1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <linux/refcount.h>
48 #include <linux/uio.h>
50 #include <linux/sched/signal.h>
52 #include <linux/file.h>
53 #include <linux/fdtable.h>
55 #include <linux/mman.h>
56 #include <linux/mmu_context.h>
57 #include <linux/percpu.h>
58 #include <linux/slab.h>
59 #include <linux/kthread.h>
60 #include <linux/blkdev.h>
61 #include <linux/bvec.h>
62 #include <linux/net.h>
64 #include <net/af_unix.h>
66 #include <linux/anon_inodes.h>
67 #include <linux/sched/mm.h>
68 #include <linux/uaccess.h>
69 #include <linux/nospec.h>
70 #include <linux/sizes.h>
71 #include <linux/hugetlb.h>
72 #include <linux/highmem.h>
74 #define CREATE_TRACE_POINTS
75 #include <trace/events/io_uring.h>
77 #include <uapi/linux/io_uring.h>
82 #define IORING_MAX_ENTRIES 32768
83 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
86 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
88 #define IORING_FILE_TABLE_SHIFT 9
89 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
90 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
91 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
94 u32 head ____cacheline_aligned_in_smp;
95 u32 tail ____cacheline_aligned_in_smp;
99 * This data is shared with the application through the mmap at offsets
100 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
102 * The offsets to the member fields are published through struct
103 * io_sqring_offsets when calling io_uring_setup.
107 * Head and tail offsets into the ring; the offsets need to be
108 * masked to get valid indices.
110 * The kernel controls head of the sq ring and the tail of the cq ring,
111 * and the application controls tail of the sq ring and the head of the
114 struct io_uring sq, cq;
116 * Bitmasks to apply to head and tail offsets (constant, equals
119 u32 sq_ring_mask, cq_ring_mask;
120 /* Ring sizes (constant, power of 2) */
121 u32 sq_ring_entries, cq_ring_entries;
123 * Number of invalid entries dropped by the kernel due to
124 * invalid index stored in array
126 * Written by the kernel, shouldn't be modified by the
127 * application (i.e. get number of "new events" by comparing to
130 * After a new SQ head value was read by the application this
131 * counter includes all submissions that were dropped reaching
132 * the new SQ head (and possibly more).
138 * Written by the kernel, shouldn't be modified by the
141 * The application needs a full memory barrier before checking
142 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
146 * Number of completion events lost because the queue was full;
147 * this should be avoided by the application by making sure
148 * there are not more requests pending than there is space in
149 * the completion queue.
151 * Written by the kernel, shouldn't be modified by the
152 * application (i.e. get number of "new events" by comparing to
155 * As completion events come in out of order this counter is not
156 * ordered with any other data.
160 * Ring buffer of completion events.
162 * The kernel writes completion events fresh every time they are
163 * produced, so the application is allowed to modify pending
166 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
169 struct io_mapped_ubuf {
172 struct bio_vec *bvec;
173 unsigned int nr_bvecs;
176 struct fixed_file_table {
182 struct percpu_ref refs;
183 } ____cacheline_aligned_in_smp;
189 bool cq_overflow_flushed;
193 * Ring buffer of indices into array of io_uring_sqe, which is
194 * mmapped by the application using the IORING_OFF_SQES offset.
196 * This indirection could e.g. be used to assign fixed
197 * io_uring_sqe entries to operations and only submit them to
198 * the queue when needed.
200 * The kernel modifies neither the indices array nor the entries
204 unsigned cached_sq_head;
207 unsigned sq_thread_idle;
208 unsigned cached_sq_dropped;
209 atomic_t cached_cq_overflow;
210 struct io_uring_sqe *sq_sqes;
212 struct list_head defer_list;
213 struct list_head timeout_list;
214 struct list_head cq_overflow_list;
216 wait_queue_head_t inflight_wait;
217 } ____cacheline_aligned_in_smp;
219 struct io_rings *rings;
223 struct task_struct *sqo_thread; /* if using sq thread polling */
224 struct mm_struct *sqo_mm;
225 wait_queue_head_t sqo_wait;
228 * If used, fixed file set. Writers must ensure that ->refs is dead,
229 * readers must ensure that ->refs is alive as long as the file* is
230 * used. Only updated through io_uring_register(2).
232 struct fixed_file_table *file_table;
233 unsigned nr_user_files;
235 /* if used, fixed mapped user buffers */
236 unsigned nr_user_bufs;
237 struct io_mapped_ubuf *user_bufs;
239 struct user_struct *user;
241 const struct cred *creds;
243 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
244 struct completion *completions;
246 /* if all else fails... */
247 struct io_kiocb *fallback_req;
249 #if defined(CONFIG_UNIX)
250 struct socket *ring_sock;
254 unsigned cached_cq_tail;
257 atomic_t cq_timeouts;
258 struct wait_queue_head cq_wait;
259 struct fasync_struct *cq_fasync;
260 struct eventfd_ctx *cq_ev_fd;
261 } ____cacheline_aligned_in_smp;
264 struct mutex uring_lock;
265 wait_queue_head_t wait;
266 } ____cacheline_aligned_in_smp;
269 spinlock_t completion_lock;
270 bool poll_multi_file;
272 * ->poll_list is protected by the ctx->uring_lock for
273 * io_uring instances that don't use IORING_SETUP_SQPOLL.
274 * For SQPOLL, only the single threaded io_sq_thread() will
275 * manipulate the list, hence no extra locking is needed there.
277 struct list_head poll_list;
278 struct hlist_head *cancel_hash;
279 unsigned cancel_hash_bits;
281 spinlock_t inflight_lock;
282 struct list_head inflight_list;
283 } ____cacheline_aligned_in_smp;
287 * First field must be the file pointer in all the
288 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
290 struct io_poll_iocb {
292 struct wait_queue_head *head;
296 struct wait_queue_entry wait;
299 struct io_timeout_data {
300 struct io_kiocb *req;
301 struct hrtimer timer;
302 struct timespec64 ts;
303 enum hrtimer_mode mode;
307 struct io_async_connect {
308 struct sockaddr_storage address;
311 struct io_async_msghdr {
312 struct iovec fast_iov[UIO_FASTIOV];
314 struct sockaddr __user *uaddr;
319 struct iovec fast_iov[UIO_FASTIOV];
325 struct io_async_ctx {
326 struct io_uring_sqe sqe;
328 struct io_async_rw rw;
329 struct io_async_msghdr msg;
330 struct io_async_connect connect;
331 struct io_timeout_data timeout;
336 * NOTE! Each of the iocb union members has the file pointer
337 * as the first entry in their struct definition. So you can
338 * access the file pointer through any of the sub-structs,
339 * or directly as just 'ki_filp' in this struct.
345 struct io_poll_iocb poll;
348 const struct io_uring_sqe *sqe;
349 struct io_async_ctx *io;
350 struct file *ring_file;
354 bool needs_fixed_file;
356 struct io_ring_ctx *ctx;
358 struct list_head list;
359 struct hlist_node hash_node;
361 struct list_head link_list;
364 #define REQ_F_NOWAIT 1 /* must not punt to workers */
365 #define REQ_F_IOPOLL_COMPLETED 2 /* polled IO has completed */
366 #define REQ_F_FIXED_FILE 4 /* ctx owns file */
367 #define REQ_F_LINK_NEXT 8 /* already grabbed next link */
368 #define REQ_F_IO_DRAIN 16 /* drain existing IO first */
369 #define REQ_F_IO_DRAINED 32 /* drain done */
370 #define REQ_F_LINK 64 /* linked sqes */
371 #define REQ_F_LINK_TIMEOUT 128 /* has linked timeout */
372 #define REQ_F_FAIL_LINK 256 /* fail rest of links */
373 #define REQ_F_DRAIN_LINK 512 /* link should be fully drained */
374 #define REQ_F_TIMEOUT 1024 /* timeout request */
375 #define REQ_F_ISREG 2048 /* regular file */
376 #define REQ_F_MUST_PUNT 4096 /* must be punted even for NONBLOCK */
377 #define REQ_F_TIMEOUT_NOSEQ 8192 /* no timeout sequence */
378 #define REQ_F_INFLIGHT 16384 /* on inflight list */
379 #define REQ_F_COMP_LOCKED 32768 /* completion under lock */
380 #define REQ_F_HARDLINK 65536 /* doesn't sever on completion < 0 */
385 struct list_head inflight_entry;
387 struct io_wq_work work;
390 #define IO_PLUG_THRESHOLD 2
391 #define IO_IOPOLL_BATCH 8
393 struct io_submit_state {
394 struct blk_plug plug;
397 * io_kiocb alloc cache
399 void *reqs[IO_IOPOLL_BATCH];
400 unsigned int free_reqs;
401 unsigned int cur_req;
404 * File reference cache
408 unsigned int has_refs;
409 unsigned int used_refs;
410 unsigned int ios_left;
413 static void io_wq_submit_work(struct io_wq_work **workptr);
414 static void io_cqring_fill_event(struct io_kiocb *req, long res);
415 static void __io_free_req(struct io_kiocb *req);
416 static void io_put_req(struct io_kiocb *req);
417 static void io_double_put_req(struct io_kiocb *req);
418 static void __io_double_put_req(struct io_kiocb *req);
419 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
420 static void io_queue_linked_timeout(struct io_kiocb *req);
422 static struct kmem_cache *req_cachep;
424 static const struct file_operations io_uring_fops;
426 struct sock *io_uring_get_socket(struct file *file)
428 #if defined(CONFIG_UNIX)
429 if (file->f_op == &io_uring_fops) {
430 struct io_ring_ctx *ctx = file->private_data;
432 return ctx->ring_sock->sk;
437 EXPORT_SYMBOL(io_uring_get_socket);
439 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
441 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
443 complete(&ctx->completions[0]);
446 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
448 struct io_ring_ctx *ctx;
451 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
455 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
456 if (!ctx->fallback_req)
459 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
460 if (!ctx->completions)
464 * Use 5 bits less than the max cq entries, that should give us around
465 * 32 entries per hash list if totally full and uniformly spread.
467 hash_bits = ilog2(p->cq_entries);
471 ctx->cancel_hash_bits = hash_bits;
472 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
474 if (!ctx->cancel_hash)
476 __hash_init(ctx->cancel_hash, 1U << hash_bits);
478 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
479 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
482 ctx->flags = p->flags;
483 init_waitqueue_head(&ctx->cq_wait);
484 INIT_LIST_HEAD(&ctx->cq_overflow_list);
485 init_completion(&ctx->completions[0]);
486 init_completion(&ctx->completions[1]);
487 mutex_init(&ctx->uring_lock);
488 init_waitqueue_head(&ctx->wait);
489 spin_lock_init(&ctx->completion_lock);
490 INIT_LIST_HEAD(&ctx->poll_list);
491 INIT_LIST_HEAD(&ctx->defer_list);
492 INIT_LIST_HEAD(&ctx->timeout_list);
493 init_waitqueue_head(&ctx->inflight_wait);
494 spin_lock_init(&ctx->inflight_lock);
495 INIT_LIST_HEAD(&ctx->inflight_list);
498 if (ctx->fallback_req)
499 kmem_cache_free(req_cachep, ctx->fallback_req);
500 kfree(ctx->completions);
501 kfree(ctx->cancel_hash);
506 static inline bool __req_need_defer(struct io_kiocb *req)
508 struct io_ring_ctx *ctx = req->ctx;
510 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
511 + atomic_read(&ctx->cached_cq_overflow);
514 static inline bool req_need_defer(struct io_kiocb *req)
516 if ((req->flags & (REQ_F_IO_DRAIN|REQ_F_IO_DRAINED)) == REQ_F_IO_DRAIN)
517 return __req_need_defer(req);
522 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
524 struct io_kiocb *req;
526 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
527 if (req && !req_need_defer(req)) {
528 list_del_init(&req->list);
535 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
537 struct io_kiocb *req;
539 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
541 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
543 if (!__req_need_defer(req)) {
544 list_del_init(&req->list);
552 static void __io_commit_cqring(struct io_ring_ctx *ctx)
554 struct io_rings *rings = ctx->rings;
556 if (ctx->cached_cq_tail != READ_ONCE(rings->cq.tail)) {
557 /* order cqe stores with ring update */
558 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
560 if (wq_has_sleeper(&ctx->cq_wait)) {
561 wake_up_interruptible(&ctx->cq_wait);
562 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
567 static inline bool io_sqe_needs_user(const struct io_uring_sqe *sqe)
569 u8 opcode = READ_ONCE(sqe->opcode);
571 return !(opcode == IORING_OP_READ_FIXED ||
572 opcode == IORING_OP_WRITE_FIXED);
575 static inline bool io_prep_async_work(struct io_kiocb *req,
576 struct io_kiocb **link)
578 bool do_hashed = false;
581 switch (req->sqe->opcode) {
582 case IORING_OP_WRITEV:
583 case IORING_OP_WRITE_FIXED:
584 /* only regular files should be hashed for writes */
585 if (req->flags & REQ_F_ISREG)
588 case IORING_OP_READV:
589 case IORING_OP_READ_FIXED:
590 case IORING_OP_SENDMSG:
591 case IORING_OP_RECVMSG:
592 case IORING_OP_ACCEPT:
593 case IORING_OP_POLL_ADD:
594 case IORING_OP_CONNECT:
596 * We know REQ_F_ISREG is not set on some of these
597 * opcodes, but this enables us to keep the check in
600 if (!(req->flags & REQ_F_ISREG))
601 req->work.flags |= IO_WQ_WORK_UNBOUND;
604 if (io_sqe_needs_user(req->sqe))
605 req->work.flags |= IO_WQ_WORK_NEEDS_USER;
608 *link = io_prep_linked_timeout(req);
612 static inline void io_queue_async_work(struct io_kiocb *req)
614 struct io_ring_ctx *ctx = req->ctx;
615 struct io_kiocb *link;
618 do_hashed = io_prep_async_work(req, &link);
620 trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work,
623 io_wq_enqueue(ctx->io_wq, &req->work);
625 io_wq_enqueue_hashed(ctx->io_wq, &req->work,
626 file_inode(req->file));
630 io_queue_linked_timeout(link);
633 static void io_kill_timeout(struct io_kiocb *req)
637 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
639 atomic_inc(&req->ctx->cq_timeouts);
640 list_del_init(&req->list);
641 io_cqring_fill_event(req, 0);
646 static void io_kill_timeouts(struct io_ring_ctx *ctx)
648 struct io_kiocb *req, *tmp;
650 spin_lock_irq(&ctx->completion_lock);
651 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
652 io_kill_timeout(req);
653 spin_unlock_irq(&ctx->completion_lock);
656 static void io_commit_cqring(struct io_ring_ctx *ctx)
658 struct io_kiocb *req;
660 while ((req = io_get_timeout_req(ctx)) != NULL)
661 io_kill_timeout(req);
663 __io_commit_cqring(ctx);
665 while ((req = io_get_deferred_req(ctx)) != NULL) {
666 req->flags |= REQ_F_IO_DRAINED;
667 io_queue_async_work(req);
671 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
673 struct io_rings *rings = ctx->rings;
676 tail = ctx->cached_cq_tail;
678 * writes to the cq entry need to come after reading head; the
679 * control dependency is enough as we're using WRITE_ONCE to
682 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
685 ctx->cached_cq_tail++;
686 return &rings->cqes[tail & ctx->cq_mask];
689 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
691 if (waitqueue_active(&ctx->wait))
693 if (waitqueue_active(&ctx->sqo_wait))
694 wake_up(&ctx->sqo_wait);
696 eventfd_signal(ctx->cq_ev_fd, 1);
699 /* Returns true if there are no backlogged entries after the flush */
700 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
702 struct io_rings *rings = ctx->rings;
703 struct io_uring_cqe *cqe;
704 struct io_kiocb *req;
709 if (list_empty_careful(&ctx->cq_overflow_list))
711 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
712 rings->cq_ring_entries))
716 spin_lock_irqsave(&ctx->completion_lock, flags);
718 /* if force is set, the ring is going away. always drop after that */
720 ctx->cq_overflow_flushed = true;
723 while (!list_empty(&ctx->cq_overflow_list)) {
724 cqe = io_get_cqring(ctx);
728 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
730 list_move(&req->list, &list);
732 WRITE_ONCE(cqe->user_data, req->user_data);
733 WRITE_ONCE(cqe->res, req->result);
734 WRITE_ONCE(cqe->flags, 0);
736 WRITE_ONCE(ctx->rings->cq_overflow,
737 atomic_inc_return(&ctx->cached_cq_overflow));
741 io_commit_cqring(ctx);
742 spin_unlock_irqrestore(&ctx->completion_lock, flags);
743 io_cqring_ev_posted(ctx);
745 while (!list_empty(&list)) {
746 req = list_first_entry(&list, struct io_kiocb, list);
747 list_del(&req->list);
754 static void io_cqring_fill_event(struct io_kiocb *req, long res)
756 struct io_ring_ctx *ctx = req->ctx;
757 struct io_uring_cqe *cqe;
759 trace_io_uring_complete(ctx, req->user_data, res);
762 * If we can't get a cq entry, userspace overflowed the
763 * submission (by quite a lot). Increment the overflow count in
766 cqe = io_get_cqring(ctx);
768 WRITE_ONCE(cqe->user_data, req->user_data);
769 WRITE_ONCE(cqe->res, res);
770 WRITE_ONCE(cqe->flags, 0);
771 } else if (ctx->cq_overflow_flushed) {
772 WRITE_ONCE(ctx->rings->cq_overflow,
773 atomic_inc_return(&ctx->cached_cq_overflow));
775 refcount_inc(&req->refs);
777 list_add_tail(&req->list, &ctx->cq_overflow_list);
781 static void io_cqring_add_event(struct io_kiocb *req, long res)
783 struct io_ring_ctx *ctx = req->ctx;
786 spin_lock_irqsave(&ctx->completion_lock, flags);
787 io_cqring_fill_event(req, res);
788 io_commit_cqring(ctx);
789 spin_unlock_irqrestore(&ctx->completion_lock, flags);
791 io_cqring_ev_posted(ctx);
794 static inline bool io_is_fallback_req(struct io_kiocb *req)
796 return req == (struct io_kiocb *)
797 ((unsigned long) req->ctx->fallback_req & ~1UL);
800 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
802 struct io_kiocb *req;
804 req = ctx->fallback_req;
805 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
811 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
812 struct io_submit_state *state)
814 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
815 struct io_kiocb *req;
817 if (!percpu_ref_tryget(&ctx->refs))
821 req = kmem_cache_alloc(req_cachep, gfp);
824 } else if (!state->free_reqs) {
828 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
829 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
832 * Bulk alloc is all-or-nothing. If we fail to get a batch,
833 * retry single alloc to be on the safe side.
