are published through struct
103 * io_sqring_offsets when calling io_uring_setup.
107 * Head and tail offsets into the ring; the offsets need to be
108 * masked to get valid indices.
110 * The kernel controls head of the sq ring and the tail of the cq ring,
111 * and the application controls tail of the sq ring and the head of the
114 struct io_uring sq, cq;
116 * Bitmasks to apply to head and tail offsets (constant, equals
119 u32 sq_ring_mask, cq_ring_mask;
120 /* Ring sizes (constant, power of 2) */
121 u32 sq_ring_entries, cq_ring_entries;
123 * Number of invalid entries dropped by the kernel due to
124 * invalid index stored in array
126 * Written by the kernel, shouldn't be modified by the
127 * application (i.e. get number of "new events" by comparing to
130 * After a new SQ head value was read by the application this
131 * counter includes all submissions that were dropped reaching
132 * the new SQ head (and possibly more).
138 * Written by the kernel, shouldn't be modified by the
141 * The application needs a full memory barrier before checking
142 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
146 * Number of completion events lost because the queue was full;
147 * this should be avoided by the application by making sure
148 * there are not more requests pending thatn there is space in
149 * the completion queue.
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * As completion events come in out of order this counter is not
156 * ordered with any other data.
160 * Ring buffer of completion events.
162 * The kernel writes completion events fresh every time they are
163 * produced, so the application is allowed to modify pending
166 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
169 struct io_mapped_ubuf {
172 struct bio_vec *bvec;
173 unsigned int nr_bvecs;
176 struct fixed_file_table {
182 struct percpu_ref refs;
183 } ____cacheline_aligned_in_smp;
189 bool cq_overflow_flushed;
193 * Ring buffer of indices into array of io_uring_sqe, which is
194 * mmapped by the application using the IORING_OFF_SQES offset.
196 * This indirection could e.g. be used to assign fixed
197 * io_uring_sqe entries to operations and only submit them to
198 * the queue when needed.
200 * The kernel modifies neither the indices array nor the entries
204 unsigned cached_sq_head;
207 unsigned sq_thread_idle;
208 unsigned cached_sq_dropped;
209 atomic_t cached_cq_overflow;
210 struct io_uring_sqe *sq_sqes;
212 struct list_head defer_list;
213 struct list_head timeout_list;
214 struct list_head cq_overflow_list;
216 wait_queue_head_t inflight_wait;
217 } ____cacheline_aligned_in_smp;
219 struct io_rings *rings;
223 struct task_struct *sqo_thread; /* if using sq thread polling */
224 struct mm_struct *sqo_mm;
225 wait_queue_head_t sqo_wait;
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
232 struct fixed_file_table *file_table;
233 unsigned nr_user_files;
235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
239 struct user_struct *user;
241 const struct cred *creds;
243 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
244 struct completion *completions;
246 /* if all else fails... */
247 struct io_kiocb *fallback_req;
249 #if defined(CONFIG_UNIX)
250 struct socket *ring_sock;
254 unsigned cached_cq_tail;
257 atomic_t cq_timeouts;
258 struct wait_queue_head cq_wait;
259 struct fasync_struct *cq_fasync;
260 struct eventfd_ctx *cq_ev_fd;
261 } ____cacheline_aligned_in_smp;
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
269 spinlock_t completion_lock;
270 bool poll_multi_file;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list;
278 struct rb_root cancel_tree;
280 spinlock_t inflight_lock;
281 struct list_head inflight_list;
282 } ____cacheline_aligned_in_smp;
286 * First field must be the file pointer in all the
287 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
289 struct io_poll_iocb {
291 struct wait_queue_head *head;
295 struct wait_queue_entry *wait;
298 struct io_timeout_data {
299 struct io_kiocb *req;
300 struct hrtimer timer;
301 struct timespec64 ts;
302 enum hrtimer_mode mode;
308 struct io_timeout_data *data;
312 struct iovec fast_iov[UIO_FASTIOV];
318 struct io_async_ctx {
319 struct io_uring_sqe sqe;
321 struct io_async_rw rw;
326 * NOTE! Each of the iocb union members has the file pointer
327 * as the first entry in their struct definition. So you can
328 * access the file pointer through any of the sub-structs,
329 * or directly as just 'ki_filp' in this struct.
335 struct io_poll_iocb poll;
336 struct io_timeout timeout;
339 const struct io_uring_sqe *sqe;
340 struct io_async_ctx *io;
341 struct file *ring_file;
345 bool needs_fixed_file;
347 struct io_ring_ctx *ctx;
349 struct list_head list;
350 struct rb_node rb_node;
352 struct list_head link_list;
355 #define REQ_F_NOWAIT 1 /* must not punt to workers */
356 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
357 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
358 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
359 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
360 #define REQ_F_IO_DRAINED 32 /* drain done */
361 #define REQ_F_LINK 64 /* linked sqes */
362 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
363 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
364 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
365 #define REQ_F_TIMEOUT 1024 /* timeout request */
366 #define REQ_F_ISREG 2048 /* regular file */
367 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
368 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
369 #define REQ_F_INFLIGHT 16384 /* on inflight list */
370 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
375 struct list_head inflight_entry;
377 struct io_wq_work work;
380 #define IO_PLUG_THRESHOLD 2
381 #define IO_IOPOLL_BATCH 8
383 struct io_submit_state {
384 struct blk_plug plug;
387 * io_kiocb alloc cache
389 void *reqs[IO_IOPOLL_BATCH];
390 unsigned int free_reqs;
391 unsigned int cur_req;
394 * File reference cache
398 unsigned int has_refs;
399 unsigned int used_refs;
400 unsigned int ios_left;
403 static void io_wq_submit_work(struct io_wq_work **workptr);
404 static void io_cqring_fill_event(struct io_kiocb *req, long res);
405 static void __io_free_req(struct io_kiocb *req);
406 static void io_put_req(struct io_kiocb *req);
407 static void io_double_put_req(struct io_kiocb *req);
408 static void __io_double_put_req(struct io_kiocb *req);
409 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
410 static void io_queue_linked_timeout(struct io_kiocb *req);
412 static struct kmem_cache *req_cachep;
414 static const struct file_operations io_uring_fops;
416 struct sock *io_uring_get_socket(struct file *file)
418 #if defined(CONFIG_UNIX)
419 if (file->f_op == &io_uring_fops) {
420 struct io_ring_ctx *ctx = file->private_data;
422 return ctx->ring_sock->sk;
427 EXPORT_SYMBOL(io_uring_get_socket);
429 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
431 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
433 complete(&ctx->completions[0]);
436 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
438 struct io_ring_ctx *ctx;
440 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
444 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
445 if (!ctx->fallback_req)
448 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
449 if (!ctx->completions)
452 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
453 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
456 ctx->flags = p->flags;
457 init_waitqueue_head(&ctx->cq_wait);
458 INIT_LIST_HEAD(&ctx->cq_overflow_list);
459 init_completion(&ctx->completions[0]);
460 init_completion(&ctx->completions[1]);
461 mutex_init(&ctx->uring_lock);
462 init_waitqueue_head(&ctx->wait);
463 spin_lock_init(&ctx->completion_lock);
464 INIT_LIST_HEAD(&ctx->poll_list);
465 ctx->cancel_tree = RB_ROOT;
466 INIT_LIST_HEAD(&ctx->defer_list);
467 INIT_LIST_HEAD(&ctx->timeout_list);
468 init_waitqueue_head(&ctx->inflight_wait);
469 spin_lock_init(&ctx->inflight_lock);
470 INIT_LIST_HEAD(&ctx->inflight_list);
473 if (ctx->fallback_req)
474 kmem_cache_free(req_cachep, ctx->fallback_req);
475 kfree(ctx->completions);
480 static inline bool __req_need_defer(struct io_kiocb *req)
482 struct io_ring_ctx *ctx = req->ctx;
484 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
485 + atomic_read(&ctx->cached_cq_overflow);
488 static inline bool req_need_defer(struct io_kiocb *req)
490 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
491 return __req_need_defer(req);
496 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
498 struct io_kiocb *req;
500 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
501 if (req && !req_need_defer(req)) {
502 list_del_init(&req->list);
509 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
511 struct io_kiocb *req;
513 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
515 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
517 if (!__req_need_defer(req)) {
518 list_del_init(&req->list);
526 static void __io_commit_cqring(struct io_ring_ctx *ctx)
528 struct io_rings *rings = ctx->rings;
530 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
531 /* order cqe stores with ring update */
532 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
534 if (wq_has_sleeper(&ctx->cq_wait)) {
535 wake_up_interruptible(&ctx->cq_wait);
536 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
541 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
543 u8 opcode = READ_ONCE(sqe->opcode);
545 return !(opcode == IORING_OP_READ_FIXED ||
546 opcode == IORING_OP_WRITE_FIXED);
549 static inline bool io_prep_async_work(struct io_kiocb *req,
550 struct io_kiocb **link)
552 bool do_hashed = false;
555 switch (req->sqe->opcode) {
556 case IORING_OP_WRITEV:
557 case IORING_OP_WRITE_FIXED:
560 case IORING_OP_READV:
561 case IORING_OP_READ_FIXED:
562 case IORING_OP_SENDMSG:
563 case IORING_OP_RECVMSG:
564 case IORING_OP_ACCEPT:
565 case IORING_OP_POLL_ADD:
566 case IORING_OP_CONNECT:
568 * We know REQ_F_ISREG is not set on some of these
569 * opcodes, but this enables us to keep the check in
572 if (!(req->flags & REQ_F_ISREG))
573 req->work.flags |= IO_WQ_WORK_UNBOUND;
576 if (io_sqe_needs_user(req->sqe))
577 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
580 *link = io_prep_linked_timeout(req);
584 static inline void io_queue_async_work(struct io_kiocb *req)
586 struct io_ring_ctx *ctx = req->ctx;
587 struct io_kiocb *link;
590 do_hashed = io_prep_async_work(req, &link);
592 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
595 io_wq_enqueue(ctx->io_wq, &req->work);
597 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
598 file_inode(req->file));
602 io_queue_linked_timeout(link);
605 static void io_kill_timeout(struct io_kiocb *req)
609 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
611 atomic_inc(&req->ctx->cq_timeouts);
612 list_del_init(&req->list);
613 io_cqring_fill_event(req, 0);
618 static void io_kill_timeouts(struct io_ring_ctx *ctx)
620 struct io_kiocb *req, *tmp;
622 spin_lock_irq(&ctx->completion_lock);
623 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
624 io_kill_timeout(req);
625 spin_unlock_irq(&ctx->completion_lock);
628 static void io_commit_cqring(struct io_ring_ctx *ctx)
630 struct io_kiocb *req;
632 while ((req = io_get_timeout_req(ctx)) != NULL)
633 io_kill_timeout(req);
635 __io_commit_cqring(ctx);
637 while ((req = io_get_deferred_req(ctx)) != NULL) {
638 req->flags |= REQ_F_IO_DRAINED;
639 io_queue_async_work(req);
643 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
645 struct io_rings *rings = ctx->rings;
648 tail = ctx->cached_cq_tail;
650 * writes to the cq entry need to come after reading head; the
651 * control dependency is enough as we're using WRITE_ONCE to
654 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
657 ctx->cached_cq_tail++;
658 return &rings->cqes[tail & ctx->cq_mask];
661 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
663 if (waitqueue_active(&ctx->wait))
665 if (waitqueue_active(&ctx->sqo_wait))
666 wake_up(&ctx->sqo_wait);
668 eventfd_signal(ctx->cq_ev_fd, 1);
671 /* Returns true if there are no backlogged entries after the flush */
672 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
674 struct io_rings *rings = ctx->rings;
675 struct io_uring_cqe *cqe;
676 struct io_kiocb *req;
681 if (list_empty_careful(&ctx->cq_overflow_list))
683 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
684 rings->cq_ring_entries))
688 spin_lock_irqsave(&ctx->completion_lock, flags);
690 /* if force is set, the ring is going away. always drop after that */
692 ctx->cq_overflow_flushed = true;
695 while (!list_empty(&ctx->cq_overflow_list)) {
696 cqe = io_get_cqring(ctx);
700 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
702 list_move(&req->list, &list);
704 WRITE_ONCE(cqe->user_data, req->user_data);
705 WRITE_ONCE(cqe->res, req->result);
706 WRITE_ONCE(cqe->flags, 0);
708 WRITE_ONCE(ctx->rings->cq_overflow,
709 atomic_inc_return(&ctx->cached_cq_overflow));
713 io_commit_cqring(ctx);
714 spin_unlock_irqrestore(&ctx->completion_lock, flags);
715 io_cqring_ev_posted(ctx);
717 while (!list_empty(&list)) {
718 req = list_first_entry(&list, struct io_kiocb, list);
719 list_del(&req->list);
726 static void io_cqring_fill_event(struct io_kiocb *req, long res)
728 struct io_ring_ctx *ctx = req->ctx;
729 struct io_uring_cqe *cqe;
731 trace_io_uring_complete(ctx, req->user_data, res);
734 * If we can't get a cq entry, userspace overflowed the
735 * submission (by quite a lot). Increment the overflow count in
738 cqe = io_get_cqring(ctx);
740 WRITE_ONCE(cqe->user_data, req->user_data);
741 WRITE_ONCE(cqe->res, res);
742 WRITE_ONCE(cqe->flags, 0);
743 } else if (ctx->cq_overflow_flushed) {
744 WRITE_ONCE(ctx->rings->cq_overflow,
745 atomic_inc_return(&ctx->cached_cq_overflow));
747 refcount_inc(&req->refs);
749 list_add_tail(&req->list, &ctx->cq_overflow_list);
753 static void io_cqring_add_event(struct io_kiocb *req, long res)
755 struct io_ring_ctx *ctx = req->ctx;
758 spin_lock_irqsave(&ctx->completion_lock, flags);
759 io_cqring_fill_event(req, res);
760 io_commit_cqring(ctx);
761 spin_unlock_irqrestore(&ctx->completion_lock, flags);
763 io_cqring_ev_posted(ctx);
766 static inline bool io_is_fallback_req(struct io_kiocb *req)
768 return req == (struct io_kiocb *)
769 ((unsigned long) req->ctx->fallback_req & ~1UL);
772 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
774 struct io_kiocb *req;
776 req = ctx->fallback_req;
777 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
783 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
784 struct io_submit_state *state)
786 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
787 struct io_kiocb *req;
789 if (!percpu_ref_tryget(&ctx->refs))
793 req = kmem_cache_alloc(req_cachep, gfp);
796 } else if (!state->free_reqs) {
800 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
801 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
804 * Bulk alloc is all-or-nothing. If we fail to get a batch,
805 * retry single alloc to be on the safe side.
