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;
311 struct io_async_connect {
312 struct sockaddr_storage address;
315 struct io_async_msghdr {
316 struct iovec fast_iov[UIO_FASTIOV];
318 struct sockaddr __user *uaddr;
323 struct iovec fast_iov[UIO_FASTIOV];
329 struct io_async_ctx {
330 struct io_uring_sqe sqe;
332 struct io_async_rw rw;
333 struct io_async_msghdr msg;
334 struct io_async_connect connect;
339 * NOTE! Each of the iocb union members has the file pointer
340 * as the first entry in their struct definition. So you can
341 * access the file pointer through any of the sub-structs,
342 * or directly as just 'ki_filp' in this struct.
348 struct io_poll_iocb poll;
349 struct io_timeout timeout;
352 const struct io_uring_sqe *sqe;
353 struct io_async_ctx *io;
354 struct file *ring_file;
358 bool needs_fixed_file;
360 struct io_ring_ctx *ctx;
362 struct list_head list;
363 struct rb_node rb_node;
365 struct list_head link_list;
368 #define REQ_F_NOWAIT 1 /* must not punt to workers */
369 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
370 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
371 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
372 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
373 #define REQ_F_IO_DRAINED 32 /* drain done */
374 #define REQ_F_LINK 64 /* linked sqes */
375 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
376 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
377 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
378 #define REQ_F_TIMEOUT 1024 /* timeout request */
379 #define REQ_F_ISREG 2048 /* regular file */
380 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
381 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
382 #define REQ_F_INFLIGHT 16384 /* on inflight list */
383 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
388 struct list_head inflight_entry;
390 struct io_wq_work work;
393 #define IO_PLUG_THRESHOLD 2
394 #define IO_IOPOLL_BATCH 8
396 struct io_submit_state {
397 struct blk_plug plug;
400 * io_kiocb alloc cache
402 void *reqs[IO_IOPOLL_BATCH];
403 unsigned int free_reqs;
404 unsigned int cur_req;
407 * File reference cache
411 unsigned int has_refs;
412 unsigned int used_refs;
413 unsigned int ios_left;
416 static void io_wq_submit_work(struct io_wq_work **workptr);
417 static void io_cqring_fill_event(struct io_kiocb *req, long res);
418 static void __io_free_req(struct io_kiocb *req);
419 static void io_put_req(struct io_kiocb *req);
420 static void io_double_put_req(struct io_kiocb *req);
421 static void __io_double_put_req(struct io_kiocb *req);
422 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
423 static void io_queue_linked_timeout(struct io_kiocb *req);
425 static struct kmem_cache *req_cachep;
427 static const struct file_operations io_uring_fops;
429 struct sock *io_uring_get_socket(struct file *file)
431 #if defined(CONFIG_UNIX)
432 if (file->f_op == &io_uring_fops) {
433 struct io_ring_ctx *ctx = file->private_data;
435 return ctx->ring_sock->sk;
440 EXPORT_SYMBOL(io_uring_get_socket);
442 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
444 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
446 complete(&ctx->completions[0]);
449 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
451 struct io_ring_ctx *ctx;
453 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
457 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
458 if (!ctx->fallback_req)
461 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
462 if (!ctx->completions)
465 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
466 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
469 ctx->flags = p->flags;
470 init_waitqueue_head(&ctx->cq_wait);
471 INIT_LIST_HEAD(&ctx->cq_overflow_list);
472 init_completion(&ctx->completions[0]);
473 init_completion(&ctx->completions[1]);
474 mutex_init(&ctx->uring_lock);
475 init_waitqueue_head(&ctx->wait);
476 spin_lock_init(&ctx->completion_lock);
477 INIT_LIST_HEAD(&ctx->poll_list);
478 ctx->cancel_tree = RB_ROOT;
479 INIT_LIST_HEAD(&ctx->defer_list);
480 INIT_LIST_HEAD(&ctx->timeout_list);
481 init_waitqueue_head(&ctx->inflight_wait);
482 spin_lock_init(&ctx->inflight_lock);
483 INIT_LIST_HEAD(&ctx->inflight_list);
486 if (ctx->fallback_req)
487 kmem_cache_free(req_cachep, ctx->fallback_req);
488 kfree(ctx->completions);
493 static inline bool __req_need_defer(struct io_kiocb *req)
495 struct io_ring_ctx *ctx = req->ctx;
497 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
498 + atomic_read(&ctx->cached_cq_overflow);
501 static inline bool req_need_defer(struct io_kiocb *req)
503 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
504 return __req_need_defer(req);
509 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
511 struct io_kiocb *req;
513 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
514 if (req && !req_need_defer(req)) {
515 list_del_init(&req->list);
522 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
524 struct io_kiocb *req;
526 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
528 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
530 if (!__req_need_defer(req)) {
531 list_del_init(&req->list);
539 static void __io_commit_cqring(struct io_ring_ctx *ctx)
541 struct io_rings *rings = ctx->rings;
543 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
544 /* order cqe stores with ring update */
545 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
547 if (wq_has_sleeper(&ctx->cq_wait)) {
548 wake_up_interruptible(&ctx->cq_wait);
549 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
554 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
556 u8 opcode = READ_ONCE(sqe->opcode);
558 return !(opcode == IORING_OP_READ_FIXED ||
559 opcode == IORING_OP_WRITE_FIXED);
562 static inline bool io_prep_async_work(struct io_kiocb *req,
563 struct io_kiocb **link)
565 bool do_hashed = false;
568 switch (req->sqe->opcode) {
569 case IORING_OP_WRITEV:
570 case IORING_OP_WRITE_FIXED:
573 case IORING_OP_READV:
574 case IORING_OP_READ_FIXED:
575 case IORING_OP_SENDMSG:
576 case IORING_OP_RECVMSG:
577 case IORING_OP_ACCEPT:
578 case IORING_OP_POLL_ADD:
579 case IORING_OP_CONNECT:
581 * We know REQ_F_ISREG is not set on some of these
582 * opcodes, but this enables us to keep the check in
585 if (!(req->flags & REQ_F_ISREG))
586 req->work.flags |= IO_WQ_WORK_UNBOUND;
589 if (io_sqe_needs_user(req->sqe))
590 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
593 *link = io_prep_linked_timeout(req);
597 static inline void io_queue_async_work(struct io_kiocb *req)
599 struct io_ring_ctx *ctx = req->ctx;
600 struct io_kiocb *link;
603 do_hashed = io_prep_async_work(req, &link);
605 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
608 io_wq_enqueue(ctx->io_wq, &req->work);
610 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
611 file_inode(req->file));
615 io_queue_linked_timeout(link);
618 static void io_kill_timeout(struct io_kiocb *req)
622 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
624 atomic_inc(&req->ctx->cq_timeouts);
625 list_del_init(&req->list);
626 io_cqring_fill_event(req, 0);
631 static void io_kill_timeouts(struct io_ring_ctx *ctx)
633 struct io_kiocb *req, *tmp;
635 spin_lock_irq(&ctx->completion_lock);
636 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
637 io_kill_timeout(req);
638 spin_unlock_irq(&ctx->completion_lock);
641 static void io_commit_cqring(struct io_ring_ctx *ctx)
643 struct io_kiocb *req;
645 while ((req = io_get_timeout_req(ctx)) != NULL)
646 io_kill_timeout(req);
648 __io_commit_cqring(ctx);
650 while ((req = io_get_deferred_req(ctx)) != NULL) {
651 req->flags |= REQ_F_IO_DRAINED;
652 io_queue_async_work(req);
656 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
658 struct io_rings *rings = ctx->rings;
661 tail = ctx->cached_cq_tail;
663 * writes to the cq entry need to come after reading head; the
664 * control dependency is enough as we're using WRITE_ONCE to
667 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
670 ctx->cached_cq_tail++;
671 return &rings->cqes[tail & ctx->cq_mask];
674 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
676 if (waitqueue_active(&ctx->wait))
678 if (waitqueue_active(&ctx->sqo_wait))
679 wake_up(&ctx->sqo_wait);
681 eventfd_signal(ctx->cq_ev_fd, 1);
684 /* Returns true if there are no backlogged entries after the flush */
685 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
687 struct io_rings *rings = ctx->rings;
688 struct io_uring_cqe *cqe;
689 struct io_kiocb *req;
694 if (list_empty_careful(&ctx->cq_overflow_list))
696 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
697 rings->cq_ring_entries))
701 spin_lock_irqsave(&ctx->completion_lock, flags);
703 /* if force is set, the ring is going away. always drop after that */
705 ctx->cq_overflow_flushed = true;
708 while (!list_empty(&ctx->cq_overflow_list)) {
709 cqe = io_get_cqring(ctx);
713 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
715 list_move(&req->list, &list);
717 WRITE_ONCE(cqe->user_data, req->user_data);
718 WRITE_ONCE(cqe->res, req->result);
719 WRITE_ONCE(cqe->flags, 0);
721 WRITE_ONCE(ctx->rings->cq_overflow,
722 atomic_inc_return(&ctx->cached_cq_overflow));
726 io_commit_cqring(ctx);
727 spin_unlock_irqrestore(&ctx->completion_lock, flags);
728 io_cqring_ev_posted(ctx);
730 while (!list_empty(&list)) {
731 req = list_first_entry(&list, struct io_kiocb, list);
732 list_del(&req->list);
739 static void io_cqring_fill_event(struct io_kiocb *req, long res)
741 struct io_ring_ctx *ctx = req->ctx;
742 struct io_uring_cqe *cqe;
744 trace_io_uring_complete(ctx, req->user_data, res);
747 * If we can't get a cq entry, userspace overflowed the
748 * submission (by quite a lot). Increment the overflow count in
751 cqe = io_get_cqring(ctx);
753 WRITE_ONCE(cqe->user_data, req->user_data);
754 WRITE_ONCE(cqe->res, res);
755 WRITE_ONCE(cqe->flags, 0);
756 } else if (ctx->cq_overflow_flushed) {
757 WRITE_ONCE(ctx->rings->cq_overflow,
758 atomic_inc_return(&ctx->cached_cq_overflow));
760 refcount_inc(&req->refs);
762 list_add_tail(&req->list, &ctx->cq_overflow_list);
766 static void io_cqring_add_event(struct io_kiocb *req, long res)
768 struct io_ring_ctx *ctx = req->ctx;
771 spin_lock_irqsave(&ctx->completion_lock, flags);
772 io_cqring_fill_event(req, res);
773 io_commit_cqring(ctx);
774 spin_unlock_irqrestore(&ctx->completion_lock, flags);
776 io_cqring_ev_posted(ctx);
779 static inline bool io_is_fallback_req(struct io_kiocb *req)
781 return req == (struct io_kiocb *)
782 ((unsigned long) req->ctx->fallback_req & ~1UL);
785 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
787 struct io_kiocb *req;
789 req = ctx->fallback_req;
790 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
796 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
797 struct io_submit_state *state)
799 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
800 struct io_kiocb *req;
802 if (!percpu_ref_tryget(&ctx->refs))
806 req = kmem_cache_alloc(req_cachep, gfp);
809 } else if (!state->free_reqs) {
813 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
814 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
817 * Bulk alloc is all-or-nothing. If we fail to get a batch,
818 * retry single alloc to be on the safe side.