835 if (unlikely(ret <= 0)) {
836 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
841 state->free_reqs = ret - 1;
843 req = state->reqs[0];
845 req = state->reqs[state->cur_req];
852 req->ring_file = NULL;
856 /* one is dropped after submission, the other at completion */
857 refcount_set(&req->refs, 2);
859 INIT_IO_WORK(&req->work, io_wq_submit_work);
862 req = io_get_fallback_req(ctx);
865 percpu_ref_put(&ctx->refs);
869 static void io_free_req_many(struct io_ring_ctx *ctx, void **reqs, int *nr)
872 kmem_cache_free_bulk(req_cachep, *nr, reqs);
873 percpu_ref_put_many(&ctx->refs, *nr);
878 static void __io_free_req(struct io_kiocb *req)
880 struct io_ring_ctx *ctx = req->ctx;
884 if (req->file && !(req->flags & REQ_F_FIXED_FILE))
886 if (req->flags & REQ_F_INFLIGHT) {
889 spin_lock_irqsave(&ctx->inflight_lock, flags);
890 list_del(&req->inflight_entry);
891 if (waitqueue_active(&ctx->inflight_wait))
892 wake_up(&ctx->inflight_wait);
893 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
895 percpu_ref_put(&ctx->refs);
896 if (likely(!io_is_fallback_req(req)))
897 kmem_cache_free(req_cachep, req);
899 clear_bit_unlock(0, (unsigned long *) ctx->fallback_req);
902 static bool io_link_cancel_timeout(struct io_kiocb *req)
904 struct io_ring_ctx *ctx = req->ctx;
907 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
909 io_cqring_fill_event(req, -ECANCELED);
910 io_commit_cqring(ctx);
911 req->flags &= ~REQ_F_LINK;
919 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
921 struct io_ring_ctx *ctx = req->ctx;
922 bool wake_ev = false;
924 /* Already got next link */
925 if (req->flags & REQ_F_LINK_NEXT)
929 * The list should never be empty when we are called here. But could
930 * potentially happen if the chain is messed up, check to be on the
933 while (!list_empty(&req->link_list)) {
934 struct io_kiocb *nxt = list_first_entry(&req->link_list,
935 struct io_kiocb, link_list);
937 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
938 (nxt->flags & REQ_F_TIMEOUT))) {
939 list_del_init(&nxt->link_list);
940 wake_ev |= io_link_cancel_timeout(nxt);
941 req->flags &= ~REQ_F_LINK_TIMEOUT;
945 list_del_init(&req->link_list);
946 if (!list_empty(&nxt->link_list))
947 nxt->flags |= REQ_F_LINK;
952 req->flags |= REQ_F_LINK_NEXT;
954 io_cqring_ev_posted(ctx);
958 * Called if REQ_F_LINK is set, and we fail the head request
960 static void io_fail_links(struct io_kiocb *req)
962 struct io_ring_ctx *ctx = req->ctx;
965 spin_lock_irqsave(&ctx->completion_lock, flags);
967 while (!list_empty(&req->link_list)) {
968 struct io_kiocb *link = list_first_entry(&req->link_list,
969 struct io_kiocb, link_list);
971 list_del_init(&link->link_list);
972 trace_io_uring_fail_link(req, link);
974 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
975 link->sqe->opcode == IORING_OP_LINK_TIMEOUT) {
976 io_link_cancel_timeout(link);
978 io_cqring_fill_event(link, -ECANCELED);
979 __io_double_put_req(link);
981 req->flags &= ~REQ_F_LINK_TIMEOUT;
984 io_commit_cqring(ctx);
985 spin_unlock_irqrestore(&ctx->completion_lock, flags);
986 io_cqring_ev_posted(ctx);
989 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
991 if (likely(!(req->flags & REQ_F_LINK)))
995 * If LINK is set, we have dependent requests in this chain. If we
996 * didn't fail this request, queue the first one up, moving any other
997 * dependencies to the next request. In case of failure, fail the rest
1000 if (req->flags & REQ_F_FAIL_LINK) {
1002 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1003 REQ_F_LINK_TIMEOUT) {
1004 struct io_ring_ctx *ctx = req->ctx;
1005 unsigned long flags;
1008 * If this is a timeout link, we could be racing with the
1009 * timeout timer. Grab the completion lock for this case to
1010 * protect against that.
1012 spin_lock_irqsave(&ctx->completion_lock, flags);
1013 io_req_link_next(req, nxt);
1014 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1016 io_req_link_next(req, nxt);
1020 static void io_free_req(struct io_kiocb *req)
1022 struct io_kiocb *nxt = NULL;
1024 io_req_find_next(req, &nxt);
1028 io_queue_async_work(nxt);
1032 * Drop reference to request, return next in chain (if there is one) if this
1033 * was the last reference to this request.
1035 __attribute__((nonnull))
1036 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1038 io_req_find_next(req, nxtptr);
1040 if (refcount_dec_and_test(&req->refs))
1044 static void io_put_req(struct io_kiocb *req)
1046 if (refcount_dec_and_test(&req->refs))
1051 * Must only be used if we don't need to care about links, usually from
1052 * within the completion handling itself.
1054 static void __io_double_put_req(struct io_kiocb *req)
1056 /* drop both submit and complete references */
1057 if (refcount_sub_and_test(2, &req->refs))
1061 static void io_double_put_req(struct io_kiocb *req)
1063 /* drop both submit and complete references */
1064 if (refcount_sub_and_test(2, &req->refs))
1068 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1070 struct io_rings *rings = ctx->rings;
1073 * noflush == true is from the waitqueue handler, just ensure we wake
1074 * up the task, and the next invocation will flush the entries. We
1075 * cannot safely to it from here.
1077 if (noflush && !list_empty(&ctx->cq_overflow_list))
1080 io_cqring_overflow_flush(ctx, false);
1082 /* See comment at the top of this file */
1084 return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
1087 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1089 struct io_rings *rings = ctx->rings;
1091 /* make sure SQ entry isn't read before tail */
1092 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1096 * Find and free completed poll iocbs
1098 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1099 struct list_head *done)
1101 void *reqs[IO_IOPOLL_BATCH];
1102 struct io_kiocb *req;
1106 while (!list_empty(done)) {
1107 req = list_first_entry(done, struct io_kiocb, list);
1108 list_del(&req->list);
1110 io_cqring_fill_event(req, req->result);
1113 if (refcount_dec_and_test(&req->refs)) {
1114 /* If we're not using fixed files, we have to pair the
1115 * completion part with the file put. Use regular
1116 * completions for those, only batch free for fixed
1117 * file and non-linked commands.
1119 if (((req->flags & (REQ_F_FIXED_FILE|REQ_F_LINK)) ==
1120 REQ_F_FIXED_FILE) && !io_is_fallback_req(req) &&
1122 reqs[to_free++] = req;
1123 if (to_free == ARRAY_SIZE(reqs))
1124 io_free_req_many(ctx, reqs, &to_free);
1131 io_commit_cqring(ctx);
1132 io_free_req_many(ctx, reqs, &to_free);
1135 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1138 struct io_kiocb *req, *tmp;
1144 * Only spin for completions if we don't have multiple devices hanging
1145 * off our complete list, and we're under the requested amount.
1147 spin = !ctx->poll_multi_file && *nr_events < min;
1150 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1151 struct kiocb *kiocb = &req->rw;
1154 * Move completed entries to our local list. If we find a
1155 * request that requires polling, break out and complete
1156 * the done list first, if we have entries there.
1158 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1159 list_move_tail(&req->list, &done);
1162 if (!list_empty(&done))
1165 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1174 if (!list_empty(&done))
1175 io_iopoll_complete(ctx, nr_events, &done);
1181 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1182 * non-spinning poll check - we'll still enter the driver poll loop, but only
1183 * as a non-spinning completion check.
1185 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1188 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1191 ret = io_do_iopoll(ctx, nr_events, min);
1194 if (!min || *nr_events >= min)
1202 * We can't just wait for polled events to come to us, we have to actively
1203 * find and complete them.
1205 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1207 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1210 mutex_lock(&ctx->uring_lock);
1211 while (!list_empty(&ctx->poll_list)) {
1212 unsigned int nr_events = 0;
1214 io_iopoll_getevents(ctx, &nr_events, 1);
1217 * Ensure we allow local-to-the-cpu processing to take place,
1218 * in this case we need to ensure that we reap all events.
1222 mutex_unlock(&ctx->uring_lock);
1225 static int __io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1228 int iters = 0, ret = 0;
1234 * Don't enter poll loop if we already have events pending.
1235 * If we do, we can potentially be spinning for commands that
1236 * already triggered a CQE (eg in error).
1238 if (io_cqring_events(ctx, false))
1242 * If a submit got punted to a workqueue, we can have the
1243 * application entering polling for a command before it gets
1244 * issued. That app will hold the uring_lock for the duration
1245 * of the poll right here, so we need to take a breather every
1246 * now and then to ensure that the issue has a chance to add
1247 * the poll to the issued list. Otherwise we can spin here
1248 * forever, while the workqueue is stuck trying to acquire the
1251 if (!(++iters & 7)) {
1252 mutex_unlock(&ctx->uring_lock);
1253 mutex_lock(&ctx->uring_lock);
1256 if (*nr_events < min)
1257 tmin = min - *nr_events;
1259 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1263 } while (min && !*nr_events && !need_resched());
1268 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1274 * We disallow the app entering submit/complete with polling, but we
1275 * still need to lock the ring to prevent racing with polled issue
1276 * that got punted to a workqueue.
1278 mutex_lock(&ctx->uring_lock);
1279 ret = __io_iopoll_check(ctx, nr_events, min);
1280 mutex_unlock(&ctx->uring_lock);
1284 static void kiocb_end_write(struct io_kiocb *req)
1287 * Tell lockdep we inherited freeze protection from submission
1290 if (req->flags & REQ_F_ISREG) {
1291 struct inode *inode = file_inode(req->file);
1293 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1295 file_end_write(req->file);
1298 static inline void req_set_fail_links(struct io_kiocb *req)
1300 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1301 req->flags |= REQ_F_FAIL_LINK;
1304 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1306 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1308 if (kiocb->ki_flags & IOCB_WRITE)
1309 kiocb_end_write(req);
1311 if (res != req->result)
1312 req_set_fail_links(req);
1313 io_cqring_add_event(req, res);
1316 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1318 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1320 io_complete_rw_common(kiocb, res);
1324 static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res)
1326 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1327 struct io_kiocb *nxt = NULL;
1329 io_complete_rw_common(kiocb, res);
1330 io_put_req_find_next(req, &nxt);
1335 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1337 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw);
1339 if (kiocb->ki_flags & IOCB_WRITE)
1340 kiocb_end_write(req);
1342 if (res != req->result)
1343 req_set_fail_links(req);
1346 req->flags |= REQ_F_IOPOLL_COMPLETED;
1350 * After the iocb has been issued, it's safe to be found on the poll list.
