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 <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
178 struct io_mapped_ubuf {
181 struct bio_vec *bvec;
182 unsigned int nr_bvecs;
185 struct fixed_file_table {
189 struct fixed_file_ref_node {
190 struct percpu_ref refs;
191 struct list_head node;
192 struct list_head file_list;
193 struct fixed_file_data *file_data;
194 struct work_struct work;
197 struct fixed_file_data {
198 struct fixed_file_table *table;
199 struct io_ring_ctx *ctx;
201 struct percpu_ref *cur_refs;
202 struct percpu_ref refs;
203 struct completion done;
204 struct list_head ref_list;
209 struct list_head list;
217 struct percpu_ref refs;
218 } ____cacheline_aligned_in_smp;
222 unsigned int compat: 1;
223 unsigned int account_mem: 1;
224 unsigned int cq_overflow_flushed: 1;
225 unsigned int drain_next: 1;
226 unsigned int eventfd_async: 1;
229 * Ring buffer of indices into array of io_uring_sqe, which is
230 * mmapped by the application using the IORING_OFF_SQES offset.
232 * This indirection could e.g. be used to assign fixed
233 * io_uring_sqe entries to operations and only submit them to
234 * the queue when needed.
236 * The kernel modifies neither the indices array nor the entries
240 unsigned cached_sq_head;
243 unsigned sq_thread_idle;
244 unsigned cached_sq_dropped;
245 atomic_t cached_cq_overflow;
246 unsigned long sq_check_overflow;
248 struct list_head defer_list;
249 struct list_head timeout_list;
250 struct list_head cq_overflow_list;
252 wait_queue_head_t inflight_wait;
253 struct io_uring_sqe *sq_sqes;
254 } ____cacheline_aligned_in_smp;
256 struct io_rings *rings;
260 struct task_struct *sqo_thread; /* if using sq thread polling */
261 struct mm_struct *sqo_mm;
262 wait_queue_head_t sqo_wait;
265 * If used, fixed file set. Writers must ensure that ->refs is dead,
266 * readers must ensure that ->refs is alive as long as the file* is
267 * used. Only updated through io_uring_register(2).
269 struct fixed_file_data *file_data;
270 unsigned nr_user_files;
272 struct file *ring_file;
274 /* if used, fixed mapped user buffers */
275 unsigned nr_user_bufs;
276 struct io_mapped_ubuf *user_bufs;
278 struct user_struct *user;
280 const struct cred *creds;
282 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
283 struct completion *completions;
285 /* if all else fails... */
286 struct io_kiocb *fallback_req;
288 #if defined(CONFIG_UNIX)
289 struct socket *ring_sock;
292 struct idr io_buffer_idr;
294 struct idr personality_idr;
297 unsigned cached_cq_tail;
300 atomic_t cq_timeouts;
301 unsigned long cq_check_overflow;
302 struct wait_queue_head cq_wait;
303 struct fasync_struct *cq_fasync;
304 struct eventfd_ctx *cq_ev_fd;
305 } ____cacheline_aligned_in_smp;
308 struct mutex uring_lock;
309 wait_queue_head_t wait;
310 } ____cacheline_aligned_in_smp;
313 spinlock_t completion_lock;
316 * ->poll_list is protected by the ctx->uring_lock for
317 * io_uring instances that don't use IORING_SETUP_SQPOLL.
318 * For SQPOLL, only the single threaded io_sq_thread() will
319 * manipulate the list, hence no extra locking is needed there.
321 struct list_head poll_list;
322 struct hlist_head *cancel_hash;
323 unsigned cancel_hash_bits;
324 bool poll_multi_file;
326 spinlock_t inflight_lock;
327 struct list_head inflight_list;
328 } ____cacheline_aligned_in_smp;
330 struct work_struct exit_work;
334 * First field must be the file pointer in all the
335 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
337 struct io_poll_iocb {
340 struct wait_queue_head *head;
346 struct wait_queue_entry wait;
351 struct file *put_file;
355 struct io_timeout_data {
356 struct io_kiocb *req;
357 struct hrtimer timer;
358 struct timespec64 ts;
359 enum hrtimer_mode mode;
364 struct sockaddr __user *addr;
365 int __user *addr_len;
367 unsigned long nofile;
391 /* NOTE: kiocb has the file as the first member, so don't do it here */
399 struct sockaddr __user *addr;
406 struct user_msghdr __user *msg;
412 struct io_buffer *kbuf;
421 struct filename *filename;
422 struct statx __user *buffer;
424 unsigned long nofile;
427 struct io_files_update {
453 struct epoll_event event;
457 struct file *file_out;
458 struct file *file_in;
465 struct io_provide_buf {
474 struct io_async_connect {
475 struct sockaddr_storage address;
478 struct io_async_msghdr {
479 struct iovec fast_iov[UIO_FASTIOV];
481 struct sockaddr __user *uaddr;
483 struct sockaddr_storage addr;
487 struct iovec fast_iov[UIO_FASTIOV];
493 struct io_async_ctx {
495 struct io_async_rw rw;
496 struct io_async_msghdr msg;
497 struct io_async_connect connect;
498 struct io_timeout_data timeout;
503 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
504 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
505 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
506 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
507 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
508 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
516 REQ_F_IOPOLL_COMPLETED_BIT,
517 REQ_F_LINK_TIMEOUT_BIT,
521 REQ_F_TIMEOUT_NOSEQ_BIT,
522 REQ_F_COMP_LOCKED_BIT,
523 REQ_F_NEED_CLEANUP_BIT,
526 REQ_F_BUFFER_SELECTED_BIT,
527 REQ_F_NO_FILE_TABLE_BIT,
529 /* not a real bit, just to check we're not overflowing the space */
535 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
536 /* drain existing IO first */
537 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
539 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
540 /* doesn't sever on completion < 0 */
541 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
543 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
544 /* IOSQE_BUFFER_SELECT */
545 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
548 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
549 /* already grabbed next link */
550 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
551 /* fail rest of links */
552 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
553 /* on inflight list */
554 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
555 /* read/write uses file position */
556 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
557 /* must not punt to workers */
558 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
559 /* polled IO has completed */
560 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
561 /* has linked timeout */
562 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
563 /* timeout request */
564 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
566 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
567 /* must be punted even for NONBLOCK */
568 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
569 /* no timeout sequence */
570 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
571 /* completion under lock */
572 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
574 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
575 /* in overflow list */
576 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
577 /* already went through poll handler */
578 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
579 /* buffer already selected */
580 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
581 /* doesn't need file table for this request */
582 REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT),
586 struct io_poll_iocb poll;
587 struct io_wq_work work;
591 * NOTE! Each of the iocb union members has the file pointer
592 * as the first entry in their struct definition. So you can
593 * access the file pointer through any of the sub-structs,
594 * or directly as just 'ki_filp' in this struct.
600 struct io_poll_iocb poll;
601 struct io_accept accept;
603 struct io_cancel cancel;
604 struct io_timeout timeout;
605 struct io_connect connect;
606 struct io_sr_msg sr_msg;
608 struct io_close close;
609 struct io_files_update files_update;
610 struct io_fadvise fadvise;
611 struct io_madvise madvise;
612 struct io_epoll epoll;
613 struct io_splice splice;
614 struct io_provide_buf pbuf;
617 struct io_async_ctx *io;
619 bool needs_fixed_file;
622 struct io_ring_ctx *ctx;
623 struct list_head list;
626 struct task_struct *task;
632 struct list_head link_list;
634 struct list_head inflight_entry;
636 struct percpu_ref *fixed_file_refs;
640 * Only commands that never go async can use the below fields,
641 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
642 * async armed poll handlers for regular commands. The latter
643 * restore the work, if needed.
646 struct callback_head task_work;
647 struct hlist_node hash_node;
648 struct async_poll *apoll;
650 struct io_wq_work work;
654 #define IO_PLUG_THRESHOLD 2
655 #define IO_IOPOLL_BATCH 8
657 struct io_submit_state {
658 struct blk_plug plug;
661 * io_kiocb alloc cache
663 void *reqs[IO_IOPOLL_BATCH];
664 unsigned int free_reqs;
667 * File reference cache
671 unsigned int has_refs;
672 unsigned int used_refs;
673 unsigned int ios_left;
677 /* needs req->io allocated for deferral/async */
678 unsigned async_ctx : 1;
679 /* needs current->mm setup, does mm access */
680 unsigned needs_mm : 1;
681 /* needs req->file assigned */
682 unsigned needs_file : 1;
683 /* needs req->file assigned IFF fd is >= 0 */
684 unsigned fd_non_neg : 1;
685 /* hash wq insertion if file is a regular file */
686 unsigned hash_reg_file : 1;
687 /* unbound wq insertion if file is a non-regular file */
688 unsigned unbound_nonreg_file : 1;
689 /* opcode is not supported by this kernel */
690 unsigned not_supported : 1;
691 /* needs file table */
692 unsigned file_table : 1;
694 unsigned needs_fs : 1;
695 /* set if opcode supports polled "wait" */
697 unsigned pollout : 1;
698 /* op supports buffer selection */
699 unsigned buffer_select : 1;
702 static const struct io_op_def io_op_defs[] = {
703 [IORING_OP_NOP] = {},
704 [IORING_OP_READV] = {
708 .unbound_nonreg_file = 1,
712 [IORING_OP_WRITEV] = {
717 .unbound_nonreg_file = 1,
720 [IORING_OP_FSYNC] = {
723 [IORING_OP_READ_FIXED] = {
725 .unbound_nonreg_file = 1,
728 [IORING_OP_WRITE_FIXED] = {
731 .unbound_nonreg_file = 1,
734 [IORING_OP_POLL_ADD] = {
736 .unbound_nonreg_file = 1,
738 [IORING_OP_POLL_REMOVE] = {},
739 [IORING_OP_SYNC_FILE_RANGE] = {
742 [IORING_OP_SENDMSG] = {
746 .unbound_nonreg_file = 1,
750 [IORING_OP_RECVMSG] = {
754 .unbound_nonreg_file = 1,
759 [IORING_OP_TIMEOUT] = {
763 [IORING_OP_TIMEOUT_REMOVE] = {},
764 [IORING_OP_ACCEPT] = {
767 .unbound_nonreg_file = 1,
771 [IORING_OP_ASYNC_CANCEL] = {},
772 [IORING_OP_LINK_TIMEOUT] = {
776 [IORING_OP_CONNECT] = {
780 .unbound_nonreg_file = 1,
783 [IORING_OP_FALLOCATE] = {
786 [IORING_OP_OPENAT] = {
792 [IORING_OP_CLOSE] = {
796 [IORING_OP_FILES_UPDATE] = {
800 [IORING_OP_STATX] = {
810 .unbound_nonreg_file = 1,
814 [IORING_OP_WRITE] = {
817 .unbound_nonreg_file = 1,
820 [IORING_OP_FADVISE] = {
823 [IORING_OP_MADVISE] = {
829 .unbound_nonreg_file = 1,
835 .unbound_nonreg_file = 1,
839 [IORING_OP_OPENAT2] = {
845 [IORING_OP_EPOLL_CTL] = {
846 .unbound_nonreg_file = 1,
849 [IORING_OP_SPLICE] = {
852 .unbound_nonreg_file = 1,
854 [IORING_OP_PROVIDE_BUFFERS] = {},
855 [IORING_OP_REMOVE_BUFFERS] = {},
858 static void io_wq_submit_work(struct io_wq_work **workptr);
859 static void io_cqring_fill_event(struct io_kiocb *req, long res);
860 static void io_put_req(struct io_kiocb *req);
861 static void __io_double_put_req(struct io_kiocb *req);
862 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
863 static void io_queue_linked_timeout(struct io_kiocb *req);
864 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
865 struct io_uring_files_update *ip,
867 static int io_grab_files(struct io_kiocb *req);
868 static void io_cleanup_req(struct io_kiocb *req);
869 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
870 int fd, struct file **out_file, bool fixed);
871 static void __io_queue_sqe(struct io_kiocb *req,
872 const struct io_uring_sqe *sqe);
874 static struct kmem_cache *req_cachep;
876 static const struct file_operations io_uring_fops;
878 struct sock *io_uring_get_socket(struct file *file)
880 #if defined(CONFIG_UNIX)
881 if (file->f_op == &io_uring_fops) {
882 struct io_ring_ctx *ctx = file->private_data;
884 return ctx->ring_sock->sk;
889 EXPORT_SYMBOL(io_uring_get_socket);
891 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
893 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
895 complete(&ctx->completions[0]);
898 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
900 struct io_ring_ctx *ctx;
903 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
907 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
908 if (!ctx->fallback_req)
911 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
912 if (!ctx->completions)
916 * Use 5 bits less than the max cq entries, that should give us around
917 * 32 entries per hash list if totally full and uniformly spread.
919 hash_bits = ilog2(p->cq_entries);
923 ctx->cancel_hash_bits = hash_bits;
924 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
926 if (!ctx->cancel_hash)
928 __hash_init(ctx->cancel_hash, 1U << hash_bits);
930 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
931 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
934 ctx->flags = p->flags;
935 init_waitqueue_head(&ctx->cq_wait);
936 INIT_LIST_HEAD(&ctx->cq_overflow_list);
937 init_completion(&ctx->completions[0]);
938 init_completion(&ctx->completions[1]);
939 idr_init(&ctx->io_buffer_idr);
940 idr_init(&ctx->personality_idr);
941 mutex_init(&ctx->uring_lock);
942 init_waitqueue_head(&ctx->wait);
943 spin_lock_init(&ctx->completion_lock);
944 INIT_LIST_HEAD(&ctx->poll_list);
945 INIT_LIST_HEAD(&ctx->defer_list);
946 INIT_LIST_HEAD(&ctx->timeout_list);
947 init_waitqueue_head(&ctx->inflight_wait);
948 spin_lock_init(&ctx->inflight_lock);
949 INIT_LIST_HEAD(&ctx->inflight_list);
952 if (ctx->fallback_req)
953 kmem_cache_free(req_cachep, ctx->fallback_req);
954 kfree(ctx->completions);
955 kfree(ctx->cancel_hash);
960 static inline bool __req_need_defer(struct io_kiocb *req)
962 struct io_ring_ctx *ctx = req->ctx;
964 return req->sequence != ctx->cached_cq_tail
965 + atomic_read(&ctx->cached_cq_overflow);
968 static inline bool req_need_defer(struct io_kiocb *req)
970 if (unlikely(req->flags & REQ_F_IO_DRAIN))
971 return __req_need_defer(req);
976 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
978 struct io_kiocb *req;
980 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
981 if (req && !req_need_defer(req)) {
982 list_del_init(&req->list);
989 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
991 struct io_kiocb *req;
993 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
995 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
997 if (!__req_need_defer(req)) {
998 list_del_init(&req->list);
1006 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1008 struct io_rings *rings = ctx->rings;
1010 /* order cqe stores with ring update */
1011 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1013 if (wq_has_sleeper(&ctx->cq_wait)) {
1014 wake_up_interruptible(&ctx->cq_wait);
1015 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1019 static inline void io_req_work_grab_env(struct io_kiocb *req,
1020 const struct io_op_def *def)
1022 if (!req->work.mm && def->needs_mm) {
1023 mmgrab(current->mm);
1024 req->work.mm = current->mm;
1026 if (!req->work.creds)
1027 req->work.creds = get_current_cred();
1028 if (!req->work.fs && def->needs_fs) {
1029 spin_lock(¤t->fs->lock);
1030 if (!current->fs->in_exec) {
1031 req->work.fs = current->fs;
1032 req->work.fs->users++;
1034 req->work.flags |= IO_WQ_WORK_CANCEL;
1036 spin_unlock(¤t->fs->lock);
1038 if (!req->work.task_pid)
1039 req->work.task_pid = task_pid_vnr(current);
1042 static inline void io_req_work_drop_env(struct io_kiocb *req)
1045 mmdrop(req->work.mm);
1046 req->work.mm = NULL;
1048 if (req->work.creds) {
1049 put_cred(req->work.creds);
1050 req->work.creds = NULL;
1053 struct fs_struct *fs = req->work.fs;
1055 spin_lock(&req->work.fs->lock);
1058 spin_unlock(&req->work.fs->lock);
1064 static inline void io_prep_async_work(struct io_kiocb *req,
1065 struct io_kiocb **link)
1067 const struct io_op_def *def = &io_op_defs[req->opcode];
1069 if (req->flags & REQ_F_ISREG) {
1070 if (def->hash_reg_file)
1071 io_wq_hash_work(&req->work, file_inode(req->file));
1073 if (def->unbound_nonreg_file)
1074 req->work.flags |= IO_WQ_WORK_UNBOUND;
1077 io_req_work_grab_env(req, def);
1079 *link = io_prep_linked_timeout(req);
1082 static inline void io_queue_async_work(struct io_kiocb *req)
1084 struct io_ring_ctx *ctx = req->ctx;
1085 struct io_kiocb *link;
1087 io_prep_async_work(req, &link);
1089 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1090 &req->work, req->flags);
1091 io_wq_enqueue(ctx->io_wq, &req->work);
1094 io_queue_linked_timeout(link);
1097 static void io_kill_timeout(struct io_kiocb *req)
1101 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1103 atomic_inc(&req->ctx->cq_timeouts);
1104 list_del_init(&req->list);
1105 req->flags |= REQ_F_COMP_LOCKED;
1106 io_cqring_fill_event(req, 0);
1111 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1113 struct io_kiocb *req, *tmp;
1115 spin_lock_irq(&ctx->completion_lock);
1116 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1117 io_kill_timeout(req);
1118 spin_unlock_irq(&ctx->completion_lock);
1121 static void io_commit_cqring(struct io_ring_ctx *ctx)
1123 struct io_kiocb *req;
1125 while ((req = io_get_timeout_req(ctx)) != NULL)
1126 io_kill_timeout(req);
1128 __io_commit_cqring(ctx);
1130 while ((req = io_get_deferred_req(ctx)) != NULL)
1131 io_queue_async_work(req);
1134 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1136 struct io_rings *rings = ctx->rings;
1139 tail = ctx->cached_cq_tail;
1141 * writes to the cq entry need to come after reading head; the
1142 * control dependency is enough as we're using WRITE_ONCE to
1145 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1148 ctx->cached_cq_tail++;
1149 return &rings->cqes[tail & ctx->cq_mask];
1152 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1156 if (!ctx->eventfd_async)
1158 return io_wq_current_is_worker();
1161 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1163 if (waitqueue_active(&ctx->wait))
1164 wake_up(&ctx->wait);
1165 if (waitqueue_active(&ctx->sqo_wait))
1166 wake_up(&ctx->sqo_wait);
1167 if (io_should_trigger_evfd(ctx))
1168 eventfd_signal(ctx->cq_ev_fd, 1);
1171 /* Returns true if there are no backlogged entries after the flush */
1172 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1174 struct io_rings *rings = ctx->rings;
1175 struct io_uring_cqe *cqe;
1176 struct io_kiocb *req;
1177 unsigned long flags;
1181 if (list_empty_careful(&ctx->cq_overflow_list))
1183 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1184 rings->cq_ring_entries))
1188 spin_lock_irqsave(&ctx->completion_lock, flags);
1190 /* if force is set, the ring is going away. always drop after that */
1192 ctx->cq_overflow_flushed = 1;
1195 while (!list_empty(&ctx->cq_overflow_list)) {
1196 cqe = io_get_cqring(ctx);
1200 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1202 list_move(&req->list, &list);
1203 req->flags &= ~REQ_F_OVERFLOW;
1205 WRITE_ONCE(cqe->user_data, req->user_data);
1206 WRITE_ONCE(cqe->res, req->result);
1207 WRITE_ONCE(cqe->flags, req->cflags);
1209 WRITE_ONCE(ctx->rings->cq_overflow,
1210 atomic_inc_return(&ctx->cached_cq_overflow));
1214 io_commit_cqring(ctx);
1216 clear_bit(0, &ctx->sq_check_overflow);
1217 clear_bit(0, &ctx->cq_check_overflow);
1219 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1220 io_cqring_ev_posted(ctx);
1222 while (!list_empty(&list)) {
1223 req = list_first_entry(&list, struct io_kiocb, list);
1224 list_del(&req->list);
1231 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1233 struct io_ring_ctx *ctx = req->ctx;
1234 struct io_uring_cqe *cqe;
1236 trace_io_uring_complete(ctx, req->user_data, res);
1239 * If we can't get a cq entry, userspace overflowed the
1240 * submission (by quite a lot). Increment the overflow count in
1243 cqe = io_get_cqring(ctx);
1245 WRITE_ONCE(cqe->user_data, req->user_data);
1246 WRITE_ONCE(cqe->res, res);
1247 WRITE_ONCE(cqe->flags, cflags);
1248 } else if (ctx->cq_overflow_flushed) {
1249 WRITE_ONCE(ctx->rings->cq_overflow,
1250 atomic_inc_return(&ctx->cached_cq_overflow));
1252 if (list_empty(&ctx->cq_overflow_list)) {
1253 set_bit(0, &ctx->sq_check_overflow);
1254 set_bit(0, &ctx->cq_check_overflow);
1256 req->flags |= REQ_F_OVERFLOW;
1257 refcount_inc(&req->refs);
1259 req->cflags = cflags;
1260 list_add_tail(&req->list, &ctx->cq_overflow_list);
1264 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1266 __io_cqring_fill_event(req, res, 0);
1269 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1271 struct io_ring_ctx *ctx = req->ctx;
1272 unsigned long flags;
1274 spin_lock_irqsave(&ctx->completion_lock, flags);
1275 __io_cqring_fill_event(req, res, cflags);
1276 io_commit_cqring(ctx);
1277 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1279 io_cqring_ev_posted(ctx);
1282 static void io_cqring_add_event(struct io_kiocb *req, long res)
1284 __io_cqring_add_event(req, res, 0);
1287 static inline bool io_is_fallback_req(struct io_kiocb *req)
1289 return req == (struct io_kiocb *)
1290 ((unsigned long) req->ctx->fallback_req & ~1UL);
1293 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1295 struct io_kiocb *req;
1297 req = ctx->fallback_req;
1298 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1304 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1305 struct io_submit_state *state)
1307 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1308 struct io_kiocb *req;
1311 req = kmem_cache_alloc(req_cachep, gfp);
1314 } else if (!state->free_reqs) {
1318 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1319 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1322 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1323 * retry single alloc to be on the safe side.