820 if (unlikely(ret <= 0)) {
821 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
826 state->free_reqs = ret - 1;
828 req = state->reqs[0];
830 req = state->reqs[state->cur_req];
837 req->ring_file = NULL;
841 /* one is dropped after submission, the other at completion */
842 refcount_set(&req->refs, 2);
844 INIT_IO_WORK(&req->work, io_wq_submit_work);
847 req = io_get_fallback_req(ctx);
850 percpu_ref_put(&ctx->refs);
854 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
857 kmem_cache_free_bulk(req_cachep, *nr, reqs);
858 percpu_ref_put_many(&ctx->refs, *nr);
863 static void __io_free_req(struct io_kiocb *req)
865 struct io_ring_ctx *ctx = req->ctx;
869 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
871 if (req->flags & REQ_F_INFLIGHT) {
874 spin_lock_irqsave(&ctx->inflight_lock, flags);
875 list_del(&req->inflight_entry);
876 if (waitqueue_active(&ctx->inflight_wait))
877 wake_up(&ctx->inflight_wait);
878 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
880 if (req->flags & REQ_F_TIMEOUT)
881 kfree(req->timeout.data);
882 percpu_ref_put(&ctx->refs);
883 if (likely(!io_is_fallback_req(req)))
884 kmem_cache_free(req_cachep, req);
886 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
889 static bool io_link_cancel_timeout(struct io_kiocb *req)
891 struct io_ring_ctx *ctx = req->ctx;
894 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
896 io_cqring_fill_event(req, -ECANCELED);
897 io_commit_cqring(ctx);
898 req->flags &= ~REQ_F_LINK;
906 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
908 struct io_ring_ctx *ctx = req->ctx;
909 struct io_kiocb *nxt;
910 bool wake_ev = false;
912 /* Already got next link */
913 if (req->flags & REQ_F_LINK_NEXT)
917 * The list should never be empty when we are called here. But could
918 * potentially happen if the chain is messed up, check to be on the
921 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
923 list_del_init(&nxt->list);
925 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
926 (nxt->flags & REQ_F_TIMEOUT)) {
927 wake_ev |= io_link_cancel_timeout(nxt);
928 nxt = list_first_entry_or_null(&req->link_list,
929 struct io_kiocb, list);
930 req->flags &= ~REQ_F_LINK_TIMEOUT;
933 if (!list_empty(&req->link_list)) {
934 INIT_LIST_HEAD(&nxt->link_list);
935 list_splice(&req->link_list, &nxt->link_list);
936 nxt->flags |= REQ_F_LINK;
943 req->flags |= REQ_F_LINK_NEXT;
945 io_cqring_ev_posted(ctx);
949 * Called if REQ_F_LINK is set, and we fail the head request
951 static void io_fail_links(struct io_kiocb *req)
953 struct io_ring_ctx *ctx = req->ctx;
954 struct io_kiocb *link;
957 spin_lock_irqsave(&ctx->completion_lock, flags);
959 while (!list_empty(&req->link_list)) {
960 link = list_first_entry(&req->link_list, struct io_kiocb, list);
961 list_del_init(&link->list);
963 trace_io_uring_fail_link(req, link);
965 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
966 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
967 io_link_cancel_timeout(link);
969 io_cqring_fill_event(link, -ECANCELED);
970 __io_double_put_req(link);
972 req->flags &= ~REQ_F_LINK_TIMEOUT;
975 io_commit_cqring(ctx);
976 spin_unlock_irqrestore(&ctx->completion_lock, flags);
977 io_cqring_ev_posted(ctx);
980 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
982 if (likely(!(req->flags & REQ_F_LINK)))
986 * If LINK is set, we have dependent requests in this chain. If we
987 * didn't fail this request, queue the first one up, moving any other
988 * dependencies to the next request. In case of failure, fail the rest
991 if (req->flags & REQ_F_FAIL_LINK) {
993 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
994 REQ_F_LINK_TIMEOUT) {
995 struct io_ring_ctx *ctx = req->ctx;
999 * If this is a timeout link, we could be racing with the
1000 * timeout timer. Grab the completion lock for this case to
1001 * protect against that.
1003 spin_lock_irqsave(&ctx->completion_lock, flags);
1004 io_req_link_next(req, nxt);
1005 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1007 io_req_link_next(req, nxt);
1011 static void io_free_req(struct io_kiocb *req)
1013 struct io_kiocb *nxt = NULL;
1015 io_req_find_next(req, &nxt);
1019 io_queue_async_work(nxt);
1023 * Drop reference to request, return next in chain (if there is one) if this
1024 * was the last reference to this request.
1026 __attribute__((nonnull))
1027 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1029 io_req_find_next(req, nxtptr);
1031 if (refcount_dec_and_test(&req->refs))
1035 static void io_put_req(struct io_kiocb *req)
1037 if (refcount_dec_and_test(&req->refs))
1042 * Must only be used if we don't need to care about links, usually from
1043 * within the completion handling itself.
1045 static void __io_double_put_req(struct io_kiocb *req)
1047 /* drop both submit and complete references */
1048 if (refcount_sub_and_test(2, &req->refs))
1052 static void io_double_put_req(struct io_kiocb *req)
1054 /* drop both submit and complete references */
1055 if (refcount_sub_and_test(2, &req->refs))
1059 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1061 struct io_rings *rings = ctx->rings;
1064 * noflush == true is from the waitqueue handler, just ensure we wake
1065 * up the task, and the next invocation will flush the entries. We
1066 * cannot safely to it from here.
1068 if (noflush && !list_empty(&ctx->cq_overflow_list))
1071 io_cqring_overflow_flush(ctx, false);
1073 /* See comment at the top of this file */
1075 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1078 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1080 struct io_rings *rings = ctx->rings;
1082 /* make sure SQ entry isn't read before tail */
1083 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1087 * Find and free completed poll iocbs
1089 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1090 struct list_head *done)
1092 void *reqs[IO_IOPOLL_BATCH];
1093 struct io_kiocb *req;
1097 while (!list_empty(done)) {
1098 req = list_first_entry(done, struct io_kiocb, list);
1099 list_del(&req->list);
1101 io_cqring_fill_event(req, req->result);
1104 if (refcount_dec_and_test(&req->refs)) {
1105 /* If we're not using fixed files, we have to pair the
1106 * completion part with the file put. Use regular
1107 * completions for those, only batch free for fixed
1108 * file and non-linked commands.
1110 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1111 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1113 reqs[to_free++] = req;
1114 if (to_free == ARRAY_SIZE(reqs))
1115 io_free_req_many(ctx, reqs, &to_free);
1122 io_commit_cqring(ctx);
1123 io_free_req_many(ctx, reqs, &to_free);
1126 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1129 struct io_kiocb *req, *tmp;
1135 * Only spin for completions if we don't have multiple devices hanging
1136 * off our complete list, and we're under the requested amount.
1138 spin = !ctx->poll_multi_file && *nr_events < min;
1141 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1142 struct kiocb *kiocb = &req->rw;
1145 * Move completed entries to our local list. If we find a
1146 * request that requires polling, break out and complete
1147 * the done list first, if we have entries there.
1149 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1150 list_move_tail(&req->list, &done);
1153 if (!list_empty(&done))
1156 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1165 if (!list_empty(&done))
1166 io_iopoll_complete(ctx, nr_events, &done);
1172 * Poll for a mininum of 'min' events. Note that if min == 0 we consider that a
1173 * non-spinning poll check - we'll still enter the driver poll loop, but only
1174 * as a non-spinning completion check.
1176 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1179 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1182 ret = io_do_iopoll(ctx, nr_events, min);
1185 if (!min || *nr_events >= min)
1193 * We can't just wait for polled events to come to us, we have to actively
1194 * find and complete them.
1196 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1198 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1201 mutex_lock(&ctx->uring_lock);
1202 while (!list_empty(&ctx->poll_list)) {
1203 unsigned int nr_events = 0;
1205 io_iopoll_getevents(ctx, &nr_events, 1);
1208 * Ensure we allow local-to-the-cpu processing to take place,
1209 * in this case we need to ensure that we reap all events.
1213 mutex_unlock(&ctx->uring_lock);
1216 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1219 int iters = 0, ret = 0;
1225 * Don't enter poll loop if we already have events pending.
1226 * If we do, we can potentially be spinning for commands that
1227 * already triggered a CQE (eg in error).
1229 if (io_cqring_events(ctx, false))
1233 * If a submit got punted to a workqueue, we can have the
1234 * application entering polling for a command before it gets
1235 * issued. That app will hold the uring_lock for the duration
1236 * of the poll right here, so we need to take a breather every
1237 * now and then to ensure that the issue has a chance to add
1238 * the poll to the issued list. Otherwise we can spin here
1239 * forever, while the workqueue is stuck trying to acquire the
1242 if (!(++iters & 7)) {
1243 mutex_unlock(&ctx->uring_lock);
1244 mutex_lock(&ctx->uring_lock);
1247 if (*nr_events < min)
1248 tmin = min - *nr_events;
1250 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1254 } while (min && !*nr_events && !need_resched());
1259 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1265 * We disallow the app entering submit/complete with polling, but we
1266 * still need to lock the ring to prevent racing with polled issue
1267 * that got punted to a workqueue.
1269 mutex_lock(&ctx->uring_lock);
1270 ret = __io_iopoll_check(ctx, nr_events, min);
1271 mutex_unlock(&ctx->uring_lock);
1275 static void kiocb_end_write(struct io_kiocb *req)
1278 * Tell lockdep we inherited freeze protection from submission
1281 if (req->flags & REQ_F_ISREG) {
1282 struct inode *inode = file_inode(req->file);
1284 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1286 file_end_write(req->file);
1289 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1291 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1293 if (kiocb->ki_flags & IOCB_WRITE)
1294 kiocb_end_write(req);
1296 if ((req->flags & REQ_F_LINK) && res != req->result)
1297 req->flags |= REQ_F_FAIL_LINK;
1298 io_cqring_add_event(req, res);
1301 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1303 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1305 io_complete_rw_common(kiocb, res);
1309 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1311 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1312 struct io_kiocb *nxt = NULL;
1314 io_complete_rw_common(kiocb, res);
1315 io_put_req_find_next(req, &nxt);
1320 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1322 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1324 if (kiocb->ki_flags & IOCB_WRITE)
1325 kiocb_end_write(req);
1327 if ((req->flags & REQ_F_LINK) && res != req->result)
1328 req->flags |= REQ_F_FAIL_LINK;
1331 req->flags |= REQ_F_IOPOLL_COMPLETED;
1335 * After the iocb has been issued, it's safe to be found on the poll list.
1336 * Adding the kiocb to the list AFTER submission ensures that we don't
1337 * find it from a io_iopoll_getevents() thread before the issuer is done
1338 * accessing the kiocb cookie.
1340 static void io_iopoll_req_issued(struct io_kiocb *req)
1342 struct io_ring_ctx *ctx = req->ctx;
1345 * Track whether we have multiple files in our lists. This will impact
1346 * how we do polling eventually, not spinning if we're on potentially
1347 * different devices.
1349 if (list_empty(&ctx->poll_list)) {
1350 ctx->poll_multi_file = false;
1351 } else if (!ctx->poll_multi_file) {
1352 struct io_kiocb *list_req;
1354 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1356 if (list_req->rw.ki_filp != req->rw.ki_filp)
1357 ctx->poll_multi_file = true;
1361 * For fast devices, IO may have already completed. If it has, add
1362 * it to the front so we find it first.
1364 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1365 list_add(&req->list, &ctx->poll_list);
1367 list_add_tail(&req->list, &ctx->poll_list);
1370 static void io_file_put(struct io_submit_state *state)
1373 int diff = state->has_refs - state->used_refs;
1376 fput_many(state->file, diff);
1382 * Get as many references to a file as we have IOs left in this submission,
1383 * assuming most submissions are for one file, or at least that each file
1384 * has more than one submission.
1386 static struct file *io_file_get(struct io_submit_state *state, int fd)
1392 if (state->fd == fd) {
1399 state->file = fget_many(fd, state->ios_left);
1404 state->has_refs = state->ios_left;
1405 state->used_refs = 1;
1411 * If we tracked the file through the SCM inflight mechanism, we could support
1412 * any file. For now, just ensure that anything potentially problematic is done
1415 static bool io_file_supports_async(struct file *file)
1417 umode_t mode = file_inode(file)->i_mode;
1419 if (S_ISBLK(mode) || S_ISCHR(mode))
1421 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1427 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1429 const struct io_uring_sqe *sqe = req->sqe;
1430 struct io_ring_ctx *ctx = req->ctx;
1431 struct kiocb *kiocb = &req->rw;
1438 if (S_ISREG(file_inode(req->file)->i_mode))
1439 req->flags |= REQ_F_ISREG;
1441 kiocb->ki_pos = READ_ONCE(sqe->off);
1442 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1443 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1445 ioprio = READ_ONCE(sqe->ioprio);
1447 ret = ioprio_check_cap(ioprio);
1451 kiocb->ki_ioprio = ioprio;
1453 kiocb->ki_ioprio = get_current_ioprio();
1455 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1459 /* don't allow async punt if RWF_NOWAIT was requested */
1460 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1461 (req->file->f_flags & O_NONBLOCK))
1462 req->flags |= REQ_F_NOWAIT;
1465 kiocb->ki_flags |= IOCB_NOWAIT;
1467 if (ctx->flags & IORING_SETUP_IOPOLL) {
1468 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1469 !kiocb->ki_filp->f_op->iopoll)
1472 kiocb->ki_flags |= IOCB_HIPRI;
1473 kiocb->ki_complete = io_complete_rw_iopoll;
1476 if (kiocb->ki_flags & IOCB_HIPRI)
1478 kiocb->ki_complete = io_complete_rw;
1483 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1489 case -ERESTARTNOINTR:
1490 case -ERESTARTNOHAND:
1491 case -ERESTART_RESTARTBLOCK:
1493 * We can't just restart the syscall, since previously
1494 * submitted sqes may already be in progress. Just fail this
1500 kiocb->ki_complete(kiocb, ret, 0);
1504 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1507 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1508 *nxt = __io_complete_rw(kiocb, ret);
1510 io_rw_done(kiocb, ret);
1513 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1514 const struct io_uring_sqe *sqe,
1515 struct iov_iter *iter)
1517 size_t len = READ_ONCE(sqe->len);
1518 struct io_mapped_ubuf *imu;
1519 unsigned index, buf_index;
1523 /* attempt to use fixed buffers without having provided iovecs */
1524 if (unlikely(!ctx->user_bufs))
1527 buf_index = READ_ONCE(sqe->buf_index);
1528 if (unlikely(buf_index >= ctx->nr_user_bufs))
1531 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1532 imu = &ctx->user_bufs[index];
1533 buf_addr = READ_ONCE(sqe->addr);
1536 if (buf_addr + len < buf_addr)
1538 /* not inside the mapped region */
1539 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1543 * May not be a start of buffer, set size appropriately
1544 * and advance us to the beginning.