1351 * Adding the kiocb to the list AFTER submission ensures that we don't
1352 * find it from a io_iopoll_getevents() thread before the issuer is done
1353 * accessing the kiocb cookie.
1355 static void io_iopoll_req_issued(struct io_kiocb *req)
1357 struct io_ring_ctx *ctx = req->ctx;
1360 * Track whether we have multiple files in our lists. This will impact
1361 * how we do polling eventually, not spinning if we're on potentially
1362 * different devices.
1364 if (list_empty(&ctx->poll_list)) {
1365 ctx->poll_multi_file = false;
1366 } else if (!ctx->poll_multi_file) {
1367 struct io_kiocb *list_req;
1369 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1371 if (list_req->rw.ki_filp != req->rw.ki_filp)
1372 ctx->poll_multi_file = true;
1376 * For fast devices, IO may have already completed. If it has, add
1377 * it to the front so we find it first.
1379 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1380 list_add(&req->list, &ctx->poll_list);
1382 list_add_tail(&req->list, &ctx->poll_list);
1385 static void io_file_put(struct io_submit_state *state)
1388 int diff = state->has_refs - state->used_refs;
1391 fput_many(state->file, diff);
1397 * Get as many references to a file as we have IOs left in this submission,
1398 * assuming most submissions are for one file, or at least that each file
1399 * has more than one submission.
1401 static struct file *io_file_get(struct io_submit_state *state, int fd)
1407 if (state->fd == fd) {
1414 state->file = fget_many(fd, state->ios_left);
1419 state->has_refs = state->ios_left;
1420 state->used_refs = 1;
1426 * If we tracked the file through the SCM inflight mechanism, we could support
1427 * any file. For now, just ensure that anything potentially problematic is done
1430 static bool io_file_supports_async(struct file *file)
1432 umode_t mode = file_inode(file)->i_mode;
1434 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
1436 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
1442 static int io_prep_rw(struct io_kiocb *req, bool force_nonblock)
1444 const struct io_uring_sqe *sqe = req->sqe;
1445 struct io_ring_ctx *ctx = req->ctx;
1446 struct kiocb *kiocb = &req->rw;
1453 if (S_ISREG(file_inode(req->file)->i_mode))
1454 req->flags |= REQ_F_ISREG;
1456 kiocb->ki_pos = READ_ONCE(sqe->off);
1457 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
1458 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
1460 ioprio = READ_ONCE(sqe->ioprio);
1462 ret = ioprio_check_cap(ioprio);
1466 kiocb->ki_ioprio = ioprio;
1468 kiocb->ki_ioprio = get_current_ioprio();
1470 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
1474 /* don't allow async punt if RWF_NOWAIT was requested */
1475 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
1476 (req->file->f_flags & O_NONBLOCK))
1477 req->flags |= REQ_F_NOWAIT;
1480 kiocb->ki_flags |= IOCB_NOWAIT;
1482 if (ctx->flags & IORING_SETUP_IOPOLL) {
1483 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
1484 !kiocb->ki_filp->f_op->iopoll)
1487 kiocb->ki_flags |= IOCB_HIPRI;
1488 kiocb->ki_complete = io_complete_rw_iopoll;
1491 if (kiocb->ki_flags & IOCB_HIPRI)
1493 kiocb->ki_complete = io_complete_rw;
1498 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
1504 case -ERESTARTNOINTR:
1505 case -ERESTARTNOHAND:
1506 case -ERESTART_RESTARTBLOCK:
1508 * We can't just restart the syscall, since previously
1509 * submitted sqes may already be in progress. Just fail this
1515 kiocb->ki_complete(kiocb, ret, 0);
1519 static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt,
1522 if (in_async && ret >= 0 && kiocb->ki_complete == io_complete_rw)
1523 *nxt = __io_complete_rw(kiocb, ret);
1525 io_rw_done(kiocb, ret);
1528 static ssize_t io_import_fixed(struct io_ring_ctx *ctx, int rw,
1529 const struct io_uring_sqe *sqe,
1530 struct iov_iter *iter)
1532 size_t len = READ_ONCE(sqe->len);
1533 struct io_mapped_ubuf *imu;
1534 unsigned index, buf_index;
1538 /* attempt to use fixed buffers without having provided iovecs */
1539 if (unlikely(!ctx->user_bufs))
1542 buf_index = READ_ONCE(sqe->buf_index);
1543 if (unlikely(buf_index >= ctx->nr_user_bufs))
1546 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
1547 imu = &ctx->user_bufs[index];
1548 buf_addr = READ_ONCE(sqe->addr);
1551 if (buf_addr + len < buf_addr)
1553 /* not inside the mapped region */
1554 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
1558 * May not be a start of buffer, set size appropriately
1559 * and advance us to the beginning.
1561 offset = buf_addr - imu->ubuf;
1562 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
1566 * Don't use iov_iter_advance() here, as it's really slow for
1567 * using the latter parts of a big fixed buffer - it iterates
1568 * over each segment manually. We can cheat a bit here, because
1571 * 1) it's a BVEC iter, we set it up
1572 * 2) all bvecs are PAGE_SIZE in size, except potentially the
1573 * first and last bvec
1575 * So just find our index, and adjust the iterator afterwards.
1576 * If the offset is within the first bvec (or the whole first
1577 * bvec, just use iov_iter_advance(). This makes it easier
1578 * since we can just skip the first segment, which may not
1579 * be PAGE_SIZE aligned.
1581 const struct bio_vec *bvec = imu->bvec;
1583 if (offset <= bvec->bv_len) {
1584 iov_iter_advance(iter, offset);
1586 unsigned long seg_skip;
1588 /* skip first vec */
1589 offset -= bvec->bv_len;
1590 seg_skip = 1 + (offset >> PAGE_SHIFT);
1592 iter->bvec = bvec + seg_skip;
1593 iter->nr_segs -= seg_skip;
1594 iter->count -= bvec->bv_len + offset;
1595 iter->iov_offset = offset & ~PAGE_MASK;
1602 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
1603 struct iovec **iovec, struct iov_iter *iter)
1605 const struct io_uring_sqe *sqe = req->sqe;
1606 void __user *buf = u64_to_user_ptr(READ_ONCE(sqe->addr));
1607 size_t sqe_len = READ_ONCE(sqe->len);
1611 * We're reading ->opcode for the second time, but the first read
1612 * doesn't care whether it's _FIXED or not, so it doesn't matter
1613 * whether ->opcode changes concurrently. The first read does care
1614 * about whether it is a READ or a WRITE, so we don't trust this read
1615 * for that purpose and instead let the caller pass in the read/write
1618 opcode = READ_ONCE(sqe->opcode);
1619 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
1621 return io_import_fixed(req->ctx, rw, sqe, iter);
1625 struct io_async_rw *iorw = &req->io->rw;
1628 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
1629 if (iorw->iov == iorw->fast_iov)
1637 #ifdef CONFIG_COMPAT
1638 if (req->ctx->compat)
1639 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
1643 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
1647 * For files that don't have ->read_iter() and ->write_iter(), handle them
1648 * by looping over ->read() or ->write() manually.
1650 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
1651 struct iov_iter *iter)
1656 * Don't support polled IO through this interface, and we can't
1657 * support non-blocking either. For the latter, this just causes
1658 * the kiocb to be handled from an async context.
1660 if (kiocb->ki_flags & IOCB_HIPRI)
1662 if (kiocb->ki_flags & IOCB_NOWAIT)
1665 while (iov_iter_count(iter)) {
1669 if (!iov_iter_is_bvec(iter)) {
1670 iovec = iov_iter_iovec(iter);
1672 /* fixed buffers import bvec */
1673 iovec.iov_base = kmap(iter->bvec->bv_page)
1675 iovec.iov_len = min(iter->count,
1676 iter->bvec->bv_len - iter->iov_offset);
1680 nr = file->f_op->read(file, iovec.iov_base,
1681 iovec.iov_len, &kiocb->ki_pos);
1683 nr = file->f_op->write(file, iovec.iov_base,
1684 iovec.iov_len, &kiocb->ki_pos);
1687 if (iov_iter_is_bvec(iter))
1688 kunmap(iter->bvec->bv_page);
1696 if (nr != iovec.iov_len)
1698 iov_iter_advance(iter, nr);
1704 static void io_req_map_io(struct io_kiocb *req, ssize_t io_size,
1705 struct iovec *iovec, struct iovec *fast_iov,
1706 struct iov_iter *iter)
1708 req->io->rw.nr_segs = iter->nr_segs;
1709 req->io->rw.size = io_size;
1710 req->io->rw.iov = iovec;
1711 if (!req->io->rw.iov) {
1712 req->io->rw.iov = req->io->rw.fast_iov;
1713 memcpy(req->io->rw.iov, fast_iov,
1714 sizeof(struct iovec) * iter->nr_segs);
1718 static int io_setup_async_io(struct io_kiocb *req, ssize_t io_size,
1719 struct iovec *iovec, struct iovec *fast_iov,
1720 struct iov_iter *iter)
1722 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
1724 io_req_map_io(req, io_size, iovec, fast_iov, iter);
1725 memcpy(&req->io->sqe, req->sqe, sizeof(req->io->sqe));
1726 req->sqe = &req->io->sqe;
1733 static int io_read_prep(struct io_kiocb *req, struct iovec **iovec,
1734 struct iov_iter *iter, bool force_nonblock)
1738 ret = io_prep_rw(req, force_nonblock);
1742 if (unlikely(!(req->file->f_mode & FMODE_READ)))
1745 return io_import_iovec(READ, req, iovec, iter);
1748 static int io_read(struct io_kiocb *req, struct io_kiocb **nxt,
1749 bool force_nonblock)
1751 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1752 struct kiocb *kiocb = &req->rw;
1753 struct iov_iter iter;
1756 ssize_t io_size, ret;
1759 ret = io_read_prep(req, &iovec, &iter, force_nonblock);
1763 ret = io_import_iovec(READ, req, &iovec, &iter);
1770 if (req->flags & REQ_F_LINK)
1771 req->result = io_size;
1774 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1775 * we know to async punt it even if it was opened O_NONBLOCK
1777 if (force_nonblock && !io_file_supports_async(file)) {
1778 req->flags |= REQ_F_MUST_PUNT;
1782 iov_count = iov_iter_count(&iter);
1783 ret = rw_verify_area(READ, file, &kiocb->ki_pos, iov_count);
1787 if (file->f_op->read_iter)
1788 ret2 = call_read_iter(file, kiocb, &iter);
1790 ret2 = loop_rw_iter(READ, file, kiocb, &iter);
1793 * In case of a short read, punt to async. This can happen
1794 * if we have data partially cached. Alternatively we can
1795 * return the short read, in which case the application will
1796 * need to issue another SQE and wait for it. That SQE will
1797 * need async punt anyway, so it's more efficient to do it
1800 if (force_nonblock && !(req->flags & REQ_F_NOWAIT) &&
1801 (req->flags & REQ_F_ISREG) &&
1802 ret2 > 0 && ret2 < io_size)
1804 /* Catch -EAGAIN return for forced non-blocking submission */
1805 if (!force_nonblock || ret2 != -EAGAIN) {
1806 kiocb_done(kiocb, ret2, nxt, req->in_async);
1809 ret = io_setup_async_io(req, io_size, iovec,
1810 inline_vecs, &iter);
1821 static int io_write_prep(struct io_kiocb *req, struct iovec **iovec,
1822 struct iov_iter *iter, bool force_nonblock)
1826 ret = io_prep_rw(req, force_nonblock);
1830 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
1833 return io_import_iovec(WRITE, req, iovec, iter);
1836 static int io_write(struct io_kiocb *req, struct io_kiocb **nxt,
1837 bool force_nonblock)
1839 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1840 struct kiocb *kiocb = &req->rw;
1841 struct iov_iter iter;
1844 ssize_t ret, io_size;
1847 ret = io_write_prep(req, &iovec, &iter, force_nonblock);
1851 ret = io_import_iovec(WRITE, req, &iovec, &iter);
1856 file = kiocb->ki_filp;
1858 if (req->flags & REQ_F_LINK)
1859 req->result = io_size;
1862 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
1863 * we know to async punt it even if it was opened O_NONBLOCK
1865 if (force_nonblock && !io_file_supports_async(req->file)) {
1866 req->flags |= REQ_F_MUST_PUNT;
1870 /* file path doesn't support NOWAIT for non-direct_IO */
1871 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
1872 (req->flags & REQ_F_ISREG))
1875 iov_count = iov_iter_count(&iter);
1876 ret = rw_verify_area(WRITE, file, &kiocb->ki_pos, iov_count);
1881 * Open-code file_start_write here to grab freeze protection,
1882 * which will be released by another thread in
1883 * io_complete_rw(). Fool lockdep by telling it the lock got
1884 * released so that it doesn't complain about the held lock when
1885 * we return to userspace.
1887 if (req->flags & REQ_F_ISREG) {
1888 __sb_start_write(file_inode(file)->i_sb,
1889 SB_FREEZE_WRITE, true);
1890 __sb_writers_release(file_inode(file)->i_sb,
1893 kiocb->ki_flags |= IOCB_WRITE;
1895 if (file->f_op->write_iter)
1896 ret2 = call_write_iter(file, kiocb, &iter);
1898 ret2 = loop_rw_iter(WRITE, file, kiocb, &iter);
1899 if (!force_nonblock || ret2 != -EAGAIN) {
1900 kiocb_done(kiocb, ret2, nxt, req->in_async);
1903 ret = io_setup_async_io(req, io_size, iovec,
1904 inline_vecs, &iter);
1916 * IORING_OP_NOP just posts a completion event, nothing else.