1325 if (unlikely(ret <= 0)) {
1326 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1327 if (!state->reqs[0])
1331 state->free_reqs = ret - 1;
1332 req = state->reqs[ret - 1];
1335 req = state->reqs[state->free_reqs];
1340 return io_get_fallback_req(ctx);
1343 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1347 percpu_ref_put(req->fixed_file_refs);
1352 static void __io_req_aux_free(struct io_kiocb *req)
1354 if (req->flags & REQ_F_NEED_CLEANUP)
1355 io_cleanup_req(req);
1359 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1361 put_task_struct(req->task);
1363 io_req_work_drop_env(req);
1366 static void __io_free_req(struct io_kiocb *req)
1368 __io_req_aux_free(req);
1370 if (req->flags & REQ_F_INFLIGHT) {
1371 struct io_ring_ctx *ctx = req->ctx;
1372 unsigned long flags;
1374 spin_lock_irqsave(&ctx->inflight_lock, flags);
1375 list_del(&req->inflight_entry);
1376 if (waitqueue_active(&ctx->inflight_wait))
1377 wake_up(&ctx->inflight_wait);
1378 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1381 percpu_ref_put(&req->ctx->refs);
1382 if (likely(!io_is_fallback_req(req)))
1383 kmem_cache_free(req_cachep, req);
1385 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1389 void *reqs[IO_IOPOLL_BATCH];
1394 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1398 if (rb->need_iter) {
1399 int i, inflight = 0;
1400 unsigned long flags;
1402 for (i = 0; i < rb->to_free; i++) {
1403 struct io_kiocb *req = rb->reqs[i];
1405 if (req->flags & REQ_F_FIXED_FILE) {
1407 percpu_ref_put(req->fixed_file_refs);
1409 if (req->flags & REQ_F_INFLIGHT)
1411 __io_req_aux_free(req);
1416 spin_lock_irqsave(&ctx->inflight_lock, flags);
1417 for (i = 0; i < rb->to_free; i++) {
1418 struct io_kiocb *req = rb->reqs[i];
1420 if (req->flags & REQ_F_INFLIGHT) {
1421 list_del(&req->inflight_entry);
1426 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1428 if (waitqueue_active(&ctx->inflight_wait))
1429 wake_up(&ctx->inflight_wait);
1432 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1433 percpu_ref_put_many(&ctx->refs, rb->to_free);
1434 rb->to_free = rb->need_iter = 0;
1437 static bool io_link_cancel_timeout(struct io_kiocb *req)
1439 struct io_ring_ctx *ctx = req->ctx;
1442 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1444 io_cqring_fill_event(req, -ECANCELED);
1445 io_commit_cqring(ctx);
1446 req->flags &= ~REQ_F_LINK_HEAD;
1454 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1456 struct io_ring_ctx *ctx = req->ctx;
1457 bool wake_ev = false;
1459 /* Already got next link */
1460 if (req->flags & REQ_F_LINK_NEXT)
1464 * The list should never be empty when we are called here. But could
1465 * potentially happen if the chain is messed up, check to be on the
1468 while (!list_empty(&req->link_list)) {
1469 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1470 struct io_kiocb, link_list);
1472 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1473 (nxt->flags & REQ_F_TIMEOUT))) {
1474 list_del_init(&nxt->link_list);
1475 wake_ev |= io_link_cancel_timeout(nxt);
1476 req->flags &= ~REQ_F_LINK_TIMEOUT;
1480 list_del_init(&req->link_list);
1481 if (!list_empty(&nxt->link_list))
1482 nxt->flags |= REQ_F_LINK_HEAD;
1487 req->flags |= REQ_F_LINK_NEXT;
1489 io_cqring_ev_posted(ctx);
1493 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1495 static void io_fail_links(struct io_kiocb *req)
1497 struct io_ring_ctx *ctx = req->ctx;
1498 unsigned long flags;
1500 spin_lock_irqsave(&ctx->completion_lock, flags);
1502 while (!list_empty(&req->link_list)) {
1503 struct io_kiocb *link = list_first_entry(&req->link_list,
1504 struct io_kiocb, link_list);
1506 list_del_init(&link->link_list);
1507 trace_io_uring_fail_link(req, link);
1509 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1510 link->opcode == IORING_OP_LINK_TIMEOUT) {
1511 io_link_cancel_timeout(link);
1513 io_cqring_fill_event(link, -ECANCELED);
1514 __io_double_put_req(link);
1516 req->flags &= ~REQ_F_LINK_TIMEOUT;
1519 io_commit_cqring(ctx);
1520 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1521 io_cqring_ev_posted(ctx);
1524 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1526 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1530 * If LINK is set, we have dependent requests in this chain. If we
1531 * didn't fail this request, queue the first one up, moving any other
1532 * dependencies to the next request. In case of failure, fail the rest
1535 if (req->flags & REQ_F_FAIL_LINK) {
1537 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1538 REQ_F_LINK_TIMEOUT) {
1539 struct io_ring_ctx *ctx = req->ctx;
1540 unsigned long flags;
1543 * If this is a timeout link, we could be racing with the
1544 * timeout timer. Grab the completion lock for this case to
1545 * protect against that.
1547 spin_lock_irqsave(&ctx->completion_lock, flags);
1548 io_req_link_next(req, nxt);
1549 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1551 io_req_link_next(req, nxt);
1555 static void io_free_req(struct io_kiocb *req)
1557 struct io_kiocb *nxt = NULL;
1559 io_req_find_next(req, &nxt);
1563 io_queue_async_work(nxt);
1566 static void io_link_work_cb(struct io_wq_work **workptr)
1568 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1569 struct io_kiocb *link;
1571 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1572 io_queue_linked_timeout(link);
1573 io_wq_submit_work(workptr);
1576 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1578 struct io_kiocb *link;
1579 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1581 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1582 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1584 *workptr = &nxt->work;
1585 link = io_prep_linked_timeout(nxt);
1587 nxt->work.func = io_link_work_cb;
1591 * Drop reference to request, return next in chain (if there is one) if this
1592 * was the last reference to this request.
1594 __attribute__((nonnull))
1595 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1597 if (refcount_dec_and_test(&req->refs)) {
1598 io_req_find_next(req, nxtptr);
1603 static void io_put_req(struct io_kiocb *req)
1605 if (refcount_dec_and_test(&req->refs))
1609 static void io_steal_work(struct io_kiocb *req,
1610 struct io_wq_work **workptr)
1613 * It's in an io-wq worker, so there always should be at least
1614 * one reference, which will be dropped in io_put_work() just
1615 * after the current handler returns.
1617 * It also means, that if the counter dropped to 1, then there is
1618 * no asynchronous users left, so it's safe to steal the next work.
1620 if (refcount_read(&req->refs) == 1) {
1621 struct io_kiocb *nxt = NULL;
1623 io_req_find_next(req, &nxt);
1625 io_wq_assign_next(workptr, nxt);
1630 * Must only be used if we don't need to care about links, usually from
1631 * within the completion handling itself.
1633 static void __io_double_put_req(struct io_kiocb *req)
1635 /* drop both submit and complete references */
1636 if (refcount_sub_and_test(2, &req->refs))
1640 static void io_double_put_req(struct io_kiocb *req)
1642 /* drop both submit and complete references */
1643 if (refcount_sub_and_test(2, &req->refs))
1647 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1649 struct io_rings *rings = ctx->rings;
1651 if (test_bit(0, &ctx->cq_check_overflow)) {
1653 * noflush == true is from the waitqueue handler, just ensure
1654 * we wake up the task, and the next invocation will flush the
1655 * entries. We cannot safely to it from here.
1657 if (noflush && !list_empty(&ctx->cq_overflow_list))
1660 io_cqring_overflow_flush(ctx, false);
1663 /* See comment at the top of this file */
1665 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1668 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1670 struct io_rings *rings = ctx->rings;
1672 /* make sure SQ entry isn't read before tail */
1673 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1676 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1678 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1681 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1684 rb->reqs[rb->to_free++] = req;
1685 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1686 io_free_req_many(req->ctx, rb);
1690 static int io_put_kbuf(struct io_kiocb *req)
1692 struct io_buffer *kbuf;
1695 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1696 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1697 cflags |= IORING_CQE_F_BUFFER;
1704 * Find and free completed poll iocbs
1706 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1707 struct list_head *done)
1709 struct req_batch rb;
1710 struct io_kiocb *req;
1712 rb.to_free = rb.need_iter = 0;
1713 while (!list_empty(done)) {
1716 req = list_first_entry(done, struct io_kiocb, list);
1717 list_del(&req->list);
1719 if (req->flags & REQ_F_BUFFER_SELECTED)
1720 cflags = io_put_kbuf(req);
1722 __io_cqring_fill_event(req, req->result, cflags);
1725 if (refcount_dec_and_test(&req->refs) &&
1726 !io_req_multi_free(&rb, req))
1730 io_commit_cqring(ctx);
1731 if (ctx->flags & IORING_SETUP_SQPOLL)
1732 io_cqring_ev_posted(ctx);
1733 io_free_req_many(ctx, &rb);
1736 static void io_iopoll_queue(struct list_head *again)
1738 struct io_kiocb *req;
1741 req = list_first_entry(again, struct io_kiocb, list);
1742 list_del(&req->list);
1743 refcount_inc(&req->refs);
1744 io_queue_async_work(req);
1745 } while (!list_empty(again));
1748 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1751 struct io_kiocb *req, *tmp;
1758 * Only spin for completions if we don't have multiple devices hanging
1759 * off our complete list, and we're under the requested amount.
1761 spin = !ctx->poll_multi_file && *nr_events < min;
1764 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1765 struct kiocb *kiocb = &req->rw.kiocb;
1768 * Move completed and retryable entries to our local lists.
1769 * If we find a request that requires polling, break out
1770 * and complete those lists first, if we have entries there.
1772 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1773 list_move_tail(&req->list, &done);
1776 if (!list_empty(&done))
1779 if (req->result == -EAGAIN) {
1780 list_move_tail(&req->list, &again);
1783 if (!list_empty(&again))
1786 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1795 if (!list_empty(&done))
1796 io_iopoll_complete(ctx, nr_events, &done);
1798 if (!list_empty(&again))
1799 io_iopoll_queue(&again);
1805 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1806 * non-spinning poll check - we'll still enter the driver poll loop, but only
1807 * as a non-spinning completion check.
1809 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1812 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1815 ret = io_do_iopoll(ctx, nr_events, min);
1818 if (!min || *nr_events >= min)
1826 * We can't just wait for polled events to come to us, we have to actively
1827 * find and complete them.
1829 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1831 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1834 mutex_lock(&ctx->uring_lock);
1835 while (!list_empty(&ctx->poll_list)) {
1836 unsigned int nr_events = 0;
1838 io_iopoll_getevents(ctx, &nr_events, 1);
1841 * Ensure we allow local-to-the-cpu processing to take place,
1842 * in this case we need to ensure that we reap all events.
1846 mutex_unlock(&ctx->uring_lock);
1849 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1852 int iters = 0, ret = 0;
1855 * We disallow the app entering submit/complete with polling, but we
1856 * still need to lock the ring to prevent racing with polled issue
1857 * that got punted to a workqueue.
1859 mutex_lock(&ctx->uring_lock);
1864 * Don't enter poll loop if we already have events pending.
1865 * If we do, we can potentially be spinning for commands that
1866 * already triggered a CQE (eg in error).
1868 if (io_cqring_events(ctx, false))
1872 * If a submit got punted to a workqueue, we can have the
1873 * application entering polling for a command before it gets
1874 * issued. That app will hold the uring_lock for the duration
1875 * of the poll right here, so we need to take a breather every
1876 * now and then to ensure that the issue has a chance to add
1877 * the poll to the issued list. Otherwise we can spin here
1878 * forever, while the workqueue is stuck trying to acquire the
1881 if (!(++iters & 7)) {
1882 mutex_unlock(&ctx->uring_lock);
1883 mutex_lock(&ctx->uring_lock);
1886 if (*nr_events < min)
1887 tmin = min - *nr_events;
1889 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1893 } while (min && !*nr_events && !need_resched());
1895 mutex_unlock(&ctx->uring_lock);
1899 static void kiocb_end_write(struct io_kiocb *req)
1902 * Tell lockdep we inherited freeze protection from submission
1905 if (req->flags & REQ_F_ISREG) {
1906 struct inode *inode = file_inode(req->file);
1908 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1910 file_end_write(req->file);
1913 static inline void req_set_fail_links(struct io_kiocb *req)
1915 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1916 req->flags |= REQ_F_FAIL_LINK;
1919 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1921 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1924 if (kiocb->ki_flags & IOCB_WRITE)
1925 kiocb_end_write(req);
1927 if (res != req->result)
1928 req_set_fail_links(req);
1929 if (req->flags & REQ_F_BUFFER_SELECTED)
1930 cflags = io_put_kbuf(req);
1931 __io_cqring_add_event(req, res, cflags);
1934 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1936 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1938 io_complete_rw_common(kiocb, res);
1942 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1944 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1946 if (kiocb->ki_flags & IOCB_WRITE)
1947 kiocb_end_write(req);
1949 if (res != req->result)
1950 req_set_fail_links(req);
1953 req->flags |= REQ_F_IOPOLL_COMPLETED;
1957 * After the iocb has been issued, it's safe to be found on the poll list.
1958 * Adding the kiocb to the list AFTER submission ensures that we don't
1959 * find it from a io_iopoll_getevents() thread before the issuer is done
1960 * accessing the kiocb cookie.
1962 static void io_iopoll_req_issued(struct io_kiocb *req)
1964 struct io_ring_ctx *ctx = req->ctx;
1967 * Track whether we have multiple files in our lists. This will impact
1968 * how we do polling eventually, not spinning if we're on potentially
1969 * different devices.
1971 if (list_empty(&ctx->poll_list)) {
1972 ctx->poll_multi_file = false;
1973 } else if (!ctx->poll_multi_file) {
1974 struct io_kiocb *list_req;
1976 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1978 if (list_req->file != req->file)
1979 ctx->poll_multi_file = true;
1983 * For fast devices, IO may have already completed. If it has, add
1984 * it to the front so we find it first.
1986 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1987 list_add(&req->list, &ctx->poll_list);
1989 list_add_tail(&req->list, &ctx->poll_list);
1991 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1992 wq_has_sleeper(&ctx->sqo_wait))
1993 wake_up(&ctx->sqo_wait);
1996 static void io_file_put(struct io_submit_state *state)
1999 int diff = state->has_refs - state->used_refs;
2002 fput_many(state->file, diff);
2008 * Get as many references to a file as we have IOs left in this submission,
2009 * assuming most submissions are for one file, or at least that each file
2010 * has more than one submission.
2012 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2018 if (state->fd == fd) {
2025 state->file = fget_many(fd, state->ios_left);
2030 state->has_refs = state->ios_left;
2031 state->used_refs = 1;
2037 * If we tracked the file through the SCM inflight mechanism, we could support
2038 * any file. For now, just ensure that anything potentially problematic is done
2041 static bool io_file_supports_async(struct file *file, int rw)
2043 umode_t mode = file_inode(file)->i_mode;
2045 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2047 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2050 if (!(file->f_mode & FMODE_NOWAIT))
2054 return file->f_op->read_iter != NULL;
2056 return file->f_op->write_iter != NULL;
2059 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2060 bool force_nonblock)
2062 struct io_ring_ctx *ctx = req->ctx;
2063 struct kiocb *kiocb = &req->rw.kiocb;
2067 if (S_ISREG(file_inode(req->file)->i_mode))
2068 req->flags |= REQ_F_ISREG;
2070 kiocb->ki_pos = READ_ONCE(sqe->off);
2071 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2072 req->flags |= REQ_F_CUR_POS;
2073 kiocb->ki_pos = req->file->f_pos;
2075 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2076 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2077 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2081 ioprio = READ_ONCE(sqe->ioprio);
2083 ret = ioprio_check_cap(ioprio);
2087 kiocb->ki_ioprio = ioprio;
2089 kiocb->ki_ioprio = get_current_ioprio();
2091 /* don't allow async punt if RWF_NOWAIT was requested */
2092 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2093 (req->file->f_flags & O_NONBLOCK))
2094 req->flags |= REQ_F_NOWAIT;
2097 kiocb->ki_flags |= IOCB_NOWAIT;
2099 if (ctx->flags & IORING_SETUP_IOPOLL) {
2100 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2101 !kiocb->ki_filp->f_op->iopoll)
2104 kiocb->ki_flags |= IOCB_HIPRI;
2105 kiocb->ki_complete = io_complete_rw_iopoll;
2108 if (kiocb->ki_flags & IOCB_HIPRI)
2110 kiocb->ki_complete = io_complete_rw;
2113 req->rw.addr = READ_ONCE(sqe->addr);
2114 req->rw.len = READ_ONCE(sqe->len);
2115 /* we own ->private, reuse it for the buffer index / buffer ID */
2116 req->rw.kiocb.private = (void *) (unsigned long)
2117 READ_ONCE(sqe->buf_index);
2121 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2127 case -ERESTARTNOINTR:
2128 case -ERESTARTNOHAND:
2129 case -ERESTART_RESTARTBLOCK:
2131 * We can't just restart the syscall, since previously
2132 * submitted sqes may already be in progress. Just fail this
2138 kiocb->ki_complete(kiocb, ret, 0);
2142 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2144 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2146 if (req->flags & REQ_F_CUR_POS)
2147 req->file->f_pos = kiocb->ki_pos;
2148 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2149 io_complete_rw(kiocb, ret, 0);
2151 io_rw_done(kiocb, ret);
2154 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2155 struct iov_iter *iter)
2157 struct io_ring_ctx *ctx = req->ctx;
2158 size_t len = req->rw.len;
2159 struct io_mapped_ubuf *imu;
2160 unsigned index, buf_index;
2164 /* attempt to use fixed buffers without having provided iovecs */
2165 if (unlikely(!ctx->user_bufs))
2168 buf_index = (unsigned long) req->rw.kiocb.private;
2169 if (unlikely(buf_index >= ctx->nr_user_bufs))
2172 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2173 imu = &ctx->user_bufs[index];
2174 buf_addr = req->rw.addr;
2177 if (buf_addr + len < buf_addr)
2179 /* not inside the mapped region */
2180 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2184 * May not be a start of buffer, set size appropriately
2185 * and advance us to the beginning.