1546 offset = buf_addr - imu->ubuf;
1547 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1551 * Don't use iov_iter_advance() here, as it's really slow for
1552 * using the latter parts of a big fixed buffer - it iterates
1553 * over each segment manually. We can cheat a bit here, because
1556 * 1) it's a BVEC iter, we set it up
1557 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1558 * first and last bvec
1560 * So just find our index, and adjust the iterator afterwards.
1561 * If the offset is within the first bvec (or the whole first
1562 * bvec, just use iov_iter_advance(). This makes it easier
1563 * since we can just skip the first segment, which may not
1564 * be PAGE_SIZE aligned.
1566 const struct bio_vec *bvec = imu->bvec;
1568 if (offset <= bvec->bv_len) {
1569 iov_iter_advance(iter, offset);
1571 unsigned long seg_skip;
1573 /* skip first vec */
1574 offset -= bvec->bv_len;
1575 seg_skip = 1 + (offset >> PAGE_SHIFT);
1577 iter->bvec = bvec + seg_skip;
1578 iter->nr_segs -= seg_skip;
1579 iter->count -= bvec->bv_len + offset;
1580 iter->iov_offset = offset & ~PAGE_MASK;
1587 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1588 struct iovec **iovec, struct iov_iter *iter)
1590 const struct io_uring_sqe *sqe = req->sqe;
1591 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1592 size_t sqe_len = READ_ONCE(sqe->len);
1596 * We're reading ->opcode for the second time, but the first read
1597 * doesn't care whether it's _FIXED or not, so it doesn't matter
1598 * whether ->opcode changes concurrently. The first read does care
1599 * about whether it is a READ or a WRITE, so we don't trust this read
1600 * for that purpose and instead let the caller pass in the read/write
1603 opcode = READ_ONCE(sqe->opcode);
1604 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1606 return io_import_fixed(req->ctx, rw, sqe, iter);
1610 struct io_async_rw *iorw = &req->io->rw;
1613 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1614 if (iorw->iov == iorw->fast_iov)
1622 #ifdef CONFIG_COMPAT
1623 if (req->ctx->compat)
1624 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1628 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1632 * For files that don't have ->read_iter() and ->write_iter(), handle them
1633 * by looping over ->read() or ->write() manually.
1635 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1636 struct iov_iter *iter)
1641 * Don't support polled IO through this interface, and we can't
1642 * support non-blocking either. For the latter, this just causes
1643 * the kiocb to be handled from an async context.
1645 if (kiocb->ki_flags & IOCB_HIPRI)
1647 if (kiocb->ki_flags & IOCB_NOWAIT)
1650 while (iov_iter_count(iter)) {
1654 if (!iov_iter_is_bvec(iter)) {
1655 iovec = iov_iter_iovec(iter);
1657 /* fixed buffers import bvec */
1658 iovec.iov_base = kmap(iter->bvec->bv_page)
1660 iovec.iov_len = min(iter->count,
1661 iter->bvec->bv_len - iter->iov_offset);
1665 nr = file->f_op->read(file, iovec.iov_base,
1666 iovec.iov_len, &kiocb->ki_pos);
1668 nr = file->f_op->write(file, iovec.iov_base,
1669 iovec.iov_len, &kiocb->ki_pos);
1672 if (iov_iter_is_bvec(iter))
1673 kunmap(iter->bvec->bv_page);
1681 if (nr != iovec.iov_len)
1683 iov_iter_advance(iter, nr);
1689 static void io_req_map_io(struct io_kiocb *req, ssize_t io_size,
1690 struct iovec *iovec, struct iovec *fast_iov,
1691 struct iov_iter *iter)
1693 req->io->rw.nr_segs = iter->nr_segs;
1694 req->io->rw.size = io_size;
1695 req->io->rw.iov = iovec;
1696 if (!req->io->rw.iov) {
1697 req->io->rw.iov = req->io->rw.fast_iov;
1698 memcpy(req->io->rw.iov, fast_iov,
1699 sizeof(struct iovec) * iter->nr_segs);
1703 static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size,
1704 struct iovec *iovec, struct iovec *fast_iov,
1705 struct iov_iter *iter)
1707 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1709 io_req_map_io(req, io_size, iovec, fast_iov, iter);
1710 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1711 req->sqe = &req->io->sqe;
1718 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1719 struct iov_iter *iter, bool force_nonblock)
1723 ret = io_prep_rw(req, force_nonblock);
1727 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1730 return io_import_iovec(READ, req, iovec, iter);
1733 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1734 bool force_nonblock)
1736 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1737 struct kiocb *kiocb = &req->rw;
1738 struct iov_iter iter;
1741 ssize_t io_size, ret;
1744 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1748 ret = io_import_iovec(READ, req, &iovec, &iter);
1755 if (req->flags & REQ_F_LINK)
1756 req->result = io_size;
1759 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1760 * we know to async punt it even if it was opened O_NONBLOCK
1762 if (force_nonblock && !io_file_supports_async(file)) {
1763 req->flags |= REQ_F_MUST_PUNT;
1767 iov_count = iov_iter_count(&iter);
1768 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1772 if (file->f_op->read_iter)
1773 ret2 = call_read_iter(file, kiocb, &iter);
1775 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1778 * In case of a short read, punt to async. This can happen
1779 * if we have data partially cached. Alternatively we can
1780 * return the short read, in which case the application will
1781 * need to issue another SQE and wait for it. That SQE will
1782 * need async punt anyway, so it's more efficient to do it
1785 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1786 (req->flags & REQ_F_ISREG) &&
1787 ret2 > 0 && ret2 < io_size)
1789 /* Catch -EAGAIN return for forced non-blocking submission */
1790 if (!force_nonblock || ret2 != -EAGAIN) {
1791 kiocb_done(kiocb, ret2, nxt, req->in_async);
1794 ret = io_setup_async_io(req, io_size, iovec,
1795 inline_vecs, &iter);
1806 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1807 struct iov_iter *iter, bool force_nonblock)
1811 ret = io_prep_rw(req, force_nonblock);
1815 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1818 return io_import_iovec(WRITE, req, iovec, iter);
1821 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1822 bool force_nonblock)
1824 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1825 struct kiocb *kiocb = &req->rw;
1826 struct iov_iter iter;
1829 ssize_t ret, io_size;
1832 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1836 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1841 file = kiocb->ki_filp;
1843 if (req->flags & REQ_F_LINK)
1844 req->result = io_size;
1847 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1848 * we know to async punt it even if it was opened O_NONBLOCK
1850 if (force_nonblock && !io_file_supports_async(req->file)) {
1851 req->flags |= REQ_F_MUST_PUNT;
1855 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT))
1858 iov_count = iov_iter_count(&iter);
1859 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1864 * Open-code file_start_write here to grab freeze protection,
1865 * which will be released by another thread in
1866 * io_complete_rw(). Fool lockdep by telling it the lock got
1867 * released so that it doesn't complain about the held lock when
1868 * we return to userspace.
1870 if (req->flags & REQ_F_ISREG) {
1871 __sb_start_write(file_inode(file)->i_sb,
1872 SB_FREEZE_WRITE, true);
1873 __sb_writers_release(file_inode(file)->i_sb,
1876 kiocb->ki_flags |= IOCB_WRITE;
1878 if (file->f_op->write_iter)
1879 ret2 = call_write_iter(file, kiocb, &iter);
1881 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1882 if (!force_nonblock || ret2 != -EAGAIN) {
1883 kiocb_done(kiocb, ret2, nxt, req->in_async);
1886 ret = io_setup_async_io(req, io_size, iovec,
1887 inline_vecs, &iter);
1899 * IORING_OP_NOP just posts a completion event, nothing else.
1901 static int io_nop(struct io_kiocb *req)
1903 struct io_ring_ctx *ctx = req->ctx;
1905 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1908 io_cqring_add_event(req, 0);
1913 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1915 struct io_ring_ctx *ctx = req->ctx;
1920 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1922 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1928 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1929 struct io_kiocb **nxt, bool force_nonblock)
1931 loff_t sqe_off = READ_ONCE(sqe->off);
1932 loff_t sqe_len = READ_ONCE(sqe->len);
1933 loff_t end = sqe_off + sqe_len;
1934 unsigned fsync_flags;
1937 fsync_flags = READ_ONCE(sqe->fsync_flags);
1938 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1941 ret = io_prep_fsync(req, sqe);
1945 /* fsync always requires a blocking context */
1949 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1950 end > 0 ? end : LLONG_MAX,
1951 fsync_flags & IORING_FSYNC_DATASYNC);
1953 if (ret < 0 && (req->flags & REQ_F_LINK))
1954 req->flags |= REQ_F_FAIL_LINK;
1955 io_cqring_add_event(req, ret);
1956 io_put_req_find_next(req, nxt);
1960 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1962 struct io_ring_ctx *ctx = req->ctx;
1968 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1970 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1976 static int io_sync_file_range(struct io_kiocb *req,
1977 const struct io_uring_sqe *sqe,
1978 struct io_kiocb **nxt,
1979 bool force_nonblock)
1986 ret = io_prep_sfr(req, sqe);
1990 /* sync_file_range always requires a blocking context */
1994 sqe_off = READ_ONCE(sqe->off);
1995 sqe_len = READ_ONCE(sqe->len);
1996 flags = READ_ONCE(sqe->sync_range_flags);
1998 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
2000 if (ret < 0 && (req->flags & REQ_F_LINK))
2001 req->flags |= REQ_F_FAIL_LINK;
2002 io_cqring_add_event(req, ret);
2003 io_put_req_find_next(req, nxt);
2007 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2009 #if defined(CONFIG_NET)
2010 const struct io_uring_sqe *sqe = req->sqe;
2011 struct user_msghdr __user *msg;
2014 flags = READ_ONCE(sqe->msg_flags);
2015 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2016 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2022 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2023 struct io_kiocb **nxt, bool force_nonblock)
2025 #if defined(CONFIG_NET)
2026 struct socket *sock;
2029 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2032 sock = sock_from_file(req->file, &ret);
2034 struct io_async_ctx io, *copy;
2035 struct sockaddr_storage addr;
2036 struct msghdr *kmsg;
2039 flags = READ_ONCE(sqe->msg_flags);
2040 if (flags & MSG_DONTWAIT)
2041 req->flags |= REQ_F_NOWAIT;
2042 else if (force_nonblock)
2043 flags |= MSG_DONTWAIT;
2046 kmsg = &req->io->msg.msg;
2047 kmsg->msg_name = &addr;
2050 kmsg->msg_name = &addr;
2051 io.msg.iov = io.msg.fast_iov;
2052 ret = io_sendmsg_prep(req, &io);
2057 ret = __sys_sendmsg_sock(sock, kmsg, flags);
2058 if (force_nonblock && ret == -EAGAIN) {
2059 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2064 memcpy(©->msg, &io.msg, sizeof(copy->msg));
2066 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2067 req->sqe = &req->io->sqe;
2070 if (ret == -ERESTARTSYS)
2075 io_cqring_add_event(req, ret);
2076 if (ret < 0 && (req->flags & REQ_F_LINK))
2077 req->flags |= REQ_F_FAIL_LINK;
2078 io_put_req_find_next(req, nxt);
2085 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2087 #if defined(CONFIG_NET)
2088 const struct io_uring_sqe *sqe = req->sqe;
2089 struct user_msghdr __user *msg;
2092 flags = READ_ONCE(sqe->msg_flags);
2093 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2094 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2101 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2102 struct io_kiocb **nxt, bool force_nonblock)
2104 #if defined(CONFIG_NET)
2105 struct socket *sock;
2108 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2111 sock = sock_from_file(req->file, &ret);
2113 struct user_msghdr __user *msg;
2114 struct io_async_ctx io, *copy;
2115 struct sockaddr_storage addr;
2116 struct msghdr *kmsg;
2119 flags = READ_ONCE(sqe->msg_flags);
2120 if (flags & MSG_DONTWAIT)
2121 req->flags |= REQ_F_NOWAIT;
2122 else if (force_nonblock)
2123 flags |= MSG_DONTWAIT;
2125 msg = (struct user_msghdr __user *) (unsigned long)
2126 READ_ONCE(sqe->addr);
2128 kmsg = &req->io->msg.msg;
2129 kmsg->msg_name = &addr;
2132 kmsg->msg_name = &addr;
2133 io.msg.iov = io.msg.fast_iov;
2134 ret = io_recvmsg_prep(req, &io);
2139 ret = __sys_recvmsg_sock(sock, kmsg, msg, io.msg.uaddr, flags);
2140 if (force_nonblock && ret == -EAGAIN) {
2141 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2146 memcpy(copy, &io, sizeof(*copy));