1918 static int io_nop(struct io_kiocb *req)
1920 struct io_ring_ctx *ctx = req->ctx;
1922 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1925 io_cqring_add_event(req, 0);
1930 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1932 struct io_ring_ctx *ctx = req->ctx;
1937 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1939 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1945 static int io_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe,
1946 struct io_kiocb **nxt, bool force_nonblock)
1948 loff_t sqe_off = READ_ONCE(sqe->off);
1949 loff_t sqe_len = READ_ONCE(sqe->len);
1950 loff_t end = sqe_off + sqe_len;
1951 unsigned fsync_flags;
1954 fsync_flags = READ_ONCE(sqe->fsync_flags);
1955 if (unlikely(fsync_flags & ~IORING_FSYNC_DATASYNC))
1958 ret = io_prep_fsync(req, sqe);
1962 /* fsync always requires a blocking context */
1966 ret = vfs_fsync_range(req->rw.ki_filp, sqe_off,
1967 end > 0 ? end : LLONG_MAX,
1968 fsync_flags & IORING_FSYNC_DATASYNC);
1971 req_set_fail_links(req);
1972 io_cqring_add_event(req, ret);
1973 io_put_req_find_next(req, nxt);
1977 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
1979 struct io_ring_ctx *ctx = req->ctx;
1985 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
1987 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
1993 static int io_sync_file_range(struct io_kiocb *req,
1994 const struct io_uring_sqe *sqe,
1995 struct io_kiocb **nxt,
1996 bool force_nonblock)
2003 ret = io_prep_sfr(req, sqe);
2007 /* sync_file_range always requires a blocking context */
2011 sqe_off = READ_ONCE(sqe->off);
2012 sqe_len = READ_ONCE(sqe->len);
2013 flags = READ_ONCE(sqe->sync_range_flags);
2015 ret = sync_file_range(req->rw.ki_filp, sqe_off, sqe_len, flags);
2018 req_set_fail_links(req);
2019 io_cqring_add_event(req, ret);
2020 io_put_req_find_next(req, nxt);
2024 static int io_sendmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2026 #if defined(CONFIG_NET)
2027 const struct io_uring_sqe *sqe = req->sqe;
2028 struct user_msghdr __user *msg;
2031 flags = READ_ONCE(sqe->msg_flags);
2032 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2033 io->msg.iov = io->msg.fast_iov;
2034 return sendmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.iov);
2040 static int io_sendmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2041 struct io_kiocb **nxt, bool force_nonblock)
2043 #if defined(CONFIG_NET)
2044 struct io_async_msghdr *kmsg = NULL;
2045 struct socket *sock;
2048 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2051 sock = sock_from_file(req->file, &ret);
2053 struct io_async_ctx io, *copy;
2054 struct sockaddr_storage addr;
2057 flags = READ_ONCE(sqe->msg_flags);
2058 if (flags & MSG_DONTWAIT)
2059 req->flags |= REQ_F_NOWAIT;
2060 else if (force_nonblock)
2061 flags |= MSG_DONTWAIT;
2064 kmsg = &req->io->msg;
2065 kmsg->msg.msg_name = &addr;
2066 /* if iov is set, it's allocated already */
2068 kmsg->iov = kmsg->fast_iov;
2069 kmsg->msg.msg_iter.iov = kmsg->iov;
2072 kmsg->msg.msg_name = &addr;
2073 ret = io_sendmsg_prep(req, &io);
2078 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
2079 if (force_nonblock && ret == -EAGAIN) {
2080 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2085 memcpy(©->msg, &io.msg, sizeof(copy->msg));
2087 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2088 req->sqe = &req->io->sqe;
2091 if (ret == -ERESTARTSYS)
2096 if (kmsg && kmsg->iov != kmsg->fast_iov)
2098 io_cqring_add_event(req, ret);
2100 req_set_fail_links(req);
2101 io_put_req_find_next(req, nxt);
2108 static int io_recvmsg_prep(struct io_kiocb *req, struct io_async_ctx *io)
2110 #if defined(CONFIG_NET)
2111 const struct io_uring_sqe *sqe = req->sqe;
2112 struct user_msghdr __user *msg;
2115 flags = READ_ONCE(sqe->msg_flags);
2116 msg = (struct user_msghdr __user *)(unsigned long) READ_ONCE(sqe->addr);
2117 io->msg.iov = io->msg.fast_iov;
2118 return recvmsg_copy_msghdr(&io->msg.msg, msg, flags, &io->msg.uaddr,
2125 static int io_recvmsg(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2126 struct io_kiocb **nxt, bool force_nonblock)
2128 #if defined(CONFIG_NET)
2129 struct io_async_msghdr *kmsg = NULL;
2130 struct socket *sock;
2133 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2136 sock = sock_from_file(req->file, &ret);
2138 struct user_msghdr __user *msg;
2139 struct io_async_ctx io, *copy;
2140 struct sockaddr_storage addr;
2143 flags = READ_ONCE(sqe->msg_flags);
2144 if (flags & MSG_DONTWAIT)
2145 req->flags |= REQ_F_NOWAIT;
2146 else if (force_nonblock)
2147 flags |= MSG_DONTWAIT;
2149 msg = (struct user_msghdr __user *) (unsigned long)
2150 READ_ONCE(sqe->addr);
2152 kmsg = &req->io->msg;
2153 kmsg->msg.msg_name = &addr;
2154 /* if iov is set, it's allocated already */
2156 kmsg->iov = kmsg->fast_iov;
2157 kmsg->msg.msg_iter.iov = kmsg->iov;
2160 kmsg->msg.msg_name = &addr;
2161 ret = io_recvmsg_prep(req, &io);
2166 ret = __sys_recvmsg_sock(sock, &kmsg->msg, msg, kmsg->uaddr, flags);
2167 if (force_nonblock && ret == -EAGAIN) {
2168 copy = kmalloc(sizeof(*copy), GFP_KERNEL);
2173 memcpy(copy, &io, sizeof(*copy));
2175 memcpy(&req->io->sqe, req->sqe, sizeof(*req->sqe));
2176 req->sqe = &req->io->sqe;
2179 if (ret == -ERESTARTSYS)
2184 if (kmsg && kmsg->iov != kmsg->fast_iov)
2186 io_cqring_add_event(req, ret);
2188 req_set_fail_links(req);
2189 io_put_req_find_next(req, nxt);
2196 static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2197 struct io_kiocb **nxt, bool force_nonblock)
2199 #if defined(CONFIG_NET)
2200 struct sockaddr __user *addr;
2201 int __user *addr_len;
2202 unsigned file_flags;
2205 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2207 if (sqe->ioprio || sqe->len || sqe->buf_index)
2210 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2211 addr_len = (int __user *) (unsigned long) READ_ONCE(sqe->addr2);
2212 flags = READ_ONCE(sqe->accept_flags);
2213 file_flags = force_nonblock ? O_NONBLOCK : 0;
2215 ret = __sys_accept4_file(req->file, file_flags, addr, addr_len, flags);
2216 if (ret == -EAGAIN && force_nonblock) {
2217 req->work.flags |= IO_WQ_WORK_NEEDS_FILES;
2220 if (ret == -ERESTARTSYS)
2223 req_set_fail_links(req);
2224 io_cqring_add_event(req, ret);
2225 io_put_req_find_next(req, nxt);
2232 static int io_connect_prep(struct io_kiocb *req, struct io_async_ctx *io)
2234 #if defined(CONFIG_NET)
2235 const struct io_uring_sqe *sqe = req->sqe;
2236 struct sockaddr __user *addr;
2239 addr = (struct sockaddr __user *) (unsigned long) READ_ONCE(sqe->addr);
2240 addr_len = READ_ONCE(sqe->addr2);
2241 return move_addr_to_kernel(addr, addr_len, &io->connect.address);
2247 static int io_connect(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2248 struct io_kiocb **nxt, bool force_nonblock)
2250 #if defined(CONFIG_NET)
2251 struct io_async_ctx __io, *io;
2252 unsigned file_flags;
2255 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
2257 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
2260 addr_len = READ_ONCE(sqe->addr2);
2261 file_flags = force_nonblock ? O_NONBLOCK : 0;
2266 ret = io_connect_prep(req, &__io);
2272 ret = __sys_connect_file(req->file, &io->connect.address, addr_len,
2274 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
2275 io = kmalloc(sizeof(*io), GFP_KERNEL);
2280 memcpy(&io->connect, &__io.connect, sizeof(io->connect));
2282 memcpy(&io->sqe, req->sqe, sizeof(*req->sqe));
2283 req->sqe = &io->sqe;
2286 if (ret == -ERESTARTSYS)
2290 req_set_fail_links(req);
2291 io_cqring_add_event(req, ret);
2292 io_put_req_find_next(req, nxt);
2299 static void io_poll_remove_one(struct io_kiocb *req)
2301 struct io_poll_iocb *poll = &req->poll;
2303 spin_lock(&poll->head->lock);
2304 WRITE_ONCE(poll->canceled, true);
2305 if (!list_empty(&poll->wait.entry)) {
2306 list_del_init(&poll->wait.entry);
2307 io_queue_async_work(req);
2309 spin_unlock(&poll->head->lock);
2310 hash_del(&req->hash_node);
2313 static void io_poll_remove_all(struct io_ring_ctx *ctx)
2315 struct hlist_node *tmp;
2316 struct io_kiocb *req;
2319 spin_lock_irq(&ctx->completion_lock);
2320 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
2321 struct hlist_head *list;
2323 list = &ctx->cancel_hash[i];
2324 hlist_for_each_entry_safe(req, tmp, list, hash_node)
2325 io_poll_remove_one(req);
2327 spin_unlock_irq(&ctx->completion_lock);
2330 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
2332 struct hlist_head *list;
2333 struct io_kiocb *req;
2335 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
2336 hlist_for_each_entry(req, list, hash_node) {
2337 if (sqe_addr == req->user_data) {
2338 io_poll_remove_one(req);
2347 * Find a running poll command that matches one specified in sqe->addr,
2348 * and remove it if found.
2350 static int io_poll_remove(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2352 struct io_ring_ctx *ctx = req->ctx;
2355 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2357 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
2361 spin_lock_irq(&ctx->completion_lock);
2362 ret = io_poll_cancel(ctx, READ_ONCE(sqe->addr));
2363 spin_unlock_irq(&ctx->completion_lock);
2365 io_cqring_add_event(req, ret);
2367 req_set_fail_links(req);
2372 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
2374 struct io_ring_ctx *ctx = req->ctx;
2376 req->poll.done = true;
2378 io_cqring_fill_event(req, error);
2380 io_cqring_fill_event(req, mangle_poll(mask));
2381 io_commit_cqring(ctx);
2384 static void io_poll_complete_work(struct io_wq_work **workptr)
2386 struct io_wq_work *work = *workptr;
2387 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2388 struct io_poll_iocb *poll = &req->poll;
2389 struct poll_table_struct pt = { ._key = poll->events };
2390 struct io_ring_ctx *ctx = req->ctx;
2391 struct io_kiocb *nxt = NULL;
2395 if (work->flags & IO_WQ_WORK_CANCEL) {
2396 WRITE_ONCE(poll->canceled, true);
2398 } else if (READ_ONCE(poll->canceled)) {
2402 if (ret != -ECANCELED)
2403 mask = vfs_poll(poll->file, &pt) & poll->events;
2406 * Note that ->ki_cancel callers also delete iocb from active_reqs after
2407 * calling ->ki_cancel. We need the ctx_lock roundtrip here to
2408 * synchronize with them. In the cancellation case the list_del_init
2409 * itself is not actually needed, but harmless so we keep it in to
2410 * avoid further branches in the fast path.
2412 spin_lock_irq(&ctx->completion_lock);
2413 if (!mask && ret != -ECANCELED) {
2414 add_wait_queue(poll->head, &poll->wait);
2415 spin_unlock_irq(&ctx->completion_lock);
2418 hash_del(&req->hash_node);
2419 io_poll_complete(req, mask, ret);
2420 spin_unlock_irq(&ctx->completion_lock);
2422 io_cqring_ev_posted(ctx);
2425 req_set_fail_links(req);
2426 io_put_req_find_next(req, &nxt);
2428 *workptr = &nxt->work;
2431 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
2434 struct io_poll_iocb *poll = wait->private;
2435 struct io_kiocb *req = container_of(poll, struct io_kiocb, poll);
2436 struct io_ring_ctx *ctx = req->ctx;
2437 __poll_t mask = key_to_poll(key);
2438 unsigned long flags;
2440 /* for instances that support it check for an event match first: */
2441 if (mask && !(mask & poll->events))
2444 list_del_init(&poll->wait.entry);
2447 * Run completion inline if we can. We're using trylock here because
2448 * we are violating the completion_lock -> poll wq lock ordering.
2449 * If we have a link timeout we're going to need the completion_lock
2450 * for finalizing the request, mark us as having grabbed that already.
2452 if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
2453 hash_del(&req->hash_node);
2454 io_poll_complete(req, mask, 0);
2455 req->flags |= REQ_F_COMP_LOCKED;
2457 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2459 io_cqring_ev_posted(ctx);
2461 io_queue_async_work(req);
2467 struct io_poll_table {
2468 struct poll_table_struct pt;
2469 struct io_kiocb *req;
2473 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
2474 struct poll_table_struct *p)
2476 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
2478 if (unlikely(pt->req->poll.head)) {
2479 pt->error = -EINVAL;
2484 pt->req->poll.head = head;
2485 add_wait_queue(head, &pt->req->poll.wait);
2488 static void io_poll_req_insert(struct io_kiocb *req)
2490 struct io_ring_ctx *ctx = req->ctx;
2491 struct hlist_head *list;
2493 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
2494 hlist_add_head(&req->hash_node, list);
2497 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2498 struct io_kiocb **nxt)
2500 struct io_poll_iocb *poll = &req->poll;
2501 struct io_ring_ctx *ctx = req->ctx;
2502 struct io_poll_table ipt;
2503 bool cancel = false;
2507 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2509 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
2515 INIT_IO_WORK(&req->work, io_poll_complete_work);
2516 events = READ_ONCE(sqe->poll_events);
2517 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
2518 INIT_HLIST_NODE(&req->hash_node);
2522 poll->canceled = false;
2524 ipt.pt._qproc = io_poll_queue_proc;
2525 ipt.pt._key = poll->events;
2527 ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
2529 /* initialized the list so that we can do list_empty checks */
2530 INIT_LIST_HEAD(&poll->wait.entry);
2531 init_waitqueue_func_entry(&poll->wait, io_poll_wake);
2532 poll->wait.private = poll;
2534 INIT_LIST_HEAD(&req->list);
2536 mask = vfs_poll(poll->file, &ipt.pt) & poll->events;
2538 spin_lock_irq(&ctx->completion_lock);
2539 if (likely(poll->head)) {
2540 spin_lock(&poll->head->lock);
2541 if (unlikely(list_empty(&poll->wait.entry))) {
2547 if (mask || ipt.error)
2548 list_del_init(&poll->wait.entry);
2550 WRITE_ONCE(poll->canceled, true);
2551 else if (!poll->done) /* actually waiting for an event */
2552 io_poll_req_insert(req);
2553 spin_unlock(&poll->head->lock);
2555 if (mask) { /* no async, we'd stolen it */
2557 io_poll_complete(req, mask, 0);
2559 spin_unlock_irq(&ctx->completion_lock);
2562 io_cqring_ev_posted(ctx);
2563 io_put_req_find_next(req, nxt);
2568 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
2570 struct io_timeout_data *data = container_of(timer,
2571 struct io_timeout_data, timer);
2572 struct io_kiocb *req = data->req;
2573 struct io_ring_ctx *ctx = req->ctx;
2574 unsigned long flags;
2576 atomic_inc(&ctx->cq_timeouts);
2578 spin_lock_irqsave(&ctx->completion_lock, flags);
2580 * We could be racing with timeout deletion. If the list is empty,
2581 * then timeout lookup already found it and will be handling it.