2187 offset = buf_addr - imu->ubuf;
2188 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2192 * Don't use iov_iter_advance() here, as it's really slow for
2193 * using the latter parts of a big fixed buffer - it iterates
2194 * over each segment manually. We can cheat a bit here, because
2197 * 1) it's a BVEC iter, we set it up
2198 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2199 * first and last bvec
2201 * So just find our index, and adjust the iterator afterwards.
2202 * If the offset is within the first bvec (or the whole first
2203 * bvec, just use iov_iter_advance(). This makes it easier
2204 * since we can just skip the first segment, which may not
2205 * be PAGE_SIZE aligned.
2207 const struct bio_vec *bvec = imu->bvec;
2209 if (offset <= bvec->bv_len) {
2210 iov_iter_advance(iter, offset);
2212 unsigned long seg_skip;
2214 /* skip first vec */
2215 offset -= bvec->bv_len;
2216 seg_skip = 1 + (offset >> PAGE_SHIFT);
2218 iter->bvec = bvec + seg_skip;
2219 iter->nr_segs -= seg_skip;
2220 iter->count -= bvec->bv_len + offset;
2221 iter->iov_offset = offset & ~PAGE_MASK;
2228 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2231 mutex_unlock(&ctx->uring_lock);
2234 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2237 * "Normal" inline submissions always hold the uring_lock, since we
2238 * grab it from the system call. Same is true for the SQPOLL offload.
2239 * The only exception is when we've detached the request and issue it
2240 * from an async worker thread, grab the lock for that case.
2243 mutex_lock(&ctx->uring_lock);
2246 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2247 int bgid, struct io_buffer *kbuf,
2250 struct io_buffer *head;
2252 if (req->flags & REQ_F_BUFFER_SELECTED)
2255 io_ring_submit_lock(req->ctx, needs_lock);
2257 lockdep_assert_held(&req->ctx->uring_lock);
2259 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2261 if (!list_empty(&head->list)) {
2262 kbuf = list_last_entry(&head->list, struct io_buffer,
2264 list_del(&kbuf->list);
2267 idr_remove(&req->ctx->io_buffer_idr, bgid);
2269 if (*len > kbuf->len)
2272 kbuf = ERR_PTR(-ENOBUFS);
2275 io_ring_submit_unlock(req->ctx, needs_lock);
2280 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2283 struct io_buffer *kbuf;
2286 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2287 bgid = (int) (unsigned long) req->rw.kiocb.private;
2288 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2291 req->rw.addr = (u64) (unsigned long) kbuf;
2292 req->flags |= REQ_F_BUFFER_SELECTED;
2293 return u64_to_user_ptr(kbuf->addr);
2296 #ifdef CONFIG_COMPAT
2297 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2300 struct compat_iovec __user *uiov;
2301 compat_ssize_t clen;
2305 uiov = u64_to_user_ptr(req->rw.addr);
2306 if (!access_ok(uiov, sizeof(*uiov)))
2308 if (__get_user(clen, &uiov->iov_len))
2314 buf = io_rw_buffer_select(req, &len, needs_lock);
2316 return PTR_ERR(buf);
2317 iov[0].iov_base = buf;
2318 iov[0].iov_len = (compat_size_t) len;
2323 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2326 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2330 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2333 len = iov[0].iov_len;
2336 buf = io_rw_buffer_select(req, &len, needs_lock);
2338 return PTR_ERR(buf);
2339 iov[0].iov_base = buf;
2340 iov[0].iov_len = len;
2344 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2347 if (req->flags & REQ_F_BUFFER_SELECTED)
2351 else if (req->rw.len > 1)
2354 #ifdef CONFIG_COMPAT
2355 if (req->ctx->compat)
2356 return io_compat_import(req, iov, needs_lock);
2359 return __io_iov_buffer_select(req, iov, needs_lock);
2362 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2363 struct iovec **iovec, struct iov_iter *iter,
2366 void __user *buf = u64_to_user_ptr(req->rw.addr);
2367 size_t sqe_len = req->rw.len;
2371 opcode = req->opcode;
2372 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2374 return io_import_fixed(req, rw, iter);
2377 /* buffer index only valid with fixed read/write, or buffer select */
2378 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2381 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2382 if (req->flags & REQ_F_BUFFER_SELECT) {
2383 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2386 return PTR_ERR(buf);
2388 req->rw.len = sqe_len;
2391 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2393 return ret < 0 ? ret : sqe_len;
2397 struct io_async_rw *iorw = &req->io->rw;
2400 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2401 if (iorw->iov == iorw->fast_iov)
2406 if (req->flags & REQ_F_BUFFER_SELECT) {
2407 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2409 ret = (*iovec)->iov_len;
2410 iov_iter_init(iter, rw, *iovec, 1, ret);
2416 #ifdef CONFIG_COMPAT
2417 if (req->ctx->compat)
2418 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2422 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2426 * For files that don't have ->read_iter() and ->write_iter(), handle them
2427 * by looping over ->read() or ->write() manually.
2429 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2430 struct iov_iter *iter)
2435 * Don't support polled IO through this interface, and we can't
2436 * support non-blocking either. For the latter, this just causes
2437 * the kiocb to be handled from an async context.
2439 if (kiocb->ki_flags & IOCB_HIPRI)
2441 if (kiocb->ki_flags & IOCB_NOWAIT)
2444 while (iov_iter_count(iter)) {
2448 if (!iov_iter_is_bvec(iter)) {
2449 iovec = iov_iter_iovec(iter);
2451 /* fixed buffers import bvec */
2452 iovec.iov_base = kmap(iter->bvec->bv_page)
2454 iovec.iov_len = min(iter->count,
2455 iter->bvec->bv_len - iter->iov_offset);
2459 nr = file->f_op->read(file, iovec.iov_base,
2460 iovec.iov_len, &kiocb->ki_pos);
2462 nr = file->f_op->write(file, iovec.iov_base,
2463 iovec.iov_len, &kiocb->ki_pos);
2466 if (iov_iter_is_bvec(iter))
2467 kunmap(iter->bvec->bv_page);
2475 if (nr != iovec.iov_len)
2477 iov_iter_advance(iter, nr);
2483 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2484 struct iovec *iovec, struct iovec *fast_iov,
2485 struct iov_iter *iter)
2487 req->io->rw.nr_segs = iter->nr_segs;
2488 req->io->rw.size = io_size;
2489 req->io->rw.iov = iovec;
2490 if (!req->io->rw.iov) {
2491 req->io->rw.iov = req->io->rw.fast_iov;
2492 if (req->io->rw.iov != fast_iov)
2493 memcpy(req->io->rw.iov, fast_iov,
2494 sizeof(struct iovec) * iter->nr_segs);
2496 req->flags |= REQ_F_NEED_CLEANUP;
2500 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2502 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2503 return req->io == NULL;
2506 static int io_alloc_async_ctx(struct io_kiocb *req)
2508 if (!io_op_defs[req->opcode].async_ctx)
2511 return __io_alloc_async_ctx(req);
2514 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2515 struct iovec *iovec, struct iovec *fast_iov,
2516 struct iov_iter *iter)
2518 if (!io_op_defs[req->opcode].async_ctx)
2521 if (__io_alloc_async_ctx(req))
2524 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2529 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2530 bool force_nonblock)
2532 struct io_async_ctx *io;
2533 struct iov_iter iter;
2536 ret = io_prep_rw(req, sqe, force_nonblock);
2540 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2543 /* either don't need iovec imported or already have it */
2544 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2548 io->rw.iov = io->rw.fast_iov;
2550 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2555 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2559 static int io_read(struct io_kiocb *req, bool force_nonblock)
2561 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2562 struct kiocb *kiocb = &req->rw.kiocb;
2563 struct iov_iter iter;
2565 ssize_t io_size, ret;
2567 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2571 /* Ensure we clear previously set non-block flag */
2572 if (!force_nonblock)
2573 kiocb->ki_flags &= ~IOCB_NOWAIT;
2577 if (req->flags & REQ_F_LINK_HEAD)
2578 req->result = io_size;
2581 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2582 * we know to async punt it even if it was opened O_NONBLOCK
2584 if (force_nonblock && !io_file_supports_async(req->file, READ))
2587 iov_count = iov_iter_count(&iter);
2588 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2592 if (req->file->f_op->read_iter)
2593 ret2 = call_read_iter(req->file, kiocb, &iter);
2595 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2597 /* Catch -EAGAIN return for forced non-blocking submission */
2598 if (!force_nonblock || ret2 != -EAGAIN) {
2599 kiocb_done(kiocb, ret2);
2602 ret = io_setup_async_rw(req, io_size, iovec,
2603 inline_vecs, &iter);
2606 /* any defer here is final, must blocking retry */
2607 if (!(req->flags & REQ_F_NOWAIT) &&
2608 !file_can_poll(req->file))
2609 req->flags |= REQ_F_MUST_PUNT;
2615 req->flags &= ~REQ_F_NEED_CLEANUP;
2619 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2620 bool force_nonblock)
2622 struct io_async_ctx *io;
2623 struct iov_iter iter;
2626 ret = io_prep_rw(req, sqe, force_nonblock);
2630 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2633 req->fsize = rlimit(RLIMIT_FSIZE);
2635 /* either don't need iovec imported or already have it */
2636 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2640 io->rw.iov = io->rw.fast_iov;
2642 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2647 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2651 static int io_write(struct io_kiocb *req, bool force_nonblock)
2653 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2654 struct kiocb *kiocb = &req->rw.kiocb;
2655 struct iov_iter iter;
2657 ssize_t ret, io_size;
2659 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2663 /* Ensure we clear previously set non-block flag */
2664 if (!force_nonblock)
2665 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2669 if (req->flags & REQ_F_LINK_HEAD)
2670 req->result = io_size;
2673 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2674 * we know to async punt it even if it was opened O_NONBLOCK
2676 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2679 /* file path doesn't support NOWAIT for non-direct_IO */
2680 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2681 (req->flags & REQ_F_ISREG))
2684 iov_count = iov_iter_count(&iter);
2685 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2690 * Open-code file_start_write here to grab freeze protection,
2691 * which will be released by another thread in
2692 * io_complete_rw(). Fool lockdep by telling it the lock got
2693 * released so that it doesn't complain about the held lock when
2694 * we return to userspace.
2696 if (req->flags & REQ_F_ISREG) {
2697 __sb_start_write(file_inode(req->file)->i_sb,
2698 SB_FREEZE_WRITE, true);
2699 __sb_writers_release(file_inode(req->file)->i_sb,
2702 kiocb->ki_flags |= IOCB_WRITE;
2704 if (!force_nonblock)
2705 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2707 if (req->file->f_op->write_iter)
2708 ret2 = call_write_iter(req->file, kiocb, &iter);
2710 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2712 if (!force_nonblock)
2713 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2716 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2717 * retry them without IOCB_NOWAIT.
2719 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2721 if (!force_nonblock || ret2 != -EAGAIN) {
2722 kiocb_done(kiocb, ret2);
2725 ret = io_setup_async_rw(req, io_size, iovec,
2726 inline_vecs, &iter);
2729 /* any defer here is final, must blocking retry */
2730 if (!file_can_poll(req->file))
2731 req->flags |= REQ_F_MUST_PUNT;
2736 req->flags &= ~REQ_F_NEED_CLEANUP;
2741 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2743 struct io_splice* sp = &req->splice;
2744 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2747 if (req->flags & REQ_F_NEED_CLEANUP)
2751 sp->off_in = READ_ONCE(sqe->splice_off_in);
2752 sp->off_out = READ_ONCE(sqe->off);
2753 sp->len = READ_ONCE(sqe->len);
2754 sp->flags = READ_ONCE(sqe->splice_flags);
2756 if (unlikely(sp->flags & ~valid_flags))
2759 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2760 (sp->flags & SPLICE_F_FD_IN_FIXED));
2763 req->flags |= REQ_F_NEED_CLEANUP;
2765 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2766 req->work.flags |= IO_WQ_WORK_UNBOUND;
2771 static bool io_splice_punt(struct file *file, int rw)
2773 if (get_pipe_info(file))
2775 if (!io_file_supports_async(file, rw))
2777 return !(file->f_flags & O_NONBLOCK);
2780 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2782 struct io_splice *sp = &req->splice;
2783 struct file *in = sp->file_in;
2784 struct file *out = sp->file_out;
2785 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2786 loff_t *poff_in, *poff_out;
2789 if (force_nonblock) {
2790 if (io_splice_punt(in, READ) || io_splice_punt(out, WRITE))
2792 flags |= SPLICE_F_NONBLOCK;
2795 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2796 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2797 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2798 if (force_nonblock && ret == -EAGAIN)
2801 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2802 req->flags &= ~REQ_F_NEED_CLEANUP;
2804 io_cqring_add_event(req, ret);
2806 req_set_fail_links(req);
2812 * IORING_OP_NOP just posts a completion event, nothing else.
2814 static int io_nop(struct io_kiocb *req)
2816 struct io_ring_ctx *ctx = req->ctx;
2818 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2821 io_cqring_add_event(req, 0);
2826 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2828 struct io_ring_ctx *ctx = req->ctx;
2833 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2835 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2838 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2839 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2842 req->sync.off = READ_ONCE(sqe->off);
2843 req->sync.len = READ_ONCE(sqe->len);
2847 static bool io_req_cancelled(struct io_kiocb *req)
2849 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2850 req_set_fail_links(req);
2851 io_cqring_add_event(req, -ECANCELED);
2859 static void __io_fsync(struct io_kiocb *req)
2861 loff_t end = req->sync.off + req->sync.len;
2864 ret = vfs_fsync_range(req->file, req->sync.off,
2865 end > 0 ? end : LLONG_MAX,
2866 req->sync.flags & IORING_FSYNC_DATASYNC);
2868 req_set_fail_links(req);
2869 io_cqring_add_event(req, ret);
2873 static void io_fsync_finish(struct io_wq_work **workptr)
2875 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2877 if (io_req_cancelled(req))
2880 io_steal_work(req, workptr);
2883 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2885 /* fsync always requires a blocking context */
2886 if (force_nonblock) {
2887 req->work.func = io_fsync_finish;
2894 static void __io_fallocate(struct io_kiocb *req)
2898 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2899 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2901 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2903 req_set_fail_links(req);
2904 io_cqring_add_event(req, ret);
2908 static void io_fallocate_finish(struct io_wq_work **workptr)
2910 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2912 if (io_req_cancelled(req))
2914 __io_fallocate(req);
2915 io_steal_work(req, workptr);
2918 static int io_fallocate_prep(struct io_kiocb *req,
2919 const struct io_uring_sqe *sqe)
2921 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2924 req->sync.off = READ_ONCE(sqe->off);
2925 req->sync.len = READ_ONCE(sqe->addr);
2926 req->sync.mode = READ_ONCE(sqe->len);
2927 req->fsize = rlimit(RLIMIT_FSIZE);
2931 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2933 /* fallocate always requiring blocking context */
2934 if (force_nonblock) {
2935 req->work.func = io_fallocate_finish;
2939 __io_fallocate(req);
2943 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2945 const char __user *fname;
2948 if (sqe->ioprio || sqe->buf_index)
2950 if (req->flags & REQ_F_FIXED_FILE)
2952 if (req->flags & REQ_F_NEED_CLEANUP)
2955 req->open.dfd = READ_ONCE(sqe->fd);
2956 req->open.how.mode = READ_ONCE(sqe->len);
2957 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2958 req->open.how.flags = READ_ONCE(sqe->open_flags);
2959 if (force_o_largefile())
2960 req->open.how.flags |= O_LARGEFILE;
2962 req->open.filename = getname(fname);
2963 if (IS_ERR(req->open.filename)) {
2964 ret = PTR_ERR(req->open.filename);
2965 req->open.filename = NULL;
2969 req->open.nofile = rlimit(RLIMIT_NOFILE);
2970 req->flags |= REQ_F_NEED_CLEANUP;
2974 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2976 struct open_how __user *how;
2977 const char __user *fname;
2981 if (sqe->ioprio || sqe->buf_index)
2983 if (req->flags & REQ_F_FIXED_FILE)
2985 if (req->flags & REQ_F_NEED_CLEANUP)
2988 req->open.dfd = READ_ONCE(sqe->fd);
2989 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2990 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2991 len = READ_ONCE(sqe->len);
2993 if (len < OPEN_HOW_SIZE_VER0)
2996 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3001 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3002 req->open.how.flags |= O_LARGEFILE;
3004 req->open.filename = getname(fname);
3005 if (IS_ERR(req->open.filename)) {
3006 ret = PTR_ERR(req->open.filename);
3007 req->open.filename = NULL;
3011 req->open.nofile = rlimit(RLIMIT_NOFILE);
3012 req->flags |= REQ_F_NEED_CLEANUP;
3016 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3018 struct open_flags op;
3025 ret = build_open_flags(&req->open.how, &op);
3029 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3033 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3036 ret = PTR_ERR(file);
3038 fsnotify_open(file);
3039 fd_install(ret, file);
3042 putname(req->open.filename);
3043 req->flags &= ~REQ_F_NEED_CLEANUP;
3045 req_set_fail_links(req);
3046 io_cqring_add_event(req, ret);
3051 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3053 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3054 return io_openat2(req, force_nonblock);
3057 static int io_remove_buffers_prep(struct io_kiocb *req,
3058 const struct io_uring_sqe *sqe)
3060 struct io_provide_buf *p = &req->pbuf;
3063 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3066 tmp = READ_ONCE(sqe->fd);
3067 if (!tmp || tmp > USHRT_MAX)
3070 memset(p, 0, sizeof(*p));
3072 p->bgid = READ_ONCE(sqe->buf_group);
3076 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3077 int bgid, unsigned nbufs)
3081 /* shouldn't happen */
3085 /* the head kbuf is the list itself */
3086 while (!list_empty(&buf->list)) {
3087 struct io_buffer *nxt;
3089 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3090 list_del(&nxt->list);
3097 idr_remove(&ctx->io_buffer_idr, bgid);
3102 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3104 struct io_provide_buf *p = &req->pbuf;
3105 struct io_ring_ctx *ctx = req->ctx;
3106 struct io_buffer *head;
3109 io_ring_submit_lock(ctx, !force_nonblock);
3111 lockdep_assert_held(&ctx->uring_lock);
3114 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3116 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3118 io_ring_submit_lock(ctx, !force_nonblock);
3120 req_set_fail_links(req);
3121 io_cqring_add_event(req, ret);
3126 static int io_provide_buffers_prep(struct io_kiocb *req,
3127 const struct io_uring_sqe *sqe)
3129 struct io_provide_buf *p = &req->pbuf;
3132 if (sqe->ioprio || sqe->rw_flags)
3135 tmp = READ_ONCE(sqe->fd);
3136 if (!tmp || tmp > USHRT_MAX)
3139 p->addr = READ_ONCE(sqe->addr);
3140 p->len = READ_ONCE(sqe->len);
3142 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3145 p->bgid = READ_ONCE(sqe->buf_group);
3146 tmp = READ_ONCE(sqe->off);
3147 if (tmp > USHRT_MAX)
3153 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3155 struct io_buffer *buf;
3156 u64 addr = pbuf->addr;
3157 int i, bid = pbuf->bid;
3159 for (i = 0; i < pbuf->nbufs; i++) {
3160 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3165 buf->len = pbuf->len;
3170 INIT_LIST_HEAD(&buf->list);
3173 list_add_tail(&buf->list, &(*head)->list);
3177 return i ? i : -ENOMEM;
3180 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3182 struct io_provide_buf *p = &req->pbuf;
3183 struct io_ring_ctx *ctx = req->ctx;
3184 struct io_buffer *head, *list;
3187 io_ring_submit_lock(ctx, !force_nonblock);
3189 lockdep_assert_held(&ctx->uring_lock);
3191 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3193 ret = io_add_buffers(p, &head);
3198 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3201 __io_remove_buffers(ctx, head, p->bgid, -1U);
3206 io_ring_submit_unlock(ctx, !force_nonblock);
3208 req_set_fail_links(req);
3209 io_cqring_add_event(req, ret);
3214 static int io_epoll_ctl_prep(struct io_kiocb *req,
3215 const struct io_uring_sqe *sqe)
3217 #if defined(CONFIG_EPOLL)
3218 if (sqe->ioprio || sqe->buf_index)
3221 req->epoll.epfd = READ_ONCE(sqe->fd);
3222 req->epoll.op = READ_ONCE(sqe->len);
3223 req->epoll.fd = READ_ONCE(sqe->off);
3225 if (ep_op_has_event(req->epoll.op)) {
3226 struct epoll_event __user *ev;
3228 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3229 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3239 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3241 #if defined(CONFIG_EPOLL)
3242 struct io_epoll *ie = &req->epoll;
3245 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3246 if (force_nonblock && ret == -EAGAIN)
3250 req_set_fail_links(req);
3251 io_cqring_add_event(req, ret);
3259 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3261 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3262 if (sqe->ioprio || sqe->buf_index || sqe->off)
3265 req->madvise.addr = READ_ONCE(sqe->addr);
3266 req->madvise.len = READ_ONCE(sqe->len);
3267 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3274 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3276 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3277 struct io_madvise *ma = &req->madvise;
3283 ret = do_madvise(ma->addr, ma->len, ma->advice);
3285 req_set_fail_links(req);
3286 io_cqring_add_event(req, ret);
3294 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3296 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3299 req->fadvise.offset = READ_ONCE(sqe->off);
3300 req->fadvise.len = READ_ONCE(sqe->len);
3301 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3305 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3307 struct io_fadvise *fa = &req->fadvise;
3310 if (force_nonblock) {
3311 switch (fa->advice) {
3312 case POSIX_FADV_NORMAL:
3313 case POSIX_FADV_RANDOM:
3314 case POSIX_FADV_SEQUENTIAL:
3321 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3323 req_set_fail_links(req);
3324 io_cqring_add_event(req, ret);
3329 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3331 const char __user *fname;
3332 unsigned lookup_flags;
3335 if (sqe->ioprio || sqe->buf_index)
3337 if (req->flags & REQ_F_FIXED_FILE)
3339 if (req->flags & REQ_F_NEED_CLEANUP)
3342 req->open.dfd = READ_ONCE(sqe->fd);
3343 req->open.mask = READ_ONCE(sqe->len);
3344 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3345 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3346 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3348 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3351 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3352 if (IS_ERR(req->open.filename)) {
3353 ret = PTR_ERR(req->open.filename);
3354 req->open.filename = NULL;
3358 req->flags |= REQ_F_NEED_CLEANUP;
3362 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3364 struct io_open *ctx = &req->open;
3365 unsigned lookup_flags;
3370 if (force_nonblock) {
3371 /* only need file table for an actual valid fd */
3372 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3373 req->flags |= REQ_F_NO_FILE_TABLE;
3377 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3381 /* filename_lookup() drops it, keep a reference */
3382 ctx->filename->refcnt++;
3384 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3389 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3391 if (retry_estale(ret, lookup_flags)) {
3392 lookup_flags |= LOOKUP_REVAL;
3396 ret = cp_statx(&stat, ctx->buffer);
3398 putname(ctx->filename);
3399 req->flags &= ~REQ_F_NEED_CLEANUP;
3401 req_set_fail_links(req);
3402 io_cqring_add_event(req, ret);
3407 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3410 * If we queue this for async, it must not be cancellable. That would
3411 * leave the 'file' in an undeterminate state.