807 if (unlikely(ret <= 0)) {
808 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
813 state->free_reqs = ret - 1;
815 req = state->reqs[0];
817 req = state->reqs[state->cur_req];
824 req->ring_file = NULL;
828 /* one is dropped after submission, the other at completion */
829 refcount_set(&req->refs, 2);
831 INIT_IO_WORK(&req->work, io_wq_submit_work);
834 req = io_get_fallback_req(ctx);
837 percpu_ref_put(&ctx->refs);
841 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
844 kmem_cache_free_bulk(req_cachep, *nr, reqs);
845 percpu_ref_put_many(&ctx->refs, *nr);
850 static void __io_free_req(struct io_kiocb *req)
852 struct io_ring_ctx *ctx = req->ctx;
856 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
858 if (req->flags & REQ_F_INFLIGHT) {
861 spin_lock_irqsave(&ctx->inflight_lock, flags);
862 list_del(&req->inflight_entry);
863 if (waitqueue_active(&ctx->inflight_wait))
864 wake_up(&ctx->inflight_wait);
865 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
867 if (req->flags & REQ_F_TIMEOUT)
868 kfree(req->timeout.data);
869 percpu_ref_put(&ctx->refs);
870 if (likely(!io_is_fallback_req(req)))
871 kmem_cache_free(req_cachep, req);
873 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
876 static bool io_link_cancel_timeout(struct io_kiocb *req)
878 struct io_ring_ctx *ctx = req->ctx;
881 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
883 io_cqring_fill_event(req, -ECANCELED);
884 io_commit_cqring(ctx);
885 req->flags &= ~REQ_F_LINK;
893 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
895 struct io_ring_ctx *ctx = req->ctx;
896 struct io_kiocb *nxt;
897 bool wake_ev = false;
899 /* Already got next link */
900 if (req->flags & REQ_F_LINK_NEXT)
904 * The list should never be empty when we are called here. But could
905 * potentially happen if the chain is messed up, check to be on the
908 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
910 list_del_init(&nxt->list);
912 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
913 (nxt->flags & REQ_F_TIMEOUT)) {
914 wake_ev |= io_link_cancel_timeout(nxt);
915 nxt = list_first_entry_or_null(&req->link_list,
916 struct io_kiocb, list);
917 req->flags &= ~REQ_F_LINK_TIMEOUT;
920 if (!list_empty(&req->link_list)) {
921 INIT_LIST_HEAD(&nxt->link_list);
922 list_splice(&req->link_list, &nxt->link_list);
923 nxt->flags |= REQ_F_LINK;
930 req->flags |= REQ_F_LINK_NEXT;
932 io_cqring_ev_posted(ctx);
936 * Called if REQ_F_LINK is set, and we fail the head request
938 static void io_fail_links(struct io_kiocb *req)
940 struct io_ring_ctx *ctx = req->ctx;
941 struct io_kiocb *link;
944 spin_lock_irqsave(&ctx->completion_lock, flags);
946 while (!list_empty(&req->link_list)) {
947 link = list_first_entry(&req->link_list, struct io_kiocb, list);
948 list_del_init(&link->list);
950 trace_io_uring_fail_link(req, link);
952 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
953 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
954 io_link_cancel_timeout(link);
956 io_cqring_fill_event(link, -ECANCELED);
957 __io_double_put_req(link);
959 req->flags &= ~REQ_F_LINK_TIMEOUT;
962 io_commit_cqring(ctx);
963 spin_unlock_irqrestore(&ctx->completion_lock, flags);
964 io_cqring_ev_posted(ctx);
967 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
969 if (likely(!(req->flags & REQ_F_LINK)))
973 * If LINK is set, we have dependent requests in this chain. If we
974 * didn't fail this request, queue the first one up, moving any other
975 * dependencies to the next request. In case of failure, fail the rest
978 if (req->flags & REQ_F_FAIL_LINK) {
980 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
981 REQ_F_LINK_TIMEOUT) {
982 struct io_ring_ctx *ctx = req->ctx;
986 * If this is a timeout link, we could be racing with the
987 * timeout timer. Grab the completion lock for this case to
988 * protect against that.
990 spin_lock_irqsave(&ctx->completion_lock, flags);
991 io_req_link_next(req, nxt);
992 spin_unlock_irqrestore(&ctx->completion_lock, flags);
994 io_req_link_next(req, nxt);
998 static void io_free_req(struct io_kiocb *req)
1000 struct io_kiocb *nxt = NULL;
1002 io_req_find_next(req, &nxt);
1006 io_queue_async_work(nxt);
1010 * Drop reference to request, return next in chain (if there is one) if this
1011 * was the last reference to this request.
1013 __attribute__((nonnull))
1014 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1016 io_req_find_next(req, nxtptr);
1018 if (refcount_dec_and_test(&req->refs))
1022 static void io_put_req(struct io_kiocb *req)
1024 if (refcount_dec_and_test(&req->refs))
1029 * Must only be used if we don't need to care about links, usually from
1030 * within the completion handling itself.
1032 static void __io_double_put_req(struct io_kiocb *req)
1034 /* drop both submit and complete references */
1035 if (refcount_sub_and_test(2, &req->refs))
1039 static void io_double_put_req(struct io_kiocb *req)
1041 /* drop both submit and complete references */
1042 if (refcount_sub_and_test(2, &req->refs))
1046 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1048 struct io_rings *rings = ctx->rings;
1051 * noflush == true is from the waitqueue handler, just ensure we wake
1052 * up the task, and the next invocation will flush the entries. We
1053 * cannot safely to it from here.
1055 if (noflush && !list_empty(&ctx->cq_overflow_list))
1058 io_cqring_overflow_flush(ctx, false);
1060 /* See comment at the top of this file */
1062 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1065 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1067 struct io_rings *rings = ctx->rings;
1069 /* make sure SQ entry isn't read before tail */
1070 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1074 * Find and free completed poll iocbs
1076 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1077 struct list_head *done)
1079 void *reqs[IO_IOPOLL_BATCH];
1080 struct io_kiocb *req;
1084 while (!list_empty(done)) {
1085 req = list_first_entry(done, struct io_kiocb, list);
1086 list_del(&req->list);
1088 io_cqring_fill_event(req, req->result);
1091 if (refcount_dec_and_test(&req->refs)) {
1092 /* If we're not using fixed files, we have to pair the
1093 * completion part with the file put. Use regular
1094 * completions for those, only batch free for fixed
1095 * file and non-linked commands.
1097 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1098 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1100 reqs[to_free++] = req;
1101 if (to_free == ARRAY_SIZE(reqs))
1102 io_free_req_many(ctx, reqs, &to_free);
1109 io_commit_cqring(ctx);
1110 io_free_req_many(ctx, reqs, &to_free);
1113 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1116 struct io_kiocb *req, *tmp;
1122 * Only spin for completions if we don't have multiple devices hanging
1123 * off our complete list, and we're under the requested amount.
1125 spin = !ctx->poll_multi_file && *nr_events < min;
1128 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1129 struct kiocb *kiocb = &req->rw;
1132 * Move completed entries to our local list. If we find a
1133 * request that requires polling, break out and complete
1134 * the done list first, if we have entries there.
1136 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1137 list_move_tail(&req->list, &done);
1140 if (!list_empty(&done))
1143 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1152 if (!list_empty(&done))
1153 io_iopoll_complete(ctx, nr_events, &done);
1159 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
1160 * non-spinning poll check - we'll still enter the driver poll loop, but only
1161 * as a non-spinning completion check.
1163 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1166 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1169 ret = io_do_iopoll(ctx, nr_events, min);
1172 if (!min || *nr_events >= min)
1180 * We can't just wait for polled events to come to us, we have to actively
1181 * find and complete them.
1183 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1185 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1188 mutex_lock(&ctx->uring_lock);
1189 while (!list_empty(&ctx->poll_list)) {
1190 unsigned int nr_events = 0;
1192 io_iopoll_getevents(ctx, &nr_events, 1);
1195 * Ensure we allow local-to-the-cpu processing to take place,
1196 * in this case we need to ensure that we reap all events.
1200 mutex_unlock(&ctx->uring_lock);
1203 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1206 int iters = 0, ret = 0;
1212 * Don't enter poll loop if we already have events pending.
1213 * If we do, we can potentially be spinning for commands that
1214 * already triggered a CQE (eg in error).
1216 if (io_cqring_events(ctx, false))
1220 * If a submit got punted to a workqueue, we can have the
1221 * application entering polling for a command before it gets
1222 * issued. That app will hold the uring_lock for the duration
1223 * of the poll right here, so we need to take a breather every
1224 * now and then to ensure that the issue has a chance to add
1225 * the poll to the issued list. Otherwise we can spin here
1226 * forever, while the workqueue is stuck trying to acquire the
1229 if (!(++iters & 7)) {
1230 mutex_unlock(&ctx->uring_lock);
1231 mutex_lock(&ctx->uring_lock);
1234 if (*nr_events < min)
1235 tmin = min - *nr_events;
1237 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1241 } while (min && !*nr_events && !need_resched());
1246 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1252 * We disallow the app entering submit/complete with polling, but we
1253 * still need to lock the ring to prevent racing with polled issue
1254 * that got punted to a workqueue.
1256 mutex_lock(&ctx->uring_lock);
1257 ret = __io_iopoll_check(ctx, nr_events, min);
1258 mutex_unlock(&ctx->uring_lock);
1262 static void kiocb_end_write(struct io_kiocb *req)
1265 * Tell lockdep we inherited freeze protection from submission
1268 if (req->flags & REQ_F_ISREG) {
1269 struct inode *inode = file_inode(req->file);
1271 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1273 file_end_write(req->file);
1276 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1278 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1280 if (kiocb->ki_flags & IOCB_WRITE)
1281 kiocb_end_write(req);
1283 if ((req->flags & REQ_F_LINK) && res != req->result)
1284 req->flags |= REQ_F_FAIL_LINK;
1285 io_cqring_add_event(req, res);
1288 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1290 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1292 io_complete_rw_common(kiocb, res);
1296 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1298 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1299 struct io_kiocb *nxt = NULL;
1301 io_complete_rw_common(kiocb, res);
1302 io_put_req_find_next(req, &nxt);
1307 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1309 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1311 if (kiocb->ki_flags & IOCB_WRITE)
1312 kiocb_end_write(req);
1314 if ((req->flags & REQ_F_LINK) && res != req->result)
1315 req->flags |= REQ_F_FAIL_LINK;
1318 req->flags |= REQ_F_IOPOLL_COMPLETED;
1322 * After the iocb has been issued, it's safe to be found on the poll list.
1323 * Adding the kiocb to the list AFTER submission ensures that we don't
1324 * find it from a io_iopoll_getevents() thread before the issuer is done
1325 * accessing the kiocb cookie.
1327 static void io_iopoll_req_issued(struct io_kiocb *req)
1329 struct io_ring_ctx *ctx = req->ctx;
1332 * Track whether we have multiple files in our lists. This will impact
1333 * how we do polling eventually, not spinning if we're on potentially
1334 * different devices.
1336 if (list_empty(&ctx->poll_list)) {
1337 ctx->poll_multi_file = false;
1338 } else if (!ctx->poll_multi_file) {
1339 struct io_kiocb *list_req;
1341 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1343 if (list_req->rw.ki_filp != req->rw.ki_filp)
1344 ctx->poll_multi_file = true;
1348 * For fast devices, IO may have already completed. If it has, add
1349 * it to the front so we find it first.