2583 if (!list_empty(&req->list)) {
2584 struct io_kiocb *prev;
2587 * Adjust the reqs sequence before the current one because it
2588 * will consume a slot in the cq_ring and the cq_tail
2589 * pointer will be increased, otherwise other timeout reqs may
2590 * return in advance without waiting for enough wait_nr.
2593 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
2595 list_del_init(&req->list);
2598 io_cqring_fill_event(req, -ETIME);
2599 io_commit_cqring(ctx);
2600 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2602 io_cqring_ev_posted(ctx);
2603 req_set_fail_links(req);
2605 return HRTIMER_NORESTART;
2608 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
2610 struct io_kiocb *req;
2613 list_for_each_entry(req, &ctx->timeout_list, list) {
2614 if (user_data == req->user_data) {
2615 list_del_init(&req->list);
2624 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
2628 req_set_fail_links(req);
2629 io_cqring_fill_event(req, -ECANCELED);
2635 * Remove or update an existing timeout command
2637 static int io_timeout_remove(struct io_kiocb *req,
2638 const struct io_uring_sqe *sqe)
2640 struct io_ring_ctx *ctx = req->ctx;
2644 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2646 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
2648 flags = READ_ONCE(sqe->timeout_flags);
2652 spin_lock_irq(&ctx->completion_lock);
2653 ret = io_timeout_cancel(ctx, READ_ONCE(sqe->addr));
2655 io_cqring_fill_event(req, ret);
2656 io_commit_cqring(ctx);
2657 spin_unlock_irq(&ctx->completion_lock);
2658 io_cqring_ev_posted(ctx);
2660 req_set_fail_links(req);
2665 static int io_timeout_prep(struct io_kiocb *req, struct io_async_ctx *io,
2666 bool is_timeout_link)
2668 const struct io_uring_sqe *sqe = req->sqe;
2669 struct io_timeout_data *data;
2672 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
2674 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
2676 if (sqe->off && is_timeout_link)
2678 flags = READ_ONCE(sqe->timeout_flags);
2679 if (flags & ~IORING_TIMEOUT_ABS)
2682 data = &io->timeout;
2684 req->flags |= REQ_F_TIMEOUT;
2686 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
2689 if (flags & IORING_TIMEOUT_ABS)
2690 data->mode = HRTIMER_MODE_ABS;
2692 data->mode = HRTIMER_MODE_REL;
2694 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
2699 static int io_timeout(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2702 struct io_ring_ctx *ctx = req->ctx;
2703 struct io_timeout_data *data;
2704 struct io_async_ctx *io;
2705 struct list_head *entry;
2712 io = kmalloc(sizeof(*io), GFP_KERNEL);
2715 ret = io_timeout_prep(req, io, false);
2721 data = &req->io->timeout;
2724 * sqe->off holds how many events that need to occur for this
2725 * timeout event to be satisfied. If it isn't set, then this is
2726 * a pure timeout request, sequence isn't used.
2728 count = READ_ONCE(sqe->off);
2730 req->flags |= REQ_F_TIMEOUT_NOSEQ;
2731 spin_lock_irq(&ctx->completion_lock);
2732 entry = ctx->timeout_list.prev;
2736 req->sequence = ctx->cached_sq_head + count - 1;
2737 data->seq_offset = count;
2740 * Insertion sort, ensuring the first entry in the list is always
2741 * the one we need first.
2743 spin_lock_irq(&ctx->completion_lock);
2744 list_for_each_prev(entry, &ctx->timeout_list) {
2745 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
2746 unsigned nxt_sq_head;
2747 long long tmp, tmp_nxt;
2748 u32 nxt_offset = nxt->io->timeout.seq_offset;
2750 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
2754 * Since cached_sq_head + count - 1 can overflow, use type long
2757 tmp = (long long)ctx->cached_sq_head + count - 1;
2758 nxt_sq_head = nxt->sequence - nxt_offset + 1;
2759 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
2762 * cached_sq_head may overflow, and it will never overflow twice
2763 * once there is some timeout req still be valid.
2765 if (ctx->cached_sq_head < nxt_sq_head)
2772 * Sequence of reqs after the insert one and itself should
2773 * be adjusted because each timeout req consumes a slot.
2778 req->sequence -= span;
2780 list_add(&req->list, entry);
2781 data->timer.function = io_timeout_fn;
2782 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
2783 spin_unlock_irq(&ctx->completion_lock);
2787 static bool io_cancel_cb(struct io_wq_work *work, void *data)
2789 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
2791 return req->user_data == (unsigned long) data;
2794 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
2796 enum io_wq_cancel cancel_ret;
2799 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
2800 switch (cancel_ret) {
2801 case IO_WQ_CANCEL_OK:
2804 case IO_WQ_CANCEL_RUNNING:
2807 case IO_WQ_CANCEL_NOTFOUND:
2815 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
2816 struct io_kiocb *req, __u64 sqe_addr,
2817 struct io_kiocb **nxt, int success_ret)
2819 unsigned long flags;
2822 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
2823 if (ret != -ENOENT) {
2824 spin_lock_irqsave(&ctx->completion_lock, flags);
2828 spin_lock_irqsave(&ctx->completion_lock, flags);
2829 ret = io_timeout_cancel(ctx, sqe_addr);
2832 ret = io_poll_cancel(ctx, sqe_addr);
2836 io_cqring_fill_event(req, ret);
2837 io_commit_cqring(ctx);
2838 spin_unlock_irqrestore(&ctx->completion_lock, flags);
2839 io_cqring_ev_posted(ctx);
2842 req_set_fail_links(req);
2843 io_put_req_find_next(req, nxt);
2846 static int io_async_cancel(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2847 struct io_kiocb **nxt)
2849 struct io_ring_ctx *ctx = req->ctx;
2851 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2853 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
2857 io_async_find_and_cancel(ctx, req, READ_ONCE(sqe->addr), nxt, 0);
2861 static int io_req_defer_prep(struct io_kiocb *req, struct io_async_ctx *io)
2863 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2864 struct iov_iter iter;
2867 memcpy(&io->sqe, req->sqe, sizeof(io->sqe));
2868 req->sqe = &io->sqe;
2870 switch (io->sqe.opcode) {
2871 case IORING_OP_READV:
2872 case IORING_OP_READ_FIXED:
2873 ret = io_read_prep(req, &iovec, &iter, true);
2875 case IORING_OP_WRITEV:
2876 case IORING_OP_WRITE_FIXED:
2877 ret = io_write_prep(req, &iovec, &iter, true);
2879 case IORING_OP_SENDMSG:
2880 ret = io_sendmsg_prep(req, io);
2882 case IORING_OP_RECVMSG:
2883 ret = io_recvmsg_prep(req, io);
2885 case IORING_OP_CONNECT:
2886 ret = io_connect_prep(req, io);
2888 case IORING_OP_TIMEOUT:
2889 return io_timeout_prep(req, io, false);
2890 case IORING_OP_LINK_TIMEOUT:
2891 return io_timeout_prep(req, io, true);
2901 io_req_map_io(req, ret, iovec, inline_vecs, &iter);
2905 static int io_req_defer(struct io_kiocb *req)
2907 struct io_ring_ctx *ctx = req->ctx;
2908 struct io_async_ctx *io;
2911 /* Still need defer if there is pending req in defer list. */
2912 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
2915 io = kmalloc(sizeof(*io), GFP_KERNEL);
2919 ret = io_req_defer_prep(req, io);
2925 spin_lock_irq(&ctx->completion_lock);
2926 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
2927 spin_unlock_irq(&ctx->completion_lock);
2931 trace_io_uring_defer(ctx, req, req->user_data);
2932 list_add_tail(&req->list, &ctx->defer_list);
2933 spin_unlock_irq(&ctx->completion_lock);
2934 return -EIOCBQUEUED;
2937 __attribute__((nonnull))
2938 static int io_issue_sqe(struct io_kiocb *req, struct io_kiocb **nxt,
2939 bool force_nonblock)
2942 struct io_ring_ctx *ctx = req->ctx;
2944 opcode = READ_ONCE(req->sqe->opcode);
2949 case IORING_OP_READV:
2950 if (unlikely(req->sqe->buf_index))
2952 ret = io_read(req, nxt, force_nonblock);
2954 case IORING_OP_WRITEV:
2955 if (unlikely(req->sqe->buf_index))
2957 ret = io_write(req, nxt, force_nonblock);
2959 case IORING_OP_READ_FIXED:
2960 ret = io_read(req, nxt, force_nonblock);
2962 case IORING_OP_WRITE_FIXED:
2963 ret = io_write(req, nxt, force_nonblock);
2965 case IORING_OP_FSYNC:
2966 ret = io_fsync(req, req->sqe, nxt, force_nonblock);
2968 case IORING_OP_POLL_ADD:
2969 ret = io_poll_add(req, req->sqe, nxt);
2971 case IORING_OP_POLL_REMOVE:
2972 ret = io_poll_remove(req, req->sqe);
2974 case IORING_OP_SYNC_FILE_RANGE:
2975 ret = io_sync_file_range(req, req->sqe, nxt, force_nonblock);
2977 case IORING_OP_SENDMSG:
2978 ret = io_sendmsg(req, req->sqe, nxt, force_nonblock);
2980 case IORING_OP_RECVMSG:
2981 ret = io_recvmsg(req, req->sqe, nxt, force_nonblock);
2983 case IORING_OP_TIMEOUT:
2984 ret = io_timeout(req, req->sqe);
2986 case IORING_OP_TIMEOUT_REMOVE:
2987 ret = io_timeout_remove(req, req->sqe);
2989 case IORING_OP_ACCEPT:
2990 ret = io_accept(req, req->sqe, nxt, force_nonblock);
2992 case IORING_OP_CONNECT:
2993 ret = io_connect(req, req->sqe, nxt, force_nonblock);
2995 case IORING_OP_ASYNC_CANCEL:
2996 ret = io_async_cancel(req, req->sqe, nxt);
3006 if (ctx->flags & IORING_SETUP_IOPOLL) {
3007 if (req->result == -EAGAIN)
3010 io_iopoll_req_issued(req);
3016 static void io_link_work_cb(struct io_wq_work **workptr)
3018 struct io_wq_work *work = *workptr;
3019 struct io_kiocb *link = work->data;
3021 io_queue_linked_timeout(link);
3022 work->func = io_wq_submit_work;
3025 static void io_wq_submit_work(struct io_wq_work **workptr)
3027 struct io_wq_work *work = *workptr;
3028 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
3029 struct io_kiocb *nxt = NULL;
3032 /* Ensure we clear previously set non-block flag */
3033 req->rw.ki_flags &= ~IOCB_NOWAIT;
3035 if (work->flags & IO_WQ_WORK_CANCEL)
3039 req->has_user = (work->flags & IO_WQ_WORK_HAS_MM) != 0;
3040 req->in_async = true;
3042 ret = io_issue_sqe(req, &nxt, false);
3044 * We can get EAGAIN for polled IO even though we're
3045 * forcing a sync submission from here, since we can't
3046 * wait for request slots on the block side.
3054 /* drop submission reference */
3058 req_set_fail_links(req);
3059 io_cqring_add_event(req, ret);
3063 /* if a dependent link is ready, pass it back */
3065 struct io_kiocb *link;
3067 io_prep_async_work(nxt, &link);
3068 *workptr = &nxt->work;
3070 nxt->work.flags |= IO_WQ_WORK_CB;
3071 nxt->work.func = io_link_work_cb;
3072 nxt->work.data = link;
3077 static bool io_req_op_valid(int op)
3079 return op >= IORING_OP_NOP && op < IORING_OP_LAST;
3082 static int io_op_needs_file(const struct io_uring_sqe *sqe)
3084 int op = READ_ONCE(sqe->opcode);
3088 case IORING_OP_POLL_REMOVE:
3089 case IORING_OP_TIMEOUT:
3090 case IORING_OP_TIMEOUT_REMOVE:
3091 case IORING_OP_ASYNC_CANCEL:
3092 case IORING_OP_LINK_TIMEOUT:
3095 if (io_req_op_valid(op))
3101 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
3104 struct fixed_file_table *table;
3106 table = &ctx->file_table[index >> IORING_FILE_TABLE_SHIFT];
3107 return table->files[index & IORING_FILE_TABLE_MASK];
3110 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req)
3112 struct io_ring_ctx *ctx = req->ctx;
3116 flags = READ_ONCE(req->sqe->flags);
3117 fd = READ_ONCE(req->sqe->fd);
3119 if (flags & IOSQE_IO_DRAIN)
3120 req->flags |= REQ_F_IO_DRAIN;
3122 ret = io_op_needs_file(req->sqe);
3126 if (flags & IOSQE_FIXED_FILE) {
3127 if (unlikely(!ctx->file_table ||
3128 (unsigned) fd >= ctx->nr_user_files))
3130 fd = array_index_nospec(fd, ctx->nr_user_files);
3131 req->file = io_file_from_index(ctx, fd);
3134 req->flags |= REQ_F_FIXED_FILE;
3136 if (req->needs_fixed_file)
3138 trace_io_uring_file_get(ctx, fd);
3139 req->file = io_file_get(state, fd);
3140 if (unlikely(!req->file))
3147 static int io_grab_files(struct io_kiocb *req)
3150 struct io_ring_ctx *ctx = req->ctx;
3153 spin_lock_irq(&ctx->inflight_lock);
3155 * We use the f_ops->flush() handler to ensure that we can flush
3156 * out work accessing these files if the fd is closed. Check if
3157 * the fd has changed since we started down this path, and disallow
3158 * this operation if it has.
3160 if (fcheck(req->ring_fd) == req->ring_file) {
3161 list_add(&req->inflight_entry, &ctx->inflight_list);
3162 req->flags |= REQ_F_INFLIGHT;
3163 req->work.files = current->files;
3166 spin_unlock_irq(&ctx->inflight_lock);
3172 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
3174 struct io_timeout_data *data = container_of(timer,
3175 struct io_timeout_data, timer);
3176 struct io_kiocb *req = data->req;
3177 struct io_ring_ctx *ctx = req->ctx;
3178 struct io_kiocb *prev = NULL;
3179 unsigned long flags;
3181 spin_lock_irqsave(&ctx->completion_lock, flags);
3184 * We don't expect the list to be empty, that will only happen if we
3185 * race with the completion of the linked work.