2148 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2149 req->sqe = &req->io->sqe;
2152 if (ret == -ERESTARTSYS)
2157 io_cqring_add_event(req, ret);
2158 if (ret < 0 && (req->flags & REQ_F_LINK))
2159 req->flags |= REQ_F_FAIL_LINK;
2160 io_put_req_find_next(req, nxt);
2167 static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2168 struct io_kiocb **nxt, bool force_nonblock)
2170 #if defined(CONFIG_NET)
2171 struct sockaddr __user *addr;
2172 int __user *addr_len;
2173 unsigned file_flags;
2176 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2178 if (sqe->ioprio || sqe->len || sqe->buf_index)
2181 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2182 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2183 flags = READ_ONCE(sqe->accept_flags);
2184 file_flags = force_nonblock ? O_NONBLOCK : 0;
2186 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
2187 if (ret == -EAGAIN && force_nonblock) {
2188 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2191 if (ret == -ERESTARTSYS)
2193 if (ret < 0 && (req->flags & REQ_F_LINK))
2194 req->flags |= REQ_F_FAIL_LINK;
2195 io_cqring_add_event(req, ret);
2196 io_put_req_find_next(req, nxt);
2203 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2205 #if defined(CONFIG_NET)
2206 const struct io_uring_sqe *sqe = req->sqe;
2207 struct sockaddr __user *addr;
2210 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2211 addr_len = READ_ONCE(sqe->addr2);
2212 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2218 static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2219 struct io_kiocb **nxt, bool force_nonblock)
2221 #if defined(CONFIG_NET)
2222 struct io_async_ctx __io, *io;
2223 unsigned file_flags;
2226 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2228 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2231 addr_len = READ_ONCE(sqe->addr2);
2232 file_flags = force_nonblock ? O_NONBLOCK : 0;
2237 ret = io_connect_prep(req, &__io);
2243 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2245 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2246 io = kmalloc(sizeof(*io), GFP_KERNEL);
2251 memcpy(&io->connect, &__io.connect, sizeof(io->connect));
2253 memcpy(&io->sqe, req->sqe, sizeof(*req->sqe));
2254 req->sqe = &io->sqe;
2257 if (ret == -ERESTARTSYS)
2260 if (ret < 0 && (req->flags & REQ_F_LINK))
2261 req->flags |= REQ_F_FAIL_LINK;
2262 io_cqring_add_event(req, ret);
2263 io_put_req_find_next(req, nxt);
2270 static inline void io_poll_remove_req(struct io_kiocb *req)
2272 if (!RB_EMPTY_NODE(&req->rb_node)) {
2273 rb_erase(&req->rb_node, &req->ctx->cancel_tree);
2274 RB_CLEAR_NODE(&req->rb_node);
2278 static void io_poll_remove_one(struct io_kiocb *req)
2280 struct io_poll_iocb *poll = &req->poll;
2282 spin_lock(&poll->head->lock);
2283 WRITE_ONCE(poll->canceled, true);
2284 if (!list_empty(&poll->wait->entry)) {
2285 list_del_init(&poll->wait->entry);
2286 io_queue_async_work(req);
2288 spin_unlock(&poll->head->lock);
2289 io_poll_remove_req(req);
2292 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2294 struct rb_node *node;
2295 struct io_kiocb *req;
2297 spin_lock_irq(&ctx->completion_lock);
2298 while ((node = rb_first(&ctx->cancel_tree)) != NULL) {
2299 req = rb_entry(node, struct io_kiocb, rb_node);
2300 io_poll_remove_one(req);
2302 spin_unlock_irq(&ctx->completion_lock);
2305 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2307 struct rb_node *p, *parent = NULL;
2308 struct io_kiocb *req;
2310 p = ctx->cancel_tree.rb_node;
2313 req = rb_entry(parent, struct io_kiocb, rb_node);
2314 if (sqe_addr < req->user_data) {
2316 } else if (sqe_addr > req->user_data) {
2319 io_poll_remove_one(req);
2328 * Find a running poll command that matches one specified in sqe->addr,
2329 * and remove it if found.
2331 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2333 struct io_ring_ctx *ctx = req->ctx;
2336 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2338 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2342 spin_lock_irq(&ctx->completion_lock);
2343 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
2344 spin_unlock_irq(&ctx->completion_lock);
2346 io_cqring_add_event(req, ret);
2347 if (ret < 0 && (req->flags & REQ_F_LINK))
2348 req->flags |= REQ_F_FAIL_LINK;
2353 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2355 struct io_ring_ctx *ctx = req->ctx;
2357 req->poll.done = true;
2358 kfree(req->poll.wait);
2360 io_cqring_fill_event(req, error);
2362 io_cqring_fill_event(req, mangle_poll(mask));
2363 io_commit_cqring(ctx);
2366 static void io_poll_complete_work(struct io_wq_work **workptr)
2368 struct io_wq_work *work = *workptr;
2369 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2370 struct io_poll_iocb *poll = &req->poll;
2371 struct poll_table_struct pt = { ._key = poll->events };
2372 struct io_ring_ctx *ctx = req->ctx;
2373 struct io_kiocb *nxt = NULL;
2377 if (work->flags & IO_WQ_WORK_CANCEL) {
2378 WRITE_ONCE(poll->canceled, true);
2380 } else if (READ_ONCE(poll->canceled)) {
2384 if (ret != -ECANCELED)
2385 mask = vfs_poll(poll->file, &pt) & poll->events;
2388 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2389 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2390 * synchronize with them. In the cancellation case the list_del_init
2391 * itself is not actually needed, but harmless so we keep it in to
2392 * avoid further branches in the fast path.
2394 spin_lock_irq(&ctx->completion_lock);
2395 if (!mask && ret != -ECANCELED) {
2396 add_wait_queue(poll->head, poll->wait);
2397 spin_unlock_irq(&ctx->completion_lock);
2400 io_poll_remove_req(req);
2401 io_poll_complete(req, mask, ret);
2402 spin_unlock_irq(&ctx->completion_lock);
2404 io_cqring_ev_posted(ctx);
2406 if (ret < 0 && req->flags & REQ_F_LINK)
2407 req->flags |= REQ_F_FAIL_LINK;
2408 io_put_req_find_next(req, &nxt);
2410 *workptr = &nxt->work;
2413 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2416 struct io_poll_iocb *poll = wait->private;
2417 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2418 struct io_ring_ctx *ctx = req->ctx;
2419 __poll_t mask = key_to_poll(key);
2420 unsigned long flags;
2422 /* for instances that support it check for an event match first: */
2423 if (mask && !(mask & poll->events))
2426 list_del_init(&poll->wait->entry);
2429 * Run completion inline if we can. We're using trylock here because
2430 * we are violating the completion_lock -> poll wq lock ordering.
2431 * If we have a link timeout we're going to need the completion_lock
2432 * for finalizing the request, mark us as having grabbed that already.
2434 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2435 io_poll_remove_req(req);
2436 io_poll_complete(req, mask, 0);
2437 req->flags |= REQ_F_COMP_LOCKED;
2439 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2441 io_cqring_ev_posted(ctx);
2443 io_queue_async_work(req);
2449 struct io_poll_table {
2450 struct poll_table_struct pt;
2451 struct io_kiocb *req;
2455 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2456 struct poll_table_struct *p)
2458 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2460 if (unlikely(pt->req->poll.head)) {
2461 pt->error = -EINVAL;
2466 pt->req->poll.head = head;
2467 add_wait_queue(head, pt->req->poll.wait);
2470 static void io_poll_req_insert(struct io_kiocb *req)
2472 struct io_ring_ctx *ctx = req->ctx;
2473 struct rb_node **p = &ctx->cancel_tree.rb_node;
2474 struct rb_node *parent = NULL;
2475 struct io_kiocb *tmp;
2479 tmp = rb_entry(parent, struct io_kiocb, rb_node);
2480 if (req->user_data < tmp->user_data)
2483 p = &(*p)->rb_right;
2485 rb_link_node(&req->rb_node, parent, p);
2486 rb_insert_color(&req->rb_node, &ctx->cancel_tree);
2489 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2490 struct io_kiocb **nxt)
2492 struct io_poll_iocb *poll = &req->poll;
2493 struct io_ring_ctx *ctx = req->ctx;
2494 struct io_poll_table ipt;
2495 bool cancel = false;
2499 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2501 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2506 poll->wait = kmalloc(sizeof(*poll->wait), GFP_KERNEL);
2511 INIT_IO_WORK(&req->work, io_poll_complete_work);
2512 events = READ_ONCE(sqe->poll_events);
2513 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2514 RB_CLEAR_NODE(&req->rb_node);
2518 poll->canceled = false;
2520 ipt.pt._qproc = io_poll_queue_proc;
2521 ipt.pt._key = poll->events;
2523 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2525 /* initialized the list so that we can do list_empty checks */
2526 INIT_LIST_HEAD(&poll->wait->entry);
2527 init_waitqueue_func_entry(poll->wait, io_poll_wake);
2528 poll->wait->private = poll;
2530 INIT_LIST_HEAD(&req->list);
2532 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2534 spin_lock_irq(&ctx->completion_lock);
2535 if (likely(poll->head)) {
2536 spin_lock(&poll->head->lock);
2537 if (unlikely(list_empty(&poll->wait->entry))) {
2543 if (mask || ipt.error)
2544 list_del_init(&poll->wait->entry);
2546 WRITE_ONCE(poll->canceled, true);
2547 else if (!poll->done) /* actually waiting for an event */
2548 io_poll_req_insert(req);
2549 spin_unlock(&poll->head->lock);
2551 if (mask) { /* no async, we'd stolen it */
2553 io_poll_complete(req, mask, 0);
2555 spin_unlock_irq(&ctx->completion_lock);
2558 io_cqring_ev_posted(ctx);
2559 io_put_req_find_next(req, nxt);
2564 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2566 struct io_timeout_data *data = container_of(timer,
2567 struct io_timeout_data, timer);
2568 struct io_kiocb *req = data->req;
2569 struct io_ring_ctx *ctx = req->ctx;
2570 unsigned long flags;
2572 atomic_inc(&ctx->cq_timeouts);
2574 spin_lock_irqsave(&ctx->completion_lock, flags);
2576 * We could be racing with timeout deletion. If the list is empty,
2577 * then timeout lookup already found it and will be handling it.
2579 if (!list_empty(&req->list)) {
2580 struct io_kiocb *prev;
2583 * Adjust the reqs sequence before the current one because it
2584 * will consume a slot in the cq_ring and the the cq_tail
2585 * pointer will be increased, otherwise other timeout reqs may
2586 * return in advance without waiting for enough wait_nr.
2589 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2591 list_del_init(&req->list);
2594 io_cqring_fill_event(req, -ETIME);
2595 io_commit_cqring(ctx);
2596 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2598 io_cqring_ev_posted(ctx);
2599 if (req->flags & REQ_F_LINK)
2600 req->flags |= REQ_F_FAIL_LINK;
2602 return HRTIMER_NORESTART;
2605 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2607 struct io_kiocb *req;
2610 list_for_each_entry(req, &ctx->timeout_list, list) {
2611 if (user_data == req->user_data) {
2612 list_del_init(&req->list);
2621 ret = hrtimer_try_to_cancel(&req->timeout.data->timer);
2625 if (req->flags & REQ_F_LINK)
2626 req->flags |= REQ_F_FAIL_LINK;
2627 io_cqring_fill_event(req, -ECANCELED);
2633 * Remove or update an existing timeout command
2635 static int io_timeout_remove(struct io_kiocb *req,
2636 const struct io_uring_sqe *sqe)
2638 struct io_ring_ctx *ctx = req->ctx;
2642 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2644 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2646 flags = READ_ONCE(sqe->timeout_flags);
2650 spin_lock_irq(&ctx->completion_lock);
2651 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
2653 io_cqring_fill_event(req, ret);
2654 io_commit_cqring(ctx);
2655 spin_unlock_irq(&ctx->completion_lock);
2656 io_cqring_ev_posted(ctx);
2657 if (ret < 0 && req->flags & REQ_F_LINK)
2658 req->flags |= REQ_F_FAIL_LINK;
2663 static int io_timeout_setup(struct io_kiocb *req)
2665 const struct io_uring_sqe *sqe = req->sqe;
2666 struct io_timeout_data *data;
2669 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2671 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2673 flags = READ_ONCE(sqe->timeout_flags);
2674 if (flags & ~IORING_TIMEOUT_ABS)
2677 data = kzalloc(sizeof(struct io_timeout_data), GFP_KERNEL);
2681 req->timeout.data = data;
2682 req->flags |= REQ_F_TIMEOUT;
2684 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2687 if (flags & IORING_TIMEOUT_ABS)
2688 data->mode = HRTIMER_MODE_ABS;
2690 data->mode = HRTIMER_MODE_REL;
2692 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2696 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2699 struct io_ring_ctx *ctx = req->ctx;
2700 struct io_timeout_data *data;
2701 struct list_head *entry;
2705 ret = io_timeout_setup(req);
2710 * sqe->off holds how many events that need to occur for this
2711 * timeout event to be satisfied. If it isn't set, then this is
2712 * a pure timeout request, sequence isn't used.
