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 /* hash wq insertion if file is a regular file */
684 unsigned hash_reg_file : 1;
685 /* unbound wq insertion if file is a non-regular file */
686 unsigned unbound_nonreg_file : 1;
687 /* opcode is not supported by this kernel */
688 unsigned not_supported : 1;
689 /* needs file table */
690 unsigned file_table : 1;
692 unsigned needs_fs : 1;
693 /* set if opcode supports polled "wait" */
695 unsigned pollout : 1;
696 /* op supports buffer selection */
697 unsigned buffer_select : 1;
700 static const struct io_op_def io_op_defs[] = {
701 [IORING_OP_NOP] = {},
702 [IORING_OP_READV] = {
706 .unbound_nonreg_file = 1,
710 [IORING_OP_WRITEV] = {
715 .unbound_nonreg_file = 1,
718 [IORING_OP_FSYNC] = {
721 [IORING_OP_READ_FIXED] = {
723 .unbound_nonreg_file = 1,
726 [IORING_OP_WRITE_FIXED] = {
729 .unbound_nonreg_file = 1,
732 [IORING_OP_POLL_ADD] = {
734 .unbound_nonreg_file = 1,
736 [IORING_OP_POLL_REMOVE] = {},
737 [IORING_OP_SYNC_FILE_RANGE] = {
740 [IORING_OP_SENDMSG] = {
744 .unbound_nonreg_file = 1,
748 [IORING_OP_RECVMSG] = {
752 .unbound_nonreg_file = 1,
757 [IORING_OP_TIMEOUT] = {
761 [IORING_OP_TIMEOUT_REMOVE] = {},
762 [IORING_OP_ACCEPT] = {
765 .unbound_nonreg_file = 1,
769 [IORING_OP_ASYNC_CANCEL] = {},
770 [IORING_OP_LINK_TIMEOUT] = {
774 [IORING_OP_CONNECT] = {
778 .unbound_nonreg_file = 1,
781 [IORING_OP_FALLOCATE] = {
784 [IORING_OP_OPENAT] = {
788 [IORING_OP_CLOSE] = {
792 [IORING_OP_FILES_UPDATE] = {
796 [IORING_OP_STATX] = {
804 .unbound_nonreg_file = 1,
808 [IORING_OP_WRITE] = {
811 .unbound_nonreg_file = 1,
814 [IORING_OP_FADVISE] = {
817 [IORING_OP_MADVISE] = {
823 .unbound_nonreg_file = 1,
829 .unbound_nonreg_file = 1,
833 [IORING_OP_OPENAT2] = {
837 [IORING_OP_EPOLL_CTL] = {
838 .unbound_nonreg_file = 1,
841 [IORING_OP_SPLICE] = {
844 .unbound_nonreg_file = 1,
846 [IORING_OP_PROVIDE_BUFFERS] = {},
847 [IORING_OP_REMOVE_BUFFERS] = {},
850 static void io_wq_submit_work(struct io_wq_work **workptr);
851 static void io_cqring_fill_event(struct io_kiocb *req, long res);
852 static void io_put_req(struct io_kiocb *req);
853 static void __io_double_put_req(struct io_kiocb *req);
854 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
855 static void io_queue_linked_timeout(struct io_kiocb *req);
856 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
857 struct io_uring_files_update *ip,
859 static int io_grab_files(struct io_kiocb *req);
860 static void io_cleanup_req(struct io_kiocb *req);
861 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
862 int fd, struct file **out_file, bool fixed);
863 static void __io_queue_sqe(struct io_kiocb *req,
864 const struct io_uring_sqe *sqe);
866 static struct kmem_cache *req_cachep;
868 static const struct file_operations io_uring_fops;
870 struct sock *io_uring_get_socket(struct file *file)
872 #if defined(CONFIG_UNIX)
873 if (file->f_op == &io_uring_fops) {
874 struct io_ring_ctx *ctx = file->private_data;
876 return ctx->ring_sock->sk;
881 EXPORT_SYMBOL(io_uring_get_socket);
883 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
885 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
887 complete(&ctx->completions[0]);
890 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
892 struct io_ring_ctx *ctx;
895 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
899 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
900 if (!ctx->fallback_req)
903 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
904 if (!ctx->completions)
908 * Use 5 bits less than the max cq entries, that should give us around
909 * 32 entries per hash list if totally full and uniformly spread.