3187 if (!list_empty(&req->link_list)) {
3188 prev = list_entry(req->link_list.prev, struct io_kiocb,
3190 if (refcount_inc_not_zero(&prev->refs)) {
3191 list_del_init(&req->link_list);
3192 prev->flags &= ~REQ_F_LINK_TIMEOUT;
3197 spin_unlock_irqrestore(&ctx->completion_lock, flags);
3200 req_set_fail_links(prev);
3201 io_async_find_and_cancel(ctx, req, prev->user_data, NULL,
3205 io_cqring_add_event(req, -ETIME);
3208 return HRTIMER_NORESTART;
3211 static void io_queue_linked_timeout(struct io_kiocb *req)
3213 struct io_ring_ctx *ctx = req->ctx;
3216 * If the list is now empty, then our linked request finished before
3217 * we got a chance to setup the timer
3219 spin_lock_irq(&ctx->completion_lock);
3220 if (!list_empty(&req->link_list)) {
3221 struct io_timeout_data *data = &req->io->timeout;
3223 data->timer.function = io_link_timeout_fn;
3224 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
3227 spin_unlock_irq(&ctx->completion_lock);
3229 /* drop submission reference */
3233 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
3235 struct io_kiocb *nxt;
3237 if (!(req->flags & REQ_F_LINK))
3240 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
3242 if (!nxt || nxt->sqe->opcode != IORING_OP_LINK_TIMEOUT)
3245 req->flags |= REQ_F_LINK_TIMEOUT;
3249 static void __io_queue_sqe(struct io_kiocb *req)
3251 struct io_kiocb *linked_timeout;
3252 struct io_kiocb *nxt = NULL;
3256 linked_timeout = io_prep_linked_timeout(req);
3258 ret = io_issue_sqe(req, &nxt, true);
3261 * We async punt it if the file wasn't marked NOWAIT, or if the file
3262 * doesn't support non-blocking read/write attempts
3264 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
3265 (req->flags & REQ_F_MUST_PUNT))) {
3266 if (req->work.flags & IO_WQ_WORK_NEEDS_FILES) {
3267 ret = io_grab_files(req);
3273 * Queued up for async execution, worker will release
3274 * submit reference when the iocb is actually submitted.
3276 io_queue_async_work(req);
3281 /* drop submission reference */
3284 if (linked_timeout) {
3286 io_queue_linked_timeout(linked_timeout);
3288 io_put_req(linked_timeout);
3291 /* and drop final reference, if we failed */
3293 io_cqring_add_event(req, ret);
3294 req_set_fail_links(req);
3305 static void io_queue_sqe(struct io_kiocb *req)
3309 if (unlikely(req->ctx->drain_next)) {
3310 req->flags |= REQ_F_IO_DRAIN;
3311 req->ctx->drain_next = false;
3313 req->ctx->drain_next = (req->flags & REQ_F_DRAIN_LINK);
3315 ret = io_req_defer(req);
3317 if (ret != -EIOCBQUEUED) {
3318 io_cqring_add_event(req, ret);
3319 req_set_fail_links(req);
3320 io_double_put_req(req);
3323 __io_queue_sqe(req);
3326 static inline void io_queue_link_head(struct io_kiocb *req)
3328 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
3329 io_cqring_add_event(req, -ECANCELED);
3330 io_double_put_req(req);
3335 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
3338 static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
3339 struct io_kiocb **link)
3341 struct io_ring_ctx *ctx = req->ctx;
3344 req->user_data = req->sqe->user_data;
3346 /* enforce forwards compatibility on users */
3347 if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
3352 ret = io_req_set_file(state, req);
3353 if (unlikely(ret)) {
3355 io_cqring_add_event(req, ret);
3356 io_double_put_req(req);
3361 * If we already have a head request, queue this one for async
3362 * submittal once the head completes. If we don't have a head but
3363 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
3364 * submitted sync once the chain is complete. If none of those
3365 * conditions are true (normal request), then just queue it.
3368 struct io_kiocb *prev = *link;
3369 struct io_async_ctx *io;
3371 if (req->sqe->flags & IOSQE_IO_DRAIN)
3372 (*link)->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
3374 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3375 req->flags |= REQ_F_HARDLINK;
3377 io = kmalloc(sizeof(*io), GFP_KERNEL);
3383 ret = io_req_defer_prep(req, io);
3386 /* fail even hard links since we don't submit */
3387 prev->flags |= REQ_F_FAIL_LINK;
3390 trace_io_uring_link(ctx, req, prev);
3391 list_add_tail(&req->link_list, &prev->link_list);
3392 } else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
3393 req->flags |= REQ_F_LINK;
3394 if (req->sqe->flags & IOSQE_IO_HARDLINK)
3395 req->flags |= REQ_F_HARDLINK;
3397 INIT_LIST_HEAD(&req->link_list);
3407 * Batched submission is done, ensure local IO is flushed out.
3409 static void io_submit_state_end(struct io_submit_state *state)
3411 blk_finish_plug(&state->plug);
3413 if (state->free_reqs)
3414 kmem_cache_free_bulk(req_cachep, state->free_reqs,
3415 &state->reqs[state->cur_req]);
3419 * Start submission side cache.
3421 static void io_submit_state_start(struct io_submit_state *state,
3422 unsigned int max_ios)
3424 blk_start_plug(&state->plug);
3425 state->free_reqs = 0;
3427 state->ios_left = max_ios;
3430 static void io_commit_sqring(struct io_ring_ctx *ctx)
3432 struct io_rings *rings = ctx->rings;
3434 if (ctx->cached_sq_head != READ_ONCE(rings->sq.head)) {
3436 * Ensure any loads from the SQEs are done at this point,
3437 * since once we write the new head, the application could
3438 * write new data to them.
3440 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
3445 * Fetch an sqe, if one is available. Note that req->sqe will point to memory
3446 * that is mapped by userspace. This means that care needs to be taken to
3447 * ensure that reads are stable, as we cannot rely on userspace always
3448 * being a good citizen. If members of the sqe are validated and then later
3449 * used, it's important that those reads are done through READ_ONCE() to
3450 * prevent a re-load down the line.
3452 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req)
3454 struct io_rings *rings = ctx->rings;
3455 u32 *sq_array = ctx->sq_array;
3459 * The cached sq head (or cq tail) serves two purposes:
3461 * 1) allows us to batch the cost of updating the user visible
3463 * 2) allows the kernel side to track the head on its own, even
3464 * though the application is the one updating it.
3466 head = ctx->cached_sq_head;
3467 /* make sure SQ entry isn't read before tail */
3468 if (unlikely(head == smp_load_acquire(&rings->sq.tail)))
3471 head = READ_ONCE(sq_array[head & ctx->sq_mask]);
3472 if (likely(head < ctx->sq_entries)) {
3474 * All io need record the previous position, if LINK vs DARIN,
3475 * it can be used to mark the position of the first IO in the
3478 req->sequence = ctx->cached_sq_head;
3479 req->sqe = &ctx->sq_sqes[head];
3480 ctx->cached_sq_head++;
3484 /* drop invalid entries */
3485 ctx->cached_sq_head++;
3486 ctx->cached_sq_dropped++;
3487 WRITE_ONCE(rings->sq_dropped, ctx->cached_sq_dropped);
3491 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
3492 struct file *ring_file, int ring_fd,
3493 struct mm_struct **mm, bool async)
3495 struct io_submit_state state, *statep = NULL;
3496 struct io_kiocb *link = NULL;
3497 int i, submitted = 0;
3498 bool mm_fault = false;
3500 /* if we have a backlog and couldn't flush it all, return BUSY */
3501 if (!list_empty(&ctx->cq_overflow_list) &&
3502 !io_cqring_overflow_flush(ctx, false))
3505 if (nr > IO_PLUG_THRESHOLD) {
3506 io_submit_state_start(&state, nr);
3510 for (i = 0; i < nr; i++) {
3511 struct io_kiocb *req;
3512 unsigned int sqe_flags;
3514 req = io_get_req(ctx, statep);
3515 if (unlikely(!req)) {
3517 submitted = -EAGAIN;
3520 if (!io_get_sqring(ctx, req)) {
3525 if (io_sqe_needs_user(req->sqe) && !*mm) {
3526 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
3528 use_mm(ctx->sqo_mm);
3534 sqe_flags = req->sqe->flags;
3536 req->ring_file = ring_file;
3537 req->ring_fd = ring_fd;
3538 req->has_user = *mm != NULL;
3539 req->in_async = async;
3540 req->needs_fixed_file = async;
3541 trace_io_uring_submit_sqe(ctx, req->sqe->user_data,
3543 if (!io_submit_sqe(req, statep, &link))
3546 * If previous wasn't linked and we have a linked command,
3547 * that's the end of the chain. Submit the previous link.
3549 if (!(sqe_flags & IOSQE_IO_LINK) && link) {
3550 io_queue_link_head(link);
3556 io_queue_link_head(link);
3558 io_submit_state_end(&state);
3560 /* Commit SQ ring head once we've consumed and submitted all SQEs */
3561 io_commit_sqring(ctx);
3566 static int io_sq_thread(void *data)
3568 struct io_ring_ctx *ctx = data;
3569 struct mm_struct *cur_mm = NULL;
3570 const struct cred *old_cred;
3571 mm_segment_t old_fs;
3574 unsigned long timeout;
3577 complete(&ctx->completions[1]);
3581 old_cred = override_creds(ctx->creds);
3583 ret = timeout = inflight = 0;
3584 while (!kthread_should_park()) {
3585 unsigned int to_submit;
3588 unsigned nr_events = 0;
3590 if (ctx->flags & IORING_SETUP_IOPOLL) {
3592 * inflight is the count of the maximum possible
3593 * entries we submitted, but it can be smaller
3594 * if we dropped some of them. If we don't have
3595 * poll entries available, then we know that we
3596 * have nothing left to poll for. Reset the
3597 * inflight count to zero in that case.
3599 mutex_lock(&ctx->uring_lock);
3600 if (!list_empty(&ctx->poll_list))
3601 __io_iopoll_check(ctx, &nr_events, 0);
3604 mutex_unlock(&ctx->uring_lock);
3607 * Normal IO, just pretend everything completed.
3608 * We don't have to poll completions for that.
3610 nr_events = inflight;
3613 inflight -= nr_events;
3615 timeout = jiffies + ctx->sq_thread_idle;
3618 to_submit = io_sqring_entries(ctx);
3621 * If submit got -EBUSY, flag us as needing the application
3622 * to enter the kernel to reap and flush events.
3624 if (!to_submit || ret == -EBUSY) {
3626 * We're polling. If we're within the defined idle
3627 * period, then let us spin without work before going
3628 * to sleep. The exception is if we got EBUSY doing
3629 * more IO, we should wait for the application to
3630 * reap events and wake us up.
3633 (!time_after(jiffies, timeout) && ret != -EBUSY)) {
3639 * Drop cur_mm before scheduling, we can't hold it for
3640 * long periods (or over schedule()). Do this before
3641 * adding ourselves to the waitqueue, as the unuse/drop
3650 prepare_to_wait(&ctx->sqo_wait, &wait,
3651 TASK_INTERRUPTIBLE);
3653 /* Tell userspace we may need a wakeup call */
3654 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
3655 /* make sure to read SQ tail after writing flags */
3658 to_submit = io_sqring_entries(ctx);
3659 if (!to_submit || ret == -EBUSY) {
3660 if (kthread_should_park()) {
3661 finish_wait(&ctx->sqo_wait, &wait);
3664 if (signal_pending(current))
3665 flush_signals(current);
3667 finish_wait(&ctx->sqo_wait, &wait);
3669 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3672 finish_wait(&ctx->sqo_wait, &wait);
3674 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
3677 to_submit = min(to_submit, ctx->sq_entries);
3678 mutex_lock(&ctx->uring_lock);
3679 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
3680 mutex_unlock(&ctx->uring_lock);
3690 revert_creds(old_cred);
3697 struct io_wait_queue {
3698 struct wait_queue_entry wq;
3699 struct io_ring_ctx *ctx;
3701 unsigned nr_timeouts;
3704 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
3706 struct io_ring_ctx *ctx = iowq->ctx;
3709 * Wake up if we have enough events, or if a timeout occurred since we
3710 * started waiting. For timeouts, we always want to return to userspace,
3711 * regardless of event count.
3713 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
3714 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
3717 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
3718 int wake_flags, void *key)
3720 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
3723 /* use noflush == true, as we can't safely rely on locking context */
3724 if (!io_should_wake(iowq, true))
3727 return autoremove_wake_function(curr, mode, wake_flags, key);
3731 * Wait until events become available, if we don't already have some. The
3732 * application must reap them itself, as they reside on the shared cq ring.
3734 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
3735 const sigset_t __user *sig, size_t sigsz)
3737 struct io_wait_queue iowq = {
3740 .func = io_wake_function,
3741 .entry = LIST_HEAD_INIT(iowq.wq.entry),
3744 .to_wait = min_events,
3746 struct io_rings *rings = ctx->rings;
3749 if (io_cqring_events(ctx, false) >= min_events)
3753 #ifdef CONFIG_COMPAT
3754 if (in_compat_syscall())
3755 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
3759 ret = set_user_sigmask(sig, sigsz);
3765 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
3766 trace_io_uring_cqring_wait(ctx, min_events);
3768 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
3769 TASK_INTERRUPTIBLE);
3770 if (io_should_wake(&iowq, false))
3773 if (signal_pending(current)) {
3778 finish_wait(&ctx->wait, &iowq.wq);
3780 restore_saved_sigmask_unless(ret == -EINTR);
3782 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
3785 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
3787 #if defined(CONFIG_UNIX)
3788 if (ctx->ring_sock) {
3789 struct sock *sock = ctx->ring_sock->sk;
3790 struct sk_buff *skb;
3792 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
3798 for (i = 0; i < ctx->nr_user_files; i++) {
3801 file = io_file_from_index(ctx, i);
3808 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
3810 unsigned nr_tables, i;
3812 if (!ctx->file_table)
3815 __io_sqe_files_unregister(ctx);
3816 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
3817 for (i = 0; i < nr_tables; i++)
3818 kfree(ctx->file_table[i].files);
3819 kfree(ctx->file_table);
3820 ctx->file_table = NULL;
3821 ctx->nr_user_files = 0;
3825 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
3827 if (ctx->sqo_thread) {
3828 wait_for_completion(&ctx->completions[1]);
3830 * The park is a bit of a work-around, without it we get
3831 * warning spews on shutdown with SQPOLL set and affinity
3832 * set to a single CPU.