3413 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3415 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3416 sqe->rw_flags || sqe->buf_index)
3418 if (req->flags & REQ_F_FIXED_FILE)
3421 req->close.fd = READ_ONCE(sqe->fd);
3422 if (req->file->f_op == &io_uring_fops ||
3423 req->close.fd == req->ctx->ring_fd)
3429 /* only called when __close_fd_get_file() is done */
3430 static void __io_close_finish(struct io_kiocb *req)
3434 ret = filp_close(req->close.put_file, req->work.files);
3436 req_set_fail_links(req);
3437 io_cqring_add_event(req, ret);
3438 fput(req->close.put_file);
3442 static void io_close_finish(struct io_wq_work **workptr)
3444 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3446 /* not cancellable, don't do io_req_cancelled() */
3447 __io_close_finish(req);
3448 io_steal_work(req, workptr);
3451 static int io_close(struct io_kiocb *req, bool force_nonblock)
3455 req->close.put_file = NULL;
3456 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3460 /* if the file has a flush method, be safe and punt to async */
3461 if (req->close.put_file->f_op->flush && force_nonblock) {
3462 /* submission ref will be dropped, take it for async */
3463 refcount_inc(&req->refs);
3465 req->work.func = io_close_finish;
3467 * Do manual async queue here to avoid grabbing files - we don't
3468 * need the files, and it'll cause io_close_finish() to close
3469 * the file again and cause a double CQE entry for this request
3471 io_queue_async_work(req);
3476 * No ->flush(), safely close from here and just punt the
3477 * fput() to async context.
3479 __io_close_finish(req);
3483 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3485 struct io_ring_ctx *ctx = req->ctx;
3490 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3492 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3495 req->sync.off = READ_ONCE(sqe->off);
3496 req->sync.len = READ_ONCE(sqe->len);
3497 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3501 static void __io_sync_file_range(struct io_kiocb *req)
3505 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3508 req_set_fail_links(req);
3509 io_cqring_add_event(req, ret);
3514 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3516 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3518 if (io_req_cancelled(req))
3520 __io_sync_file_range(req);
3521 io_put_req(req); /* put submission ref */
3524 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3526 /* sync_file_range always requires a blocking context */
3527 if (force_nonblock) {
3528 req->work.func = io_sync_file_range_finish;
3532 __io_sync_file_range(req);
3536 #if defined(CONFIG_NET)
3537 static int io_setup_async_msg(struct io_kiocb *req,
3538 struct io_async_msghdr *kmsg)
3542 if (io_alloc_async_ctx(req)) {
3543 if (kmsg->iov != kmsg->fast_iov)
3547 req->flags |= REQ_F_NEED_CLEANUP;
3548 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3552 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3554 struct io_sr_msg *sr = &req->sr_msg;
3555 struct io_async_ctx *io = req->io;
3558 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3559 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3560 sr->len = READ_ONCE(sqe->len);
3562 #ifdef CONFIG_COMPAT
3563 if (req->ctx->compat)
3564 sr->msg_flags |= MSG_CMSG_COMPAT;
3567 if (!io || req->opcode == IORING_OP_SEND)
3569 /* iovec is already imported */
3570 if (req->flags & REQ_F_NEED_CLEANUP)
3573 io->msg.iov = io->msg.fast_iov;
3574 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3577 req->flags |= REQ_F_NEED_CLEANUP;
3581 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3583 struct io_async_msghdr *kmsg = NULL;
3584 struct socket *sock;
3587 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3590 sock = sock_from_file(req->file, &ret);
3592 struct io_async_ctx io;
3596 kmsg = &req->io->msg;
3597 kmsg->msg.msg_name = &req->io->msg.addr;
3598 /* if iov is set, it's allocated already */
3600 kmsg->iov = kmsg->fast_iov;
3601 kmsg->msg.msg_iter.iov = kmsg->iov;
3603 struct io_sr_msg *sr = &req->sr_msg;
3606 kmsg->msg.msg_name = &io.msg.addr;
3608 io.msg.iov = io.msg.fast_iov;
3609 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3610 sr->msg_flags, &io.msg.iov);
3615 flags = req->sr_msg.msg_flags;
3616 if (flags & MSG_DONTWAIT)
3617 req->flags |= REQ_F_NOWAIT;
3618 else if (force_nonblock)
3619 flags |= MSG_DONTWAIT;
3621 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3622 if (force_nonblock && ret == -EAGAIN)
3623 return io_setup_async_msg(req, kmsg);
3624 if (ret == -ERESTARTSYS)
3628 if (kmsg && kmsg->iov != kmsg->fast_iov)
3630 req->flags &= ~REQ_F_NEED_CLEANUP;
3631 io_cqring_add_event(req, ret);
3633 req_set_fail_links(req);
3638 static int io_send(struct io_kiocb *req, bool force_nonblock)
3640 struct socket *sock;
3643 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3646 sock = sock_from_file(req->file, &ret);
3648 struct io_sr_msg *sr = &req->sr_msg;
3653 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3658 msg.msg_name = NULL;
3659 msg.msg_control = NULL;
3660 msg.msg_controllen = 0;
3661 msg.msg_namelen = 0;
3663 flags = req->sr_msg.msg_flags;
3664 if (flags & MSG_DONTWAIT)
3665 req->flags |= REQ_F_NOWAIT;
3666 else if (force_nonblock)
3667 flags |= MSG_DONTWAIT;
3669 msg.msg_flags = flags;
3670 ret = sock_sendmsg(sock, &msg);
3671 if (force_nonblock && ret == -EAGAIN)
3673 if (ret == -ERESTARTSYS)
3677 io_cqring_add_event(req, ret);
3679 req_set_fail_links(req);
3684 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3686 struct io_sr_msg *sr = &req->sr_msg;
3687 struct iovec __user *uiov;
3691 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3696 if (req->flags & REQ_F_BUFFER_SELECT) {
3699 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3701 sr->len = io->msg.iov[0].iov_len;
3702 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3706 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3707 &io->msg.iov, &io->msg.msg.msg_iter);
3715 #ifdef CONFIG_COMPAT
3716 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3717 struct io_async_ctx *io)
3719 struct compat_msghdr __user *msg_compat;
3720 struct io_sr_msg *sr = &req->sr_msg;
3721 struct compat_iovec __user *uiov;
3726 msg_compat = (struct compat_msghdr __user *) sr->msg;
3727 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3732 uiov = compat_ptr(ptr);
3733 if (req->flags & REQ_F_BUFFER_SELECT) {
3734 compat_ssize_t clen;
3738 if (!access_ok(uiov, sizeof(*uiov)))
3740 if (__get_user(clen, &uiov->iov_len))
3744 sr->len = io->msg.iov[0].iov_len;
3747 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3749 &io->msg.msg.msg_iter);
3758 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3760 io->msg.iov = io->msg.fast_iov;
3762 #ifdef CONFIG_COMPAT
3763 if (req->ctx->compat)
3764 return __io_compat_recvmsg_copy_hdr(req, io);
3767 return __io_recvmsg_copy_hdr(req, io);
3770 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3771 int *cflags, bool needs_lock)
3773 struct io_sr_msg *sr = &req->sr_msg;
3774 struct io_buffer *kbuf;
3776 if (!(req->flags & REQ_F_BUFFER_SELECT))
3779 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3784 req->flags |= REQ_F_BUFFER_SELECTED;
3786 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3787 *cflags |= IORING_CQE_F_BUFFER;
3791 static int io_recvmsg_prep(struct io_kiocb *req,
3792 const struct io_uring_sqe *sqe)
3794 struct io_sr_msg *sr = &req->sr_msg;
3795 struct io_async_ctx *io = req->io;
3798 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3799 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3800 sr->len = READ_ONCE(sqe->len);
3801 sr->bgid = READ_ONCE(sqe->buf_group);
3803 #ifdef CONFIG_COMPAT
3804 if (req->ctx->compat)
3805 sr->msg_flags |= MSG_CMSG_COMPAT;
3808 if (!io || req->opcode == IORING_OP_RECV)
3810 /* iovec is already imported */
3811 if (req->flags & REQ_F_NEED_CLEANUP)
3814 ret = io_recvmsg_copy_hdr(req, io);
3816 req->flags |= REQ_F_NEED_CLEANUP;
3820 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3822 struct io_async_msghdr *kmsg = NULL;
3823 struct socket *sock;
3824 int ret, cflags = 0;
3826 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3829 sock = sock_from_file(req->file, &ret);
3831 struct io_buffer *kbuf;
3832 struct io_async_ctx io;
3836 kmsg = &req->io->msg;
3837 kmsg->msg.msg_name = &req->io->msg.addr;
3838 /* if iov is set, it's allocated already */
3840 kmsg->iov = kmsg->fast_iov;
3841 kmsg->msg.msg_iter.iov = kmsg->iov;
3844 kmsg->msg.msg_name = &io.msg.addr;
3846 ret = io_recvmsg_copy_hdr(req, &io);
3851 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3853 return PTR_ERR(kbuf);
3855 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3856 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3857 1, req->sr_msg.len);
3860 flags = req->sr_msg.msg_flags;
3861 if (flags & MSG_DONTWAIT)
3862 req->flags |= REQ_F_NOWAIT;
3863 else if (force_nonblock)
3864 flags |= MSG_DONTWAIT;
3866 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3867 kmsg->uaddr, flags);
3868 if (force_nonblock && ret == -EAGAIN)
3869 return io_setup_async_msg(req, kmsg);
3870 if (ret == -ERESTARTSYS)
3874 if (kmsg && kmsg->iov != kmsg->fast_iov)
3876 req->flags &= ~REQ_F_NEED_CLEANUP;
3877 __io_cqring_add_event(req, ret, cflags);
3879 req_set_fail_links(req);
3884 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3886 struct io_buffer *kbuf = NULL;
3887 struct socket *sock;
3888 int ret, cflags = 0;
3890 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3893 sock = sock_from_file(req->file, &ret);
3895 struct io_sr_msg *sr = &req->sr_msg;
3896 void __user *buf = sr->buf;
3901 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3903 return PTR_ERR(kbuf);
3905 buf = u64_to_user_ptr(kbuf->addr);
3907 ret = import_single_range(READ, buf, sr->len, &iov,
3914 req->flags |= REQ_F_NEED_CLEANUP;
3915 msg.msg_name = NULL;
3916 msg.msg_control = NULL;
3917 msg.msg_controllen = 0;
3918 msg.msg_namelen = 0;
3919 msg.msg_iocb = NULL;
3922 flags = req->sr_msg.msg_flags;
3923 if (flags & MSG_DONTWAIT)
3924 req->flags |= REQ_F_NOWAIT;
3925 else if (force_nonblock)
3926 flags |= MSG_DONTWAIT;
3928 ret = sock_recvmsg(sock, &msg, flags);
3929 if (force_nonblock && ret == -EAGAIN)
3931 if (ret == -ERESTARTSYS)
3936 req->flags &= ~REQ_F_NEED_CLEANUP;
3937 __io_cqring_add_event(req, ret, cflags);
3939 req_set_fail_links(req);
3944 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3946 struct io_accept *accept = &req->accept;
3948 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3950 if (sqe->ioprio || sqe->len || sqe->buf_index)
3953 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3954 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3955 accept->flags = READ_ONCE(sqe->accept_flags);
3956 accept->nofile = rlimit(RLIMIT_NOFILE);
3960 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3962 struct io_accept *accept = &req->accept;
3963 unsigned file_flags;
3966 file_flags = force_nonblock ? O_NONBLOCK : 0;
3967 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3968 accept->addr_len, accept->flags,
3970 if (ret == -EAGAIN && force_nonblock)
3972 if (ret == -ERESTARTSYS)
3975 req_set_fail_links(req);
3976 io_cqring_add_event(req, ret);
3981 static void io_accept_finish(struct io_wq_work **workptr)
3983 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3985 if (io_req_cancelled(req))
3987 __io_accept(req, false);
3988 io_steal_work(req, workptr);
3991 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3995 ret = __io_accept(req, force_nonblock);
3996 if (ret == -EAGAIN && force_nonblock) {
3997 req->work.func = io_accept_finish;
4003 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4005 struct io_connect *conn = &req->connect;
4006 struct io_async_ctx *io = req->io;
4008 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4010 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4013 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4014 conn->addr_len = READ_ONCE(sqe->addr2);
4019 return move_addr_to_kernel(conn->addr, conn->addr_len,
4020 &io->connect.address);
4023 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4025 struct io_async_ctx __io, *io;
4026 unsigned file_flags;
4032 ret = move_addr_to_kernel(req->connect.addr,
4033 req->connect.addr_len,
4034 &__io.connect.address);
4040 file_flags = force_nonblock ? O_NONBLOCK : 0;
4042 ret = __sys_connect_file(req->file, &io->connect.address,
4043 req->connect.addr_len, file_flags);
4044 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4047 if (io_alloc_async_ctx(req)) {
4051 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4054 if (ret == -ERESTARTSYS)
4058 req_set_fail_links(req);
4059 io_cqring_add_event(req, ret);
4063 #else /* !CONFIG_NET */
4064 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4069 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4074 static int io_send(struct io_kiocb *req, bool force_nonblock)
4079 static int io_recvmsg_prep(struct io_kiocb *req,
4080 const struct io_uring_sqe *sqe)
4085 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4090 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4095 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4100 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4105 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4110 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4114 #endif /* CONFIG_NET */
4116 struct io_poll_table {
4117 struct poll_table_struct pt;
4118 struct io_kiocb *req;
4122 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4123 struct wait_queue_head *head)
4125 if (unlikely(poll->head)) {
4126 pt->error = -EINVAL;
4132 add_wait_queue(head, &poll->wait);
4135 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4136 struct poll_table_struct *p)
4138 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4140 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4143 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4144 __poll_t mask, task_work_func_t func)
4146 struct task_struct *tsk;
4149 /* for instances that support it check for an event match first: */
4150 if (mask && !(mask & poll->events))
4153 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4155 list_del_init(&poll->wait.entry);
4159 init_task_work(&req->task_work, func);
4161 * If this fails, then the task is exiting. Punt to one of the io-wq
4162 * threads to ensure the work gets run, we can't always rely on exit
4163 * cancelation taking care of this.