1351 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1352 list_add(&req->list, &ctx->poll_list);
1354 list_add_tail(&req->list, &ctx->poll_list);
1357 static void io_file_put(struct io_submit_state *state)
1360 int diff = state->has_refs - state->used_refs;
1363 fput_many(state->file, diff);
1369 * Get as many references to a file as we have IOs left in this submission,
1370 * assuming most submissions are for one file, or at least that each file
1371 * has more than one submission.
1373 static struct file *io_file_get(struct io_submit_state *state, int fd)
1379 if (state->fd == fd) {
1386 state->file = fget_many(fd, state->ios_left);
1391 state->has_refs = state->ios_left;
1392 state->used_refs = 1;
1398 * If we tracked the file through the SCM inflight mechanism, we could support
1399 * any file. For now, just ensure that anything potentially problematic is done
1402 static bool io_file_supports_async(struct file *file)
1404 umode_t mode = file_inode(file)->i_mode;
1406 if (S_ISBLK(mode) || S_ISCHR(mode))
1408 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1414 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1416 const struct io_uring_sqe *sqe = req->sqe;
1417 struct io_ring_ctx *ctx = req->ctx;
1418 struct kiocb *kiocb = &req->rw;
1425 if (S_ISREG(file_inode(req->file)->i_mode))
1426 req->flags |= REQ_F_ISREG;
1428 kiocb->ki_pos = READ_ONCE(sqe->off);
1429 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1430 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1432 ioprio = READ_ONCE(sqe->ioprio);
1434 ret = ioprio_check_cap(ioprio);
1438 kiocb->ki_ioprio = ioprio;
1440 kiocb->ki_ioprio = get_current_ioprio();
1442 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1446 /* don't allow async punt if RWF_NOWAIT was requested */
1447 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1448 (req->file->f_flags & O_NONBLOCK))
1449 req->flags |= REQ_F_NOWAIT;
1452 kiocb->ki_flags |= IOCB_NOWAIT;
1454 if (ctx->flags & IORING_SETUP_IOPOLL) {
1455 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1456 !kiocb->ki_filp->f_op->iopoll)
1459 kiocb->ki_flags |= IOCB_HIPRI;
1460 kiocb->ki_complete = io_complete_rw_iopoll;
1463 if (kiocb->ki_flags & IOCB_HIPRI)
1465 kiocb->ki_complete = io_complete_rw;
1470 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1476 case -ERESTARTNOINTR:
1477 case -ERESTARTNOHAND:
1478 case -ERESTART_RESTARTBLOCK:
1480 * We can't just restart the syscall, since previously
1481 * submitted sqes may already be in progress. Just fail this
1487 kiocb->ki_complete(kiocb, ret, 0);
1491 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1494 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1495 *nxt = __io_complete_rw(kiocb, ret);
1497 io_rw_done(kiocb, ret);
1500 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1501 const struct io_uring_sqe *sqe,
1502 struct iov_iter *iter)
1504 size_t len = READ_ONCE(sqe->len);
1505 struct io_mapped_ubuf *imu;
1506 unsigned index, buf_index;
1510 /* attempt to use fixed buffers without having provided iovecs */
1511 if (unlikely(!ctx->user_bufs))
1514 buf_index = READ_ONCE(sqe->buf_index);
1515 if (unlikely(buf_index >= ctx->nr_user_bufs))
1518 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1519 imu = &ctx->user_bufs[index];
1520 buf_addr = READ_ONCE(sqe->addr);
1523 if (buf_addr + len < buf_addr)
1525 /* not inside the mapped region */
1526 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1530 * May not be a start of buffer, set size appropriately
1531 * and advance us to the beginning.
1533 offset = buf_addr - imu->ubuf;
1534 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1538 * Don't use iov_iter_advance() here, as it's really slow for
1539 * using the latter parts of a big fixed buffer - it iterates
1540 * over each segment manually. We can cheat a bit here, because
1543 * 1) it's a BVEC iter, we set it up
1544 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1545 * first and last bvec
1547 * So just find our index, and adjust the iterator afterwards.
1548 * If the offset is within the first bvec (or the whole first
1549 * bvec, just use iov_iter_advance(). This makes it easier
1550 * since we can just skip the first segment, which may not
1551 * be PAGE_SIZE aligned.
1553 const struct bio_vec *bvec = imu->bvec;
1555 if (offset <= bvec->bv_len) {
1556 iov_iter_advance(iter, offset);
1558 unsigned long seg_skip;
1560 /* skip first vec */
1561 offset -= bvec->bv_len;
1562 seg_skip = 1 + (offset >> PAGE_SHIFT);
1564 iter->bvec = bvec + seg_skip;
1565 iter->nr_segs -= seg_skip;
1566 iter->count -= bvec->bv_len + offset;
1567 iter->iov_offset = offset & ~PAGE_MASK;
1574 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1575 struct iovec **iovec, struct iov_iter *iter)
1577 const struct io_uring_sqe *sqe = req->sqe;
1578 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1579 size_t sqe_len = READ_ONCE(sqe->len);
1583 * We're reading ->opcode for the second time, but the first read
1584 * doesn't care whether it's _FIXED or not, so it doesn't matter
1585 * whether ->opcode changes concurrently. The first read does care
1586 * about whether it is a READ or a WRITE, so we don't trust this read
1587 * for that purpose and instead let the caller pass in the read/write
1590 opcode = READ_ONCE(sqe->opcode);
1591 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1593 return io_import_fixed(req->ctx, rw, sqe, iter);
1597 struct io_async_rw *iorw = &req->io->rw;
1600 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1601 if (iorw->iov == iorw->fast_iov)
1609 #ifdef CONFIG_COMPAT
1610 if (req->ctx->compat)
1611 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1615 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1619 * For files that don't have ->read_iter() and ->write_iter(), handle them
1620 * by looping over ->read() or ->write() manually.
1622 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1623 struct iov_iter *iter)
1628 * Don't support polled IO through this interface, and we can't
1629 * support non-blocking either. For the latter, this just causes
1630 * the kiocb to be handled from an async context.
1632 if (kiocb->ki_flags & IOCB_HIPRI)
1634 if (kiocb->ki_flags & IOCB_NOWAIT)
1637 while (iov_iter_count(iter)) {
1641 if (!iov_iter_is_bvec(iter)) {
1642 iovec = iov_iter_iovec(iter);
1644 /* fixed buffers import bvec */
1645 iovec.iov_base = kmap(iter->bvec->bv_page)
1647 iovec.iov_len = min(iter->count,
1648 iter->bvec->bv_len - iter->iov_offset);
1652 nr = file->f_op->read(file, iovec.iov_base,
1653 iovec.iov_len, &kiocb->ki_pos);
1655 nr = file->f_op->write(file, iovec.iov_base,
1656 iovec.iov_len, &kiocb->ki_pos);
1659 if (iov_iter_is_bvec(iter))
1660 kunmap(iter->bvec->bv_page);
1668 if (nr != iovec.iov_len)
1670 iov_iter_advance(iter, nr);
1676 static void io_req_map_io(struct io_kiocb *req, ssize_t io_size,
1677 struct iovec *iovec, struct iovec *fast_iov,
1678 struct iov_iter *iter)
1680 req->io->rw.nr_segs = iter->nr_segs;
1681 req->io->rw.size = io_size;
1682 req->io->rw.iov = iovec;
1683 if (!req->io->rw.iov) {
1684 req->io->rw.iov = req->io->rw.fast_iov;
1685 memcpy(req->io->rw.iov, fast_iov,
1686 sizeof(struct iovec) * iter->nr_segs);
1690 static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size,
1691 struct iovec *iovec, struct iovec *fast_iov,
1692 struct iov_iter *iter)
1694 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1696 io_req_map_io(req, io_size, iovec, fast_iov, iter);
1697 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1698 req->sqe = &req->io->sqe;
1705 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1706 struct iov_iter *iter, bool force_nonblock)
1710 ret = io_prep_rw(req, force_nonblock);
1714 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1717 return io_import_iovec(READ, req, iovec, iter);
1720 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1721 bool force_nonblock)
1723 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1724 struct kiocb *kiocb = &req->rw;
1725 struct iov_iter iter;
1728 ssize_t io_size, ret;
1731 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1735 ret = io_import_iovec(READ, req, &iovec, &iter);
1742 if (req->flags & REQ_F_LINK)
1743 req->result = io_size;
1746 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1747 * we know to async punt it even if it was opened O_NONBLOCK
1749 if (force_nonblock && !io_file_supports_async(file)) {
1750 req->flags |= REQ_F_MUST_PUNT;
1754 iov_count = iov_iter_count(&iter);
1755 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1759 if (file->f_op->read_iter)
1760 ret2 = call_read_iter(file, kiocb, &iter);
1762 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1765 * In case of a short read, punt to async. This can happen
1766 * if we have data partially cached. Alternatively we can
1767 * return the short read, in which case the application will
1768 * need to issue another SQE and wait for it. That SQE will
1769 * need async punt anyway, so it's more efficient to do it
1772 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1773 (req->flags & REQ_F_ISREG) &&
1774 ret2 > 0 && ret2 < io_size)
1776 /* Catch -EAGAIN return for forced non-blocking submission */
1777 if (!force_nonblock || ret2 != -EAGAIN) {
1778 kiocb_done(kiocb, ret2, nxt, req->in_async);
1781 ret = io_setup_async_io(req, io_size, iovec,
1782 inline_vecs, &iter);
1793 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1794 struct iov_iter *iter, bool force_nonblock)
1798 ret = io_prep_rw(req, force_nonblock);
1802 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1805 return io_import_iovec(WRITE, req, iovec, iter);
1808 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1809 bool force_nonblock)
1811 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1812 struct kiocb *kiocb = &req->rw;
1813 struct iov_iter iter;
1816 ssize_t ret, io_size;
1819 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1823 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1828 file = kiocb->ki_filp;
1830 if (req->flags & REQ_F_LINK)
1831 req->result = io_size;
1834 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1835 * we know to async punt it even if it was opened O_NONBLOCK
1837 if (force_nonblock && !io_file_supports_async(req->file)) {
1838 req->flags |= REQ_F_MUST_PUNT;
1842 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
1845 iov_count = iov_iter_count(&iter);
1846 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1851 * Open-code file_start_write here to grab freeze protection,
1852 * which will be released by another thread in
1853 * io_complete_rw(). Fool lockdep by telling it the lock got
1854 * released so that it doesn't complain about the held lock when
1855 * we return to userspace.
1857 if (req->flags & REQ_F_ISREG) {
1858 __sb_start_write(file_inode(file)->i_sb,
1859 SB_FREEZE_WRITE, true);
1860 __sb_writers_release(file_inode(file)->i_sb,
1863 kiocb->ki_flags |= IOCB_WRITE;
1865 if (file->f_op->write_iter)
1866 ret2 = call_write_iter(file, kiocb, &iter);
1868 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1869 if (!force_nonblock || ret2 != -EAGAIN) {
1870 kiocb_done(kiocb, ret2, nxt, req->in_async);
1873 ret = io_setup_async_io(req, io_size, iovec,
1874 inline_vecs, &iter);
1886 * IORING_OP_NOP just posts a completion event, nothing else.