3834 kthread_park(ctx->sqo_thread);
3835 kthread_stop(ctx->sqo_thread);
3836 ctx->sqo_thread = NULL;
3840 static void io_finish_async(struct io_ring_ctx *ctx)
3842 io_sq_thread_stop(ctx);
3845 io_wq_destroy(ctx->io_wq);
3850 #if defined(CONFIG_UNIX)
3851 static void io_destruct_skb(struct sk_buff *skb)
3853 struct io_ring_ctx *ctx = skb->sk->sk_user_data;
3856 io_wq_flush(ctx->io_wq);
3858 unix_destruct_scm(skb);
3862 * Ensure the UNIX gc is aware of our file set, so we are certain that
3863 * the io_uring can be safely unregistered on process exit, even if we have
3864 * loops in the file referencing.
3866 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
3868 struct sock *sk = ctx->ring_sock->sk;
3869 struct scm_fp_list *fpl;
3870 struct sk_buff *skb;
3873 if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) {
3874 unsigned long inflight = ctx->user->unix_inflight + nr;
3876 if (inflight > task_rlimit(current, RLIMIT_NOFILE))
3880 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
3884 skb = alloc_skb(0, GFP_KERNEL);
3893 fpl->user = get_uid(ctx->user);
3894 for (i = 0; i < nr; i++) {
3895 struct file *file = io_file_from_index(ctx, i + offset);
3899 fpl->fp[nr_files] = get_file(file);
3900 unix_inflight(fpl->user, fpl->fp[nr_files]);
3905 fpl->max = SCM_MAX_FD;
3906 fpl->count = nr_files;
3907 UNIXCB(skb).fp = fpl;
3908 skb->destructor = io_destruct_skb;
3909 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
3910 skb_queue_head(&sk->sk_receive_queue, skb);
3912 for (i = 0; i < nr_files; i++)
3923 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
3924 * causes regular reference counting to break down. We rely on the UNIX
3925 * garbage collection to take care of this problem for us.
3927 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3929 unsigned left, total;
3933 left = ctx->nr_user_files;
3935 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
3937 ret = __io_sqe_files_scm(ctx, this_files, total);
3941 total += this_files;
3947 while (total < ctx->nr_user_files) {
3948 struct file *file = io_file_from_index(ctx, total);
3958 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
3964 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
3969 for (i = 0; i < nr_tables; i++) {
3970 struct fixed_file_table *table = &ctx->file_table[i];
3971 unsigned this_files;
3973 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
3974 table->files = kcalloc(this_files, sizeof(struct file *),
3978 nr_files -= this_files;
3984 for (i = 0; i < nr_tables; i++) {
3985 struct fixed_file_table *table = &ctx->file_table[i];
3986 kfree(table->files);
3991 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
3994 __s32 __user *fds = (__s32 __user *) arg;
3999 if (ctx->file_table)
4003 if (nr_args > IORING_MAX_FIXED_FILES)
4006 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
4007 ctx->file_table = kcalloc(nr_tables, sizeof(struct fixed_file_table),
4009 if (!ctx->file_table)
4012 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
4013 kfree(ctx->file_table);
4014 ctx->file_table = NULL;
4018 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
4019 struct fixed_file_table *table;
4023 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
4025 /* allow sparse sets */
4031 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4032 index = i & IORING_FILE_TABLE_MASK;
4033 table->files[index] = fget(fd);
4036 if (!table->files[index])
4039 * Don't allow io_uring instances to be registered. If UNIX
4040 * isn't enabled, then this causes a reference cycle and this
4041 * instance can never get freed. If UNIX is enabled we'll
4042 * handle it just fine, but there's still no point in allowing
4043 * a ring fd as it doesn't support regular read/write anyway.
4045 if (table->files[index]->f_op == &io_uring_fops) {
4046 fput(table->files[index]);
4053 for (i = 0; i < ctx->nr_user_files; i++) {
4056 file = io_file_from_index(ctx, i);
4060 for (i = 0; i < nr_tables; i++)
4061 kfree(ctx->file_table[i].files);
4063 kfree(ctx->file_table);
4064 ctx->file_table = NULL;
4065 ctx->nr_user_files = 0;
4069 ret = io_sqe_files_scm(ctx);
4071 io_sqe_files_unregister(ctx);
4076 static void io_sqe_file_unregister(struct io_ring_ctx *ctx, int index)
4078 #if defined(CONFIG_UNIX)
4079 struct file *file = io_file_from_index(ctx, index);
4080 struct sock *sock = ctx->ring_sock->sk;
4081 struct sk_buff_head list, *head = &sock->sk_receive_queue;
4082 struct sk_buff *skb;
4085 __skb_queue_head_init(&list);
4088 * Find the skb that holds this file in its SCM_RIGHTS. When found,
4089 * remove this entry and rearrange the file array.
4091 skb = skb_dequeue(head);
4093 struct scm_fp_list *fp;
4095 fp = UNIXCB(skb).fp;
4096 for (i = 0; i < fp->count; i++) {
4099 if (fp->fp[i] != file)
4102 unix_notinflight(fp->user, fp->fp[i]);
4103 left = fp->count - 1 - i;
4105 memmove(&fp->fp[i], &fp->fp[i + 1],
4106 left * sizeof(struct file *));
4113 __skb_queue_tail(&list, skb);
4123 __skb_queue_tail(&list, skb);
4125 skb = skb_dequeue(head);
4128 if (skb_peek(&list)) {
4129 spin_lock_irq(&head->lock);
4130 while ((skb = __skb_dequeue(&list)) != NULL)
4131 __skb_queue_tail(head, skb);
4132 spin_unlock_irq(&head->lock);
4135 fput(io_file_from_index(ctx, index));
4139 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
4142 #if defined(CONFIG_UNIX)
4143 struct sock *sock = ctx->ring_sock->sk;
4144 struct sk_buff_head *head = &sock->sk_receive_queue;
4145 struct sk_buff *skb;
4148 * See if we can merge this file into an existing skb SCM_RIGHTS
4149 * file set. If there's no room, fall back to allocating a new skb
4150 * and filling it in.
4152 spin_lock_irq(&head->lock);
4153 skb = skb_peek(head);
4155 struct scm_fp_list *fpl = UNIXCB(skb).fp;
4157 if (fpl->count < SCM_MAX_FD) {
4158 __skb_unlink(skb, head);
4159 spin_unlock_irq(&head->lock);
4160 fpl->fp[fpl->count] = get_file(file);
4161 unix_inflight(fpl->user, fpl->fp[fpl->count]);
4163 spin_lock_irq(&head->lock);
4164 __skb_queue_head(head, skb);
4169 spin_unlock_irq(&head->lock);
4176 return __io_sqe_files_scm(ctx, 1, index);
4182 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
4185 struct io_uring_files_update up;
4190 if (!ctx->file_table)
4194 if (copy_from_user(&up, arg, sizeof(up)))
4196 if (check_add_overflow(up.offset, nr_args, &done))
4198 if (done > ctx->nr_user_files)
4202 fds = (__s32 __user *) up.fds;
4204 struct fixed_file_table *table;
4208 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
4212 i = array_index_nospec(up.offset, ctx->nr_user_files);
4213 table = &ctx->file_table[i >> IORING_FILE_TABLE_SHIFT];
4214 index = i & IORING_FILE_TABLE_MASK;
4215 if (table->files[index]) {
4216 io_sqe_file_unregister(ctx, i);
4217 table->files[index] = NULL;
4228 * Don't allow io_uring instances to be registered. If
4229 * UNIX isn't enabled, then this causes a reference
4230 * cycle and this instance can never get freed. If UNIX
4231 * is enabled we'll handle it just fine, but there's
4232 * still no point in allowing a ring fd as it doesn't
4233 * support regular read/write anyway.
4235 if (file->f_op == &io_uring_fops) {
4240 table->files[index] = file;
4241 err = io_sqe_file_register(ctx, file, i);
4250 return done ? done : err;
4253 static void io_put_work(struct io_wq_work *work)
4255 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4260 static void io_get_work(struct io_wq_work *work)
4262 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4264 refcount_inc(&req->refs);
4267 static int io_sq_offload_start(struct io_ring_ctx *ctx,
4268 struct io_uring_params *p)
4270 struct io_wq_data data;
4271 unsigned concurrency;
4274 init_waitqueue_head(&ctx->sqo_wait);
4275 mmgrab(current->mm);
4276 ctx->sqo_mm = current->mm;
4278 if (ctx->flags & IORING_SETUP_SQPOLL) {
4280 if (!capable(CAP_SYS_ADMIN))
4283 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
4284 if (!ctx->sq_thread_idle)
4285 ctx->sq_thread_idle = HZ;
4287 if (p->flags & IORING_SETUP_SQ_AFF) {
4288 int cpu = p->sq_thread_cpu;
4291 if (cpu >= nr_cpu_ids)
4293 if (!cpu_online(cpu))
4296 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
4300 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
4303 if (IS_ERR(ctx->sqo_thread)) {
4304 ret = PTR_ERR(ctx->sqo_thread);
4305 ctx->sqo_thread = NULL;
4308 wake_up_process(ctx->sqo_thread);
4309 } else if (p->flags & IORING_SETUP_SQ_AFF) {
4310 /* Can't have SQ_AFF without SQPOLL */
4315 data.mm = ctx->sqo_mm;
4316 data.user = ctx->user;
4317 data.creds = ctx->creds;
4318 data.get_work = io_get_work;
4319 data.put_work = io_put_work;
4321 /* Do QD, or 4 * CPUS, whatever is smallest */
4322 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
4323 ctx->io_wq = io_wq_create(concurrency, &data);
4324 if (IS_ERR(ctx->io_wq)) {
4325 ret = PTR_ERR(ctx->io_wq);
4332 io_finish_async(ctx);
4333 mmdrop(ctx->sqo_mm);
4338 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
4340 atomic_long_sub(nr_pages, &user->locked_vm);
4343 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
4345 unsigned long page_limit, cur_pages, new_pages;
4347 /* Don't allow more pages than we can safely lock */
4348 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
4351 cur_pages = atomic_long_read(&user->locked_vm);
4352 new_pages = cur_pages + nr_pages;
4353 if (new_pages > page_limit)
4355 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
4356 new_pages) != cur_pages);
4361 static void io_mem_free(void *ptr)
4368 page = virt_to_head_page(ptr);
4369 if (put_page_testzero(page))
4370 free_compound_page(page);
4373 static void *io_mem_alloc(size_t size)
4375 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
4378 return (void *) __get_free_pages(gfp_flags, get_order(size));
4381 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
4384 struct io_rings *rings;
4385 size_t off, sq_array_size;
4387 off = struct_size(rings, cqes, cq_entries);
4388 if (off == SIZE_MAX)
4392 off = ALIGN(off, SMP_CACHE_BYTES);
4397 sq_array_size = array_size(sizeof(u32), sq_entries);
4398 if (sq_array_size == SIZE_MAX)
4401 if (check_add_overflow(off, sq_array_size, &off))
4410 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
4414 pages = (size_t)1 << get_order(
4415 rings_size(sq_entries, cq_entries, NULL));
4416 pages += (size_t)1 << get_order(
4417 array_size(sizeof(struct io_uring_sqe), sq_entries));
4422 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
4426 if (!ctx->user_bufs)
4429 for (i = 0; i < ctx->nr_user_bufs; i++) {
4430 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4432 for (j = 0; j < imu->nr_bvecs; j++)
4433 put_user_page(imu->bvec[j].bv_page);
4435 if (ctx->account_mem)
4436 io_unaccount_mem(ctx->user, imu->nr_bvecs);
4441 kfree(ctx->user_bufs);
4442 ctx->user_bufs = NULL;
4443 ctx->nr_user_bufs = 0;
4447 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
4448 void __user *arg, unsigned index)
4450 struct iovec __user *src;
4452 #ifdef CONFIG_COMPAT
4454 struct compat_iovec __user *ciovs;
4455 struct compat_iovec ciov;
4457 ciovs = (struct compat_iovec __user *) arg;
4458 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
4461 dst->iov_base = (void __user *) (unsigned long) ciov.iov_base;
4462 dst->iov_len = ciov.iov_len;
4466 src = (struct iovec __user *) arg;
4467 if (copy_from_user(dst, &src[index], sizeof(*dst)))
4472 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
4475 struct vm_area_struct **vmas = NULL;
4476 struct page **pages = NULL;
4477 int i, j, got_pages = 0;
4482 if (!nr_args || nr_args > UIO_MAXIOV)
4485 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
4487 if (!ctx->user_bufs)
4490 for (i = 0; i < nr_args; i++) {
4491 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
4492 unsigned long off, start, end, ubuf;
4497 ret = io_copy_iov(ctx, &iov, arg, i);
4502 * Don't impose further limits on the size and buffer
4503 * constraints here, we'll -EINVAL later when IO is
4504 * submitted if they are wrong.