4165 ret = task_work_add(tsk, &req->task_work, true);
4166 if (unlikely(ret)) {
4167 tsk = io_wq_get_task(req->ctx->io_wq);
4168 task_work_add(tsk, &req->task_work, true);
4170 wake_up_process(tsk);
4174 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4175 __acquires(&req->ctx->completion_lock)
4177 struct io_ring_ctx *ctx = req->ctx;
4179 if (!req->result && !READ_ONCE(poll->canceled)) {
4180 struct poll_table_struct pt = { ._key = poll->events };
4182 req->result = vfs_poll(req->file, &pt) & poll->events;
4185 spin_lock_irq(&ctx->completion_lock);
4186 if (!req->result && !READ_ONCE(poll->canceled)) {
4187 add_wait_queue(poll->head, &poll->wait);
4194 static void io_async_task_func(struct callback_head *cb)
4196 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4197 struct async_poll *apoll = req->apoll;
4198 struct io_ring_ctx *ctx = req->ctx;
4201 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4203 if (io_poll_rewait(req, &apoll->poll)) {
4204 spin_unlock_irq(&ctx->completion_lock);
4208 if (hash_hashed(&req->hash_node))
4209 hash_del(&req->hash_node);
4211 canceled = READ_ONCE(apoll->poll.canceled);
4213 io_cqring_fill_event(req, -ECANCELED);
4214 io_commit_cqring(ctx);
4217 spin_unlock_irq(&ctx->completion_lock);
4219 /* restore ->work in case we need to retry again */
4220 memcpy(&req->work, &apoll->work, sizeof(req->work));
4224 io_cqring_ev_posted(ctx);
4225 req_set_fail_links(req);
4226 io_double_put_req(req);
4230 __set_current_state(TASK_RUNNING);
4231 mutex_lock(&ctx->uring_lock);
4232 __io_queue_sqe(req, NULL);
4233 mutex_unlock(&ctx->uring_lock);
4238 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4241 struct io_kiocb *req = wait->private;
4242 struct io_poll_iocb *poll = &req->apoll->poll;
4244 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4247 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4250 static void io_poll_req_insert(struct io_kiocb *req)
4252 struct io_ring_ctx *ctx = req->ctx;
4253 struct hlist_head *list;
4255 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4256 hlist_add_head(&req->hash_node, list);
4259 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4260 struct io_poll_iocb *poll,
4261 struct io_poll_table *ipt, __poll_t mask,
4262 wait_queue_func_t wake_func)
4263 __acquires(&ctx->completion_lock)
4265 struct io_ring_ctx *ctx = req->ctx;
4266 bool cancel = false;
4268 poll->file = req->file;
4270 poll->done = poll->canceled = false;
4271 poll->events = mask;
4273 ipt->pt._key = mask;
4275 ipt->error = -EINVAL;
4277 INIT_LIST_HEAD(&poll->wait.entry);
4278 init_waitqueue_func_entry(&poll->wait, wake_func);
4279 poll->wait.private = req;
4281 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4283 spin_lock_irq(&ctx->completion_lock);
4284 if (likely(poll->head)) {
4285 spin_lock(&poll->head->lock);
4286 if (unlikely(list_empty(&poll->wait.entry))) {
4292 if (mask || ipt->error)
4293 list_del_init(&poll->wait.entry);
4295 WRITE_ONCE(poll->canceled, true);
4296 else if (!poll->done) /* actually waiting for an event */
4297 io_poll_req_insert(req);
4298 spin_unlock(&poll->head->lock);
4304 static bool io_arm_poll_handler(struct io_kiocb *req)
4306 const struct io_op_def *def = &io_op_defs[req->opcode];
4307 struct io_ring_ctx *ctx = req->ctx;
4308 struct async_poll *apoll;
4309 struct io_poll_table ipt;
4312 if (!req->file || !file_can_poll(req->file))
4314 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4316 if (!def->pollin && !def->pollout)
4319 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4320 if (unlikely(!apoll))
4323 req->flags |= REQ_F_POLLED;
4324 memcpy(&apoll->work, &req->work, sizeof(req->work));
4326 get_task_struct(current);
4327 req->task = current;
4329 INIT_HLIST_NODE(&req->hash_node);
4333 mask |= POLLIN | POLLRDNORM;
4335 mask |= POLLOUT | POLLWRNORM;
4336 mask |= POLLERR | POLLPRI;
4338 ipt.pt._qproc = io_async_queue_proc;
4340 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4344 apoll->poll.done = true;
4345 spin_unlock_irq(&ctx->completion_lock);
4346 memcpy(&req->work, &apoll->work, sizeof(req->work));
4350 spin_unlock_irq(&ctx->completion_lock);
4351 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4352 apoll->poll.events);
4356 static bool __io_poll_remove_one(struct io_kiocb *req,
4357 struct io_poll_iocb *poll)
4359 bool do_complete = false;
4361 spin_lock(&poll->head->lock);
4362 WRITE_ONCE(poll->canceled, true);
4363 if (!list_empty(&poll->wait.entry)) {
4364 list_del_init(&poll->wait.entry);
4367 spin_unlock(&poll->head->lock);
4371 static bool io_poll_remove_one(struct io_kiocb *req)
4373 struct async_poll *apoll = NULL;
4376 if (req->opcode == IORING_OP_POLL_ADD) {
4377 do_complete = __io_poll_remove_one(req, &req->poll);
4380 /* non-poll requests have submit ref still */
4381 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4386 hash_del(&req->hash_node);
4388 if (do_complete && apoll) {
4390 * restore ->work because we need to call io_req_work_drop_env.
4392 memcpy(&req->work, &apoll->work, sizeof(req->work));
4397 io_cqring_fill_event(req, -ECANCELED);
4398 io_commit_cqring(req->ctx);
4399 req->flags |= REQ_F_COMP_LOCKED;
4406 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4408 struct hlist_node *tmp;
4409 struct io_kiocb *req;
4412 spin_lock_irq(&ctx->completion_lock);
4413 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4414 struct hlist_head *list;
4416 list = &ctx->cancel_hash[i];
4417 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4418 posted += io_poll_remove_one(req);
4420 spin_unlock_irq(&ctx->completion_lock);
4423 io_cqring_ev_posted(ctx);
4426 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4428 struct hlist_head *list;
4429 struct io_kiocb *req;
4431 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4432 hlist_for_each_entry(req, list, hash_node) {
4433 if (sqe_addr != req->user_data)
4435 if (io_poll_remove_one(req))
4443 static int io_poll_remove_prep(struct io_kiocb *req,
4444 const struct io_uring_sqe *sqe)
4446 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4448 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4452 req->poll.addr = READ_ONCE(sqe->addr);
4457 * Find a running poll command that matches one specified in sqe->addr,
4458 * and remove it if found.
4460 static int io_poll_remove(struct io_kiocb *req)
4462 struct io_ring_ctx *ctx = req->ctx;
4466 addr = req->poll.addr;
4467 spin_lock_irq(&ctx->completion_lock);
4468 ret = io_poll_cancel(ctx, addr);
4469 spin_unlock_irq(&ctx->completion_lock);
4471 io_cqring_add_event(req, ret);
4473 req_set_fail_links(req);
4478 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4480 struct io_ring_ctx *ctx = req->ctx;
4482 req->poll.done = true;
4483 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4484 io_commit_cqring(ctx);
4487 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4489 struct io_ring_ctx *ctx = req->ctx;
4490 struct io_poll_iocb *poll = &req->poll;
4492 if (io_poll_rewait(req, poll)) {
4493 spin_unlock_irq(&ctx->completion_lock);
4497 hash_del(&req->hash_node);
4498 io_poll_complete(req, req->result, 0);
4499 req->flags |= REQ_F_COMP_LOCKED;
4500 io_put_req_find_next(req, nxt);
4501 spin_unlock_irq(&ctx->completion_lock);
4503 io_cqring_ev_posted(ctx);
4506 static void io_poll_task_func(struct callback_head *cb)
4508 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4509 struct io_kiocb *nxt = NULL;
4511 io_poll_task_handler(req, &nxt);
4513 struct io_ring_ctx *ctx = nxt->ctx;
4515 mutex_lock(&ctx->uring_lock);
4516 __io_queue_sqe(nxt, NULL);
4517 mutex_unlock(&ctx->uring_lock);
4521 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4524 struct io_kiocb *req = wait->private;
4525 struct io_poll_iocb *poll = &req->poll;
4527 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4530 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4531 struct poll_table_struct *p)
4533 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4535 __io_queue_proc(&pt->req->poll, pt, head);
4538 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4540 struct io_poll_iocb *poll = &req->poll;
4543 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4545 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4550 events = READ_ONCE(sqe->poll_events);
4551 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4553 get_task_struct(current);
4554 req->task = current;
4558 static int io_poll_add(struct io_kiocb *req)
4560 struct io_poll_iocb *poll = &req->poll;
4561 struct io_ring_ctx *ctx = req->ctx;
4562 struct io_poll_table ipt;
4565 INIT_HLIST_NODE(&req->hash_node);
4566 INIT_LIST_HEAD(&req->list);
4567 ipt.pt._qproc = io_poll_queue_proc;
4569 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4572 if (mask) { /* no async, we'd stolen it */
4574 io_poll_complete(req, mask, 0);
4576 spin_unlock_irq(&ctx->completion_lock);
4579 io_cqring_ev_posted(ctx);
4585 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4587 struct io_timeout_data *data = container_of(timer,
4588 struct io_timeout_data, timer);
4589 struct io_kiocb *req = data->req;
4590 struct io_ring_ctx *ctx = req->ctx;
4591 unsigned long flags;
4593 atomic_inc(&ctx->cq_timeouts);
4595 spin_lock_irqsave(&ctx->completion_lock, flags);
4597 * We could be racing with timeout deletion. If the list is empty,
4598 * then timeout lookup already found it and will be handling it.
4600 if (!list_empty(&req->list)) {
4601 struct io_kiocb *prev;
4604 * Adjust the reqs sequence before the current one because it
4605 * will consume a slot in the cq_ring and the cq_tail
4606 * pointer will be increased, otherwise other timeout reqs may
4607 * return in advance without waiting for enough wait_nr.
4610 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4612 list_del_init(&req->list);
4615 io_cqring_fill_event(req, -ETIME);
4616 io_commit_cqring(ctx);
4617 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4619 io_cqring_ev_posted(ctx);
4620 req_set_fail_links(req);
4622 return HRTIMER_NORESTART;
4625 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4627 struct io_kiocb *req;
4630 list_for_each_entry(req, &ctx->timeout_list, list) {
4631 if (user_data == req->user_data) {
4632 list_del_init(&req->list);
4641 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4645 req_set_fail_links(req);
4646 io_cqring_fill_event(req, -ECANCELED);
4651 static int io_timeout_remove_prep(struct io_kiocb *req,
4652 const struct io_uring_sqe *sqe)
4654 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4656 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4659 req->timeout.addr = READ_ONCE(sqe->addr);
4660 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4661 if (req->timeout.flags)
4668 * Remove or update an existing timeout command
4670 static int io_timeout_remove(struct io_kiocb *req)
4672 struct io_ring_ctx *ctx = req->ctx;
4675 spin_lock_irq(&ctx->completion_lock);
4676 ret = io_timeout_cancel(ctx, req->timeout.addr);
4678 io_cqring_fill_event(req, ret);
4679 io_commit_cqring(ctx);
4680 spin_unlock_irq(&ctx->completion_lock);
4681 io_cqring_ev_posted(ctx);
4683 req_set_fail_links(req);
4688 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4689 bool is_timeout_link)
4691 struct io_timeout_data *data;
4694 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4696 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4698 if (sqe->off && is_timeout_link)
4700 flags = READ_ONCE(sqe->timeout_flags);
4701 if (flags & ~IORING_TIMEOUT_ABS)
4704 req->timeout.count = READ_ONCE(sqe->off);
4706 if (!req->io && io_alloc_async_ctx(req))
4709 data = &req->io->timeout;
4711 req->flags |= REQ_F_TIMEOUT;
4713 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4716 if (flags & IORING_TIMEOUT_ABS)
4717 data->mode = HRTIMER_MODE_ABS;
4719 data->mode = HRTIMER_MODE_REL;
4721 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4725 static int io_timeout(struct io_kiocb *req)
4727 struct io_ring_ctx *ctx = req->ctx;
4728 struct io_timeout_data *data;
4729 struct list_head *entry;
4731 u32 count = req->timeout.count;
4732 u32 seq = req->sequence;
4734 data = &req->io->timeout;
4737 * sqe->off holds how many events that need to occur for this
4738 * timeout event to be satisfied. If it isn't set, then this is
4739 * a pure timeout request, sequence isn't used.
4742 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4743 spin_lock_irq(&ctx->completion_lock);
4744 entry = ctx->timeout_list.prev;
4748 req->sequence = seq + count;
4751 * Insertion sort, ensuring the first entry in the list is always
4752 * the one we need first.
4754 spin_lock_irq(&ctx->completion_lock);
4755 list_for_each_prev(entry, &ctx->timeout_list) {
4756 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4758 long long tmp, tmp_nxt;
4759 u32 nxt_offset = nxt->timeout.count;
4761 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4765 * Since seq + count can overflow, use type long
4768 tmp = (long long)seq + count;
4769 nxt_seq = nxt->sequence - nxt_offset;
4770 tmp_nxt = (long long)nxt_seq + nxt_offset;
4773 * cached_sq_head may overflow, and it will never overflow twice
4774 * once there is some timeout req still be valid.
4783 * Sequence of reqs after the insert one and itself should
4784 * be adjusted because each timeout req consumes a slot.
4789 req->sequence -= span;
4791 list_add(&req->list, entry);
4792 data->timer.function = io_timeout_fn;
4793 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4794 spin_unlock_irq(&ctx->completion_lock);
4798 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4800 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4802 return req->user_data == (unsigned long) data;
4805 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4807 enum io_wq_cancel cancel_ret;
4810 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4811 switch (cancel_ret) {
4812 case IO_WQ_CANCEL_OK:
4815 case IO_WQ_CANCEL_RUNNING:
4818 case IO_WQ_CANCEL_NOTFOUND:
4826 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4827 struct io_kiocb *req, __u64 sqe_addr,
4830 unsigned long flags;
4833 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4834 if (ret != -ENOENT) {
4835 spin_lock_irqsave(&ctx->completion_lock, flags);
4839 spin_lock_irqsave(&ctx->completion_lock, flags);
4840 ret = io_timeout_cancel(ctx, sqe_addr);
4843 ret = io_poll_cancel(ctx, sqe_addr);
4847 io_cqring_fill_event(req, ret);
4848 io_commit_cqring(ctx);
4849 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4850 io_cqring_ev_posted(ctx);
4853 req_set_fail_links(req);
4857 static int io_async_cancel_prep(struct io_kiocb *req,
4858 const struct io_uring_sqe *sqe)
4860 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4862 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4866 req->cancel.addr = READ_ONCE(sqe->addr);
4870 static int io_async_cancel(struct io_kiocb *req)
4872 struct io_ring_ctx *ctx = req->ctx;
4874 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4878 static int io_files_update_prep(struct io_kiocb *req,
4879 const struct io_uring_sqe *sqe)
4881 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4884 req->files_update.offset = READ_ONCE(sqe->off);
4885 req->files_update.nr_args = READ_ONCE(sqe->len);
4886 if (!req->files_update.nr_args)
4888 req->files_update.arg = READ_ONCE(sqe->addr);
4892 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4894 struct io_ring_ctx *ctx = req->ctx;
4895 struct io_uring_files_update up;
4901 up.offset = req->files_update.offset;
4902 up.fds = req->files_update.arg;
4904 mutex_lock(&ctx->uring_lock);
4905 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4906 mutex_unlock(&ctx->uring_lock);
4909 req_set_fail_links(req);
4910 io_cqring_add_event(req, ret);
4915 static int io_req_defer_prep(struct io_kiocb *req,
4916 const struct io_uring_sqe *sqe)
4923 if (io_op_defs[req->opcode].file_table) {
4924 ret = io_grab_files(req);
4929 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4931 switch (req->opcode) {
4934 case IORING_OP_READV:
4935 case IORING_OP_READ_FIXED:
4936 case IORING_OP_READ:
4937 ret = io_read_prep(req, sqe, true);
4939 case IORING_OP_WRITEV:
4940 case IORING_OP_WRITE_FIXED:
4941 case IORING_OP_WRITE:
4942 ret = io_write_prep(req, sqe, true);
4944 case IORING_OP_POLL_ADD:
4945 ret = io_poll_add_prep(req, sqe);
4947 case IORING_OP_POLL_REMOVE:
4948 ret = io_poll_remove_prep(req, sqe);
4950 case IORING_OP_FSYNC:
4951 ret = io_prep_fsync(req, sqe);
4953 case IORING_OP_SYNC_FILE_RANGE:
4954 ret = io_prep_sfr(req, sqe);
4956 case IORING_OP_SENDMSG:
4957 case IORING_OP_SEND:
4958 ret = io_sendmsg_prep(req, sqe);
4960 case IORING_OP_RECVMSG:
4961 case IORING_OP_RECV:
4962 ret = io_recvmsg_prep(req, sqe);
4964 case IORING_OP_CONNECT:
4965 ret = io_connect_prep(req, sqe);
4967 case IORING_OP_TIMEOUT:
4968 ret = io_timeout_prep(req, sqe, false);
4970 case IORING_OP_TIMEOUT_REMOVE:
4971 ret = io_timeout_remove_prep(req, sqe);
4973 case IORING_OP_ASYNC_CANCEL:
4974 ret = io_async_cancel_prep(req, sqe);
4976 case IORING_OP_LINK_TIMEOUT:
4977 ret = io_timeout_prep(req, sqe, true);
4979 case IORING_OP_ACCEPT:
4980 ret = io_accept_prep(req, sqe);
4982 case IORING_OP_FALLOCATE:
4983 ret = io_fallocate_prep(req, sqe);
4985 case IORING_OP_OPENAT:
4986 ret = io_openat_prep(req, sqe);
4988 case IORING_OP_CLOSE:
4989 ret = io_close_prep(req, sqe);
4991 case IORING_OP_FILES_UPDATE:
4992 ret = io_files_update_prep(req, sqe);
4994 case IORING_OP_STATX:
4995 ret = io_statx_prep(req, sqe);
4997 case IORING_OP_FADVISE:
4998 ret = io_fadvise_prep(req, sqe);
5000 case IORING_OP_MADVISE:
5001 ret = io_madvise_prep(req, sqe);
5003 case IORING_OP_OPENAT2:
5004 ret = io_openat2_prep(req, sqe);
5006 case IORING_OP_EPOLL_CTL:
5007 ret = io_epoll_ctl_prep(req, sqe);
5009 case IORING_OP_SPLICE:
5010 ret = io_splice_prep(req, sqe);
5012 case IORING_OP_PROVIDE_BUFFERS:
5013 ret = io_provide_buffers_prep(req, sqe);
5015 case IORING_OP_REMOVE_BUFFERS:
5016 ret = io_remove_buffers_prep(req, sqe);
5019 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5028 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5030 struct io_ring_ctx *ctx = req->ctx;
5033 /* Still need defer if there is pending req in defer list. */
5034 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
5037 if (!req->io && io_alloc_async_ctx(req))
5040 ret = io_req_defer_prep(req, sqe);
5044 spin_lock_irq(&ctx->completion_lock);
5045 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5046 spin_unlock_irq(&ctx->completion_lock);
5050 trace_io_uring_defer(ctx, req, req->user_data);
5051 list_add_tail(&req->list, &ctx->defer_list);
5052 spin_unlock_irq(&ctx->completion_lock);
5053 return -EIOCBQUEUED;
5056 static void io_cleanup_req(struct io_kiocb *req)
5058 struct io_async_ctx *io = req->io;
5060 switch (req->opcode) {
5061 case IORING_OP_READV:
5062 case IORING_OP_READ_FIXED:
5063 case IORING_OP_READ:
5064 if (req->flags & REQ_F_BUFFER_SELECTED)
5065 kfree((void *)(unsigned long)req->rw.addr);
5067 case IORING_OP_WRITEV:
5068 case IORING_OP_WRITE_FIXED:
5069 case IORING_OP_WRITE:
5070 if (io->rw.iov != io->rw.fast_iov)
5073 case IORING_OP_RECVMSG:
5074 if (req->flags & REQ_F_BUFFER_SELECTED)
5075 kfree(req->sr_msg.kbuf);
5077 case IORING_OP_SENDMSG:
5078 if (io->msg.iov != io->msg.fast_iov)
5081 case IORING_OP_RECV:
5082 if (req->flags & REQ_F_BUFFER_SELECTED)
5083 kfree(req->sr_msg.kbuf);
5085 case IORING_OP_OPENAT:
5086 case IORING_OP_OPENAT2:
5087 case IORING_OP_STATX:
5088 putname(req->open.filename);
5090 case IORING_OP_SPLICE:
5091 io_put_file(req, req->splice.file_in,
5092 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5096 req->flags &= ~REQ_F_NEED_CLEANUP;
5099 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5100 bool force_nonblock)
5102 struct io_ring_ctx *ctx = req->ctx;
5105 switch (req->opcode) {
5109 case IORING_OP_READV:
5110 case IORING_OP_READ_FIXED:
5111 case IORING_OP_READ:
5113 ret = io_read_prep(req, sqe, force_nonblock);
5117 ret = io_read(req, force_nonblock);
5119 case IORING_OP_WRITEV:
5120 case IORING_OP_WRITE_FIXED:
5121 case IORING_OP_WRITE:
5123 ret = io_write_prep(req, sqe, force_nonblock);
5127 ret = io_write(req, force_nonblock);
5129 case IORING_OP_FSYNC:
5131 ret = io_prep_fsync(req, sqe);
5135 ret = io_fsync(req, force_nonblock);
5137 case IORING_OP_POLL_ADD:
5139 ret = io_poll_add_prep(req, sqe);
5143 ret = io_poll_add(req);
5145 case IORING_OP_POLL_REMOVE:
5147 ret = io_poll_remove_prep(req, sqe);
5151 ret = io_poll_remove(req);
5153 case IORING_OP_SYNC_FILE_RANGE:
5155 ret = io_prep_sfr(req, sqe);
5159 ret = io_sync_file_range(req, force_nonblock);
5161 case IORING_OP_SENDMSG:
5162 case IORING_OP_SEND:
5164 ret = io_sendmsg_prep(req, sqe);
5168 if (req->opcode == IORING_OP_SENDMSG)
5169 ret = io_sendmsg(req, force_nonblock);
5171 ret = io_send(req, force_nonblock);
5173 case IORING_OP_RECVMSG:
5174 case IORING_OP_RECV:
5176 ret = io_recvmsg_prep(req, sqe);
5180 if (req->opcode == IORING_OP_RECVMSG)
5181 ret = io_recvmsg(req, force_nonblock);
5183 ret = io_recv(req, force_nonblock);
5185 case IORING_OP_TIMEOUT:
5187 ret = io_timeout_prep(req, sqe, false);
5191 ret = io_timeout(req);
5193 case IORING_OP_TIMEOUT_REMOVE:
5195 ret = io_timeout_remove_prep(req, sqe);
5199 ret = io_timeout_remove(req);
5201 case IORING_OP_ACCEPT:
5203 ret = io_accept_prep(req, sqe);
5207 ret = io_accept(req, force_nonblock);
5209 case IORING_OP_CONNECT:
5211 ret = io_connect_prep(req, sqe);
5215 ret = io_connect(req, force_nonblock);
5217 case IORING_OP_ASYNC_CANCEL:
5219 ret = io_async_cancel_prep(req, sqe);
5223 ret = io_async_cancel(req);
5225 case IORING_OP_FALLOCATE:
5227 ret = io_fallocate_prep(req, sqe);
5231 ret = io_fallocate(req, force_nonblock);
5233 case IORING_OP_OPENAT:
5235 ret = io_openat_prep(req, sqe);
5239 ret = io_openat(req, force_nonblock);
5241 case IORING_OP_CLOSE:
5243 ret = io_close_prep(req, sqe);
5247 ret = io_close(req, force_nonblock);
5249 case IORING_OP_FILES_UPDATE:
5251 ret = io_files_update_prep(req, sqe);
5255 ret = io_files_update(req, force_nonblock);
5257 case IORING_OP_STATX:
5259 ret = io_statx_prep(req, sqe);
5263 ret = io_statx(req, force_nonblock);
5265 case IORING_OP_FADVISE:
5267 ret = io_fadvise_prep(req, sqe);
5271 ret = io_fadvise(req, force_nonblock);
5273 case IORING_OP_MADVISE:
5275 ret = io_madvise_prep(req, sqe);
5279 ret = io_madvise(req, force_nonblock);
5281 case IORING_OP_OPENAT2:
5283 ret = io_openat2_prep(req, sqe);
5287 ret = io_openat2(req, force_nonblock);
5289 case IORING_OP_EPOLL_CTL:
5291 ret = io_epoll_ctl_prep(req, sqe);
5295 ret = io_epoll_ctl(req, force_nonblock);
5297 case IORING_OP_SPLICE:
5299 ret = io_splice_prep(req, sqe);
5303 ret = io_splice(req, force_nonblock);
5305 case IORING_OP_PROVIDE_BUFFERS:
5307 ret = io_provide_buffers_prep(req, sqe);
5311 ret = io_provide_buffers(req, force_nonblock);
5313 case IORING_OP_REMOVE_BUFFERS:
5315 ret = io_remove_buffers_prep(req, sqe);
5319 ret = io_remove_buffers(req, force_nonblock);
5329 if (ctx->flags & IORING_SETUP_IOPOLL) {
5330 const bool in_async = io_wq_current_is_worker();
5332 if (req->result == -EAGAIN)
5335 /* workqueue context doesn't hold uring_lock, grab it now */
5337 mutex_lock(&ctx->uring_lock);
5339 io_iopoll_req_issued(req);
5342 mutex_unlock(&ctx->uring_lock);
5348 static void io_wq_submit_work(struct io_wq_work **workptr)
5350 struct io_wq_work *work = *workptr;
5351 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5354 /* if NO_CANCEL is set, we must still run the work */
5355 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5356 IO_WQ_WORK_CANCEL) {
5362 ret = io_issue_sqe(req, NULL, false);
5364 * We can get EAGAIN for polled IO even though we're
5365 * forcing a sync submission from here, since we can't
5366 * wait for request slots on the block side.