1888 static int io_nop(struct io_kiocb *req)
1890 struct io_ring_ctx *ctx = req->ctx;
1892 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1895 io_cqring_add_event(req, 0);
1900 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1902 struct io_ring_ctx *ctx = req->ctx;
1907 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1909 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1915 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1916 struct io_kiocb **nxt, bool force_nonblock)
1918 loff_t sqe_off = READ_ONCE(sqe->off);
1919 loff_t sqe_len = READ_ONCE(sqe->len);
1920 loff_t end = sqe_off + sqe_len;
1921 unsigned fsync_flags;
1924 fsync_flags = READ_ONCE(sqe->fsync_flags);
1925 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1928 ret = io_prep_fsync(req, sqe);
1932 /* fsync always requires a blocking context */
1936 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1937 end > 0 ? end : LLONG_MAX,
1938 fsync_flags & IORING_FSYNC_DATASYNC);
1940 if (ret < 0 && (req->flags & REQ_F_LINK))
1941 req->flags |= REQ_F_FAIL_LINK;
1942 io_cqring_add_event(req, ret);
1943 io_put_req_find_next(req, nxt);
1947 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1949 struct io_ring_ctx *ctx = req->ctx;
1955 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1957 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1963 static int io_sync_file_range(struct io_kiocb *req,
1964 const struct io_uring_sqe *sqe,
1965 struct io_kiocb **nxt,
1966 bool force_nonblock)
1973 ret = io_prep_sfr(req, sqe);
1977 /* sync_file_range always requires a blocking context */
1981 sqe_off = READ_ONCE(sqe->off);
1982 sqe_len = READ_ONCE(sqe->len);
1983 flags = READ_ONCE(sqe->sync_range_flags);
1985 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
1987 if (ret < 0 && (req->flags & REQ_F_LINK))
1988 req->flags |= REQ_F_FAIL_LINK;
1989 io_cqring_add_event(req, ret);
1990 io_put_req_find_next(req, nxt);
1994 #if defined(CONFIG_NET)
1995 static int io_send_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1996 struct io_kiocb **nxt, bool force_nonblock,
1997 long (*fn)(struct socket *, struct user_msghdr __user *,
2000 struct socket *sock;
2003 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2006 sock = sock_from_file(req->file, &ret);
2008 struct user_msghdr __user *msg;
2011 flags = READ_ONCE(sqe->msg_flags);
2012 if (flags & MSG_DONTWAIT)
2013 req->flags |= REQ_F_NOWAIT;
2014 else if (force_nonblock)
2015 flags |= MSG_DONTWAIT;
2017 msg = (struct user_msghdr __user *) (unsigned long)
2018 READ_ONCE(sqe->addr);
2020 ret = fn(sock, msg, flags);
2021 if (force_nonblock && ret == -EAGAIN)
2023 if (ret == -ERESTARTSYS)
2027 io_cqring_add_event(req, ret);
2028 if (ret < 0 && (req->flags & REQ_F_LINK))
2029 req->flags |= REQ_F_FAIL_LINK;
2030 io_put_req_find_next(req, nxt);
2035 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2036 struct io_kiocb **nxt, bool force_nonblock)
2038 #if defined(CONFIG_NET)
2039 return io_send_recvmsg(req, sqe, nxt, force_nonblock,
2040 __sys_sendmsg_sock);
2046 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2047 struct io_kiocb **nxt, bool force_nonblock)
2049 #if defined(CONFIG_NET)
2050 return io_send_recvmsg(req, sqe, nxt, force_nonblock,
2051 __sys_recvmsg_sock);
2057 static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2058 struct io_kiocb **nxt, bool force_nonblock)
2060 #if defined(CONFIG_NET)
2061 struct sockaddr __user *addr;
2062 int __user *addr_len;
2063 unsigned file_flags;
2066 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2068 if (sqe->ioprio || sqe->len || sqe->buf_index)
2071 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2072 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2073 flags = READ_ONCE(sqe->accept_flags);
2074 file_flags = force_nonblock ? O_NONBLOCK : 0;
2076 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
2077 if (ret == -EAGAIN && force_nonblock) {
2078 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2081 if (ret == -ERESTARTSYS)
2083 if (ret < 0 && (req->flags & REQ_F_LINK))
2084 req->flags |= REQ_F_FAIL_LINK;
2085 io_cqring_add_event(req, ret);
2086 io_put_req_find_next(req, nxt);
2093 static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2094 struct io_kiocb **nxt, bool force_nonblock)
2096 #if defined(CONFIG_NET)
2097 struct sockaddr __user *addr;
2098 unsigned file_flags;
2101 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2103 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2106 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2107 addr_len = READ_ONCE(sqe->addr2);
2108 file_flags = force_nonblock ? O_NONBLOCK : 0;
2110 ret = __sys_connect_file(req->file, addr, addr_len, file_flags);
2111 if (ret == -EAGAIN && force_nonblock)
2113 if (ret == -ERESTARTSYS)
2115 if (ret < 0 && (req->flags & REQ_F_LINK))
2116 req->flags |= REQ_F_FAIL_LINK;
2117 io_cqring_add_event(req, ret);
2118 io_put_req_find_next(req, nxt);
2125 static inline void io_poll_remove_req(struct io_kiocb *req)
2127 if (!RB_EMPTY_NODE(&req->rb_node)) {
2128 rb_erase(&req->rb_node, &req->ctx->cancel_tree);
2129 RB_CLEAR_NODE(&req->rb_node);
2133 static void io_poll_remove_one(struct io_kiocb *req)
2135 struct io_poll_iocb *poll = &req->poll;
2137 spin_lock(&poll->head->lock);
2138 WRITE_ONCE(poll->canceled, true);
2139 if (!list_empty(&poll->wait->entry)) {
2140 list_del_init(&poll->wait->entry);
2141 io_queue_async_work(req);
2143 spin_unlock(&poll->head->lock);
2144 io_poll_remove_req(req);
2147 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2149 struct rb_node *node;
2150 struct io_kiocb *req;
2152 spin_lock_irq(&ctx->completion_lock);
2153 while ((node = rb_first(&ctx->cancel_tree)) != NULL) {
2154 req = rb_entry(node, struct io_kiocb, rb_node);
2155 io_poll_remove_one(req);
2157 spin_unlock_irq(&ctx->completion_lock);
2160 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2162 struct rb_node *p, *parent = NULL;
2163 struct io_kiocb *req;
2165 p = ctx->cancel_tree.rb_node;
2168 req = rb_entry(parent, struct io_kiocb, rb_node);
2169 if (sqe_addr < req->user_data) {
2171 } else if (sqe_addr > req->user_data) {
2174 io_poll_remove_one(req);
2183 * Find a running poll command that matches one specified in sqe->addr,
2184 * and remove it if found.
2186 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2188 struct io_ring_ctx *ctx = req->ctx;
2191 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2193 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2197 spin_lock_irq(&ctx->completion_lock);
2198 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
2199 spin_unlock_irq(&ctx->completion_lock);
2201 io_cqring_add_event(req, ret);
2202 if (ret < 0 && (req->flags & REQ_F_LINK))
2203 req->flags |= REQ_F_FAIL_LINK;
2208 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2210 struct io_ring_ctx *ctx = req->ctx;
2212 req->poll.done = true;
2213 kfree(req->poll.wait);
2215 io_cqring_fill_event(req, error);
2217 io_cqring_fill_event(req, mangle_poll(mask));
2218 io_commit_cqring(ctx);
2221 static void io_poll_complete_work(struct io_wq_work **workptr)
2223 struct io_wq_work *work = *workptr;
2224 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2225 struct io_poll_iocb *poll = &req->poll;
2226 struct poll_table_struct pt = { ._key = poll->events };
2227 struct io_ring_ctx *ctx = req->ctx;
2228 struct io_kiocb *nxt = NULL;
2232 if (work->flags & IO_WQ_WORK_CANCEL) {
2233 WRITE_ONCE(poll->canceled, true);
2235 } else if (READ_ONCE(poll->canceled)) {
2239 if (ret != -ECANCELED)
2240 mask = vfs_poll(poll->file, &pt) & poll->events;
2243 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2244 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2245 * synchronize with them. In the cancellation case the list_del_init
2246 * itself is not actually needed, but harmless so we keep it in to
2247 * avoid further branches in the fast path.
2249 spin_lock_irq(&ctx->completion_lock);
2250 if (!mask && ret != -ECANCELED) {
2251 add_wait_queue(poll->head, poll->wait);
2252 spin_unlock_irq(&ctx->completion_lock);
2255 io_poll_remove_req(req);
2256 io_poll_complete(req, mask, ret);
2257 spin_unlock_irq(&ctx->completion_lock);
2259 io_cqring_ev_posted(ctx);
2261 if (ret < 0 && req->flags & REQ_F_LINK)
2262 req->flags |= REQ_F_FAIL_LINK;
2263 io_put_req_find_next(req, &nxt);
2265 *workptr = &nxt->work;
2268 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2271 struct io_poll_iocb *poll = wait->private;
2272 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2273 struct io_ring_ctx *ctx = req->ctx;
2274 __poll_t mask = key_to_poll(key);
2275 unsigned long flags;
2277 /* for instances that support it check for an event match first: */
2278 if (mask && !(mask & poll->events))
2281 list_del_init(&poll->wait->entry);
2284 * Run completion inline if we can. We're using trylock here because
2285 * we are violating the completion_lock -> poll wq lock ordering.
2286 * If we have a link timeout we're going to need the completion_lock
2287 * for finalizing the request, mark us as having grabbed that already.
2289 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2290 io_poll_remove_req(req);
2291 io_poll_complete(req, mask, 0);
2292 req->flags |= REQ_F_COMP_LOCKED;
2294 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2296 io_cqring_ev_posted(ctx);
2298 io_queue_async_work(req);
2304 struct io_poll_table {
2305 struct poll_table_struct pt;
2306 struct io_kiocb *req;
2310 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2311 struct poll_table_struct *p)
2313 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2315 if (unlikely(pt->req->poll.head)) {
2316 pt->error = -EINVAL;
2321 pt->req->poll.head = head;
2322 add_wait_queue(head, pt->req->poll.wait);
2325 static void io_poll_req_insert(struct io_kiocb *req)
2327 struct io_ring_ctx *ctx = req->ctx;
2328 struct rb_node **p = &ctx->cancel_tree.rb_node;
2329 struct rb_node *parent = NULL;
2330 struct io_kiocb *tmp;
2334 tmp = rb_entry(parent, struct io_kiocb, rb_node);
2335 if (req->user_data < tmp->user_data)
2338 p = &(*p)->rb_right;
2340 rb_link_node(&req->rb_node, parent, p);
2341 rb_insert_color(&req->rb_node, &ctx->cancel_tree);
2344 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2345 struct io_kiocb **nxt)
2347 struct io_poll_iocb *poll = &req->poll;
2348 struct io_ring_ctx *ctx = req->ctx;
2349 struct io_poll_table ipt;
2350 bool cancel = false;
2354 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2356 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2361 poll->wait = kmalloc(sizeof(*poll->wait), GFP_KERNEL);
2366 INIT_IO_WORK(&req->work, io_poll_complete_work);
2367 events = READ_ONCE(sqe->poll_events);
2368 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2369 RB_CLEAR_NODE(&req->rb_node);
2373 poll->canceled = false;
2375 ipt.pt._qproc = io_poll_queue_proc;
2376 ipt.pt._key = poll->events;
2378 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2380 /* initialized the list so that we can do list_empty checks */
2381 INIT_LIST_HEAD(&poll->wait->entry);
2382 init_waitqueue_func_entry(poll->wait, io_poll_wake);
2383 poll->wait->private = poll;
2385 INIT_LIST_HEAD(&req->list);
2387 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2389 spin_lock_irq(&ctx->completion_lock);
2390 if (likely(poll->head)) {
2391 spin_lock(&poll->head->lock);
2392 if (unlikely(list_empty(&poll->wait->entry))) {
2398 if (mask || ipt.error)
2399 list_del_init(&poll->wait->entry);
2401 WRITE_ONCE(poll->canceled, true);
2402 else if (!poll->done) /* actually waiting for an event */
2403 io_poll_req_insert(req);
2404 spin_unlock(&poll->head->lock);
2406 if (mask) { /* no async, we'd stolen it */
2408 io_poll_complete(req, mask, 0);
2410 spin_unlock_irq(&ctx->completion_lock);
2413 io_cqring_ev_posted(ctx);
2414 io_put_req_find_next(req, nxt);
2419 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2421 struct io_timeout_data *data = container_of(timer,
2422 struct io_timeout_data, timer);
2423 struct io_kiocb *req = data->req;
2424 struct io_ring_ctx *ctx = req->ctx;
2425 unsigned long flags;
2427 atomic_inc(&ctx->cq_timeouts);
2429 spin_lock_irqsave(&ctx->completion_lock, flags);
2431 * We could be racing with timeout deletion. If the list is empty,
2432 * then timeout lookup already found it and will be handling it.
2434 if (!list_empty(&req->list)) {
2435 struct io_kiocb *prev;
2438 * Adjust the reqs sequence before the current one because it
2439 * will consume a slot in the cq_ring and the the cq_tail
2440 * pointer will be increased, otherwise other timeout reqs may
2441 * return in advance without waiting for enough wait_nr.
2444 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2446 list_del_init(&req->list);
2449 io_cqring_fill_event(req, -ETIME);
2450 io_commit_cqring(ctx);
2451 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2453 io_cqring_ev_posted(ctx);
2454 if (req->flags & REQ_F_LINK)
2455 req->flags |= REQ_F_FAIL_LINK;
2457 return HRTIMER_NORESTART;
2460 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2462 struct io_kiocb *req;
2465 list_for_each_entry(req, &ctx->timeout_list, list) {
2466 if (user_data == req->user_data) {
2467 list_del_init(&req->list);
2476 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
2480 if (req->flags & REQ_F_LINK)
2481 req->flags |= REQ_F_FAIL_LINK;
2482 io_cqring_fill_event(req, -ECANCELED);
2488 * Remove or update an existing timeout command
2490 static int io_timeout_remove(struct io_kiocb *req,
2491 const struct io_uring_sqe *sqe)
2493 struct io_ring_ctx *ctx = req->ctx;
2497 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2499 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2501 flags = READ_ONCE(sqe->timeout_flags);
2505 spin_lock_irq(&ctx->completion_lock);
2506 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
2508 io_cqring_fill_event(req, ret);
2509 io_commit_cqring(ctx);
2510 spin_unlock_irq(&ctx->completion_lock);
2511 io_cqring_ev_posted(ctx);
2512 if (ret < 0 && req->flags & REQ_F_LINK)
2513 req->flags |= REQ_F_FAIL_LINK;
2518 static int io_timeout_setup(struct io_kiocb *req)
2520 const struct io_uring_sqe *sqe = req->sqe;
2521 struct io_timeout_data *data;
2524 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2526 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2528 flags = READ_ONCE(sqe->timeout_flags);
2529 if (flags & ~IORING_TIMEOUT_ABS)
2532 data = kzalloc(sizeof(struct io_timeout_data), GFP_KERNEL);
2536 req->timeout.data = data;
2537 req->flags |= REQ_F_TIMEOUT;
2539 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2542 if (flags & IORING_TIMEOUT_ABS)
2543 data->mode = HRTIMER_MODE_ABS;
2545 data->mode = HRTIMER_MODE_REL;
2547 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2551 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2554 struct io_ring_ctx *ctx = req->ctx;
2555 struct io_timeout_data *data;
2556 struct list_head *entry;
2560 ret = io_timeout_setup(req);
2561 /* common setup allows flags (like links) set, we don't */
2562 if (!ret && sqe->flags)
2568 * sqe->off holds how many events that need to occur for this
2569 * timeout event to be satisfied. If it isn't set, then this is
2570 * a pure timeout request, sequence isn't used.