4507 if (!iov.iov_base || !iov.iov_len)
4510 /* arbitrary limit, but we need something */
4511 if (iov.iov_len > SZ_1G)
4514 ubuf = (unsigned long) iov.iov_base;
4515 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
4516 start = ubuf >> PAGE_SHIFT;
4517 nr_pages = end - start;
4519 if (ctx->account_mem) {
4520 ret = io_account_mem(ctx->user, nr_pages);
4526 if (!pages || nr_pages > got_pages) {
4529 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
4531 vmas = kvmalloc_array(nr_pages,
4532 sizeof(struct vm_area_struct *),
4534 if (!pages || !vmas) {
4536 if (ctx->account_mem)
4537 io_unaccount_mem(ctx->user, nr_pages);
4540 got_pages = nr_pages;
4543 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
4547 if (ctx->account_mem)
4548 io_unaccount_mem(ctx->user, nr_pages);
4553 down_read(¤t->mm->mmap_sem);
4554 pret = get_user_pages(ubuf, nr_pages,
4555 FOLL_WRITE | FOLL_LONGTERM,
4557 if (pret == nr_pages) {
4558 /* don't support file backed memory */
4559 for (j = 0; j < nr_pages; j++) {
4560 struct vm_area_struct *vma = vmas[j];
4563 !is_file_hugepages(vma->vm_file)) {
4569 ret = pret < 0 ? pret : -EFAULT;
4571 up_read(¤t->mm->mmap_sem);
4574 * if we did partial map, or found file backed vmas,
4575 * release any pages we did get
4578 put_user_pages(pages, pret);
4579 if (ctx->account_mem)
4580 io_unaccount_mem(ctx->user, nr_pages);
4585 off = ubuf & ~PAGE_MASK;
4587 for (j = 0; j < nr_pages; j++) {
4590 vec_len = min_t(size_t, size, PAGE_SIZE - off);
4591 imu->bvec[j].bv_page = pages[j];
4592 imu->bvec[j].bv_len = vec_len;
4593 imu->bvec[j].bv_offset = off;
4597 /* store original address for later verification */
4599 imu->len = iov.iov_len;
4600 imu->nr_bvecs = nr_pages;
4602 ctx->nr_user_bufs++;
4610 io_sqe_buffer_unregister(ctx);
4614 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
4616 __s32 __user *fds = arg;
4622 if (copy_from_user(&fd, fds, sizeof(*fds)))
4625 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
4626 if (IS_ERR(ctx->cq_ev_fd)) {
4627 int ret = PTR_ERR(ctx->cq_ev_fd);
4628 ctx->cq_ev_fd = NULL;
4635 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
4637 if (ctx->cq_ev_fd) {
4638 eventfd_ctx_put(ctx->cq_ev_fd);
4639 ctx->cq_ev_fd = NULL;
4646 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
4648 io_finish_async(ctx);
4650 mmdrop(ctx->sqo_mm);
4652 io_iopoll_reap_events(ctx);
4653 io_sqe_buffer_unregister(ctx);
4654 io_sqe_files_unregister(ctx);
4655 io_eventfd_unregister(ctx);
4657 #if defined(CONFIG_UNIX)
4658 if (ctx->ring_sock) {
4659 ctx->ring_sock->file = NULL; /* so that iput() is called */
4660 sock_release(ctx->ring_sock);
4664 io_mem_free(ctx->rings);
4665 io_mem_free(ctx->sq_sqes);
4667 percpu_ref_exit(&ctx->refs);
4668 if (ctx->account_mem)
4669 io_unaccount_mem(ctx->user,
4670 ring_pages(ctx->sq_entries, ctx->cq_entries));
4671 free_uid(ctx->user);
4672 put_cred(ctx->creds);
4673 kfree(ctx->completions);
4674 kfree(ctx->cancel_hash);
4675 kmem_cache_free(req_cachep, ctx->fallback_req);
4679 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
4681 struct io_ring_ctx *ctx = file->private_data;
4684 poll_wait(file, &ctx->cq_wait, wait);
4686 * synchronizes with barrier from wq_has_sleeper call in
4690 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
4691 ctx->rings->sq_ring_entries)
4692 mask |= EPOLLOUT | EPOLLWRNORM;
4693 if (READ_ONCE(ctx->rings->cq.head) != ctx->cached_cq_tail)
4694 mask |= EPOLLIN | EPOLLRDNORM;
4699 static int io_uring_fasync(int fd, struct file *file, int on)
4701 struct io_ring_ctx *ctx = file->private_data;
4703 return fasync_helper(fd, file, on, &ctx->cq_fasync);
4706 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
4708 mutex_lock(&ctx->uring_lock);
4709 percpu_ref_kill(&ctx->refs);
4710 mutex_unlock(&ctx->uring_lock);
4712 io_kill_timeouts(ctx);
4713 io_poll_remove_all(ctx);
4716 io_wq_cancel_all(ctx->io_wq);
4718 io_iopoll_reap_events(ctx);
4719 /* if we failed setting up the ctx, we might not have any rings */
4721 io_cqring_overflow_flush(ctx, true);
4722 wait_for_completion(&ctx->completions[0]);
4723 io_ring_ctx_free(ctx);
4726 static int io_uring_release(struct inode *inode, struct file *file)
4728 struct io_ring_ctx *ctx = file->private_data;
4730 file->private_data = NULL;
4731 io_ring_ctx_wait_and_kill(ctx);
4735 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
4736 struct files_struct *files)
4738 struct io_kiocb *req;
4741 while (!list_empty_careful(&ctx->inflight_list)) {
4742 struct io_kiocb *cancel_req = NULL;
4744 spin_lock_irq(&ctx->inflight_lock);
4745 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
4746 if (req->work.files != files)
4748 /* req is being completed, ignore */
4749 if (!refcount_inc_not_zero(&req->refs))
4755 prepare_to_wait(&ctx->inflight_wait, &wait,
4756 TASK_UNINTERRUPTIBLE);
4757 spin_unlock_irq(&ctx->inflight_lock);
4759 /* We need to keep going until we don't find a matching req */
4763 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
4764 io_put_req(cancel_req);
4767 finish_wait(&ctx->inflight_wait, &wait);
4770 static int io_uring_flush(struct file *file, void *data)
4772 struct io_ring_ctx *ctx = file->private_data;
4774 io_uring_cancel_files(ctx, data);
4775 if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) {
4776 io_cqring_overflow_flush(ctx, true);
4777 io_wq_cancel_all(ctx->io_wq);
4782 static void *io_uring_validate_mmap_request(struct file *file,
4783 loff_t pgoff, size_t sz)
4785 struct io_ring_ctx *ctx = file->private_data;
4786 loff_t offset = pgoff << PAGE_SHIFT;
4791 case IORING_OFF_SQ_RING:
4792 case IORING_OFF_CQ_RING:
4795 case IORING_OFF_SQES:
4799 return ERR_PTR(-EINVAL);
4802 page = virt_to_head_page(ptr);
4803 if (sz > page_size(page))
4804 return ERR_PTR(-EINVAL);
4811 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4813 size_t sz = vma->vm_end - vma->vm_start;
4817 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
4819 return PTR_ERR(ptr);
4821 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
4822 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
4825 #else /* !CONFIG_MMU */
4827 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
4829 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
4832 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
4834 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
4837 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
4838 unsigned long addr, unsigned long len,
4839 unsigned long pgoff, unsigned long flags)
4843 ptr = io_uring_validate_mmap_request(file, pgoff, len);
4845 return PTR_ERR(ptr);
4847 return (unsigned long) ptr;
4850 #endif /* !CONFIG_MMU */
4852 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
4853 u32, min_complete, u32, flags, const sigset_t __user *, sig,
4856 struct io_ring_ctx *ctx;
4861 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
4869 if (f.file->f_op != &io_uring_fops)
4873 ctx = f.file->private_data;
4874 if (!percpu_ref_tryget(&ctx->refs))
4878 * For SQ polling, the thread will do all submissions and completions.
4879 * Just return the requested submit count, and wake the thread if
4883 if (ctx->flags & IORING_SETUP_SQPOLL) {
4884 if (!list_empty_careful(&ctx->cq_overflow_list))
4885 io_cqring_overflow_flush(ctx, false);
4886 if (flags & IORING_ENTER_SQ_WAKEUP)
4887 wake_up(&ctx->sqo_wait);
4888 submitted = to_submit;
4889 } else if (to_submit) {
4890 struct mm_struct *cur_mm;
4892 to_submit = min(to_submit, ctx->sq_entries);
4893 mutex_lock(&ctx->uring_lock);
4894 /* already have mm, so io_submit_sqes() won't try to grab it */
4895 cur_mm = ctx->sqo_mm;
4896 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
4898 mutex_unlock(&ctx->uring_lock);
4900 if (flags & IORING_ENTER_GETEVENTS) {
4901 unsigned nr_events = 0;
4903 min_complete = min(min_complete, ctx->cq_entries);
4905 if (ctx->flags & IORING_SETUP_IOPOLL) {
4906 ret = io_iopoll_check(ctx, &nr_events, min_complete);
4908 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
4912 percpu_ref_put(&ctx->refs);
4915 return submitted ? submitted : ret;
4918 static const struct file_operations io_uring_fops = {
4919 .release = io_uring_release,
4920 .flush = io_uring_flush,
4921 .mmap = io_uring_mmap,
4923 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
4924 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
4926 .poll = io_uring_poll,
4927 .fasync = io_uring_fasync,
4930 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
4931 struct io_uring_params *p)
4933 struct io_rings *rings;
4934 size_t size, sq_array_offset;
4936 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
4937 if (size == SIZE_MAX)
4940 rings = io_mem_alloc(size);
4945 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
4946 rings->sq_ring_mask = p->sq_entries - 1;
4947 rings->cq_ring_mask = p->cq_entries - 1;
4948 rings->sq_ring_entries = p->sq_entries;
4949 rings->cq_ring_entries = p->cq_entries;
4950 ctx->sq_mask = rings->sq_ring_mask;
4951 ctx->cq_mask = rings->cq_ring_mask;
4952 ctx->sq_entries = rings->sq_ring_entries;
4953 ctx->cq_entries = rings->cq_ring_entries;
4955 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
4956 if (size == SIZE_MAX) {
4957 io_mem_free(ctx->rings);
4962 ctx->sq_sqes = io_mem_alloc(size);
4963 if (!ctx->sq_sqes) {
4964 io_mem_free(ctx->rings);
4973 * Allocate an anonymous fd, this is what constitutes the application
4974 * visible backing of an io_uring instance. The application mmaps this
4975 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
4976 * we have to tie this fd to a socket for file garbage collection purposes.
4978 static int io_uring_get_fd(struct io_ring_ctx *ctx)
4983 #if defined(CONFIG_UNIX)
4984 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
4990 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
4994 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
4995 O_RDWR | O_CLOEXEC);
4998 ret = PTR_ERR(file);
5002 #if defined(CONFIG_UNIX)
5003 ctx->ring_sock->file = file;
5004 ctx->ring_sock->sk->sk_user_data = ctx;
5006 fd_install(ret, file);
5009 #if defined(CONFIG_UNIX)
5010 sock_release(ctx->ring_sock);
5011 ctx->ring_sock = NULL;
5016 static int io_uring_create(unsigned entries, struct io_uring_params *p)
5018 struct user_struct *user = NULL;
5019 struct io_ring_ctx *ctx;
5023 if (!entries || entries > IORING_MAX_ENTRIES)
5027 * Use twice as many entries for the CQ ring. It's possible for the
5028 * application to drive a higher depth than the size of the SQ ring,
5029 * since the sqes are only used at submission time. This allows for
5030 * some flexibility in overcommitting a bit. If the application has
5031 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
5032 * of CQ ring entries manually.
5034 p->sq_entries = roundup_pow_of_two(entries);
5035 if (p->flags & IORING_SETUP_CQSIZE) {
5037 * If IORING_SETUP_CQSIZE is set, we do the same roundup
5038 * to a power-of-two, if it isn't already. We do NOT impose
5039 * any cq vs sq ring sizing.
5041 if (p->cq_entries < p->sq_entries || p->cq_entries > IORING_MAX_CQ_ENTRIES)
5043 p->cq_entries = roundup_pow_of_two(p->cq_entries);
5045 p->cq_entries = 2 * p->sq_entries;
5048 user = get_uid(current_user());
5049 account_mem = !capable(CAP_IPC_LOCK);
5052 ret = io_account_mem(user,
5053 ring_pages(p->sq_entries, p->cq_entries));
5060 ctx = io_ring_ctx_alloc(p);
5063 io_unaccount_mem(user, ring_pages(p->sq_entries,
5068 ctx->compat = in_compat_syscall();
5069 ctx->account_mem = account_mem;
5071 ctx->creds = get_current_cred();
5073 ret = io_allocate_scq_urings(ctx, p);
5077 ret = io_sq_offload_start(ctx, p);
5081 memset(&p->sq_off, 0, sizeof(p->sq_off));
5082 p->sq_off.head = offsetof(struct io_rings, sq.head);
5083 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
5084 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
5085 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
5086 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
5087 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
5088 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
5090 memset(&p->cq_off, 0, sizeof(p->cq_off));
5091 p->cq_off.head = offsetof(struct io_rings, cq.head);
5092 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
5093 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
5094 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
5095 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
5096 p->cq_off.cqes = offsetof(struct io_rings, cqes);
5099 * Install ring fd as the very last thing, so we don't risk someone
5100 * having closed it before we finish setup
5102 ret = io_uring_get_fd(ctx);
5106 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
5107 IORING_FEAT_SUBMIT_STABLE;
5108 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
5111 io_ring_ctx_wait_and_kill(ctx);
5116 * Sets up an aio uring context, and returns the fd. Applications asks for a
5117 * ring size, we return the actual sq/cq ring sizes (among other things) in the
5118 * params structure passed in.
5120 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
5122 struct io_uring_params p;
5126 if (copy_from_user(&p, params, sizeof(p)))
5128 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
5133 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
5134 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE))
5137 ret = io_uring_create(entries, &p);
5141 if (copy_to_user(params, &p, sizeof(p)))
5147 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
5148 struct io_uring_params __user *, params)
5150 return io_uring_setup(entries, params);
5153 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
5154 void __user *arg, unsigned nr_args)
5155 __releases(ctx->uring_lock)
5156 __acquires(ctx->uring_lock)
5161 * We're inside the ring mutex, if the ref is already dying, then
5162 * someone else killed the ctx or is already going through
5163 * io_uring_register().
5165 if (percpu_ref_is_dying(&ctx->refs))
5168 percpu_ref_kill(&ctx->refs);
5171 * Drop uring mutex before waiting for references to exit. If another
5172 * thread is currently inside io_uring_enter() it might need to grab
5173 * the uring_lock to make progress. If we hold it here across the drain
5174 * wait, then we can deadlock. It's safe to drop the mutex here, since
5175 * no new references will come in after we've killed the percpu ref.
5177 mutex_unlock(&ctx->uring_lock);
5178 wait_for_completion(&ctx->completions[0]);
5179 mutex_lock(&ctx->uring_lock);
5182 case IORING_REGISTER_BUFFERS:
5183 ret = io_sqe_buffer_register(ctx, arg, nr_args);
5185 case IORING_UNREGISTER_BUFFERS:
5189 ret = io_sqe_buffer_unregister(ctx);
5191 case IORING_REGISTER_FILES:
5192 ret = io_sqe_files_register(ctx, arg, nr_args);
5194 case IORING_UNREGISTER_FILES:
5198 ret = io_sqe_files_unregister(ctx);
5200 case IORING_REGISTER_FILES_UPDATE:
5201 ret = io_sqe_files_update(ctx, arg, nr_args);
5203 case IORING_REGISTER_EVENTFD:
5207 ret = io_eventfd_register(ctx, arg);
5209 case IORING_UNREGISTER_EVENTFD:
5213 ret = io_eventfd_unregister(ctx);
5220 /* bring the ctx back to life */
5221 reinit_completion(&ctx->completions[0]);
5222 percpu_ref_reinit(&ctx->refs);
5226 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
5227 void __user *, arg, unsigned int, nr_args)
5229 struct io_ring_ctx *ctx;
5238 if (f.file->f_op != &io_uring_fops)
5241 ctx = f.file->private_data;
5243 mutex_lock(&ctx->uring_lock);
5244 ret = __io_uring_register(ctx, opcode, arg, nr_args);
5245 mutex_unlock(&ctx->uring_lock);
5246 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
5247 ctx->cq_ev_fd != NULL, ret);
5253 static int __init io_uring_init(void)
5255 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
5258 __initcall(io_uring_init);