5375 req_set_fail_links(req);
5376 io_cqring_add_event(req, ret);
5380 io_steal_work(req, workptr);
5383 static int io_req_needs_file(struct io_kiocb *req, int fd)
5385 if (!io_op_defs[req->opcode].needs_file)
5387 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5392 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5395 struct fixed_file_table *table;
5397 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5398 return table->files[index & IORING_FILE_TABLE_MASK];;
5401 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5402 int fd, struct file **out_file, bool fixed)
5404 struct io_ring_ctx *ctx = req->ctx;
5408 if (unlikely(!ctx->file_data ||
5409 (unsigned) fd >= ctx->nr_user_files))
5411 fd = array_index_nospec(fd, ctx->nr_user_files);
5412 file = io_file_from_index(ctx, fd);
5415 req->fixed_file_refs = ctx->file_data->cur_refs;
5416 percpu_ref_get(req->fixed_file_refs);
5418 trace_io_uring_file_get(ctx, fd);
5419 file = __io_file_get(state, fd);
5420 if (unlikely(!file))
5428 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5429 int fd, unsigned int flags)
5433 if (!io_req_needs_file(req, fd))
5436 fixed = (flags & IOSQE_FIXED_FILE);
5437 if (unlikely(!fixed && req->needs_fixed_file))
5440 return io_file_get(state, req, fd, &req->file, fixed);
5443 static int io_grab_files(struct io_kiocb *req)
5446 struct io_ring_ctx *ctx = req->ctx;
5448 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5450 if (!ctx->ring_file)
5454 spin_lock_irq(&ctx->inflight_lock);
5456 * We use the f_ops->flush() handler to ensure that we can flush
5457 * out work accessing these files if the fd is closed. Check if
5458 * the fd has changed since we started down this path, and disallow
5459 * this operation if it has.
5461 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5462 list_add(&req->inflight_entry, &ctx->inflight_list);
5463 req->flags |= REQ_F_INFLIGHT;
5464 req->work.files = current->files;
5467 spin_unlock_irq(&ctx->inflight_lock);
5473 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5475 struct io_timeout_data *data = container_of(timer,
5476 struct io_timeout_data, timer);
5477 struct io_kiocb *req = data->req;
5478 struct io_ring_ctx *ctx = req->ctx;
5479 struct io_kiocb *prev = NULL;
5480 unsigned long flags;
5482 spin_lock_irqsave(&ctx->completion_lock, flags);
5485 * We don't expect the list to be empty, that will only happen if we
5486 * race with the completion of the linked work.
5488 if (!list_empty(&req->link_list)) {
5489 prev = list_entry(req->link_list.prev, struct io_kiocb,
5491 if (refcount_inc_not_zero(&prev->refs)) {
5492 list_del_init(&req->link_list);
5493 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5498 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5501 req_set_fail_links(prev);
5502 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5505 io_cqring_add_event(req, -ETIME);
5508 return HRTIMER_NORESTART;
5511 static void io_queue_linked_timeout(struct io_kiocb *req)
5513 struct io_ring_ctx *ctx = req->ctx;
5516 * If the list is now empty, then our linked request finished before
5517 * we got a chance to setup the timer
5519 spin_lock_irq(&ctx->completion_lock);
5520 if (!list_empty(&req->link_list)) {
5521 struct io_timeout_data *data = &req->io->timeout;
5523 data->timer.function = io_link_timeout_fn;
5524 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5527 spin_unlock_irq(&ctx->completion_lock);
5529 /* drop submission reference */
5533 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5535 struct io_kiocb *nxt;
5537 if (!(req->flags & REQ_F_LINK_HEAD))
5539 /* for polled retry, if flag is set, we already went through here */
5540 if (req->flags & REQ_F_POLLED)
5543 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5545 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5548 req->flags |= REQ_F_LINK_TIMEOUT;
5552 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5554 struct io_kiocb *linked_timeout;
5555 struct io_kiocb *nxt;
5556 const struct cred *old_creds = NULL;
5560 linked_timeout = io_prep_linked_timeout(req);
5562 if (req->work.creds && req->work.creds != current_cred()) {
5564 revert_creds(old_creds);
5565 if (old_creds == req->work.creds)
5566 old_creds = NULL; /* restored original creds */
5568 old_creds = override_creds(req->work.creds);
5571 ret = io_issue_sqe(req, sqe, true);
5574 * We async punt it if the file wasn't marked NOWAIT, or if the file
5575 * doesn't support non-blocking read/write attempts
5577 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5578 (req->flags & REQ_F_MUST_PUNT))) {
5579 if (io_arm_poll_handler(req)) {
5581 io_queue_linked_timeout(linked_timeout);
5585 if (io_op_defs[req->opcode].file_table) {
5586 ret = io_grab_files(req);
5592 * Queued up for async execution, worker will release
5593 * submit reference when the iocb is actually submitted.
5595 io_queue_async_work(req);
5601 /* drop submission reference */
5602 io_put_req_find_next(req, &nxt);
5604 if (linked_timeout) {
5606 io_queue_linked_timeout(linked_timeout);
5608 io_put_req(linked_timeout);
5611 /* and drop final reference, if we failed */
5613 io_cqring_add_event(req, ret);
5614 req_set_fail_links(req);
5620 if (req->flags & REQ_F_FORCE_ASYNC)
5626 revert_creds(old_creds);
5629 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5633 ret = io_req_defer(req, sqe);
5635 if (ret != -EIOCBQUEUED) {
5637 io_cqring_add_event(req, ret);
5638 req_set_fail_links(req);
5639 io_double_put_req(req);
5641 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5642 ret = io_req_defer_prep(req, sqe);
5643 if (unlikely(ret < 0))
5646 * Never try inline submit of IOSQE_ASYNC is set, go straight
5647 * to async execution.
5649 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5650 io_queue_async_work(req);
5652 __io_queue_sqe(req, sqe);
5656 static inline void io_queue_link_head(struct io_kiocb *req)
5658 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5659 io_cqring_add_event(req, -ECANCELED);
5660 io_double_put_req(req);
5662 io_queue_sqe(req, NULL);
5665 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5666 struct io_submit_state *state, struct io_kiocb **link)
5668 struct io_ring_ctx *ctx = req->ctx;
5672 * If we already have a head request, queue this one for async
5673 * submittal once the head completes. If we don't have a head but
5674 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5675 * submitted sync once the chain is complete. If none of those
5676 * conditions are true (normal request), then just queue it.
5679 struct io_kiocb *head = *link;
5682 * Taking sequential execution of a link, draining both sides
5683 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5684 * requests in the link. So, it drains the head and the
5685 * next after the link request. The last one is done via
5686 * drain_next flag to persist the effect across calls.
5688 if (req->flags & REQ_F_IO_DRAIN) {
5689 head->flags |= REQ_F_IO_DRAIN;
5690 ctx->drain_next = 1;
5692 if (io_alloc_async_ctx(req))
5695 ret = io_req_defer_prep(req, sqe);
5697 /* fail even hard links since we don't submit */
5698 head->flags |= REQ_F_FAIL_LINK;
5701 trace_io_uring_link(ctx, req, head);
5702 list_add_tail(&req->link_list, &head->link_list);
5704 /* last request of a link, enqueue the link */
5705 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5706 io_queue_link_head(head);
5710 if (unlikely(ctx->drain_next)) {
5711 req->flags |= REQ_F_IO_DRAIN;
5712 ctx->drain_next = 0;
5714 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5715 req->flags |= REQ_F_LINK_HEAD;
5716 INIT_LIST_HEAD(&req->link_list);
5718 if (io_alloc_async_ctx(req))
5721 ret = io_req_defer_prep(req, sqe);
5723 req->flags |= REQ_F_FAIL_LINK;
5726 io_queue_sqe(req, sqe);
5734 * Batched submission is done, ensure local IO is flushed out.
5736 static void io_submit_state_end(struct io_submit_state *state)
5738 blk_finish_plug(&state->plug);
5740 if (state->free_reqs)
5741 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5745 * Start submission side cache.
5747 static void io_submit_state_start(struct io_submit_state *state,
5748 unsigned int max_ios)
5750 blk_start_plug(&state->plug);
5751 state->free_reqs = 0;
5753 state->ios_left = max_ios;
5756 static void io_commit_sqring(struct io_ring_ctx *ctx)
5758 struct io_rings *rings = ctx->rings;
5761 * Ensure any loads from the SQEs are done at this point,
5762 * since once we write the new head, the application could
5763 * write new data to them.
5765 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5769 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5770 * that is mapped by userspace. This means that care needs to be taken to
5771 * ensure that reads are stable, as we cannot rely on userspace always
5772 * being a good citizen. If members of the sqe are validated and then later
5773 * used, it's important that those reads are done through READ_ONCE() to
5774 * prevent a re-load down the line.
5776 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5778 u32 *sq_array = ctx->sq_array;
5782 * The cached sq head (or cq tail) serves two purposes:
5784 * 1) allows us to batch the cost of updating the user visible
5786 * 2) allows the kernel side to track the head on its own, even
5787 * though the application is the one updating it.
5789 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5790 if (likely(head < ctx->sq_entries))
5791 return &ctx->sq_sqes[head];
5793 /* drop invalid entries */
5794 ctx->cached_sq_dropped++;
5795 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5799 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5801 ctx->cached_sq_head++;
5804 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5805 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5806 IOSQE_BUFFER_SELECT)
5808 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5809 const struct io_uring_sqe *sqe,
5810 struct io_submit_state *state, bool async)
5812 unsigned int sqe_flags;
5816 * All io need record the previous position, if LINK vs DARIN,
5817 * it can be used to mark the position of the first IO in the
5820 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5821 req->opcode = READ_ONCE(sqe->opcode);
5822 req->user_data = READ_ONCE(sqe->user_data);
5827 /* one is dropped after submission, the other at completion */
5828 refcount_set(&req->refs, 2);
5831 req->needs_fixed_file = async;
5832 INIT_IO_WORK(&req->work, io_wq_submit_work);
5834 if (unlikely(req->opcode >= IORING_OP_LAST))
5837 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5838 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5840 use_mm(ctx->sqo_mm);
5843 sqe_flags = READ_ONCE(sqe->flags);
5844 /* enforce forwards compatibility on users */
5845 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5848 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5849 !io_op_defs[req->opcode].buffer_select)
5852 id = READ_ONCE(sqe->personality);
5854 req->work.creds = idr_find(&ctx->personality_idr, id);
5855 if (unlikely(!req->work.creds))
5857 get_cred(req->work.creds);
5860 /* same numerical values with corresponding REQ_F_*, safe to copy */
5861 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5862 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5863 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5865 fd = READ_ONCE(sqe->fd);
5866 return io_req_set_file(state, req, fd, sqe_flags);
5869 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5870 struct file *ring_file, int ring_fd, bool async)
5872 struct io_submit_state state, *statep = NULL;
5873 struct io_kiocb *link = NULL;
5874 int i, submitted = 0;
5876 /* if we have a backlog and couldn't flush it all, return BUSY */
5877 if (test_bit(0, &ctx->sq_check_overflow)) {
5878 if (!list_empty(&ctx->cq_overflow_list) &&
5879 !io_cqring_overflow_flush(ctx, false))
5883 /* make sure SQ entry isn't read before tail */
5884 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5886 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5889 if (nr > IO_PLUG_THRESHOLD) {
5890 io_submit_state_start(&state, nr);
5894 ctx->ring_fd = ring_fd;
5895 ctx->ring_file = ring_file;
5897 for (i = 0; i < nr; i++) {
5898 const struct io_uring_sqe *sqe;
5899 struct io_kiocb *req;
5902 sqe = io_get_sqe(ctx);
5903 if (unlikely(!sqe)) {
5904 io_consume_sqe(ctx);
5907 req = io_alloc_req(ctx, statep);
5908 if (unlikely(!req)) {
5910 submitted = -EAGAIN;
5914 err = io_init_req(ctx, req, sqe, statep, async);
5915 io_consume_sqe(ctx);
5916 /* will complete beyond this point, count as submitted */
5919 if (unlikely(err)) {
5921 io_cqring_add_event(req, err);
5922 io_double_put_req(req);
5926 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5928 err = io_submit_sqe(req, sqe, statep, &link);
5933 if (unlikely(submitted != nr)) {
5934 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5936 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5939 io_queue_link_head(link);
5941 io_submit_state_end(&state);
5943 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5944 io_commit_sqring(ctx);
5949 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5951 struct mm_struct *mm = current->mm;
5959 static int io_sq_thread(void *data)
5961 struct io_ring_ctx *ctx = data;
5962 const struct cred *old_cred;
5963 mm_segment_t old_fs;
5965 unsigned long timeout;
5968 complete(&ctx->completions[1]);
5972 old_cred = override_creds(ctx->creds);
5974 timeout = jiffies + ctx->sq_thread_idle;
5975 while (!kthread_should_park()) {
5976 unsigned int to_submit;
5978 if (!list_empty(&ctx->poll_list)) {
5979 unsigned nr_events = 0;
5981 mutex_lock(&ctx->uring_lock);
5982 if (!list_empty(&ctx->poll_list))
5983 io_iopoll_getevents(ctx, &nr_events, 0);
5985 timeout = jiffies + ctx->sq_thread_idle;
5986 mutex_unlock(&ctx->uring_lock);
5989 to_submit = io_sqring_entries(ctx);
5992 * If submit got -EBUSY, flag us as needing the application
5993 * to enter the kernel to reap and flush events.
5995 if (!to_submit || ret == -EBUSY) {
5997 * Drop cur_mm before scheduling, we can't hold it for
5998 * long periods (or over schedule()). Do this before
5999 * adding ourselves to the waitqueue, as the unuse/drop
6002 io_sq_thread_drop_mm(ctx);
6005 * We're polling. If we're within the defined idle
6006 * period, then let us spin without work before going
6007 * to sleep. The exception is if we got EBUSY doing
6008 * more IO, we should wait for the application to
6009 * reap events and wake us up.
6011 if (!list_empty(&ctx->poll_list) ||
6012 (!time_after(jiffies, timeout) && ret != -EBUSY &&
6013 !percpu_ref_is_dying(&ctx->refs))) {
6014 if (current->task_works)
6020 prepare_to_wait(&ctx->sqo_wait, &wait,
6021 TASK_INTERRUPTIBLE);
6024 * While doing polled IO, before going to sleep, we need
6025 * to check if there are new reqs added to poll_list, it
6026 * is because reqs may have been punted to io worker and
6027 * will be added to poll_list later, hence check the
6030 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6031 !list_empty_careful(&ctx->poll_list)) {
6032 finish_wait(&ctx->sqo_wait, &wait);
6036 /* Tell userspace we may need a wakeup call */
6037 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6038 /* make sure to read SQ tail after writing flags */
6041 to_submit = io_sqring_entries(ctx);
6042 if (!to_submit || ret == -EBUSY) {
6043 if (kthread_should_park()) {
6044 finish_wait(&ctx->sqo_wait, &wait);
6047 if (current->task_works) {
6049 finish_wait(&ctx->sqo_wait, &wait);
6052 if (signal_pending(current))
6053 flush_signals(current);
6055 finish_wait(&ctx->sqo_wait, &wait);
6057 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6060 finish_wait(&ctx->sqo_wait, &wait);
6062 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6065 mutex_lock(&ctx->uring_lock);
6066 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6067 mutex_unlock(&ctx->uring_lock);
6068 timeout = jiffies + ctx->sq_thread_idle;
6071 if (current->task_works)
6075 io_sq_thread_drop_mm(ctx);
6076 revert_creds(old_cred);
6083 struct io_wait_queue {
6084 struct wait_queue_entry wq;
6085 struct io_ring_ctx *ctx;
6087 unsigned nr_timeouts;
6090 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6092 struct io_ring_ctx *ctx = iowq->ctx;
6095 * Wake up if we have enough events, or if a timeout occurred since we
6096 * started waiting. For timeouts, we always want to return to userspace,
6097 * regardless of event count.
6099 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6100 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6103 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6104 int wake_flags, void *key)
6106 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6109 /* use noflush == true, as we can't safely rely on locking context */
6110 if (!io_should_wake(iowq, true))
6113 return autoremove_wake_function(curr, mode, wake_flags, key);
6117 * Wait until events become available, if we don't already have some. The
6118 * application must reap them itself, as they reside on the shared cq ring.