2572 count = READ_ONCE(sqe->off);
2574 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2575 spin_lock_irq(&ctx->completion_lock);
2576 entry = ctx->timeout_list.prev;
2580 req->sequence = ctx->cached_sq_head + count - 1;
2581 req->timeout.data->seq_offset = count;
2584 * Insertion sort, ensuring the first entry in the list is always
2585 * the one we need first.
2587 spin_lock_irq(&ctx->completion_lock);
2588 list_for_each_prev(entry, &ctx->timeout_list) {
2589 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2590 unsigned nxt_sq_head;
2591 long long tmp, tmp_nxt;
2592 u32 nxt_offset = nxt->timeout.data->seq_offset;
2594 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2598 * Since cached_sq_head + count - 1 can overflow, use type long
2601 tmp = (long long)ctx->cached_sq_head + count - 1;
2602 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2603 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2606 * cached_sq_head may overflow, and it will never overflow twice
2607 * once there is some timeout req still be valid.
2609 if (ctx->cached_sq_head < nxt_sq_head)
2616 * Sequence of reqs after the insert one and itself should
2617 * be adjusted because each timeout req consumes a slot.
2622 req->sequence -= span;
2624 list_add(&req->list, entry);
2625 data = req->timeout.data;
2626 data->timer.function = io_timeout_fn;
2627 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2628 spin_unlock_irq(&ctx->completion_lock);
2632 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2634 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2636 return req->user_data == (unsigned long) data;
2639 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2641 enum io_wq_cancel cancel_ret;
2644 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2645 switch (cancel_ret) {
2646 case IO_WQ_CANCEL_OK:
2649 case IO_WQ_CANCEL_RUNNING:
2652 case IO_WQ_CANCEL_NOTFOUND:
2660 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2661 struct io_kiocb *req, __u64 sqe_addr,
2662 struct io_kiocb **nxt, int success_ret)
2664 unsigned long flags;
2667 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2668 if (ret != -ENOENT) {
2669 spin_lock_irqsave(&ctx->completion_lock, flags);
2673 spin_lock_irqsave(&ctx->completion_lock, flags);
2674 ret = io_timeout_cancel(ctx, sqe_addr);
2677 ret = io_poll_cancel(ctx, sqe_addr);
2681 io_cqring_fill_event(req, ret);
2682 io_commit_cqring(ctx);
2683 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2684 io_cqring_ev_posted(ctx);
2686 if (ret < 0 && (req->flags & REQ_F_LINK))
2687 req->flags |= REQ_F_FAIL_LINK;
2688 io_put_req_find_next(req, nxt);
2691 static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2692 struct io_kiocb **nxt)
2694 struct io_ring_ctx *ctx = req->ctx;
2696 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2698 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2702 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
2706 static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io)
2708 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2709 struct iov_iter iter;
2712 memcpy(&io->sqe, req->sqe, sizeof(io->sqe));
2713 req->sqe = &io->sqe;
2715 switch (io->sqe.opcode) {
2716 case IORING_OP_READV:
2717 case IORING_OP_READ_FIXED:
2718 ret = io_read_prep(req, &iovec, &iter, true);
2720 case IORING_OP_WRITEV:
2721 case IORING_OP_WRITE_FIXED:
2722 ret = io_write_prep(req, &iovec, &iter, true);
2733 io_req_map_io(req, ret, iovec, inline_vecs, &iter);
2737 static int io_req_defer(struct io_kiocb *req)
2739 struct io_ring_ctx *ctx = req->ctx;
2740 struct io_async_ctx *io;
2743 /* Still need defer if there is pending req in defer list. */
2744 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
2747 io = kmalloc(sizeof(*io), GFP_KERNEL);
2751 spin_lock_irq(&ctx->completion_lock);
2752 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
2753 spin_unlock_irq(&ctx->completion_lock);
2758 ret = io_req_defer_prep(req, io);
2762 trace_io_uring_defer(ctx, req, req->user_data);
2763 list_add_tail(&req->list, &ctx->defer_list);
2764 spin_unlock_irq(&ctx->completion_lock);
2765 return -EIOCBQUEUED;
2768 __attribute__((nonnull))
2769 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
2770 bool force_nonblock)
2773 struct io_ring_ctx *ctx = req->ctx;
2775 opcode = READ_ONCE(req->sqe->opcode);
2780 case IORING_OP_READV:
2781 if (unlikely(req->sqe->buf_index))
2783 ret = io_read(req, nxt, force_nonblock);
2785 case IORING_OP_WRITEV:
2786 if (unlikely(req->sqe->buf_index))
2788 ret = io_write(req, nxt, force_nonblock);
2790 case IORING_OP_READ_FIXED:
2791 ret = io_read(req, nxt, force_nonblock);
2793 case IORING_OP_WRITE_FIXED:
2794 ret = io_write(req, nxt, force_nonblock);
2796 case IORING_OP_FSYNC:
2797 ret = io_fsync(req, req->sqe, nxt, force_nonblock);
2799 case IORING_OP_POLL_ADD:
2800 ret = io_poll_add(req, req->sqe, nxt);
2802 case IORING_OP_POLL_REMOVE:
2803 ret = io_poll_remove(req, req->sqe);
2805 case IORING_OP_SYNC_FILE_RANGE:
2806 ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock);
2808 case IORING_OP_SENDMSG:
2809 ret = io_sendmsg(req, req->sqe, nxt, force_nonblock);
2811 case IORING_OP_RECVMSG:
2812 ret = io_recvmsg(req, req->sqe, nxt, force_nonblock);
2814 case IORING_OP_TIMEOUT:
2815 ret = io_timeout(req, req->sqe);
2817 case IORING_OP_TIMEOUT_REMOVE:
2818 ret = io_timeout_remove(req, req->sqe);
2820 case IORING_OP_ACCEPT:
2821 ret = io_accept(req, req->sqe, nxt, force_nonblock);
2823 case IORING_OP_CONNECT:
2824 ret = io_connect(req, req->sqe, nxt, force_nonblock);
2826 case IORING_OP_ASYNC_CANCEL:
2827 ret = io_async_cancel(req, req->sqe, nxt);
2837 if (ctx->flags & IORING_SETUP_IOPOLL) {
2838 if (req->result == -EAGAIN)
2841 /* workqueue context doesn't hold uring_lock, grab it now */
2843 mutex_lock(&ctx->uring_lock);
2844 io_iopoll_req_issued(req);
2846 mutex_unlock(&ctx->uring_lock);
2852 static void io_link_work_cb(struct io_wq_work **workptr)
2854 struct io_wq_work *work = *workptr;
2855 struct io_kiocb *link = work->data;
2857 io_queue_linked_timeout(link);
2858 work->func = io_wq_submit_work;
2861 static void io_wq_submit_work(struct io_wq_work **workptr)
2863 struct io_wq_work *work = *workptr;
2864 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2865 struct io_kiocb *nxt = NULL;
2868 /* Ensure we clear previously set non-block flag */
2869 req->rw.ki_flags &= ~IOCB_NOWAIT;
2871 if (work->flags & IO_WQ_WORK_CANCEL)
2875 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
2876 req->in_async = true;
2878 ret = io_issue_sqe(req, &nxt, false);
2880 * We can get EAGAIN for polled IO even though we're
2881 * forcing a sync submission from here, since we can't
2882 * wait for request slots on the block side.
2890 /* drop submission reference */
2894 if (req->flags & REQ_F_LINK)
2895 req->flags |= REQ_F_FAIL_LINK;
2896 io_cqring_add_event(req, ret);
2900 /* if a dependent link is ready, pass it back */
2902 struct io_kiocb *link;
2904 io_prep_async_work(nxt, &link);
2905 *workptr = &nxt->work;
2907 nxt->work.flags |= IO_WQ_WORK_CB;
2908 nxt->work.func = io_link_work_cb;
2909 nxt->work.data = link;
2914 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
2916 int op = READ_ONCE(sqe->opcode);
2920 case IORING_OP_POLL_REMOVE:
2921 case IORING_OP_TIMEOUT:
2922 case IORING_OP_TIMEOUT_REMOVE:
2923 case IORING_OP_ASYNC_CANCEL:
2924 case IORING_OP_LINK_TIMEOUT:
2931 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
2934 struct fixed_file_table *table;
2936 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
2937 return table->files[index & IORING_FILE_TABLE_MASK];
2940 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
2942 struct io_ring_ctx *ctx = req->ctx;
2946 flags = READ_ONCE(req->sqe->flags);
2947 fd = READ_ONCE(req->sqe->fd);
2949 if (flags & IOSQE_IO_DRAIN)
2950 req->flags |= REQ_F_IO_DRAIN;
2952 if (!io_op_needs_file(req->sqe))
2955 if (flags & IOSQE_FIXED_FILE) {
2956 if (unlikely(!ctx->file_table ||
2957 (unsigned) fd >= ctx->nr_user_files))
2959 fd = array_index_nospec(fd, ctx->nr_user_files);
2960 req->file = io_file_from_index(ctx, fd);
2963 req->flags |= REQ_F_FIXED_FILE;
2965 if (req->needs_fixed_file)
2967 trace_io_uring_file_get(ctx, fd);
2968 req->file = io_file_get(state, fd);
2969 if (unlikely(!req->file))
2976 static int io_grab_files(struct io_kiocb *req)
2979 struct io_ring_ctx *ctx = req->ctx;
2982 spin_lock_irq(&ctx->inflight_lock);
2984 * We use the f_ops->flush() handler to ensure that we can flush
2985 * out work accessing these files if the fd is closed. Check if
2986 * the fd has changed since we started down this path, and disallow
2987 * this operation if it has.
2989 if (fcheck(req->ring_fd) == req->ring_file) {
2990 list_add(&req->inflight_entry, &ctx->inflight_list);
2991 req->flags |= REQ_F_INFLIGHT;
2992 req->work.files = current->files;
2995 spin_unlock_irq(&ctx->inflight_lock);
3001 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3003 struct io_timeout_data *data = container_of(timer,
3004 struct io_timeout_data, timer);
3005 struct io_kiocb *req = data->req;
3006 struct io_ring_ctx *ctx = req->ctx;
3007 struct io_kiocb *prev = NULL;
3008 unsigned long flags;
3010 spin_lock_irqsave(&ctx->completion_lock, flags);
3013 * We don't expect the list to be empty, that will only happen if we
3014 * race with the completion of the linked work.
3016 if (!list_empty(&req->list)) {
3017 prev = list_entry(req->list.prev, struct io_kiocb, link_list);
3018 if (refcount_inc_not_zero(&prev->refs)) {
3019 list_del_init(&req->list);
3020 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3025 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3028 if (prev->flags & REQ_F_LINK)
3029 prev->flags |= REQ_F_FAIL_LINK;
3030 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3034 io_cqring_add_event(req, -ETIME);
3037 return HRTIMER_NORESTART;
3040 static void io_queue_linked_timeout(struct io_kiocb *req)
3042 struct io_ring_ctx *ctx = req->ctx;
3045 * If the list is now empty, then our linked request finished before
3046 * we got a chance to setup the timer
3048 spin_lock_irq(&ctx->completion_lock);
3049 if (!list_empty(&req->list)) {
3050 struct io_timeout_data *data = req->timeout.data;
3052 data->timer.function = io_link_timeout_fn;
3053 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3056 spin_unlock_irq(&ctx->completion_lock);
3058 /* drop submission reference */
3062 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3064 struct io_kiocb *nxt;
3066 if (!(req->flags & REQ_F_LINK))
3069 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
3070 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
3073 req->flags |= REQ_F_LINK_TIMEOUT;
3077 static void __io_queue_sqe(struct io_kiocb *req)
3079 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
3080 struct io_kiocb *nxt = NULL;
3083 ret = io_issue_sqe(req, &nxt, true);
3085 io_queue_async_work(nxt);
3088 * We async punt it if the file wasn't marked NOWAIT, or if the file
3089 * doesn't support non-blocking read/write attempts
3091 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3092 (req->flags & REQ_F_MUST_PUNT))) {
3093 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3094 ret = io_grab_files(req);
3100 * Queued up for async execution, worker will release
3101 * submit reference when the iocb is actually submitted.