2714 count = READ_ONCE(sqe->off);
2716 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2717 spin_lock_irq(&ctx->completion_lock);
2718 entry = ctx->timeout_list.prev;
2722 req->sequence = ctx->cached_sq_head + count - 1;
2723 req->timeout.data->seq_offset = count;
2726 * Insertion sort, ensuring the first entry in the list is always
2727 * the one we need first.
2729 spin_lock_irq(&ctx->completion_lock);
2730 list_for_each_prev(entry, &ctx->timeout_list) {
2731 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2732 unsigned nxt_sq_head;
2733 long long tmp, tmp_nxt;
2734 u32 nxt_offset = nxt->timeout.data->seq_offset;
2736 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2740 * Since cached_sq_head + count - 1 can overflow, use type long
2743 tmp = (long long)ctx->cached_sq_head + count - 1;
2744 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2745 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2748 * cached_sq_head may overflow, and it will never overflow twice
2749 * once there is some timeout req still be valid.
2751 if (ctx->cached_sq_head < nxt_sq_head)
2758 * Sequence of reqs after the insert one and itself should
2759 * be adjusted because each timeout req consumes a slot.
2764 req->sequence -= span;
2766 list_add(&req->list, entry);
2767 data = req->timeout.data;
2768 data->timer.function = io_timeout_fn;
2769 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2770 spin_unlock_irq(&ctx->completion_lock);
2774 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2776 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2778 return req->user_data == (unsigned long) data;
2781 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2783 enum io_wq_cancel cancel_ret;
2786 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2787 switch (cancel_ret) {
2788 case IO_WQ_CANCEL_OK:
2791 case IO_WQ_CANCEL_RUNNING:
2794 case IO_WQ_CANCEL_NOTFOUND:
2802 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2803 struct io_kiocb *req, __u64 sqe_addr,
2804 struct io_kiocb **nxt, int success_ret)
2806 unsigned long flags;
2809 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2810 if (ret != -ENOENT) {
2811 spin_lock_irqsave(&ctx->completion_lock, flags);
2815 spin_lock_irqsave(&ctx->completion_lock, flags);
2816 ret = io_timeout_cancel(ctx, sqe_addr);
2819 ret = io_poll_cancel(ctx, sqe_addr);
2823 io_cqring_fill_event(req, ret);
2824 io_commit_cqring(ctx);
2825 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2826 io_cqring_ev_posted(ctx);
2828 if (ret < 0 && (req->flags & REQ_F_LINK))
2829 req->flags |= REQ_F_FAIL_LINK;
2830 io_put_req_find_next(req, nxt);
2833 static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2834 struct io_kiocb **nxt)
2836 struct io_ring_ctx *ctx = req->ctx;
2838 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2840 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2844 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
2848 static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io)
2850 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2851 struct iov_iter iter;
2854 memcpy(&io->sqe, req->sqe, sizeof(io->sqe));
2855 req->sqe = &io->sqe;
2857 switch (io->sqe.opcode) {
2858 case IORING_OP_READV:
2859 case IORING_OP_READ_FIXED:
2860 ret = io_read_prep(req, &iovec, &iter, true);
2862 case IORING_OP_WRITEV:
2863 case IORING_OP_WRITE_FIXED:
2864 ret = io_write_prep(req, &iovec, &iter, true);
2866 case IORING_OP_SENDMSG:
2867 ret = io_sendmsg_prep(req, io);
2869 case IORING_OP_RECVMSG:
2870 ret = io_recvmsg_prep(req, io);
2872 case IORING_OP_CONNECT:
2873 ret = io_connect_prep(req, io);
2884 io_req_map_io(req, ret, iovec, inline_vecs, &iter);
2888 static int io_req_defer(struct io_kiocb *req)
2890 struct io_ring_ctx *ctx = req->ctx;
2891 struct io_async_ctx *io;
2894 /* Still need defer if there is pending req in defer list. */
2895 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
2898 io = kmalloc(sizeof(*io), GFP_KERNEL);
2902 spin_lock_irq(&ctx->completion_lock);
2903 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
2904 spin_unlock_irq(&ctx->completion_lock);
2909 ret = io_req_defer_prep(req, io);
2913 trace_io_uring_defer(ctx, req, req->user_data);
2914 list_add_tail(&req->list, &ctx->defer_list);
2915 spin_unlock_irq(&ctx->completion_lock);
2916 return -EIOCBQUEUED;
2919 __attribute__((nonnull))
2920 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
2921 bool force_nonblock)
2924 struct io_ring_ctx *ctx = req->ctx;
2926 opcode = READ_ONCE(req->sqe->opcode);
2931 case IORING_OP_READV:
2932 if (unlikely(req->sqe->buf_index))
2934 ret = io_read(req, nxt, force_nonblock);
2936 case IORING_OP_WRITEV:
2937 if (unlikely(req->sqe->buf_index))
2939 ret = io_write(req, nxt, force_nonblock);
2941 case IORING_OP_READ_FIXED:
2942 ret = io_read(req, nxt, force_nonblock);
2944 case IORING_OP_WRITE_FIXED:
2945 ret = io_write(req, nxt, force_nonblock);
2947 case IORING_OP_FSYNC:
2948 ret = io_fsync(req, req->sqe, nxt, force_nonblock);
2950 case IORING_OP_POLL_ADD:
2951 ret = io_poll_add(req, req->sqe, nxt);
2953 case IORING_OP_POLL_REMOVE:
2954 ret = io_poll_remove(req, req->sqe);
2956 case IORING_OP_SYNC_FILE_RANGE:
2957 ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock);
2959 case IORING_OP_SENDMSG:
2960 ret = io_sendmsg(req, req->sqe, nxt, force_nonblock);
2962 case IORING_OP_RECVMSG:
2963 ret = io_recvmsg(req, req->sqe, nxt, force_nonblock);
2965 case IORING_OP_TIMEOUT:
2966 ret = io_timeout(req, req->sqe);
2968 case IORING_OP_TIMEOUT_REMOVE:
2969 ret = io_timeout_remove(req, req->sqe);
2971 case IORING_OP_ACCEPT:
2972 ret = io_accept(req, req->sqe, nxt, force_nonblock);
2974 case IORING_OP_CONNECT:
2975 ret = io_connect(req, req->sqe, nxt, force_nonblock);
2977 case IORING_OP_ASYNC_CANCEL:
2978 ret = io_async_cancel(req, req->sqe, nxt);
2988 if (ctx->flags & IORING_SETUP_IOPOLL) {
2989 if (req->result == -EAGAIN)
2992 /* workqueue context doesn't hold uring_lock, grab it now */
2994 mutex_lock(&ctx->uring_lock);
2995 io_iopoll_req_issued(req);
2997 mutex_unlock(&ctx->uring_lock);
3003 static void io_link_work_cb(struct io_wq_work **workptr)
3005 struct io_wq_work *work = *workptr;
3006 struct io_kiocb *link = work->data;
3008 io_queue_linked_timeout(link);
3009 work->func = io_wq_submit_work;
3012 static void io_wq_submit_work(struct io_wq_work **workptr)
3014 struct io_wq_work *work = *workptr;
3015 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3016 struct io_kiocb *nxt = NULL;
3019 /* Ensure we clear previously set non-block flag */
3020 req->rw.ki_flags &= ~IOCB_NOWAIT;
3022 if (work->flags & IO_WQ_WORK_CANCEL)
3026 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3027 req->in_async = true;
3029 ret = io_issue_sqe(req, &nxt, false);
3031 * We can get EAGAIN for polled IO even though we're
3032 * forcing a sync submission from here, since we can't
3033 * wait for request slots on the block side.
3041 /* drop submission reference */
3045 if (req->flags & REQ_F_LINK)
3046 req->flags |= REQ_F_FAIL_LINK;
3047 io_cqring_add_event(req, ret);
3051 /* if a dependent link is ready, pass it back */
3053 struct io_kiocb *link;
3055 io_prep_async_work(nxt, &link);
3056 *workptr = &nxt->work;
3058 nxt->work.flags |= IO_WQ_WORK_CB;
3059 nxt->work.func = io_link_work_cb;
3060 nxt->work.data = link;
3065 static bool io_op_needs_file(const struct io_uring_sqe *sqe)
3067 int op = READ_ONCE(sqe->opcode);
3071 case IORING_OP_POLL_REMOVE:
3072 case IORING_OP_TIMEOUT:
3073 case IORING_OP_TIMEOUT_REMOVE:
3074 case IORING_OP_ASYNC_CANCEL:
3075 case IORING_OP_LINK_TIMEOUT:
3082 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3085 struct fixed_file_table *table;
3087 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3088 return table->files[index & IORING_FILE_TABLE_MASK];
3091 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3093 struct io_ring_ctx *ctx = req->ctx;
3097 flags = READ_ONCE(req->sqe->flags);
3098 fd = READ_ONCE(req->sqe->fd);
3100 if (flags & IOSQE_IO_DRAIN)
3101 req->flags |= REQ_F_IO_DRAIN;
3103 if (!io_op_needs_file(req->sqe))
3106 if (flags & IOSQE_FIXED_FILE) {
3107 if (unlikely(!ctx->file_table ||
3108 (unsigned) fd >= ctx->nr_user_files))
3110 fd = array_index_nospec(fd, ctx->nr_user_files);
3111 req->file = io_file_from_index(ctx, fd);
3114 req->flags |= REQ_F_FIXED_FILE;
3116 if (req->needs_fixed_file)
3118 trace_io_uring_file_get(ctx, fd);
3119 req->file = io_file_get(state, fd);
3120 if (unlikely(!req->file))
3127 static int io_grab_files(struct io_kiocb *req)
3130 struct io_ring_ctx *ctx = req->ctx;
3133 spin_lock_irq(&ctx->inflight_lock);
3135 * We use the f_ops->flush() handler to ensure that we can flush
3136 * out work accessing these files if the fd is closed. Check if
3137 * the fd has changed since we started down this path, and disallow
3138 * this operation if it has.
3140 if (fcheck(req->ring_fd) == req->ring_file) {
3141 list_add(&req->inflight_entry, &ctx->inflight_list);
3142 req->flags |= REQ_F_INFLIGHT;
3143 req->work.files = current->files;
3146 spin_unlock_irq(&ctx->inflight_lock);
3152 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3154 struct io_timeout_data *data = container_of(timer,
3155 struct io_timeout_data, timer);
3156 struct io_kiocb *req = data->req;
3157 struct io_ring_ctx *ctx = req->ctx;
3158 struct io_kiocb *prev = NULL;
3159 unsigned long flags;
3161 spin_lock_irqsave(&ctx->completion_lock, flags);
3164 * We don't expect the list to be empty, that will only happen if we
3165 * race with the completion of the linked work.
3167 if (!list_empty(&req->list)) {
3168 prev = list_entry(req->list.prev, struct io_kiocb, link_list);
3169 if (refcount_inc_not_zero(&prev->refs)) {
3170 list_del_init(&req->list);
3171 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3176 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3179 if (prev->flags & REQ_F_LINK)
3180 prev->flags |= REQ_F_FAIL_LINK;
3181 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3185 io_cqring_add_event(req, -ETIME);
3188 return HRTIMER_NORESTART;
3191 static void io_queue_linked_timeout(struct io_kiocb *req)
3193 struct io_ring_ctx *ctx = req->ctx;
3196 * If the list is now empty, then our linked request finished before
3197 * we got a chance to setup the timer
3199 spin_lock_irq(&ctx->completion_lock);
3200 if (!list_empty(&req->list)) {
3201 struct io_timeout_data *data = req->timeout.data;
3203 data->timer.function = io_link_timeout_fn;
3204 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3207 spin_unlock_irq(&ctx->completion_lock);
3209 /* drop submission reference */
3213 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3215 struct io_kiocb *nxt;
3217 if (!(req->flags & REQ_F_LINK))
3220 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, list);
3221 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
3224 req->flags |= REQ_F_LINK_TIMEOUT;
3228 static void __io_queue_sqe(struct io_kiocb *req)
3230 struct io_kiocb *linked_timeout = io_prep_linked_timeout(req);
3231 struct io_kiocb *nxt = NULL;
3234 ret = io_issue_sqe(req, &nxt, true);
3236 io_queue_async_work(nxt);
3239 * We async punt it if the file wasn't marked NOWAIT, or if the file
3240 * doesn't support non-blocking read/write attempts
3242 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3243 (req->flags & REQ_F_MUST_PUNT))) {
3244 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3245 ret = io_grab_files(req);
3251 * Queued up for async execution, worker will release
3252 * submit reference when the iocb is actually submitted.