6120 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6121 const sigset_t __user *sig, size_t sigsz)
6123 struct io_wait_queue iowq = {
6126 .func = io_wake_function,
6127 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6130 .to_wait = min_events,
6132 struct io_rings *rings = ctx->rings;
6136 if (io_cqring_events(ctx, false) >= min_events)
6138 if (!current->task_works)
6144 #ifdef CONFIG_COMPAT
6145 if (in_compat_syscall())
6146 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6150 ret = set_user_sigmask(sig, sigsz);
6156 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6157 trace_io_uring_cqring_wait(ctx, min_events);
6159 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6160 TASK_INTERRUPTIBLE);
6161 if (current->task_works)
6163 if (io_should_wake(&iowq, false))
6166 if (signal_pending(current)) {
6171 finish_wait(&ctx->wait, &iowq.wq);
6173 restore_saved_sigmask_unless(ret == -EINTR);
6175 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6178 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6180 #if defined(CONFIG_UNIX)
6181 if (ctx->ring_sock) {
6182 struct sock *sock = ctx->ring_sock->sk;
6183 struct sk_buff *skb;
6185 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6191 for (i = 0; i < ctx->nr_user_files; i++) {
6194 file = io_file_from_index(ctx, i);
6201 static void io_file_ref_kill(struct percpu_ref *ref)
6203 struct fixed_file_data *data;
6205 data = container_of(ref, struct fixed_file_data, refs);
6206 complete(&data->done);
6209 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6211 struct fixed_file_data *data = ctx->file_data;
6212 struct fixed_file_ref_node *ref_node = NULL;
6213 unsigned nr_tables, i;
6214 unsigned long flags;
6219 spin_lock_irqsave(&data->lock, flags);
6220 if (!list_empty(&data->ref_list))
6221 ref_node = list_first_entry(&data->ref_list,
6222 struct fixed_file_ref_node, node);
6223 spin_unlock_irqrestore(&data->lock, flags);
6225 percpu_ref_kill(&ref_node->refs);
6227 percpu_ref_kill(&data->refs);
6229 /* wait for all refs nodes to complete */
6230 wait_for_completion(&data->done);
6232 __io_sqe_files_unregister(ctx);
6233 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6234 for (i = 0; i < nr_tables; i++)
6235 kfree(data->table[i].files);
6237 percpu_ref_exit(&data->refs);
6239 ctx->file_data = NULL;
6240 ctx->nr_user_files = 0;
6244 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6246 if (ctx->sqo_thread) {
6247 wait_for_completion(&ctx->completions[1]);
6249 * The park is a bit of a work-around, without it we get
6250 * warning spews on shutdown with SQPOLL set and affinity
6251 * set to a single CPU.
6253 kthread_park(ctx->sqo_thread);
6254 kthread_stop(ctx->sqo_thread);
6255 ctx->sqo_thread = NULL;
6259 static void io_finish_async(struct io_ring_ctx *ctx)
6261 io_sq_thread_stop(ctx);
6264 io_wq_destroy(ctx->io_wq);
6269 #if defined(CONFIG_UNIX)
6271 * Ensure the UNIX gc is aware of our file set, so we are certain that
6272 * the io_uring can be safely unregistered on process exit, even if we have
6273 * loops in the file referencing.
6275 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6277 struct sock *sk = ctx->ring_sock->sk;
6278 struct scm_fp_list *fpl;
6279 struct sk_buff *skb;
6282 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6286 skb = alloc_skb(0, GFP_KERNEL);
6295 fpl->user = get_uid(ctx->user);
6296 for (i = 0; i < nr; i++) {
6297 struct file *file = io_file_from_index(ctx, i + offset);
6301 fpl->fp[nr_files] = get_file(file);
6302 unix_inflight(fpl->user, fpl->fp[nr_files]);
6307 fpl->max = SCM_MAX_FD;
6308 fpl->count = nr_files;
6309 UNIXCB(skb).fp = fpl;
6310 skb->destructor = unix_destruct_scm;
6311 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6312 skb_queue_head(&sk->sk_receive_queue, skb);
6314 for (i = 0; i < nr_files; i++)
6325 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6326 * causes regular reference counting to break down. We rely on the UNIX
6327 * garbage collection to take care of this problem for us.
6329 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6331 unsigned left, total;
6335 left = ctx->nr_user_files;
6337 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6339 ret = __io_sqe_files_scm(ctx, this_files, total);
6343 total += this_files;
6349 while (total < ctx->nr_user_files) {
6350 struct file *file = io_file_from_index(ctx, total);
6360 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6366 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6371 for (i = 0; i < nr_tables; i++) {
6372 struct fixed_file_table *table = &ctx->file_data->table[i];
6373 unsigned this_files;
6375 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6376 table->files = kcalloc(this_files, sizeof(struct file *),
6380 nr_files -= this_files;
6386 for (i = 0; i < nr_tables; i++) {
6387 struct fixed_file_table *table = &ctx->file_data->table[i];
6388 kfree(table->files);
6393 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6395 #if defined(CONFIG_UNIX)
6396 struct sock *sock = ctx->ring_sock->sk;
6397 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6398 struct sk_buff *skb;
6401 __skb_queue_head_init(&list);
6404 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6405 * remove this entry and rearrange the file array.
6407 skb = skb_dequeue(head);
6409 struct scm_fp_list *fp;
6411 fp = UNIXCB(skb).fp;
6412 for (i = 0; i < fp->count; i++) {
6415 if (fp->fp[i] != file)
6418 unix_notinflight(fp->user, fp->fp[i]);
6419 left = fp->count - 1 - i;
6421 memmove(&fp->fp[i], &fp->fp[i + 1],
6422 left * sizeof(struct file *));
6429 __skb_queue_tail(&list, skb);
6439 __skb_queue_tail(&list, skb);
6441 skb = skb_dequeue(head);
6444 if (skb_peek(&list)) {
6445 spin_lock_irq(&head->lock);
6446 while ((skb = __skb_dequeue(&list)) != NULL)
6447 __skb_queue_tail(head, skb);
6448 spin_unlock_irq(&head->lock);
6455 struct io_file_put {
6456 struct list_head list;
6460 static void io_file_put_work(struct work_struct *work)
6462 struct fixed_file_ref_node *ref_node;
6463 struct fixed_file_data *file_data;
6464 struct io_ring_ctx *ctx;
6465 struct io_file_put *pfile, *tmp;
6466 unsigned long flags;
6468 ref_node = container_of(work, struct fixed_file_ref_node, work);
6469 file_data = ref_node->file_data;
6470 ctx = file_data->ctx;
6472 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6473 list_del_init(&pfile->list);
6474 io_ring_file_put(ctx, pfile->file);
6478 spin_lock_irqsave(&file_data->lock, flags);
6479 list_del_init(&ref_node->node);
6480 spin_unlock_irqrestore(&file_data->lock, flags);
6482 percpu_ref_exit(&ref_node->refs);
6484 percpu_ref_put(&file_data->refs);
6487 static void io_file_data_ref_zero(struct percpu_ref *ref)
6489 struct fixed_file_ref_node *ref_node;
6491 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6493 queue_work(system_wq, &ref_node->work);
6496 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6497 struct io_ring_ctx *ctx)
6499 struct fixed_file_ref_node *ref_node;
6501 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6503 return ERR_PTR(-ENOMEM);
6505 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6508 return ERR_PTR(-ENOMEM);
6510 INIT_LIST_HEAD(&ref_node->node);
6511 INIT_LIST_HEAD(&ref_node->file_list);
6512 INIT_WORK(&ref_node->work, io_file_put_work);
6513 ref_node->file_data = ctx->file_data;
6518 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6520 percpu_ref_exit(&ref_node->refs);
6524 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6527 __s32 __user *fds = (__s32 __user *) arg;
6532 struct fixed_file_ref_node *ref_node;
6533 unsigned long flags;
6539 if (nr_args > IORING_MAX_FIXED_FILES)
6542 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6543 if (!ctx->file_data)
6545 ctx->file_data->ctx = ctx;
6546 init_completion(&ctx->file_data->done);
6547 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6548 spin_lock_init(&ctx->file_data->lock);
6550 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6551 ctx->file_data->table = kcalloc(nr_tables,
6552 sizeof(struct fixed_file_table),
6554 if (!ctx->file_data->table) {
6555 kfree(ctx->file_data);
6556 ctx->file_data = NULL;
6560 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6561 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6562 kfree(ctx->file_data->table);
6563 kfree(ctx->file_data);
6564 ctx->file_data = NULL;
6568 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6569 percpu_ref_exit(&ctx->file_data->refs);
6570 kfree(ctx->file_data->table);
6571 kfree(ctx->file_data);
6572 ctx->file_data = NULL;
6576 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6577 struct fixed_file_table *table;
6581 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6583 /* allow sparse sets */
6589 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6590 index = i & IORING_FILE_TABLE_MASK;
6598 * Don't allow io_uring instances to be registered. If UNIX
6599 * isn't enabled, then this causes a reference cycle and this
6600 * instance can never get freed. If UNIX is enabled we'll
6601 * handle it just fine, but there's still no point in allowing
6602 * a ring fd as it doesn't support regular read/write anyway.
6604 if (file->f_op == &io_uring_fops) {
6609 table->files[index] = file;
6613 for (i = 0; i < ctx->nr_user_files; i++) {
6614 file = io_file_from_index(ctx, i);
6618 for (i = 0; i < nr_tables; i++)
6619 kfree(ctx->file_data->table[i].files);
6621 kfree(ctx->file_data->table);
6622 kfree(ctx->file_data);
6623 ctx->file_data = NULL;
6624 ctx->nr_user_files = 0;
6628 ret = io_sqe_files_scm(ctx);
6630 io_sqe_files_unregister(ctx);
6634 ref_node = alloc_fixed_file_ref_node(ctx);
6635 if (IS_ERR(ref_node)) {
6636 io_sqe_files_unregister(ctx);
6637 return PTR_ERR(ref_node);
6640 ctx->file_data->cur_refs = &ref_node->refs;
6641 spin_lock_irqsave(&ctx->file_data->lock, flags);
6642 list_add(&ref_node->node, &ctx->file_data->ref_list);
6643 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6644 percpu_ref_get(&ctx->file_data->refs);
6648 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6651 #if defined(CONFIG_UNIX)
6652 struct sock *sock = ctx->ring_sock->sk;
6653 struct sk_buff_head *head = &sock->sk_receive_queue;
6654 struct sk_buff *skb;
6657 * See if we can merge this file into an existing skb SCM_RIGHTS
6658 * file set. If there's no room, fall back to allocating a new skb
6659 * and filling it in.
6661 spin_lock_irq(&head->lock);
6662 skb = skb_peek(head);
6664 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6666 if (fpl->count < SCM_MAX_FD) {
6667 __skb_unlink(skb, head);
6668 spin_unlock_irq(&head->lock);
6669 fpl->fp[fpl->count] = get_file(file);
6670 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6672 spin_lock_irq(&head->lock);
6673 __skb_queue_head(head, skb);
6678 spin_unlock_irq(&head->lock);
6685 return __io_sqe_files_scm(ctx, 1, index);
6691 static int io_queue_file_removal(struct fixed_file_data *data,
6694 struct io_file_put *pfile;
6695 struct percpu_ref *refs = data->cur_refs;
6696 struct fixed_file_ref_node *ref_node;
6698 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6702 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6704 list_add(&pfile->list, &ref_node->file_list);
6709 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6710 struct io_uring_files_update *up,
6713 struct fixed_file_data *data = ctx->file_data;
6714 struct fixed_file_ref_node *ref_node;
6719 unsigned long flags;
6720 bool needs_switch = false;
6722 if (check_add_overflow(up->offset, nr_args, &done))
6724 if (done > ctx->nr_user_files)
6727 ref_node = alloc_fixed_file_ref_node(ctx);
6728 if (IS_ERR(ref_node))
6729 return PTR_ERR(ref_node);
6732 fds = u64_to_user_ptr(up->fds);
6734 struct fixed_file_table *table;
6738 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6742 i = array_index_nospec(up->offset, ctx->nr_user_files);
6743 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6744 index = i & IORING_FILE_TABLE_MASK;
6745 if (table->files[index]) {
6746 file = io_file_from_index(ctx, index);
6747 err = io_queue_file_removal(data, file);
6750 table->files[index] = NULL;
6751 needs_switch = true;
6760 * Don't allow io_uring instances to be registered. If
6761 * UNIX isn't enabled, then this causes a reference
6762 * cycle and this instance can never get freed. If UNIX
6763 * is enabled we'll handle it just fine, but there's
6764 * still no point in allowing a ring fd as it doesn't
6765 * support regular read/write anyway.
6767 if (file->f_op == &io_uring_fops) {
6772 table->files[index] = file;
6773 err = io_sqe_file_register(ctx, file, i);
6783 percpu_ref_kill(data->cur_refs);
6784 spin_lock_irqsave(&data->lock, flags);
6785 list_add(&ref_node->node, &data->ref_list);
6786 data->cur_refs = &ref_node->refs;
6787 spin_unlock_irqrestore(&data->lock, flags);
6788 percpu_ref_get(&ctx->file_data->refs);
6790 destroy_fixed_file_ref_node(ref_node);
6792 return done ? done : err;
6795 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6798 struct io_uring_files_update up;
6800 if (!ctx->file_data)
6804 if (copy_from_user(&up, arg, sizeof(up)))
6809 return __io_sqe_files_update(ctx, &up, nr_args);
6812 static void io_free_work(struct io_wq_work *work)
6814 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6816 /* Consider that io_steal_work() relies on this ref */
6820 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6821 struct io_uring_params *p)
6823 struct io_wq_data data;
6825 struct io_ring_ctx *ctx_attach;
6826 unsigned int concurrency;
6829 data.user = ctx->user;
6830 data.free_work = io_free_work;
6832 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6833 /* Do QD, or 4 * CPUS, whatever is smallest */
6834 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6836 ctx->io_wq = io_wq_create(concurrency, &data);
6837 if (IS_ERR(ctx->io_wq)) {
6838 ret = PTR_ERR(ctx->io_wq);
6844 f = fdget(p->wq_fd);
6848 if (f.file->f_op != &io_uring_fops) {
6853 ctx_attach = f.file->private_data;
6854 /* @io_wq is protected by holding the fd */
6855 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6860 ctx->io_wq = ctx_attach->io_wq;
6866 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6867 struct io_uring_params *p)
6871 init_waitqueue_head(&ctx->sqo_wait);
6872 mmgrab(current->mm);
6873 ctx->sqo_mm = current->mm;
6875 if (ctx->flags & IORING_SETUP_SQPOLL) {
6877 if (!capable(CAP_SYS_ADMIN))
6880 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6881 if (!ctx->sq_thread_idle)
6882 ctx->sq_thread_idle = HZ;
6884 if (p->flags & IORING_SETUP_SQ_AFF) {
6885 int cpu = p->sq_thread_cpu;
6888 if (cpu >= nr_cpu_ids)
6890 if (!cpu_online(cpu))
6893 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6897 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6900 if (IS_ERR(ctx->sqo_thread)) {
6901 ret = PTR_ERR(ctx->sqo_thread);
6902 ctx->sqo_thread = NULL;
6905 wake_up_process(ctx->sqo_thread);
6906 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6907 /* Can't have SQ_AFF without SQPOLL */
6912 ret = io_init_wq_offload(ctx, p);
6918 io_finish_async(ctx);
6919 mmdrop(ctx->sqo_mm);
6924 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6926 atomic_long_sub(nr_pages, &user->locked_vm);
6929 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6931 unsigned long page_limit, cur_pages, new_pages;
6933 /* Don't allow more pages than we can safely lock */
6934 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6937 cur_pages = atomic_long_read(&user->locked_vm);
6938 new_pages = cur_pages + nr_pages;
6939 if (new_pages > page_limit)
6941 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6942 new_pages) != cur_pages);
6947 static void io_mem_free(void *ptr)
6954 page = virt_to_head_page(ptr);
6955 if (put_page_testzero(page))
6956 free_compound_page(page);
6959 static void *io_mem_alloc(size_t size)
6961 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6964 return (void *) __get_free_pages(gfp_flags, get_order(size));
6967 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6970 struct io_rings *rings;
6971 size_t off, sq_array_size;
6973 off = struct_size(rings, cqes, cq_entries);
6974 if (off == SIZE_MAX)
6978 off = ALIGN(off, SMP_CACHE_BYTES);
6983 sq_array_size = array_size(sizeof(u32), sq_entries);
6984 if (sq_array_size == SIZE_MAX)
6987 if (check_add_overflow(off, sq_array_size, &off))
6996 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
7000 pages = (size_t)1 << get_order(
7001 rings_size(sq_entries, cq_entries, NULL));
7002 pages += (size_t)1 << get_order(
7003 array_size(sizeof(struct io_uring_sqe), sq_entries));
7008 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
7012 if (!ctx->user_bufs)
7015 for (i = 0; i < ctx->nr_user_bufs; i++) {
7016 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7018 for (j = 0; j < imu->nr_bvecs; j++)
7019 unpin_user_page(imu->bvec[j].bv_page);
7021 if (ctx->account_mem)
7022 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7027 kfree(ctx->user_bufs);
7028 ctx->user_bufs = NULL;
7029 ctx->nr_user_bufs = 0;
7033 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7034 void __user *arg, unsigned index)
7036 struct iovec __user *src;
7038 #ifdef CONFIG_COMPAT
7040 struct compat_iovec __user *ciovs;
7041 struct compat_iovec ciov;
7043 ciovs = (struct compat_iovec __user *) arg;
7044 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7047 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7048 dst->iov_len = ciov.iov_len;
7052 src = (struct iovec __user *) arg;
7053 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7058 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7061 struct vm_area_struct **vmas = NULL;
7062 struct page **pages = NULL;
7063 int i, j, got_pages = 0;
7068 if (!nr_args || nr_args > UIO_MAXIOV)
7071 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7073 if (!ctx->user_bufs)
7076 for (i = 0; i < nr_args; i++) {
7077 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7078 unsigned long off, start, end, ubuf;
7083 ret = io_copy_iov(ctx, &iov, arg, i);
7088 * Don't impose further limits on the size and buffer
7089 * constraints here, we'll -EINVAL later when IO is
7090 * submitted if they are wrong.