3103 io_queue_async_work(req);
3108 /* drop submission reference */
3111 if (linked_timeout) {
3113 io_queue_linked_timeout(linked_timeout);
3115 io_put_req(linked_timeout);
3118 /* and drop final reference, if we failed */
3120 io_cqring_add_event(req, ret);
3121 if (req->flags & REQ_F_LINK)
3122 req->flags |= REQ_F_FAIL_LINK;
3127 static void io_queue_sqe(struct io_kiocb *req)
3131 if (unlikely(req->ctx->drain_next)) {
3132 req->flags |= REQ_F_IO_DRAIN;
3133 req->ctx->drain_next = false;
3135 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3137 ret = io_req_defer(req);
3139 if (ret != -EIOCBQUEUED) {
3140 io_cqring_add_event(req, ret);
3141 if (req->flags & REQ_F_LINK)
3142 req->flags |= REQ_F_FAIL_LINK;
3143 io_double_put_req(req);
3146 __io_queue_sqe(req);
3149 static inline void io_queue_link_head(struct io_kiocb *req)
3151 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3152 io_cqring_add_event(req, -ECANCELED);
3153 io_double_put_req(req);
3159 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
3161 static void io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3162 struct io_kiocb **link)
3164 struct io_ring_ctx *ctx = req->ctx;
3167 req->user_data = req->sqe->user_data;
3169 /* enforce forwards compatibility on users */
3170 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3175 ret = io_req_set_file(state, req);
3176 if (unlikely(ret)) {
3178 io_cqring_add_event(req, ret);
3179 io_double_put_req(req);
3184 * If we already have a head request, queue this one for async
3185 * submittal once the head completes. If we don't have a head but
3186 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3187 * submitted sync once the chain is complete. If none of those
3188 * conditions are true (normal request), then just queue it.
3191 struct io_kiocb *prev = *link;
3192 struct io_async_ctx *io;
3194 if (req->sqe->flags & IOSQE_IO_DRAIN)
3195 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3197 if (READ_ONCE(req->sqe->opcode) == IORING_OP_LINK_TIMEOUT) {
3198 ret = io_timeout_setup(req);
3199 /* common setup allows offset being set, we don't */
3200 if (!ret && req->sqe->off)
3203 prev->flags |= REQ_F_FAIL_LINK;
3208 io = kmalloc(sizeof(*io), GFP_KERNEL);
3214 ret = io_req_defer_prep(req, io);
3217 trace_io_uring_link(ctx, req, prev);
3218 list_add_tail(&req->list, &prev->link_list);
3219 } else if (req->sqe->flags & IOSQE_IO_LINK) {
3220 req->flags |= REQ_F_LINK;
3222 INIT_LIST_HEAD(&req->link_list);
3230 * Batched submission is done, ensure local IO is flushed out.
3232 static void io_submit_state_end(struct io_submit_state *state)
3234 blk_finish_plug(&state->plug);
3236 if (state->free_reqs)
3237 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3238 &state->reqs[state->cur_req]);
3242 * Start submission side cache.
3244 static void io_submit_state_start(struct io_submit_state *state,
3245 struct io_ring_ctx *ctx, unsigned max_ios)
3247 blk_start_plug(&state->plug);
3248 state->free_reqs = 0;
3250 state->ios_left = max_ios;
3253 static void io_commit_sqring(struct io_ring_ctx *ctx)
3255 struct io_rings *rings = ctx->rings;
3257 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3259 * Ensure any loads from the SQEs are done at this point,
3260 * since once we write the new head, the application could
3261 * write new data to them.
3263 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3268 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
3269 * that is mapped by userspace. This means that care needs to be taken to
3270 * ensure that reads are stable, as we cannot rely on userspace always
3271 * being a good citizen. If members of the sqe are validated and then later
3272 * used, it's important that those reads are done through READ_ONCE() to
3273 * prevent a re-load down the line.
3275 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3277 struct io_rings *rings = ctx->rings;
3278 u32 *sq_array = ctx->sq_array;
3282 * The cached sq head (or cq tail) serves two purposes:
3284 * 1) allows us to batch the cost of updating the user visible
3286 * 2) allows the kernel side to track the head on its own, even
3287 * though the application is the one updating it.
3289 head = ctx->cached_sq_head;
3290 /* make sure SQ entry isn't read before tail */
3291 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3294 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3295 if (likely(head < ctx->sq_entries)) {
3297 * All io need record the previous position, if LINK vs DARIN,
3298 * it can be used to mark the position of the first IO in the
3301 req->sequence = ctx->cached_sq_head;
3302 req->sqe = &ctx->sq_sqes[head];
3303 ctx->cached_sq_head++;
3307 /* drop invalid entries */
3308 ctx->cached_sq_head++;
3309 ctx->cached_sq_dropped++;
3310 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3314 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3315 struct file *ring_file, int ring_fd,
3316 struct mm_struct **mm, bool async)
3318 struct io_submit_state state, *statep = NULL;
3319 struct io_kiocb *link = NULL;
3320 int i, submitted = 0;
3321 bool mm_fault = false;
3323 /* if we have a backlog and couldn't flush it all, return BUSY */
3324 if (!list_empty(&ctx->cq_overflow_list) &&
3325 !io_cqring_overflow_flush(ctx, false))
3328 if (nr > IO_PLUG_THRESHOLD) {
3329 io_submit_state_start(&state, ctx, nr);
3333 for (i = 0; i < nr; i++) {
3334 struct io_kiocb *req;
3335 unsigned int sqe_flags;
3337 req = io_get_req(ctx, statep);
3338 if (unlikely(!req)) {
3340 submitted = -EAGAIN;
3343 if (!io_get_sqring(ctx, req)) {
3348 if (io_sqe_needs_user(req->sqe) && !*mm) {
3349 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3351 use_mm(ctx->sqo_mm);
3356 sqe_flags = req->sqe->flags;
3358 req->ring_file = ring_file;
3359 req->ring_fd = ring_fd;
3360 req->has_user = *mm != NULL;
3361 req->in_async = async;
3362 req->needs_fixed_file = async;
3363 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
3365 io_submit_sqe(req, statep, &link);
3369 * If previous wasn't linked and we have a linked command,
3370 * that's the end of the chain. Submit the previous link.
3372 if (!(sqe_flags & IOSQE_IO_LINK) && link) {
3373 io_queue_link_head(link);
3379 io_queue_link_head(link);
3381 io_submit_state_end(&state);
3383 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3384 io_commit_sqring(ctx);
3389 static int io_sq_thread(void *data)
3391 struct io_ring_ctx *ctx = data;
3392 struct mm_struct *cur_mm = NULL;
3393 const struct cred *old_cred;
3394 mm_segment_t old_fs;
3397 unsigned long timeout;
3400 complete(&ctx->completions[1]);
3404 old_cred = override_creds(ctx->creds);
3406 ret = timeout = inflight = 0;
3407 while (!kthread_should_park()) {
3408 unsigned int to_submit;
3411 unsigned nr_events = 0;
3413 if (ctx->flags & IORING_SETUP_IOPOLL) {
3415 * inflight is the count of the maximum possible
3416 * entries we submitted, but it can be smaller
3417 * if we dropped some of them. If we don't have
3418 * poll entries available, then we know that we
3419 * have nothing left to poll for. Reset the
3420 * inflight count to zero in that case.
3422 mutex_lock(&ctx->uring_lock);
3423 if (!list_empty(&ctx->poll_list))
3424 __io_iopoll_check(ctx, &nr_events, 0);
3427 mutex_unlock(&ctx->uring_lock);
3430 * Normal IO, just pretend everything completed.
3431 * We don't have to poll completions for that.
3433 nr_events = inflight;
3436 inflight -= nr_events;
3438 timeout = jiffies + ctx->sq_thread_idle;
3441 to_submit = io_sqring_entries(ctx);
3444 * If submit got -EBUSY, flag us as needing the application
3445 * to enter the kernel to reap and flush events.
3447 if (!to_submit || ret == -EBUSY) {
3449 * We're polling. If we're within the defined idle
3450 * period, then let us spin without work before going
3451 * to sleep. The exception is if we got EBUSY doing
3452 * more IO, we should wait for the application to
3453 * reap events and wake us up.
3456 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3462 * Drop cur_mm before scheduling, we can't hold it for
3463 * long periods (or over schedule()). Do this before
3464 * adding ourselves to the waitqueue, as the unuse/drop
3473 prepare_to_wait(&ctx->sqo_wait, &wait,
3474 TASK_INTERRUPTIBLE);
3476 /* Tell userspace we may need a wakeup call */
3477 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3478 /* make sure to read SQ tail after writing flags */
3481 to_submit = io_sqring_entries(ctx);
3482 if (!to_submit || ret == -EBUSY) {
3483 if (kthread_should_park()) {
3484 finish_wait(&ctx->sqo_wait, &wait);
3487 if (signal_pending(current))
3488 flush_signals(current);
3490 finish_wait(&ctx->sqo_wait, &wait);
3492 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3495 finish_wait(&ctx->sqo_wait, &wait);
3497 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3500 to_submit = min(to_submit, ctx->sq_entries);
3501 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3511 revert_creds(old_cred);
3518 struct io_wait_queue {
3519 struct wait_queue_entry wq;
3520 struct io_ring_ctx *ctx;
3522 unsigned nr_timeouts;
3525 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3527 struct io_ring_ctx *ctx = iowq->ctx;
3530 * Wake up if we have enough events, or if a timeout occured since we
3531 * started waiting. For timeouts, we always want to return to userspace,
3532 * regardless of event count.
3534 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3535 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3538 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3539 int wake_flags, void *key)
3541 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3544 /* use noflush == true, as we can't safely rely on locking context */
3545 if (!io_should_wake(iowq, true))
3548 return autoremove_wake_function(curr, mode, wake_flags, key);
3552 * Wait until events become available, if we don't already have some. The
3553 * application must reap them itself, as they reside on the shared cq ring.
3555 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3556 const sigset_t __user *sig, size_t sigsz)
3558 struct io_wait_queue iowq = {
3561 .func = io_wake_function,
3562 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3565 .to_wait = min_events,
3567 struct io_rings *rings = ctx->rings;
3570 if (io_cqring_events(ctx, false) >= min_events)
3574 #ifdef CONFIG_COMPAT
3575 if (in_compat_syscall())
3576 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3580 ret = set_user_sigmask(sig, sigsz);
3586 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
3587 trace_io_uring_cqring_wait(ctx, min_events);
3589 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3590 TASK_INTERRUPTIBLE);
3591 if (io_should_wake(&iowq, false))
3594 if (signal_pending(current)) {
3599 finish_wait(&ctx->wait, &iowq.wq);
3601 restore_saved_sigmask_unless(ret == -EINTR);
3603 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
3606 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3608 #if defined(CONFIG_UNIX)
3609 if (ctx->ring_sock) {
3610 struct sock *sock = ctx->ring_sock->sk;
3611 struct sk_buff *skb;
3613 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3619 for (i = 0; i < ctx->nr_user_files; i++) {
3622 file = io_file_from_index(ctx, i);
3629 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3631 unsigned nr_tables, i;
3633 if (!ctx->file_table)
3636 __io_sqe_files_unregister(ctx);
3637 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3638 for (i = 0; i < nr_tables; i++)
3639 kfree(ctx->file_table[i].files);
3640 kfree(ctx->file_table);
3641 ctx->file_table = NULL;
3642 ctx->nr_user_files = 0;
3646 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3648 if (ctx->sqo_thread) {
3649 wait_for_completion(&ctx->completions[1]);
3651 * The park is a bit of a work-around, without it we get
3652 * warning spews on shutdown with SQPOLL set and affinity
3653 * set to a single CPU.
3655 kthread_park(ctx->sqo_thread);
3656 kthread_stop(ctx->sqo_thread);
3657 ctx->sqo_thread = NULL;
3661 static void io_finish_async(struct io_ring_ctx *ctx)
3663 io_sq_thread_stop(ctx);
3666 io_wq_destroy(ctx->io_wq);
3671 #if defined(CONFIG_UNIX)
3672 static void io_destruct_skb(struct sk_buff *skb)
3674 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3677 io_wq_flush(ctx->io_wq);
3679 unix_destruct_scm(skb);
3683 * Ensure the UNIX gc is aware of our file set, so we are certain that
3684 * the io_uring can be safely unregistered on process exit, even if we have
3685 * loops in the file referencing.
3687 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3689 struct sock *sk = ctx->ring_sock->sk;
3690 struct scm_fp_list *fpl;
3691 struct sk_buff *skb;
3694 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3695 unsigned long inflight = ctx->user->unix_inflight + nr;
3697 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3701 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3705 skb = alloc_skb(0, GFP_KERNEL);
3714 fpl->user = get_uid(ctx->user);
3715 for (i = 0; i < nr; i++) {
3716 struct file *file = io_file_from_index(ctx, i + offset);
3720 fpl->fp[nr_files] = get_file(file);
3721 unix_inflight(fpl->user, fpl->fp[nr_files]);
3726 fpl->max = SCM_MAX_FD;
3727 fpl->count = nr_files;
3728 UNIXCB(skb).fp = fpl;
3729 skb->destructor = io_destruct_skb;
3730 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3731 skb_queue_head(&sk->sk_receive_queue, skb);
3733 for (i = 0; i < nr_files; i++)
3744 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3745 * causes regular reference counting to break down. We rely on the UNIX
3746 * garbage collection to take care of this problem for us.