3254 io_queue_async_work(req);
3259 /* drop submission reference */
3262 if (linked_timeout) {
3264 io_queue_linked_timeout(linked_timeout);
3266 io_put_req(linked_timeout);
3269 /* and drop final reference, if we failed */
3271 io_cqring_add_event(req, ret);
3272 if (req->flags & REQ_F_LINK)
3273 req->flags |= REQ_F_FAIL_LINK;
3278 static void io_queue_sqe(struct io_kiocb *req)
3282 if (unlikely(req->ctx->drain_next)) {
3283 req->flags |= REQ_F_IO_DRAIN;
3284 req->ctx->drain_next = false;
3286 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3288 ret = io_req_defer(req);
3290 if (ret != -EIOCBQUEUED) {
3291 io_cqring_add_event(req, ret);
3292 if (req->flags & REQ_F_LINK)
3293 req->flags |= REQ_F_FAIL_LINK;
3294 io_double_put_req(req);
3297 __io_queue_sqe(req);
3300 static inline void io_queue_link_head(struct io_kiocb *req)
3302 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3303 io_cqring_add_event(req, -ECANCELED);
3304 io_double_put_req(req);
3310 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK)
3312 static void io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3313 struct io_kiocb **link)
3315 struct io_ring_ctx *ctx = req->ctx;
3318 req->user_data = req->sqe->user_data;
3320 /* enforce forwards compatibility on users */
3321 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3326 ret = io_req_set_file(state, req);
3327 if (unlikely(ret)) {
3329 io_cqring_add_event(req, ret);
3330 io_double_put_req(req);
3335 * If we already have a head request, queue this one for async
3336 * submittal once the head completes. If we don't have a head but
3337 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3338 * submitted sync once the chain is complete. If none of those
3339 * conditions are true (normal request), then just queue it.
3342 struct io_kiocb *prev = *link;
3343 struct io_async_ctx *io;
3345 if (req->sqe->flags & IOSQE_IO_DRAIN)
3346 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3348 if (READ_ONCE(req->sqe->opcode) == IORING_OP_LINK_TIMEOUT) {
3349 ret = io_timeout_setup(req);
3350 /* common setup allows offset being set, we don't */
3351 if (!ret && req->sqe->off)
3354 prev->flags |= REQ_F_FAIL_LINK;
3359 io = kmalloc(sizeof(*io), GFP_KERNEL);
3365 ret = io_req_defer_prep(req, io);
3368 trace_io_uring_link(ctx, req, prev);
3369 list_add_tail(&req->list, &prev->link_list);
3370 } else if (req->sqe->flags & IOSQE_IO_LINK) {
3371 req->flags |= REQ_F_LINK;
3373 INIT_LIST_HEAD(&req->link_list);
3381 * Batched submission is done, ensure local IO is flushed out.
3383 static void io_submit_state_end(struct io_submit_state *state)
3385 blk_finish_plug(&state->plug);
3387 if (state->free_reqs)
3388 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3389 &state->reqs[state->cur_req]);
3393 * Start submission side cache.
3395 static void io_submit_state_start(struct io_submit_state *state,
3396 unsigned int max_ios)
3398 blk_start_plug(&state->plug);
3399 state->free_reqs = 0;
3401 state->ios_left = max_ios;
3404 static void io_commit_sqring(struct io_ring_ctx *ctx)
3406 struct io_rings *rings = ctx->rings;
3408 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3410 * Ensure any loads from the SQEs are done at this point,
3411 * since once we write the new head, the application could
3412 * write new data to them.
3414 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3419 * Fetch an sqe, if one is available. Note that s->sqe will point to memory
3420 * that is mapped by userspace. This means that care needs to be taken to
3421 * ensure that reads are stable, as we cannot rely on userspace always
3422 * being a good citizen. If members of the sqe are validated and then later
3423 * used, it's important that those reads are done through READ_ONCE() to
3424 * prevent a re-load down the line.
3426 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3428 struct io_rings *rings = ctx->rings;
3429 u32 *sq_array = ctx->sq_array;
3433 * The cached sq head (or cq tail) serves two purposes:
3435 * 1) allows us to batch the cost of updating the user visible
3437 * 2) allows the kernel side to track the head on its own, even
3438 * though the application is the one updating it.
3440 head = ctx->cached_sq_head;
3441 /* make sure SQ entry isn't read before tail */
3442 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3445 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3446 if (likely(head < ctx->sq_entries)) {
3448 * All io need record the previous position, if LINK vs DARIN,
3449 * it can be used to mark the position of the first IO in the
3452 req->sequence = ctx->cached_sq_head;
3453 req->sqe = &ctx->sq_sqes[head];
3454 ctx->cached_sq_head++;
3458 /* drop invalid entries */
3459 ctx->cached_sq_head++;
3460 ctx->cached_sq_dropped++;
3461 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3465 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3466 struct file *ring_file, int ring_fd,
3467 struct mm_struct **mm, bool async)
3469 struct io_submit_state state, *statep = NULL;
3470 struct io_kiocb *link = NULL;
3471 int i, submitted = 0;
3472 bool mm_fault = false;
3474 /* if we have a backlog and couldn't flush it all, return BUSY */
3475 if (!list_empty(&ctx->cq_overflow_list) &&
3476 !io_cqring_overflow_flush(ctx, false))
3479 if (nr > IO_PLUG_THRESHOLD) {
3480 io_submit_state_start(&state, nr);
3484 for (i = 0; i < nr; i++) {
3485 struct io_kiocb *req;
3486 unsigned int sqe_flags;
3488 req = io_get_req(ctx, statep);
3489 if (unlikely(!req)) {
3491 submitted = -EAGAIN;
3494 if (!io_get_sqring(ctx, req)) {
3499 if (io_sqe_needs_user(req->sqe) && !*mm) {
3500 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3502 use_mm(ctx->sqo_mm);
3507 sqe_flags = req->sqe->flags;
3509 req->ring_file = ring_file;
3510 req->ring_fd = ring_fd;
3511 req->has_user = *mm != NULL;
3512 req->in_async = async;
3513 req->needs_fixed_file = async;
3514 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
3516 io_submit_sqe(req, statep, &link);
3520 * If previous wasn't linked and we have a linked command,
3521 * that's the end of the chain. Submit the previous link.
3523 if (!(sqe_flags & IOSQE_IO_LINK) && link) {
3524 io_queue_link_head(link);
3530 io_queue_link_head(link);
3532 io_submit_state_end(&state);
3534 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3535 io_commit_sqring(ctx);
3540 static int io_sq_thread(void *data)
3542 struct io_ring_ctx *ctx = data;
3543 struct mm_struct *cur_mm = NULL;
3544 const struct cred *old_cred;
3545 mm_segment_t old_fs;
3548 unsigned long timeout;
3551 complete(&ctx->completions[1]);
3555 old_cred = override_creds(ctx->creds);
3557 ret = timeout = inflight = 0;
3558 while (!kthread_should_park()) {
3559 unsigned int to_submit;
3562 unsigned nr_events = 0;
3564 if (ctx->flags & IORING_SETUP_IOPOLL) {
3566 * inflight is the count of the maximum possible
3567 * entries we submitted, but it can be smaller
3568 * if we dropped some of them. If we don't have
3569 * poll entries available, then we know that we
3570 * have nothing left to poll for. Reset the
3571 * inflight count to zero in that case.
3573 mutex_lock(&ctx->uring_lock);
3574 if (!list_empty(&ctx->poll_list))
3575 __io_iopoll_check(ctx, &nr_events, 0);
3578 mutex_unlock(&ctx->uring_lock);
3581 * Normal IO, just pretend everything completed.
3582 * We don't have to poll completions for that.
3584 nr_events = inflight;
3587 inflight -= nr_events;
3589 timeout = jiffies + ctx->sq_thread_idle;
3592 to_submit = io_sqring_entries(ctx);
3595 * If submit got -EBUSY, flag us as needing the application
3596 * to enter the kernel to reap and flush events.
3598 if (!to_submit || ret == -EBUSY) {
3600 * We're polling. If we're within the defined idle
3601 * period, then let us spin without work before going
3602 * to sleep. The exception is if we got EBUSY doing
3603 * more IO, we should wait for the application to
3604 * reap events and wake us up.
3607 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3613 * Drop cur_mm before scheduling, we can't hold it for
3614 * long periods (or over schedule()). Do this before
3615 * adding ourselves to the waitqueue, as the unuse/drop
3624 prepare_to_wait(&ctx->sqo_wait, &wait,
3625 TASK_INTERRUPTIBLE);
3627 /* Tell userspace we may need a wakeup call */
3628 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3629 /* make sure to read SQ tail after writing flags */
3632 to_submit = io_sqring_entries(ctx);
3633 if (!to_submit || ret == -EBUSY) {
3634 if (kthread_should_park()) {
3635 finish_wait(&ctx->sqo_wait, &wait);
3638 if (signal_pending(current))
3639 flush_signals(current);
3641 finish_wait(&ctx->sqo_wait, &wait);
3643 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3646 finish_wait(&ctx->sqo_wait, &wait);
3648 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3651 to_submit = min(to_submit, ctx->sq_entries);
3652 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3662 revert_creds(old_cred);
3669 struct io_wait_queue {
3670 struct wait_queue_entry wq;
3671 struct io_ring_ctx *ctx;
3673 unsigned nr_timeouts;
3676 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3678 struct io_ring_ctx *ctx = iowq->ctx;
3681 * Wake up if we have enough events, or if a timeout occured since we
3682 * started waiting. For timeouts, we always want to return to userspace,
3683 * regardless of event count.
3685 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3686 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3689 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3690 int wake_flags, void *key)
3692 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3695 /* use noflush == true, as we can't safely rely on locking context */
3696 if (!io_should_wake(iowq, true))
3699 return autoremove_wake_function(curr, mode, wake_flags, key);
3703 * Wait until events become available, if we don't already have some. The
3704 * application must reap them itself, as they reside on the shared cq ring.
3706 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3707 const sigset_t __user *sig, size_t sigsz)
3709 struct io_wait_queue iowq = {
3712 .func = io_wake_function,
3713 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3716 .to_wait = min_events,
3718 struct io_rings *rings = ctx->rings;
3721 if (io_cqring_events(ctx, false) >= min_events)
3725 #ifdef CONFIG_COMPAT
3726 if (in_compat_syscall())
3727 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3731 ret = set_user_sigmask(sig, sigsz);
3737 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
3738 trace_io_uring_cqring_wait(ctx, min_events);
3740 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3741 TASK_INTERRUPTIBLE);
3742 if (io_should_wake(&iowq, false))
3745 if (signal_pending(current)) {
3750 finish_wait(&ctx->wait, &iowq.wq);
3752 restore_saved_sigmask_unless(ret == -EINTR);
3754 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
3757 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3759 #if defined(CONFIG_UNIX)
3760 if (ctx->ring_sock) {
3761 struct sock *sock = ctx->ring_sock->sk;
3762 struct sk_buff *skb;
3764 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3770 for (i = 0; i < ctx->nr_user_files; i++) {
3773 file = io_file_from_index(ctx, i);
3780 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3782 unsigned nr_tables, i;
3784 if (!ctx->file_table)
3787 __io_sqe_files_unregister(ctx);
3788 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3789 for (i = 0; i < nr_tables; i++)
3790 kfree(ctx->file_table[i].files);
3791 kfree(ctx->file_table);
3792 ctx->file_table = NULL;
3793 ctx->nr_user_files = 0;
3797 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3799 if (ctx->sqo_thread) {
3800 wait_for_completion(&ctx->completions[1]);
3802 * The park is a bit of a work-around, without it we get
3803 * warning spews on shutdown with SQPOLL set and affinity
3804 * set to a single CPU.
3806 kthread_park(ctx->sqo_thread);
3807 kthread_stop(ctx->sqo_thread);
3808 ctx->sqo_thread = NULL;
3812 static void io_finish_async(struct io_ring_ctx *ctx)
3814 io_sq_thread_stop(ctx);
3817 io_wq_destroy(ctx->io_wq);
3822 #if defined(CONFIG_UNIX)
3823 static void io_destruct_skb(struct sk_buff *skb)
3825 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3828 io_wq_flush(ctx->io_wq);
3830 unix_destruct_scm(skb);
3834 * Ensure the UNIX gc is aware of our file set, so we are certain that
3835 * the io_uring can be safely unregistered on process exit, even if we have
3836 * loops in the file referencing.
3838 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3840 struct sock *sk = ctx->ring_sock->sk;
3841 struct scm_fp_list *fpl;
3842 struct sk_buff *skb;
3845 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3846 unsigned long inflight = ctx->user->unix_inflight + nr;
3848 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3852 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3856 skb = alloc_skb(0, GFP_KERNEL);
3865 fpl->user = get_uid(ctx->user);
3866 for (i = 0; i < nr; i++) {
3867 struct file *file = io_file_from_index(ctx, i + offset);
3871 fpl->fp[nr_files] = get_file(file);
3872 unix_inflight(fpl->user, fpl->fp[nr_files]);
3877 fpl->max = SCM_MAX_FD;
3878 fpl->count = nr_files;
3879 UNIXCB(skb).fp = fpl;
3880 skb->destructor = io_destruct_skb;
3881 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3882 skb_queue_head(&sk->sk_receive_queue, skb);
3884 for (i = 0; i < nr_files; i++)
3895 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3896 * causes regular reference counting to break down. We rely on the UNIX
3897 * garbage collection to take care of this problem for us.