7093 if (!iov.iov_base || !iov.iov_len)
7096 /* arbitrary limit, but we need something */
7097 if (iov.iov_len > SZ_1G)
7100 ubuf = (unsigned long) iov.iov_base;
7101 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7102 start = ubuf >> PAGE_SHIFT;
7103 nr_pages = end - start;
7105 if (ctx->account_mem) {
7106 ret = io_account_mem(ctx->user, nr_pages);
7112 if (!pages || nr_pages > got_pages) {
7115 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7117 vmas = kvmalloc_array(nr_pages,
7118 sizeof(struct vm_area_struct *),
7120 if (!pages || !vmas) {
7122 if (ctx->account_mem)
7123 io_unaccount_mem(ctx->user, nr_pages);
7126 got_pages = nr_pages;
7129 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7133 if (ctx->account_mem)
7134 io_unaccount_mem(ctx->user, nr_pages);
7139 down_read(¤t->mm->mmap_sem);
7140 pret = pin_user_pages(ubuf, nr_pages,
7141 FOLL_WRITE | FOLL_LONGTERM,
7143 if (pret == nr_pages) {
7144 /* don't support file backed memory */
7145 for (j = 0; j < nr_pages; j++) {
7146 struct vm_area_struct *vma = vmas[j];
7149 !is_file_hugepages(vma->vm_file)) {
7155 ret = pret < 0 ? pret : -EFAULT;
7157 up_read(¤t->mm->mmap_sem);
7160 * if we did partial map, or found file backed vmas,
7161 * release any pages we did get
7164 unpin_user_pages(pages, pret);
7165 if (ctx->account_mem)
7166 io_unaccount_mem(ctx->user, nr_pages);
7171 off = ubuf & ~PAGE_MASK;
7173 for (j = 0; j < nr_pages; j++) {
7176 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7177 imu->bvec[j].bv_page = pages[j];
7178 imu->bvec[j].bv_len = vec_len;
7179 imu->bvec[j].bv_offset = off;
7183 /* store original address for later verification */
7185 imu->len = iov.iov_len;
7186 imu->nr_bvecs = nr_pages;
7188 ctx->nr_user_bufs++;
7196 io_sqe_buffer_unregister(ctx);
7200 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7202 __s32 __user *fds = arg;
7208 if (copy_from_user(&fd, fds, sizeof(*fds)))
7211 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7212 if (IS_ERR(ctx->cq_ev_fd)) {
7213 int ret = PTR_ERR(ctx->cq_ev_fd);
7214 ctx->cq_ev_fd = NULL;
7221 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7223 if (ctx->cq_ev_fd) {
7224 eventfd_ctx_put(ctx->cq_ev_fd);
7225 ctx->cq_ev_fd = NULL;
7232 static int __io_destroy_buffers(int id, void *p, void *data)
7234 struct io_ring_ctx *ctx = data;
7235 struct io_buffer *buf = p;
7237 __io_remove_buffers(ctx, buf, id, -1U);
7241 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7243 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7244 idr_destroy(&ctx->io_buffer_idr);
7247 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7249 io_finish_async(ctx);
7251 mmdrop(ctx->sqo_mm);
7253 io_iopoll_reap_events(ctx);
7254 io_sqe_buffer_unregister(ctx);
7255 io_sqe_files_unregister(ctx);
7256 io_eventfd_unregister(ctx);
7257 io_destroy_buffers(ctx);
7258 idr_destroy(&ctx->personality_idr);
7260 #if defined(CONFIG_UNIX)
7261 if (ctx->ring_sock) {
7262 ctx->ring_sock->file = NULL; /* so that iput() is called */
7263 sock_release(ctx->ring_sock);
7267 io_mem_free(ctx->rings);
7268 io_mem_free(ctx->sq_sqes);
7270 percpu_ref_exit(&ctx->refs);
7271 if (ctx->account_mem)
7272 io_unaccount_mem(ctx->user,
7273 ring_pages(ctx->sq_entries, ctx->cq_entries));
7274 free_uid(ctx->user);
7275 put_cred(ctx->creds);
7276 kfree(ctx->completions);
7277 kfree(ctx->cancel_hash);
7278 kmem_cache_free(req_cachep, ctx->fallback_req);
7282 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7284 struct io_ring_ctx *ctx = file->private_data;
7287 poll_wait(file, &ctx->cq_wait, wait);
7289 * synchronizes with barrier from wq_has_sleeper call in
7293 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7294 ctx->rings->sq_ring_entries)
7295 mask |= EPOLLOUT | EPOLLWRNORM;
7296 if (io_cqring_events(ctx, false))
7297 mask |= EPOLLIN | EPOLLRDNORM;
7302 static int io_uring_fasync(int fd, struct file *file, int on)
7304 struct io_ring_ctx *ctx = file->private_data;
7306 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7309 static int io_remove_personalities(int id, void *p, void *data)
7311 struct io_ring_ctx *ctx = data;
7312 const struct cred *cred;
7314 cred = idr_remove(&ctx->personality_idr, id);
7320 static void io_ring_exit_work(struct work_struct *work)
7322 struct io_ring_ctx *ctx;
7324 ctx = container_of(work, struct io_ring_ctx, exit_work);
7326 io_cqring_overflow_flush(ctx, true);
7328 wait_for_completion(&ctx->completions[0]);
7329 io_ring_ctx_free(ctx);
7332 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7334 mutex_lock(&ctx->uring_lock);
7335 percpu_ref_kill(&ctx->refs);
7336 mutex_unlock(&ctx->uring_lock);
7339 * Wait for sq thread to idle, if we have one. It won't spin on new
7340 * work after we've killed the ctx ref above. This is important to do
7341 * before we cancel existing commands, as the thread could otherwise
7342 * be queueing new work post that. If that's work we need to cancel,
7343 * it could cause shutdown to hang.
7345 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7348 io_kill_timeouts(ctx);
7349 io_poll_remove_all(ctx);
7352 io_wq_cancel_all(ctx->io_wq);
7354 io_iopoll_reap_events(ctx);
7355 /* if we failed setting up the ctx, we might not have any rings */
7357 io_cqring_overflow_flush(ctx, true);
7358 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7359 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7360 queue_work(system_wq, &ctx->exit_work);
7363 static int io_uring_release(struct inode *inode, struct file *file)
7365 struct io_ring_ctx *ctx = file->private_data;
7367 file->private_data = NULL;
7368 io_ring_ctx_wait_and_kill(ctx);
7372 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7373 struct files_struct *files)
7375 struct io_kiocb *req;
7378 while (!list_empty_careful(&ctx->inflight_list)) {
7379 struct io_kiocb *cancel_req = NULL;
7381 spin_lock_irq(&ctx->inflight_lock);
7382 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7383 if (req->work.files != files)
7385 /* req is being completed, ignore */
7386 if (!refcount_inc_not_zero(&req->refs))
7392 prepare_to_wait(&ctx->inflight_wait, &wait,
7393 TASK_UNINTERRUPTIBLE);
7394 spin_unlock_irq(&ctx->inflight_lock);
7396 /* We need to keep going until we don't find a matching req */
7400 if (cancel_req->flags & REQ_F_OVERFLOW) {
7401 spin_lock_irq(&ctx->completion_lock);
7402 list_del(&cancel_req->list);
7403 cancel_req->flags &= ~REQ_F_OVERFLOW;
7404 if (list_empty(&ctx->cq_overflow_list)) {
7405 clear_bit(0, &ctx->sq_check_overflow);
7406 clear_bit(0, &ctx->cq_check_overflow);
7408 spin_unlock_irq(&ctx->completion_lock);
7410 WRITE_ONCE(ctx->rings->cq_overflow,
7411 atomic_inc_return(&ctx->cached_cq_overflow));
7414 * Put inflight ref and overflow ref. If that's
7415 * all we had, then we're done with this request.
7417 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7418 io_put_req(cancel_req);
7423 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7424 io_put_req(cancel_req);
7427 finish_wait(&ctx->inflight_wait, &wait);
7430 static int io_uring_flush(struct file *file, void *data)
7432 struct io_ring_ctx *ctx = file->private_data;
7434 io_uring_cancel_files(ctx, data);
7437 * If the task is going away, cancel work it may have pending
7439 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7440 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7445 static void *io_uring_validate_mmap_request(struct file *file,
7446 loff_t pgoff, size_t sz)
7448 struct io_ring_ctx *ctx = file->private_data;
7449 loff_t offset = pgoff << PAGE_SHIFT;
7454 case IORING_OFF_SQ_RING:
7455 case IORING_OFF_CQ_RING:
7458 case IORING_OFF_SQES:
7462 return ERR_PTR(-EINVAL);
7465 page = virt_to_head_page(ptr);
7466 if (sz > page_size(page))
7467 return ERR_PTR(-EINVAL);
7474 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7476 size_t sz = vma->vm_end - vma->vm_start;
7480 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7482 return PTR_ERR(ptr);
7484 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7485 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7488 #else /* !CONFIG_MMU */
7490 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7492 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7495 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7497 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7500 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7501 unsigned long addr, unsigned long len,
7502 unsigned long pgoff, unsigned long flags)
7506 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7508 return PTR_ERR(ptr);
7510 return (unsigned long) ptr;
7513 #endif /* !CONFIG_MMU */
7515 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7516 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7519 struct io_ring_ctx *ctx;
7524 if (current->task_works)
7527 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7535 if (f.file->f_op != &io_uring_fops)
7539 ctx = f.file->private_data;
7540 if (!percpu_ref_tryget(&ctx->refs))
7544 * For SQ polling, the thread will do all submissions and completions.
7545 * Just return the requested submit count, and wake the thread if
7549 if (ctx->flags & IORING_SETUP_SQPOLL) {
7550 if (!list_empty_careful(&ctx->cq_overflow_list))
7551 io_cqring_overflow_flush(ctx, false);
7552 if (flags & IORING_ENTER_SQ_WAKEUP)
7553 wake_up(&ctx->sqo_wait);
7554 submitted = to_submit;
7555 } else if (to_submit) {
7556 mutex_lock(&ctx->uring_lock);
7557 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7558 mutex_unlock(&ctx->uring_lock);
7560 if (submitted != to_submit)
7563 if (flags & IORING_ENTER_GETEVENTS) {
7564 unsigned nr_events = 0;
7566 min_complete = min(min_complete, ctx->cq_entries);
7569 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7570 * space applications don't need to do io completion events
7571 * polling again, they can rely on io_sq_thread to do polling
7572 * work, which can reduce cpu usage and uring_lock contention.
7574 if (ctx->flags & IORING_SETUP_IOPOLL &&
7575 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7576 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7578 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7583 percpu_ref_put(&ctx->refs);
7586 return submitted ? submitted : ret;
7589 #ifdef CONFIG_PROC_FS
7590 static int io_uring_show_cred(int id, void *p, void *data)
7592 const struct cred *cred = p;
7593 struct seq_file *m = data;
7594 struct user_namespace *uns = seq_user_ns(m);
7595 struct group_info *gi;
7600 seq_printf(m, "%5d\n", id);
7601 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7602 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7603 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7604 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7605 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7606 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7607 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7608 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7609 seq_puts(m, "\n\tGroups:\t");
7610 gi = cred->group_info;
7611 for (g = 0; g < gi->ngroups; g++) {
7612 seq_put_decimal_ull(m, g ? " " : "",
7613 from_kgid_munged(uns, gi->gid[g]));
7615 seq_puts(m, "\n\tCapEff:\t");
7616 cap = cred->cap_effective;
7617 CAP_FOR_EACH_U32(__capi)
7618 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7623 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7627 mutex_lock(&ctx->uring_lock);
7628 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7629 for (i = 0; i < ctx->nr_user_files; i++) {
7630 struct fixed_file_table *table;
7633 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7634 f = table->files[i & IORING_FILE_TABLE_MASK];
7636 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7638 seq_printf(m, "%5u: <none>\n", i);
7640 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7641 for (i = 0; i < ctx->nr_user_bufs; i++) {
7642 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7644 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7645 (unsigned int) buf->len);
7647 if (!idr_is_empty(&ctx->personality_idr)) {
7648 seq_printf(m, "Personalities:\n");
7649 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7651 seq_printf(m, "PollList:\n");
7652 spin_lock_irq(&ctx->completion_lock);
7653 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7654 struct hlist_head *list = &ctx->cancel_hash[i];
7655 struct io_kiocb *req;
7657 hlist_for_each_entry(req, list, hash_node)
7658 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7659 req->task->task_works != NULL);
7661 spin_unlock_irq(&ctx->completion_lock);
7662 mutex_unlock(&ctx->uring_lock);
7665 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7667 struct io_ring_ctx *ctx = f->private_data;
7669 if (percpu_ref_tryget(&ctx->refs)) {
7670 __io_uring_show_fdinfo(ctx, m);
7671 percpu_ref_put(&ctx->refs);
7676 static const struct file_operations io_uring_fops = {
7677 .release = io_uring_release,
7678 .flush = io_uring_flush,
7679 .mmap = io_uring_mmap,
7681 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7682 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7684 .poll = io_uring_poll,
7685 .fasync = io_uring_fasync,
7686 #ifdef CONFIG_PROC_FS
7687 .show_fdinfo = io_uring_show_fdinfo,
7691 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7692 struct io_uring_params *p)
7694 struct io_rings *rings;
7695 size_t size, sq_array_offset;
7697 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7698 if (size == SIZE_MAX)
7701 rings = io_mem_alloc(size);
7706 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7707 rings->sq_ring_mask = p->sq_entries - 1;
7708 rings->cq_ring_mask = p->cq_entries - 1;
7709 rings->sq_ring_entries = p->sq_entries;
7710 rings->cq_ring_entries = p->cq_entries;
7711 ctx->sq_mask = rings->sq_ring_mask;
7712 ctx->cq_mask = rings->cq_ring_mask;
7713 ctx->sq_entries = rings->sq_ring_entries;
7714 ctx->cq_entries = rings->cq_ring_entries;
7716 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7717 if (size == SIZE_MAX) {
7718 io_mem_free(ctx->rings);
7723 ctx->sq_sqes = io_mem_alloc(size);
7724 if (!ctx->sq_sqes) {
7725 io_mem_free(ctx->rings);
7734 * Allocate an anonymous fd, this is what constitutes the application
7735 * visible backing of an io_uring instance. The application mmaps this
7736 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7737 * we have to tie this fd to a socket for file garbage collection purposes.
7739 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7744 #if defined(CONFIG_UNIX)
7745 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7751 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7755 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7756 O_RDWR | O_CLOEXEC);
7759 ret = PTR_ERR(file);
7763 #if defined(CONFIG_UNIX)
7764 ctx->ring_sock->file = file;
7766 fd_install(ret, file);
7769 #if defined(CONFIG_UNIX)
7770 sock_release(ctx->ring_sock);
7771 ctx->ring_sock = NULL;
7776 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7778 struct user_struct *user = NULL;
7779 struct io_ring_ctx *ctx;
7785 if (entries > IORING_MAX_ENTRIES) {
7786 if (!(p->flags & IORING_SETUP_CLAMP))
7788 entries = IORING_MAX_ENTRIES;
7792 * Use twice as many entries for the CQ ring. It's possible for the
7793 * application to drive a higher depth than the size of the SQ ring,
7794 * since the sqes are only used at submission time. This allows for
7795 * some flexibility in overcommitting a bit. If the application has
7796 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7797 * of CQ ring entries manually.
7799 p->sq_entries = roundup_pow_of_two(entries);
7800 if (p->flags & IORING_SETUP_CQSIZE) {
7802 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7803 * to a power-of-two, if it isn't already. We do NOT impose
7804 * any cq vs sq ring sizing.
7806 if (p->cq_entries < p->sq_entries)
7808 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7809 if (!(p->flags & IORING_SETUP_CLAMP))
7811 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7813 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7815 p->cq_entries = 2 * p->sq_entries;
7818 user = get_uid(current_user());
7819 account_mem = !capable(CAP_IPC_LOCK);
7822 ret = io_account_mem(user,
7823 ring_pages(p->sq_entries, p->cq_entries));
7830 ctx = io_ring_ctx_alloc(p);
7833 io_unaccount_mem(user, ring_pages(p->sq_entries,
7838 ctx->compat = in_compat_syscall();
7839 ctx->account_mem = account_mem;
7841 ctx->creds = get_current_cred();
7843 ret = io_allocate_scq_urings(ctx, p);
7847 ret = io_sq_offload_start(ctx, p);
7851 memset(&p->sq_off, 0, sizeof(p->sq_off));
7852 p->sq_off.head = offsetof(struct io_rings, sq.head);
7853 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7854 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7855 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7856 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7857 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7858 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7860 memset(&p->cq_off, 0, sizeof(p->cq_off));
7861 p->cq_off.head = offsetof(struct io_rings, cq.head);
7862 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7863 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7864 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7865 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7866 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7869 * Install ring fd as the very last thing, so we don't risk someone
7870 * having closed it before we finish setup
7872 ret = io_uring_get_fd(ctx);
7876 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7877 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7878 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7879 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7882 io_ring_ctx_wait_and_kill(ctx);
7887 * Sets up an aio uring context, and returns the fd. Applications asks for a
7888 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7889 * params structure passed in.
7891 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7893 struct io_uring_params p;
7897 if (copy_from_user(&p, params, sizeof(p)))
7899 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7904 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7905 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7906 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7909 ret = io_uring_create(entries, &p);
7913 if (copy_to_user(params, &p, sizeof(p)))
7919 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7920 struct io_uring_params __user *, params)
7922 return io_uring_setup(entries, params);
7925 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7927 struct io_uring_probe *p;
7931 size = struct_size(p, ops, nr_args);
7932 if (size == SIZE_MAX)
7934 p = kzalloc(size, GFP_KERNEL);
7939 if (copy_from_user(p, arg, size))
7942 if (memchr_inv(p, 0, size))
7945 p->last_op = IORING_OP_LAST - 1;
7946 if (nr_args > IORING_OP_LAST)
7947 nr_args = IORING_OP_LAST;
7949 for (i = 0; i < nr_args; i++) {
7951 if (!io_op_defs[i].not_supported)
7952 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7957 if (copy_to_user(arg, p, size))
7964 static int io_register_personality(struct io_ring_ctx *ctx)
7966 const struct cred *creds = get_current_cred();
7969 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7970 USHRT_MAX, GFP_KERNEL);
7976 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7978 const struct cred *old_creds;
7980 old_creds = idr_remove(&ctx->personality_idr, id);
7982 put_cred(old_creds);
7989 static bool io_register_op_must_quiesce(int op)
7992 case IORING_UNREGISTER_FILES:
7993 case IORING_REGISTER_FILES_UPDATE:
7994 case IORING_REGISTER_PROBE:
7995 case IORING_REGISTER_PERSONALITY:
7996 case IORING_UNREGISTER_PERSONALITY:
8003 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
8004 void __user *arg, unsigned nr_args)
8005 __releases(ctx->uring_lock)
8006 __acquires(ctx->uring_lock)
8011 * We're inside the ring mutex, if the ref is already dying, then
8012 * someone else killed the ctx or is already going through
8013 * io_uring_register().
8015 if (percpu_ref_is_dying(&ctx->refs))
8018 if (io_register_op_must_quiesce(opcode)) {
8019 percpu_ref_kill(&ctx->refs);
8022 * Drop uring mutex before waiting for references to exit. If
8023 * another thread is currently inside io_uring_enter() it might
8024 * need to grab the uring_lock to make progress. If we hold it
8025 * here across the drain wait, then we can deadlock. It's safe
8026 * to drop the mutex here, since no new references will come in
8027 * after we've killed the percpu ref.
8029 mutex_unlock(&ctx->uring_lock);
8030 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8031 mutex_lock(&ctx->uring_lock);
8033 percpu_ref_resurrect(&ctx->refs);
8040 case IORING_REGISTER_BUFFERS:
8041 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8043 case IORING_UNREGISTER_BUFFERS:
8047 ret = io_sqe_buffer_unregister(ctx);
8049 case IORING_REGISTER_FILES:
8050 ret = io_sqe_files_register(ctx, arg, nr_args);
8052 case IORING_UNREGISTER_FILES:
8056 ret = io_sqe_files_unregister(ctx);
8058 case IORING_REGISTER_FILES_UPDATE:
8059 ret = io_sqe_files_update(ctx, arg, nr_args);
8061 case IORING_REGISTER_EVENTFD:
8062 case IORING_REGISTER_EVENTFD_ASYNC:
8066 ret = io_eventfd_register(ctx, arg);
8069 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8070 ctx->eventfd_async = 1;
8072 ctx->eventfd_async = 0;
8074 case IORING_UNREGISTER_EVENTFD:
8078 ret = io_eventfd_unregister(ctx);
8080 case IORING_REGISTER_PROBE:
8082 if (!arg || nr_args > 256)
8084 ret = io_probe(ctx, arg, nr_args);
8086 case IORING_REGISTER_PERSONALITY:
8090 ret = io_register_personality(ctx);
8092 case IORING_UNREGISTER_PERSONALITY:
8096 ret = io_unregister_personality(ctx, nr_args);
8103 if (io_register_op_must_quiesce(opcode)) {
8104 /* bring the ctx back to life */
8105 percpu_ref_reinit(&ctx->refs);
8107 reinit_completion(&ctx->completions[0]);
8112 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8113 void __user *, arg, unsigned int, nr_args)
8115 struct io_ring_ctx *ctx;
8124 if (f.file->f_op != &io_uring_fops)
8127 ctx = f.file->private_data;
8129 mutex_lock(&ctx->uring_lock);
8130 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8131 mutex_unlock(&ctx->uring_lock);
8132 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8133 ctx->cq_ev_fd != NULL, ret);
8139 static int __init io_uring_init(void)
8141 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8142 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8143 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8146 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8147 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8148 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8149 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8150 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8151 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8152 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8153 BUILD_BUG_SQE_ELEM(8, __u64, off);
8154 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8155 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8156 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8157 BUILD_BUG_SQE_ELEM(24, __u32, len);
8158 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8159 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8160 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8161 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8162 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8163 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8164 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8165 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8166 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8167 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8168 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8169 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8170 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8171 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8172 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8173 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8174 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8175 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8177 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8178 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8179 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8182 __initcall(io_uring_init);