3748 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3750 unsigned left, total;
3754 left = ctx->nr_user_files;
3756 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
3758 ret = __io_sqe_files_scm(ctx, this_files, total);
3762 total += this_files;
3768 while (total < ctx->nr_user_files) {
3769 struct file *file = io_file_from_index(ctx, total);
3779 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3785 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3790 for (i = 0; i < nr_tables; i++) {
3791 struct fixed_file_table *table = &ctx->file_table[i];
3792 unsigned this_files;
3794 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3795 table->files = kcalloc(this_files, sizeof(struct file *),
3799 nr_files -= this_files;
3805 for (i = 0; i < nr_tables; i++) {
3806 struct fixed_file_table *table = &ctx->file_table[i];
3807 kfree(table->files);
3812 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3815 __s32 __user *fds = (__s32 __user *) arg;
3820 if (ctx->file_table)
3824 if (nr_args > IORING_MAX_FIXED_FILES)
3827 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
3828 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
3830 if (!ctx->file_table)
3833 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
3834 kfree(ctx->file_table);
3835 ctx->file_table = NULL;
3839 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
3840 struct fixed_file_table *table;
3844 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3846 /* allow sparse sets */
3852 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
3853 index = i & IORING_FILE_TABLE_MASK;
3854 table->files[index] = fget(fd);
3857 if (!table->files[index])
3860 * Don't allow io_uring instances to be registered. If UNIX
3861 * isn't enabled, then this causes a reference cycle and this
3862 * instance can never get freed. If UNIX is enabled we'll
3863 * handle it just fine, but there's still no point in allowing
3864 * a ring fd as it doesn't support regular read/write anyway.
3866 if (table->files[index]->f_op == &io_uring_fops) {
3867 fput(table->files[index]);
3874 for (i = 0; i < ctx->nr_user_files; i++) {
3877 file = io_file_from_index(ctx, i);
3881 for (i = 0; i < nr_tables; i++)
3882 kfree(ctx->file_table[i].files);
3884 kfree(ctx->file_table);
3885 ctx->file_table = NULL;
3886 ctx->nr_user_files = 0;
3890 ret = io_sqe_files_scm(ctx);
3892 io_sqe_files_unregister(ctx);
3897 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
3899 #if defined(CONFIG_UNIX)
3900 struct file *file = io_file_from_index(ctx, index);
3901 struct sock *sock = ctx->ring_sock->sk;
3902 struct sk_buff_head list, *head = &sock->sk_receive_queue;
3903 struct sk_buff *skb;
3906 __skb_queue_head_init(&list);
3909 * Find the skb that holds this file in its SCM_RIGHTS. When found,
3910 * remove this entry and rearrange the file array.
3912 skb = skb_dequeue(head);
3914 struct scm_fp_list *fp;
3916 fp = UNIXCB(skb).fp;
3917 for (i = 0; i < fp->count; i++) {
3920 if (fp->fp[i] != file)
3923 unix_notinflight(fp->user, fp->fp[i]);
3924 left = fp->count - 1 - i;
3926 memmove(&fp->fp[i], &fp->fp[i + 1],
3927 left * sizeof(struct file *));
3934 __skb_queue_tail(&list, skb);
3944 __skb_queue_tail(&list, skb);
3946 skb = skb_dequeue(head);
3949 if (skb_peek(&list)) {
3950 spin_lock_irq(&head->lock);
3951 while ((skb = __skb_dequeue(&list)) != NULL)
3952 __skb_queue_tail(head, skb);
3953 spin_unlock_irq(&head->lock);
3956 fput(io_file_from_index(ctx, index));
3960 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
3963 #if defined(CONFIG_UNIX)
3964 struct sock *sock = ctx->ring_sock->sk;
3965 struct sk_buff_head *head = &sock->sk_receive_queue;
3966 struct sk_buff *skb;
3969 * See if we can merge this file into an existing skb SCM_RIGHTS
3970 * file set. If there's no room, fall back to allocating a new skb
3971 * and filling it in.
3973 spin_lock_irq(&head->lock);
3974 skb = skb_peek(head);
3976 struct scm_fp_list *fpl = UNIXCB(skb).fp;
3978 if (fpl->count < SCM_MAX_FD) {
3979 __skb_unlink(skb, head);
3980 spin_unlock_irq(&head->lock);
3981 fpl->fp[fpl->count] = get_file(file);
3982 unix_inflight(fpl->user, fpl->fp[fpl->count]);
3984 spin_lock_irq(&head->lock);
3985 __skb_queue_head(head, skb);
3990 spin_unlock_irq(&head->lock);
3997 return __io_sqe_files_scm(ctx, 1, index);
4003 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4006 struct io_uring_files_update up;
4011 if (!ctx->file_table)
4015 if (copy_from_user(&up, arg, sizeof(up)))
4017 if (check_add_overflow(up.offset, nr_args, &done))
4019 if (done > ctx->nr_user_files)
4023 fds = (__s32 __user *) up.fds;
4025 struct fixed_file_table *table;
4029 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4033 i = array_index_nospec(up.offset, ctx->nr_user_files);
4034 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4035 index = i & IORING_FILE_TABLE_MASK;
4036 if (table->files[index]) {
4037 io_sqe_file_unregister(ctx, i);
4038 table->files[index] = NULL;
4049 * Don't allow io_uring instances to be registered. If
4050 * UNIX isn't enabled, then this causes a reference
4051 * cycle and this instance can never get freed. If UNIX
4052 * is enabled we'll handle it just fine, but there's
4053 * still no point in allowing a ring fd as it doesn't
4054 * support regular read/write anyway.
4056 if (file->f_op == &io_uring_fops) {
4061 table->files[index] = file;
4062 err = io_sqe_file_register(ctx, file, i);
4071 return done ? done : err;
4074 static void io_put_work(struct io_wq_work *work)
4076 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4081 static void io_get_work(struct io_wq_work *work)
4083 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4085 refcount_inc(&req->refs);
4088 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4089 struct io_uring_params *p)
4091 struct io_wq_data data;
4092 unsigned concurrency;
4095 init_waitqueue_head(&ctx->sqo_wait);
4096 mmgrab(current->mm);
4097 ctx->sqo_mm = current->mm;
4099 if (ctx->flags & IORING_SETUP_SQPOLL) {
4101 if (!capable(CAP_SYS_ADMIN))
4104 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4105 if (!ctx->sq_thread_idle)
4106 ctx->sq_thread_idle = HZ;
4108 if (p->flags & IORING_SETUP_SQ_AFF) {
4109 int cpu = p->sq_thread_cpu;
4112 if (cpu >= nr_cpu_ids)
4114 if (!cpu_online(cpu))
4117 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4121 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4124 if (IS_ERR(ctx->sqo_thread)) {
4125 ret = PTR_ERR(ctx->sqo_thread);
4126 ctx->sqo_thread = NULL;
4129 wake_up_process(ctx->sqo_thread);
4130 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4131 /* Can't have SQ_AFF without SQPOLL */
4136 data.mm = ctx->sqo_mm;
4137 data.user = ctx->user;
4138 data.creds = ctx->creds;
4139 data.get_work = io_get_work;
4140 data.put_work = io_put_work;
4142 /* Do QD, or 4 * CPUS, whatever is smallest */
4143 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4144 ctx->io_wq = io_wq_create(concurrency, &data);
4145 if (IS_ERR(ctx->io_wq)) {
4146 ret = PTR_ERR(ctx->io_wq);
4153 io_finish_async(ctx);
4154 mmdrop(ctx->sqo_mm);
4159 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4161 atomic_long_sub(nr_pages, &user->locked_vm);
4164 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4166 unsigned long page_limit, cur_pages, new_pages;
4168 /* Don't allow more pages than we can safely lock */
4169 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4172 cur_pages = atomic_long_read(&user->locked_vm);
4173 new_pages = cur_pages + nr_pages;
4174 if (new_pages > page_limit)
4176 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4177 new_pages) != cur_pages);
4182 static void io_mem_free(void *ptr)
4189 page = virt_to_head_page(ptr);
4190 if (put_page_testzero(page))
4191 free_compound_page(page);
4194 static void *io_mem_alloc(size_t size)
4196 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4199 return (void *) __get_free_pages(gfp_flags, get_order(size));
4202 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4205 struct io_rings *rings;
4206 size_t off, sq_array_size;
4208 off = struct_size(rings, cqes, cq_entries);
4209 if (off == SIZE_MAX)
4213 off = ALIGN(off, SMP_CACHE_BYTES);
4218 sq_array_size = array_size(sizeof(u32), sq_entries);
4219 if (sq_array_size == SIZE_MAX)
4222 if (check_add_overflow(off, sq_array_size, &off))
4231 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4235 pages = (size_t)1 << get_order(
4236 rings_size(sq_entries, cq_entries, NULL));
4237 pages += (size_t)1 << get_order(
4238 array_size(sizeof(struct io_uring_sqe), sq_entries));
4243 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4247 if (!ctx->user_bufs)
4250 for (i = 0; i < ctx->nr_user_bufs; i++) {
4251 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4253 for (j = 0; j < imu->nr_bvecs; j++)
4254 put_user_page(imu->bvec[j].bv_page);
4256 if (ctx->account_mem)
4257 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4262 kfree(ctx->user_bufs);
4263 ctx->user_bufs = NULL;
4264 ctx->nr_user_bufs = 0;
4268 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4269 void __user *arg, unsigned index)
4271 struct iovec __user *src;
4273 #ifdef CONFIG_COMPAT
4275 struct compat_iovec __user *ciovs;
4276 struct compat_iovec ciov;
4278 ciovs = (struct compat_iovec __user *) arg;
4279 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4282 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4283 dst->iov_len = ciov.iov_len;
4287 src = (struct iovec __user *) arg;
4288 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4293 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4296 struct vm_area_struct **vmas = NULL;
4297 struct page **pages = NULL;
4298 int i, j, got_pages = 0;
4303 if (!nr_args || nr_args > UIO_MAXIOV)
4306 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4308 if (!ctx->user_bufs)
4311 for (i = 0; i < nr_args; i++) {
4312 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4313 unsigned long off, start, end, ubuf;
4318 ret = io_copy_iov(ctx, &iov, arg, i);
4323 * Don't impose further limits on the size and buffer
4324 * constraints here, we'll -EINVAL later when IO is
4325 * submitted if they are wrong.