3899 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3901 unsigned left, total;
3905 left = ctx->nr_user_files;
3907 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
3909 ret = __io_sqe_files_scm(ctx, this_files, total);
3913 total += this_files;
3919 while (total < ctx->nr_user_files) {
3920 struct file *file = io_file_from_index(ctx, total);
3930 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3936 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3941 for (i = 0; i < nr_tables; i++) {
3942 struct fixed_file_table *table = &ctx->file_table[i];
3943 unsigned this_files;
3945 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3946 table->files = kcalloc(this_files, sizeof(struct file *),
3950 nr_files -= this_files;
3956 for (i = 0; i < nr_tables; i++) {
3957 struct fixed_file_table *table = &ctx->file_table[i];
3958 kfree(table->files);
3963 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3966 __s32 __user *fds = (__s32 __user *) arg;
3971 if (ctx->file_table)
3975 if (nr_args > IORING_MAX_FIXED_FILES)
3978 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
3979 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
3981 if (!ctx->file_table)
3984 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
3985 kfree(ctx->file_table);
3986 ctx->file_table = NULL;
3990 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
3991 struct fixed_file_table *table;
3995 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
3997 /* allow sparse sets */
4003 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4004 index = i & IORING_FILE_TABLE_MASK;
4005 table->files[index] = fget(fd);
4008 if (!table->files[index])
4011 * Don't allow io_uring instances to be registered. If UNIX
4012 * isn't enabled, then this causes a reference cycle and this
4013 * instance can never get freed. If UNIX is enabled we'll
4014 * handle it just fine, but there's still no point in allowing
4015 * a ring fd as it doesn't support regular read/write anyway.
4017 if (table->files[index]->f_op == &io_uring_fops) {
4018 fput(table->files[index]);
4025 for (i = 0; i < ctx->nr_user_files; i++) {
4028 file = io_file_from_index(ctx, i);
4032 for (i = 0; i < nr_tables; i++)
4033 kfree(ctx->file_table[i].files);
4035 kfree(ctx->file_table);
4036 ctx->file_table = NULL;
4037 ctx->nr_user_files = 0;
4041 ret = io_sqe_files_scm(ctx);
4043 io_sqe_files_unregister(ctx);
4048 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4050 #if defined(CONFIG_UNIX)
4051 struct file *file = io_file_from_index(ctx, index);
4052 struct sock *sock = ctx->ring_sock->sk;
4053 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4054 struct sk_buff *skb;
4057 __skb_queue_head_init(&list);
4060 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4061 * remove this entry and rearrange the file array.
4063 skb = skb_dequeue(head);
4065 struct scm_fp_list *fp;
4067 fp = UNIXCB(skb).fp;
4068 for (i = 0; i < fp->count; i++) {
4071 if (fp->fp[i] != file)
4074 unix_notinflight(fp->user, fp->fp[i]);
4075 left = fp->count - 1 - i;
4077 memmove(&fp->fp[i], &fp->fp[i + 1],
4078 left * sizeof(struct file *));
4085 __skb_queue_tail(&list, skb);
4095 __skb_queue_tail(&list, skb);
4097 skb = skb_dequeue(head);
4100 if (skb_peek(&list)) {
4101 spin_lock_irq(&head->lock);
4102 while ((skb = __skb_dequeue(&list)) != NULL)
4103 __skb_queue_tail(head, skb);
4104 spin_unlock_irq(&head->lock);
4107 fput(io_file_from_index(ctx, index));
4111 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4114 #if defined(CONFIG_UNIX)
4115 struct sock *sock = ctx->ring_sock->sk;
4116 struct sk_buff_head *head = &sock->sk_receive_queue;
4117 struct sk_buff *skb;
4120 * See if we can merge this file into an existing skb SCM_RIGHTS
4121 * file set. If there's no room, fall back to allocating a new skb
4122 * and filling it in.
4124 spin_lock_irq(&head->lock);
4125 skb = skb_peek(head);
4127 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4129 if (fpl->count < SCM_MAX_FD) {
4130 __skb_unlink(skb, head);
4131 spin_unlock_irq(&head->lock);
4132 fpl->fp[fpl->count] = get_file(file);
4133 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4135 spin_lock_irq(&head->lock);
4136 __skb_queue_head(head, skb);
4141 spin_unlock_irq(&head->lock);
4148 return __io_sqe_files_scm(ctx, 1, index);
4154 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4157 struct io_uring_files_update up;
4162 if (!ctx->file_table)
4166 if (copy_from_user(&up, arg, sizeof(up)))
4168 if (check_add_overflow(up.offset, nr_args, &done))
4170 if (done > ctx->nr_user_files)
4174 fds = (__s32 __user *) up.fds;
4176 struct fixed_file_table *table;
4180 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4184 i = array_index_nospec(up.offset, ctx->nr_user_files);
4185 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4186 index = i & IORING_FILE_TABLE_MASK;
4187 if (table->files[index]) {
4188 io_sqe_file_unregister(ctx, i);
4189 table->files[index] = NULL;
4200 * Don't allow io_uring instances to be registered. If
4201 * UNIX isn't enabled, then this causes a reference
4202 * cycle and this instance can never get freed. If UNIX
4203 * is enabled we'll handle it just fine, but there's
4204 * still no point in allowing a ring fd as it doesn't
4205 * support regular read/write anyway.
4207 if (file->f_op == &io_uring_fops) {
4212 table->files[index] = file;
4213 err = io_sqe_file_register(ctx, file, i);
4222 return done ? done : err;
4225 static void io_put_work(struct io_wq_work *work)
4227 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4232 static void io_get_work(struct io_wq_work *work)
4234 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4236 refcount_inc(&req->refs);
4239 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4240 struct io_uring_params *p)
4242 struct io_wq_data data;
4243 unsigned concurrency;
4246 init_waitqueue_head(&ctx->sqo_wait);
4247 mmgrab(current->mm);
4248 ctx->sqo_mm = current->mm;
4250 if (ctx->flags & IORING_SETUP_SQPOLL) {
4252 if (!capable(CAP_SYS_ADMIN))
4255 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4256 if (!ctx->sq_thread_idle)
4257 ctx->sq_thread_idle = HZ;
4259 if (p->flags & IORING_SETUP_SQ_AFF) {
4260 int cpu = p->sq_thread_cpu;
4263 if (cpu >= nr_cpu_ids)
4265 if (!cpu_online(cpu))
4268 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4272 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4275 if (IS_ERR(ctx->sqo_thread)) {
4276 ret = PTR_ERR(ctx->sqo_thread);
4277 ctx->sqo_thread = NULL;
4280 wake_up_process(ctx->sqo_thread);
4281 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4282 /* Can't have SQ_AFF without SQPOLL */
4287 data.mm = ctx->sqo_mm;
4288 data.user = ctx->user;
4289 data.creds = ctx->creds;
4290 data.get_work = io_get_work;
4291 data.put_work = io_put_work;
4293 /* Do QD, or 4 * CPUS, whatever is smallest */
4294 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4295 ctx->io_wq = io_wq_create(concurrency, &data);
4296 if (IS_ERR(ctx->io_wq)) {
4297 ret = PTR_ERR(ctx->io_wq);
4304 io_finish_async(ctx);
4305 mmdrop(ctx->sqo_mm);
4310 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4312 atomic_long_sub(nr_pages, &user->locked_vm);
4315 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4317 unsigned long page_limit, cur_pages, new_pages;
4319 /* Don't allow more pages than we can safely lock */
4320 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4323 cur_pages = atomic_long_read(&user->locked_vm);
4324 new_pages = cur_pages + nr_pages;
4325 if (new_pages > page_limit)
4327 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4328 new_pages) != cur_pages);
4333 static void io_mem_free(void *ptr)
4340 page = virt_to_head_page(ptr);
4341 if (put_page_testzero(page))
4342 free_compound_page(page);
4345 static void *io_mem_alloc(size_t size)
4347 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4350 return (void *) __get_free_pages(gfp_flags, get_order(size));
4353 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4356 struct io_rings *rings;
4357 size_t off, sq_array_size;
4359 off = struct_size(rings, cqes, cq_entries);
4360 if (off == SIZE_MAX)
4364 off = ALIGN(off, SMP_CACHE_BYTES);
4369 sq_array_size = array_size(sizeof(u32), sq_entries);
4370 if (sq_array_size == SIZE_MAX)
4373 if (check_add_overflow(off, sq_array_size, &off))
4382 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4386 pages = (size_t)1 << get_order(
4387 rings_size(sq_entries, cq_entries, NULL));
4388 pages += (size_t)1 << get_order(
4389 array_size(sizeof(struct io_uring_sqe), sq_entries));
4394 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4398 if (!ctx->user_bufs)
4401 for (i = 0; i < ctx->nr_user_bufs; i++) {
4402 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4404 for (j = 0; j < imu->nr_bvecs; j++)
4405 put_user_page(imu->bvec[j].bv_page);
4407 if (ctx->account_mem)
4408 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4413 kfree(ctx->user_bufs);
4414 ctx->user_bufs = NULL;
4415 ctx->nr_user_bufs = 0;
4419 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4420 void __user *arg, unsigned index)
4422 struct iovec __user *src;
4424 #ifdef CONFIG_COMPAT
4426 struct compat_iovec __user *ciovs;
4427 struct compat_iovec ciov;
4429 ciovs = (struct compat_iovec __user *) arg;
4430 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4433 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4434 dst->iov_len = ciov.iov_len;
4438 src = (struct iovec __user *) arg;
4439 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4444 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4447 struct vm_area_struct **vmas = NULL;
4448 struct page **pages = NULL;
4449 int i, j, got_pages = 0;
4454 if (!nr_args || nr_args > UIO_MAXIOV)
4457 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4459 if (!ctx->user_bufs)
4462 for (i = 0; i < nr_args; i++) {
4463 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4464 unsigned long off, start, end, ubuf;
4469 ret = io_copy_iov(ctx, &iov, arg, i);
4474 * Don't impose further limits on the size and buffer
4475 * constraints here, we'll -EINVAL later when IO is
4476 * submitted if they are wrong.