4328 if (!iov.iov_base || !iov.iov_len)
4331 /* arbitrary limit, but we need something */
4332 if (iov.iov_len > SZ_1G)
4335 ubuf = (unsigned long) iov.iov_base;
4336 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4337 start = ubuf >> PAGE_SHIFT;
4338 nr_pages = end - start;
4340 if (ctx->account_mem) {
4341 ret = io_account_mem(ctx->user, nr_pages);
4347 if (!pages || nr_pages > got_pages) {
4350 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4352 vmas = kvmalloc_array(nr_pages,
4353 sizeof(struct vm_area_struct *),
4355 if (!pages || !vmas) {
4357 if (ctx->account_mem)
4358 io_unaccount_mem(ctx->user, nr_pages);
4361 got_pages = nr_pages;
4364 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4368 if (ctx->account_mem)
4369 io_unaccount_mem(ctx->user, nr_pages);
4374 down_read(¤t->mm->mmap_sem);
4375 pret = get_user_pages(ubuf, nr_pages,
4376 FOLL_WRITE | FOLL_LONGTERM,
4378 if (pret == nr_pages) {
4379 /* don't support file backed memory */
4380 for (j = 0; j < nr_pages; j++) {
4381 struct vm_area_struct *vma = vmas[j];
4384 !is_file_hugepages(vma->vm_file)) {
4390 ret = pret < 0 ? pret : -EFAULT;
4392 up_read(¤t->mm->mmap_sem);
4395 * if we did partial map, or found file backed vmas,
4396 * release any pages we did get
4399 put_user_pages(pages, pret);
4400 if (ctx->account_mem)
4401 io_unaccount_mem(ctx->user, nr_pages);
4406 off = ubuf & ~PAGE_MASK;
4408 for (j = 0; j < nr_pages; j++) {
4411 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4412 imu->bvec[j].bv_page = pages[j];
4413 imu->bvec[j].bv_len = vec_len;
4414 imu->bvec[j].bv_offset = off;
4418 /* store original address for later verification */
4420 imu->len = iov.iov_len;
4421 imu->nr_bvecs = nr_pages;
4423 ctx->nr_user_bufs++;
4431 io_sqe_buffer_unregister(ctx);
4435 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4437 __s32 __user *fds = arg;
4443 if (copy_from_user(&fd, fds, sizeof(*fds)))
4446 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4447 if (IS_ERR(ctx->cq_ev_fd)) {
4448 int ret = PTR_ERR(ctx->cq_ev_fd);
4449 ctx->cq_ev_fd = NULL;
4456 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4458 if (ctx->cq_ev_fd) {
4459 eventfd_ctx_put(ctx->cq_ev_fd);
4460 ctx->cq_ev_fd = NULL;
4467 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4469 io_finish_async(ctx);
4471 mmdrop(ctx->sqo_mm);
4473 io_iopoll_reap_events(ctx);
4474 io_sqe_buffer_unregister(ctx);
4475 io_sqe_files_unregister(ctx);
4476 io_eventfd_unregister(ctx);
4478 #if defined(CONFIG_UNIX)
4479 if (ctx->ring_sock) {
4480 ctx->ring_sock->file = NULL; /* so that iput() is called */
4481 sock_release(ctx->ring_sock);
4485 io_mem_free(ctx->rings);
4486 io_mem_free(ctx->sq_sqes);
4488 percpu_ref_exit(&ctx->refs);
4489 if (ctx->account_mem)
4490 io_unaccount_mem(ctx->user,
4491 ring_pages(ctx->sq_entries, ctx->cq_entries));
4492 free_uid(ctx->user);
4493 put_cred(ctx->creds);
4494 kfree(ctx->completions);
4495 kmem_cache_free(req_cachep, ctx->fallback_req);
4499 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4501 struct io_ring_ctx *ctx = file->private_data;
4504 poll_wait(file, &ctx->cq_wait, wait);
4506 * synchronizes with barrier from wq_has_sleeper call in
4510 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4511 ctx->rings->sq_ring_entries)
4512 mask |= EPOLLOUT | EPOLLWRNORM;
4513 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4514 mask |= EPOLLIN | EPOLLRDNORM;
4519 static int io_uring_fasync(int fd, struct file *file, int on)
4521 struct io_ring_ctx *ctx = file->private_data;
4523 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4526 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4528 mutex_lock(&ctx->uring_lock);
4529 percpu_ref_kill(&ctx->refs);
4530 mutex_unlock(&ctx->uring_lock);
4532 io_kill_timeouts(ctx);
4533 io_poll_remove_all(ctx);
4536 io_wq_cancel_all(ctx->io_wq);
4538 io_iopoll_reap_events(ctx);
4539 /* if we failed setting up the ctx, we might not have any rings */
4541 io_cqring_overflow_flush(ctx, true);
4542 wait_for_completion(&ctx->completions[0]);
4543 io_ring_ctx_free(ctx);
4546 static int io_uring_release(struct inode *inode, struct file *file)
4548 struct io_ring_ctx *ctx = file->private_data;
4550 file->private_data = NULL;
4551 io_ring_ctx_wait_and_kill(ctx);
4555 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4556 struct files_struct *files)
4558 struct io_kiocb *req;
4561 while (!list_empty_careful(&ctx->inflight_list)) {
4562 struct io_kiocb *cancel_req = NULL;
4564 spin_lock_irq(&ctx->inflight_lock);
4565 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4566 if (req->work.files != files)
4568 /* req is being completed, ignore */
4569 if (!refcount_inc_not_zero(&req->refs))
4575 prepare_to_wait(&ctx->inflight_wait, &wait,
4576 TASK_UNINTERRUPTIBLE);
4577 spin_unlock_irq(&ctx->inflight_lock);
4579 /* We need to keep going until we don't find a matching req */
4583 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4584 io_put_req(cancel_req);
4587 finish_wait(&ctx->inflight_wait, &wait);
4590 static int io_uring_flush(struct file *file, void *data)
4592 struct io_ring_ctx *ctx = file->private_data;
4594 io_uring_cancel_files(ctx, data);
4595 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4596 io_cqring_overflow_flush(ctx, true);
4597 io_wq_cancel_all(ctx->io_wq);
4602 static void *io_uring_validate_mmap_request(struct file *file,
4603 loff_t pgoff, size_t sz)
4605 struct io_ring_ctx *ctx = file->private_data;
4606 loff_t offset = pgoff << PAGE_SHIFT;
4611 case IORING_OFF_SQ_RING:
4612 case IORING_OFF_CQ_RING:
4615 case IORING_OFF_SQES:
4619 return ERR_PTR(-EINVAL);
4622 page = virt_to_head_page(ptr);
4623 if (sz > page_size(page))
4624 return ERR_PTR(-EINVAL);
4631 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4633 size_t sz = vma->vm_end - vma->vm_start;
4637 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4639 return PTR_ERR(ptr);
4641 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4642 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4645 #else /* !CONFIG_MMU */
4647 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4649 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4652 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4654 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4657 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4658 unsigned long addr, unsigned long len,
4659 unsigned long pgoff, unsigned long flags)
4663 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4665 return PTR_ERR(ptr);
4667 return (unsigned long) ptr;
4670 #endif /* !CONFIG_MMU */
4672 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
4673 u32, min_complete, u32, flags, const sigset_t __user *, sig,
4676 struct io_ring_ctx *ctx;
4681 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
4689 if (f.file->f_op != &io_uring_fops)
4693 ctx = f.file->private_data;
4694 if (!percpu_ref_tryget(&ctx->refs))
4698 * For SQ polling, the thread will do all submissions and completions.
4699 * Just return the requested submit count, and wake the thread if
4703 if (ctx->flags & IORING_SETUP_SQPOLL) {
4704 if (!list_empty_careful(&ctx->cq_overflow_list))
4705 io_cqring_overflow_flush(ctx, false);
4706 if (flags & IORING_ENTER_SQ_WAKEUP)
4707 wake_up(&ctx->sqo_wait);
4708 submitted = to_submit;
4709 } else if (to_submit) {
4710 struct mm_struct *cur_mm;
4712 to_submit = min(to_submit, ctx->sq_entries);
4713 mutex_lock(&ctx->uring_lock);
4714 /* already have mm, so io_submit_sqes() won't try to grab it */
4715 cur_mm = ctx->sqo_mm;
4716 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4718 mutex_unlock(&ctx->uring_lock);
4720 if (flags & IORING_ENTER_GETEVENTS) {
4721 unsigned nr_events = 0;
4723 min_complete = min(min_complete, ctx->cq_entries);
4725 if (ctx->flags & IORING_SETUP_IOPOLL) {
4726 ret = io_iopoll_check(ctx, &nr_events, min_complete);
4728 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4732 percpu_ref_put(&ctx->refs);
4735 return submitted ? submitted : ret;
4738 static const struct file_operations io_uring_fops = {
4739 .release = io_uring_release,
4740 .flush = io_uring_flush,
4741 .mmap = io_uring_mmap,
4743 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
4744 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
4746 .poll = io_uring_poll,
4747 .fasync = io_uring_fasync,
4750 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4751 struct io_uring_params *p)
4753 struct io_rings *rings;
4754 size_t size, sq_array_offset;
4756 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4757 if (size == SIZE_MAX)
4760 rings = io_mem_alloc(size);
4765 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4766 rings->sq_ring_mask = p->sq_entries - 1;
4767 rings->cq_ring_mask = p->cq_entries - 1;
4768 rings->sq_ring_entries = p->sq_entries;
4769 rings->cq_ring_entries = p->cq_entries;
4770 ctx->sq_mask = rings->sq_ring_mask;
4771 ctx->cq_mask = rings->cq_ring_mask;
4772 ctx->sq_entries = rings->sq_ring_entries;
4773 ctx->cq_entries = rings->cq_ring_entries;
4775 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
4776 if (size == SIZE_MAX) {
4777 io_mem_free(ctx->rings);
4782 ctx->sq_sqes = io_mem_alloc(size);
4783 if (!ctx->sq_sqes) {
4784 io_mem_free(ctx->rings);
4793 * Allocate an anonymous fd, this is what constitutes the application
4794 * visible backing of an io_uring instance. The application mmaps this
4795 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4796 * we have to tie this fd to a socket for file garbage collection purposes.
4798 static int io_uring_get_fd(struct io_ring_ctx *ctx)
4803 #if defined(CONFIG_UNIX)
4804 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4810 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4814 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4815 O_RDWR | O_CLOEXEC);
4818 ret = PTR_ERR(file);
4822 #if defined(CONFIG_UNIX)
4823 ctx->ring_sock->file = file;
4824 ctx->ring_sock->sk->sk_user_data = ctx;
4826 fd_install(ret, file);
4829 #if defined(CONFIG_UNIX)
4830 sock_release(ctx->ring_sock);
4831 ctx->ring_sock = NULL;
4836 static int io_uring_create(unsigned entries, struct io_uring_params *p)
4838 struct user_struct *user = NULL;
4839 struct io_ring_ctx *ctx;
4843 if (!entries || entries > IORING_MAX_ENTRIES)
4847 * Use twice as many entries for the CQ ring. It's possible for the
4848 * application to drive a higher depth than the size of the SQ ring,
4849 * since the sqes are only used at submission time. This allows for
4850 * some flexibility in overcommitting a bit. If the application has
4851 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
4852 * of CQ ring entries manually.
4854 p->sq_entries = roundup_pow_of_two(entries);
4855 if (p->flags & IORING_SETUP_CQSIZE) {
4857 * If IORING_SETUP_CQSIZE is set, we do the same roundup
4858 * to a power-of-two, if it isn't already. We do NOT impose
4859 * any cq vs sq ring sizing.
4861 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
4863 p->cq_entries = roundup_pow_of_two(p->cq_entries);
4865 p->cq_entries = 2 * p->sq_entries;
4868 user = get_uid(current_user());
4869 account_mem = !capable(CAP_IPC_LOCK);
4872 ret = io_account_mem(user,
4873 ring_pages(p->sq_entries, p->cq_entries));
4880 ctx = io_ring_ctx_alloc(p);
4883 io_unaccount_mem(user, ring_pages(p->sq_entries,
4888 ctx->compat = in_compat_syscall();
4889 ctx->account_mem = account_mem;
4891 ctx->creds = get_current_cred();
4893 ret = io_allocate_scq_urings(ctx, p);
4897 ret = io_sq_offload_start(ctx, p);
4901 memset(&p->sq_off, 0, sizeof(p->sq_off));
4902 p->sq_off.head = offsetof(struct io_rings, sq.head);
4903 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
4904 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
4905 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
4906 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
4907 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
4908 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
4910 memset(&p->cq_off, 0, sizeof(p->cq_off));
4911 p->cq_off.head = offsetof(struct io_rings, cq.head);
4912 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
4913 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
4914 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
4915 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
4916 p->cq_off.cqes = offsetof(struct io_rings, cqes);
4919 * Install ring fd as the very last thing, so we don't risk someone
4920 * having closed it before we finish setup
4922 ret = io_uring_get_fd(ctx);
4926 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP;
4927 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
4930 io_ring_ctx_wait_and_kill(ctx);
4935 * Sets up an aio uring context, and returns the fd. Applications asks for a
4936 * ring size, we return the actual sq/cq ring sizes (among other things) in the
4937 * params structure passed in.
4939 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
4941 struct io_uring_params p;
4945 if (copy_from_user(&p, params, sizeof(p)))
4947 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
4952 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
4953 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
4956 ret = io_uring_create(entries, &p);
4960 if (copy_to_user(params, &p, sizeof(p)))
4966 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
4967 struct io_uring_params __user *, params)
4969 return io_uring_setup(entries, params);
4972 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
4973 void __user *arg, unsigned nr_args)
4974 __releases(ctx->uring_lock)
4975 __acquires(ctx->uring_lock)
4980 * We're inside the ring mutex, if the ref is already dying, then
4981 * someone else killed the ctx or is already going through
4982 * io_uring_register().
4984 if (percpu_ref_is_dying(&ctx->refs))
4987 percpu_ref_kill(&ctx->refs);
4990 * Drop uring mutex before waiting for references to exit. If another
4991 * thread is currently inside io_uring_enter() it might need to grab
4992 * the uring_lock to make progress. If we hold it here across the drain
4993 * wait, then we can deadlock. It's safe to drop the mutex here, since
4994 * no new references will come in after we've killed the percpu ref.
4996 mutex_unlock(&ctx->uring_lock);
4997 wait_for_completion(&ctx->completions[0]);
4998 mutex_lock(&ctx->uring_lock);
5001 case IORING_REGISTER_BUFFERS:
5002 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5004 case IORING_UNREGISTER_BUFFERS:
5008 ret = io_sqe_buffer_unregister(ctx);
5010 case IORING_REGISTER_FILES:
5011 ret = io_sqe_files_register(ctx, arg, nr_args);
5013 case IORING_UNREGISTER_FILES:
5017 ret = io_sqe_files_unregister(ctx);
5019 case IORING_REGISTER_FILES_UPDATE:
5020 ret = io_sqe_files_update(ctx, arg, nr_args);
5022 case IORING_REGISTER_EVENTFD:
5026 ret = io_eventfd_register(ctx, arg);
5028 case IORING_UNREGISTER_EVENTFD:
5032 ret = io_eventfd_unregister(ctx);
5039 /* bring the ctx back to life */
5040 reinit_completion(&ctx->completions[0]);
5041 percpu_ref_reinit(&ctx->refs);
5045 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5046 void __user *, arg, unsigned int, nr_args)
5048 struct io_ring_ctx *ctx;
5057 if (f.file->f_op != &io_uring_fops)
5060 ctx = f.file->private_data;
5062 mutex_lock(&ctx->uring_lock);
5063 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5064 mutex_unlock(&ctx->uring_lock);
5065 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5066 ctx->cq_ev_fd != NULL, ret);
5072 static int __init io_uring_init(void)
5074 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5077 __initcall(io_uring_init);