4479 if (!iov.iov_base || !iov.iov_len)
4482 /* arbitrary limit, but we need something */
4483 if (iov.iov_len > SZ_1G)
4486 ubuf = (unsigned long) iov.iov_base;
4487 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4488 start = ubuf >> PAGE_SHIFT;
4489 nr_pages = end - start;
4491 if (ctx->account_mem) {
4492 ret = io_account_mem(ctx->user, nr_pages);
4498 if (!pages || nr_pages > got_pages) {
4501 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4503 vmas = kvmalloc_array(nr_pages,
4504 sizeof(struct vm_area_struct *),
4506 if (!pages || !vmas) {
4508 if (ctx->account_mem)
4509 io_unaccount_mem(ctx->user, nr_pages);
4512 got_pages = nr_pages;
4515 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4519 if (ctx->account_mem)
4520 io_unaccount_mem(ctx->user, nr_pages);
4525 down_read(¤t->mm->mmap_sem);
4526 pret = get_user_pages(ubuf, nr_pages,
4527 FOLL_WRITE | FOLL_LONGTERM,
4529 if (pret == nr_pages) {
4530 /* don't support file backed memory */
4531 for (j = 0; j < nr_pages; j++) {
4532 struct vm_area_struct *vma = vmas[j];
4535 !is_file_hugepages(vma->vm_file)) {
4541 ret = pret < 0 ? pret : -EFAULT;
4543 up_read(¤t->mm->mmap_sem);
4546 * if we did partial map, or found file backed vmas,
4547 * release any pages we did get
4550 put_user_pages(pages, pret);
4551 if (ctx->account_mem)
4552 io_unaccount_mem(ctx->user, nr_pages);
4557 off = ubuf & ~PAGE_MASK;
4559 for (j = 0; j < nr_pages; j++) {
4562 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4563 imu->bvec[j].bv_page = pages[j];
4564 imu->bvec[j].bv_len = vec_len;
4565 imu->bvec[j].bv_offset = off;
4569 /* store original address for later verification */
4571 imu->len = iov.iov_len;
4572 imu->nr_bvecs = nr_pages;
4574 ctx->nr_user_bufs++;
4582 io_sqe_buffer_unregister(ctx);
4586 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4588 __s32 __user *fds = arg;
4594 if (copy_from_user(&fd, fds, sizeof(*fds)))
4597 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4598 if (IS_ERR(ctx->cq_ev_fd)) {
4599 int ret = PTR_ERR(ctx->cq_ev_fd);
4600 ctx->cq_ev_fd = NULL;
4607 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4609 if (ctx->cq_ev_fd) {
4610 eventfd_ctx_put(ctx->cq_ev_fd);
4611 ctx->cq_ev_fd = NULL;
4618 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4620 io_finish_async(ctx);
4622 mmdrop(ctx->sqo_mm);
4624 io_iopoll_reap_events(ctx);
4625 io_sqe_buffer_unregister(ctx);
4626 io_sqe_files_unregister(ctx);
4627 io_eventfd_unregister(ctx);
4629 #if defined(CONFIG_UNIX)
4630 if (ctx->ring_sock) {
4631 ctx->ring_sock->file = NULL; /* so that iput() is called */
4632 sock_release(ctx->ring_sock);
4636 io_mem_free(ctx->rings);
4637 io_mem_free(ctx->sq_sqes);
4639 percpu_ref_exit(&ctx->refs);
4640 if (ctx->account_mem)
4641 io_unaccount_mem(ctx->user,
4642 ring_pages(ctx->sq_entries, ctx->cq_entries));
4643 free_uid(ctx->user);
4644 put_cred(ctx->creds);
4645 kfree(ctx->completions);
4646 kmem_cache_free(req_cachep, ctx->fallback_req);
4650 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4652 struct io_ring_ctx *ctx = file->private_data;
4655 poll_wait(file, &ctx->cq_wait, wait);
4657 * synchronizes with barrier from wq_has_sleeper call in
4661 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4662 ctx->rings->sq_ring_entries)
4663 mask |= EPOLLOUT | EPOLLWRNORM;
4664 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4665 mask |= EPOLLIN | EPOLLRDNORM;
4670 static int io_uring_fasync(int fd, struct file *file, int on)
4672 struct io_ring_ctx *ctx = file->private_data;
4674 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4677 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4679 mutex_lock(&ctx->uring_lock);
4680 percpu_ref_kill(&ctx->refs);
4681 mutex_unlock(&ctx->uring_lock);
4683 io_kill_timeouts(ctx);
4684 io_poll_remove_all(ctx);
4687 io_wq_cancel_all(ctx->io_wq);
4689 io_iopoll_reap_events(ctx);
4690 /* if we failed setting up the ctx, we might not have any rings */
4692 io_cqring_overflow_flush(ctx, true);
4693 wait_for_completion(&ctx->completions[0]);
4694 io_ring_ctx_free(ctx);
4697 static int io_uring_release(struct inode *inode, struct file *file)
4699 struct io_ring_ctx *ctx = file->private_data;
4701 file->private_data = NULL;
4702 io_ring_ctx_wait_and_kill(ctx);
4706 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4707 struct files_struct *files)
4709 struct io_kiocb *req;
4712 while (!list_empty_careful(&ctx->inflight_list)) {
4713 struct io_kiocb *cancel_req = NULL;
4715 spin_lock_irq(&ctx->inflight_lock);
4716 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4717 if (req->work.files != files)
4719 /* req is being completed, ignore */
4720 if (!refcount_inc_not_zero(&req->refs))
4726 prepare_to_wait(&ctx->inflight_wait, &wait,
4727 TASK_UNINTERRUPTIBLE);
4728 spin_unlock_irq(&ctx->inflight_lock);
4730 /* We need to keep going until we don't find a matching req */
4734 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4735 io_put_req(cancel_req);
4738 finish_wait(&ctx->inflight_wait, &wait);
4741 static int io_uring_flush(struct file *file, void *data)
4743 struct io_ring_ctx *ctx = file->private_data;
4745 io_uring_cancel_files(ctx, data);
4746 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4747 io_cqring_overflow_flush(ctx, true);
4748 io_wq_cancel_all(ctx->io_wq);
4753 static void *io_uring_validate_mmap_request(struct file *file,
4754 loff_t pgoff, size_t sz)
4756 struct io_ring_ctx *ctx = file->private_data;
4757 loff_t offset = pgoff << PAGE_SHIFT;
4762 case IORING_OFF_SQ_RING:
4763 case IORING_OFF_CQ_RING:
4766 case IORING_OFF_SQES:
4770 return ERR_PTR(-EINVAL);
4773 page = virt_to_head_page(ptr);
4774 if (sz > page_size(page))
4775 return ERR_PTR(-EINVAL);
4782 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4784 size_t sz = vma->vm_end - vma->vm_start;
4788 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4790 return PTR_ERR(ptr);
4792 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4793 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4796 #else /* !CONFIG_MMU */
4798 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4800 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4803 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4805 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4808 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4809 unsigned long addr, unsigned long len,
4810 unsigned long pgoff, unsigned long flags)
4814 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4816 return PTR_ERR(ptr);
4818 return (unsigned long) ptr;
4821 #endif /* !CONFIG_MMU */
4823 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
4824 u32, min_complete, u32, flags, const sigset_t __user *, sig,
4827 struct io_ring_ctx *ctx;
4832 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
4840 if (f.file->f_op != &io_uring_fops)
4844 ctx = f.file->private_data;
4845 if (!percpu_ref_tryget(&ctx->refs))
4849 * For SQ polling, the thread will do all submissions and completions.
4850 * Just return the requested submit count, and wake the thread if
4854 if (ctx->flags & IORING_SETUP_SQPOLL) {
4855 if (!list_empty_careful(&ctx->cq_overflow_list))
4856 io_cqring_overflow_flush(ctx, false);
4857 if (flags & IORING_ENTER_SQ_WAKEUP)
4858 wake_up(&ctx->sqo_wait);
4859 submitted = to_submit;
4860 } else if (to_submit) {
4861 struct mm_struct *cur_mm;
4863 to_submit = min(to_submit, ctx->sq_entries);
4864 mutex_lock(&ctx->uring_lock);
4865 /* already have mm, so io_submit_sqes() won't try to grab it */
4866 cur_mm = ctx->sqo_mm;
4867 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4869 mutex_unlock(&ctx->uring_lock);
4871 if (flags & IORING_ENTER_GETEVENTS) {
4872 unsigned nr_events = 0;
4874 min_complete = min(min_complete, ctx->cq_entries);
4876 if (ctx->flags & IORING_SETUP_IOPOLL) {
4877 ret = io_iopoll_check(ctx, &nr_events, min_complete);
4879 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4883 percpu_ref_put(&ctx->refs);
4886 return submitted ? submitted : ret;
4889 static const struct file_operations io_uring_fops = {
4890 .release = io_uring_release,
4891 .flush = io_uring_flush,
4892 .mmap = io_uring_mmap,
4894 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
4895 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
4897 .poll = io_uring_poll,
4898 .fasync = io_uring_fasync,
4901 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4902 struct io_uring_params *p)
4904 struct io_rings *rings;
4905 size_t size, sq_array_offset;
4907 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4908 if (size == SIZE_MAX)
4911 rings = io_mem_alloc(size);
4916 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4917 rings->sq_ring_mask = p->sq_entries - 1;
4918 rings->cq_ring_mask = p->cq_entries - 1;
4919 rings->sq_ring_entries = p->sq_entries;
4920 rings->cq_ring_entries = p->cq_entries;
4921 ctx->sq_mask = rings->sq_ring_mask;
4922 ctx->cq_mask = rings->cq_ring_mask;
4923 ctx->sq_entries = rings->sq_ring_entries;
4924 ctx->cq_entries = rings->cq_ring_entries;
4926 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
4927 if (size == SIZE_MAX) {
4928 io_mem_free(ctx->rings);
4933 ctx->sq_sqes = io_mem_alloc(size);
4934 if (!ctx->sq_sqes) {
4935 io_mem_free(ctx->rings);
4944 * Allocate an anonymous fd, this is what constitutes the application
4945 * visible backing of an io_uring instance. The application mmaps this
4946 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4947 * we have to tie this fd to a socket for file garbage collection purposes.
4949 static int io_uring_get_fd(struct io_ring_ctx *ctx)
4954 #if defined(CONFIG_UNIX)
4955 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4961 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4965 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4966 O_RDWR | O_CLOEXEC);
4969 ret = PTR_ERR(file);
4973 #if defined(CONFIG_UNIX)
4974 ctx->ring_sock->file = file;
4975 ctx->ring_sock->sk->sk_user_data = ctx;
4977 fd_install(ret, file);
4980 #if defined(CONFIG_UNIX)
4981 sock_release(ctx->ring_sock);
4982 ctx->ring_sock = NULL;
4987 static int io_uring_create(unsigned entries, struct io_uring_params *p)
4989 struct user_struct *user = NULL;
4990 struct io_ring_ctx *ctx;
4994 if (!entries || entries > IORING_MAX_ENTRIES)
4998 * Use twice as many entries for the CQ ring. It's possible for the
4999 * application to drive a higher depth than the size of the SQ ring,
5000 * since the sqes are only used at submission time. This allows for
5001 * some flexibility in overcommitting a bit. If the application has
5002 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5003 * of CQ ring entries manually.
5005 p->sq_entries = roundup_pow_of_two(entries);
5006 if (p->flags & IORING_SETUP_CQSIZE) {
5008 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5009 * to a power-of-two, if it isn't already. We do NOT impose
5010 * any cq vs sq ring sizing.
5012 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5014 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5016 p->cq_entries = 2 * p->sq_entries;
5019 user = get_uid(current_user());
5020 account_mem = !capable(CAP_IPC_LOCK);
5023 ret = io_account_mem(user,
5024 ring_pages(p->sq_entries, p->cq_entries));
5031 ctx = io_ring_ctx_alloc(p);
5034 io_unaccount_mem(user, ring_pages(p->sq_entries,
5039 ctx->compat = in_compat_syscall();
5040 ctx->account_mem = account_mem;
5042 ctx->creds = get_current_cred();
5044 ret = io_allocate_scq_urings(ctx, p);
5048 ret = io_sq_offload_start(ctx, p);
5052 memset(&p->sq_off, 0, sizeof(p->sq_off));
5053 p->sq_off.head = offsetof(struct io_rings, sq.head);
5054 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5055 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5056 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5057 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5058 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5059 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5061 memset(&p->cq_off, 0, sizeof(p->cq_off));
5062 p->cq_off.head = offsetof(struct io_rings, cq.head);
5063 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5064 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5065 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5066 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5067 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5070 * Install ring fd as the very last thing, so we don't risk someone
5071 * having closed it before we finish setup
5073 ret = io_uring_get_fd(ctx);
5077 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5078 IORING_FEAT_SUBMIT_STABLE;
5079 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5082 io_ring_ctx_wait_and_kill(ctx);
5087 * Sets up an aio uring context, and returns the fd. Applications asks for a
5088 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5089 * params structure passed in.
5091 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5093 struct io_uring_params p;
5097 if (copy_from_user(&p, params, sizeof(p)))
5099 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5104 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5105 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5108 ret = io_uring_create(entries, &p);
5112 if (copy_to_user(params, &p, sizeof(p)))
5118 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5119 struct io_uring_params __user *, params)
5121 return io_uring_setup(entries, params);
5124 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5125 void __user *arg, unsigned nr_args)
5126 __releases(ctx->uring_lock)
5127 __acquires(ctx->uring_lock)
5132 * We're inside the ring mutex, if the ref is already dying, then
5133 * someone else killed the ctx or is already going through
5134 * io_uring_register().
5136 if (percpu_ref_is_dying(&ctx->refs))
5139 percpu_ref_kill(&ctx->refs);
5142 * Drop uring mutex before waiting for references to exit. If another
5143 * thread is currently inside io_uring_enter() it might need to grab
5144 * the uring_lock to make progress. If we hold it here across the drain
5145 * wait, then we can deadlock. It's safe to drop the mutex here, since
5146 * no new references will come in after we've killed the percpu ref.
5148 mutex_unlock(&ctx->uring_lock);
5149 wait_for_completion(&ctx->completions[0]);
5150 mutex_lock(&ctx->uring_lock);
5153 case IORING_REGISTER_BUFFERS:
5154 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5156 case IORING_UNREGISTER_BUFFERS:
5160 ret = io_sqe_buffer_unregister(ctx);
5162 case IORING_REGISTER_FILES:
5163 ret = io_sqe_files_register(ctx, arg, nr_args);
5165 case IORING_UNREGISTER_FILES:
5169 ret = io_sqe_files_unregister(ctx);
5171 case IORING_REGISTER_FILES_UPDATE:
5172 ret = io_sqe_files_update(ctx, arg, nr_args);
5174 case IORING_REGISTER_EVENTFD:
5178 ret = io_eventfd_register(ctx, arg);
5180 case IORING_UNREGISTER_EVENTFD:
5184 ret = io_eventfd_unregister(ctx);
5191 /* bring the ctx back to life */
5192 reinit_completion(&ctx->completions[0]);
5193 percpu_ref_reinit(&ctx->refs);
5197 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5198 void __user *, arg, unsigned int, nr_args)
5200 struct io_ring_ctx *ctx;
5209 if (f.file->f_op != &io_uring_fops)
5212 ctx = f.file->private_data;
5214 mutex_lock(&ctx->uring_lock);
5215 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5216 mutex_unlock(&ctx->uring_lock);
5217 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5218 ctx->cq_ev_fd != NULL, ret);
5224 static int __init io_uring_init(void)
5226 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5229 __initcall(io_uring_init);