911 hash_bits = ilog2(p->cq_entries);
915 ctx->cancel_hash_bits = hash_bits;
916 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
918 if (!ctx->cancel_hash)
920 __hash_init(ctx->cancel_hash, 1U << hash_bits);
922 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
923 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
926 ctx->flags = p->flags;
927 init_waitqueue_head(&ctx->sqo_wait);
928 init_waitqueue_head(&ctx->cq_wait);
929 INIT_LIST_HEAD(&ctx->cq_overflow_list);
930 init_completion(&ctx->completions[0]);
931 init_completion(&ctx->completions[1]);
932 idr_init(&ctx->io_buffer_idr);
933 idr_init(&ctx->personality_idr);
934 mutex_init(&ctx->uring_lock);
935 init_waitqueue_head(&ctx->wait);
936 spin_lock_init(&ctx->completion_lock);
937 INIT_LIST_HEAD(&ctx->poll_list);
938 INIT_LIST_HEAD(&ctx->defer_list);
939 INIT_LIST_HEAD(&ctx->timeout_list);
940 init_waitqueue_head(&ctx->inflight_wait);
941 spin_lock_init(&ctx->inflight_lock);
942 INIT_LIST_HEAD(&ctx->inflight_list);
945 if (ctx->fallback_req)
946 kmem_cache_free(req_cachep, ctx->fallback_req);
947 kfree(ctx->completions);
948 kfree(ctx->cancel_hash);
953 static inline bool __req_need_defer(struct io_kiocb *req)
955 struct io_ring_ctx *ctx = req->ctx;
957 return req->sequence != ctx->cached_cq_tail
958 + atomic_read(&ctx->cached_cq_overflow);
961 static inline bool req_need_defer(struct io_kiocb *req)
963 if (unlikely(req->flags & REQ_F_IO_DRAIN))
964 return __req_need_defer(req);
969 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
971 struct io_kiocb *req;
973 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
974 if (req && !req_need_defer(req)) {
975 list_del_init(&req->list);
982 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
984 struct io_kiocb *req;
986 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
988 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
990 if (!__req_need_defer(req)) {
991 list_del_init(&req->list);
999 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1001 struct io_rings *rings = ctx->rings;
1003 /* order cqe stores with ring update */
1004 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1006 if (wq_has_sleeper(&ctx->cq_wait)) {
1007 wake_up_interruptible(&ctx->cq_wait);
1008 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1012 static inline void io_req_work_grab_env(struct io_kiocb *req,
1013 const struct io_op_def *def)
1015 if (!req->work.mm && def->needs_mm) {
1016 mmgrab(current->mm);
1017 req->work.mm = current->mm;
1019 if (!req->work.creds)
1020 req->work.creds = get_current_cred();
1021 if (!req->work.fs && def->needs_fs) {
1022 spin_lock(¤t->fs->lock);
1023 if (!current->fs->in_exec) {
1024 req->work.fs = current->fs;
1025 req->work.fs->users++;
1027 req->work.flags |= IO_WQ_WORK_CANCEL;
1029 spin_unlock(¤t->fs->lock);
1031 if (!req->work.task_pid)
1032 req->work.task_pid = task_pid_vnr(current);
1035 static inline void io_req_work_drop_env(struct io_kiocb *req)
1038 mmdrop(req->work.mm);
1039 req->work.mm = NULL;
1041 if (req->work.creds) {
1042 put_cred(req->work.creds);
1043 req->work.creds = NULL;
1046 struct fs_struct *fs = req->work.fs;
1048 spin_lock(&req->work.fs->lock);
1051 spin_unlock(&req->work.fs->lock);
1057 static inline void io_prep_async_work(struct io_kiocb *req,
1058 struct io_kiocb **link)
1060 const struct io_op_def *def = &io_op_defs[req->opcode];
1062 if (req->flags & REQ_F_ISREG) {
1063 if (def->hash_reg_file)
1064 io_wq_hash_work(&req->work, file_inode(req->file));
1066 if (def->unbound_nonreg_file)
1067 req->work.flags |= IO_WQ_WORK_UNBOUND;
1070 io_req_work_grab_env(req, def);
1072 *link = io_prep_linked_timeout(req);
1075 static inline void io_queue_async_work(struct io_kiocb *req)
1077 struct io_ring_ctx *ctx = req->ctx;
1078 struct io_kiocb *link;
1080 io_prep_async_work(req, &link);
1082 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1083 &req->work, req->flags);
1084 io_wq_enqueue(ctx->io_wq, &req->work);
1087 io_queue_linked_timeout(link);
1090 static void io_kill_timeout(struct io_kiocb *req)
1094 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1096 atomic_inc(&req->ctx->cq_timeouts);
1097 list_del_init(&req->list);
1098 req->flags |= REQ_F_COMP_LOCKED;
1099 io_cqring_fill_event(req, 0);
1104 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1106 struct io_kiocb *req, *tmp;
1108 spin_lock_irq(&ctx->completion_lock);
1109 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1110 io_kill_timeout(req);
1111 spin_unlock_irq(&ctx->completion_lock);
1114 static void io_commit_cqring(struct io_ring_ctx *ctx)
1116 struct io_kiocb *req;
1118 while ((req = io_get_timeout_req(ctx)) != NULL)
1119 io_kill_timeout(req);
1121 __io_commit_cqring(ctx);
1123 while ((req = io_get_deferred_req(ctx)) != NULL)
1124 io_queue_async_work(req);
1127 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1129 struct io_rings *rings = ctx->rings;
1132 tail = ctx->cached_cq_tail;
1134 * writes to the cq entry need to come after reading head; the
1135 * control dependency is enough as we're using WRITE_ONCE to
1138 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1141 ctx->cached_cq_tail++;
1142 return &rings->cqes[tail & ctx->cq_mask];
1145 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1149 if (!ctx->eventfd_async)
1151 return io_wq_current_is_worker();
1154 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1156 if (waitqueue_active(&ctx->wait))
1157 wake_up(&ctx->wait);
1158 if (waitqueue_active(&ctx->sqo_wait))
1159 wake_up(&ctx->sqo_wait);
1160 if (io_should_trigger_evfd(ctx))
1161 eventfd_signal(ctx->cq_ev_fd, 1);
1164 /* Returns true if there are no backlogged entries after the flush */
1165 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1167 struct io_rings *rings = ctx->rings;
1168 struct io_uring_cqe *cqe;
1169 struct io_kiocb *req;
1170 unsigned long flags;
1174 if (list_empty_careful(&ctx->cq_overflow_list))
1176 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1177 rings->cq_ring_entries))
1181 spin_lock_irqsave(&ctx->completion_lock, flags);
1183 /* if force is set, the ring is going away. always drop after that */
1185 ctx->cq_overflow_flushed = 1;
1188 while (!list_empty(&ctx->cq_overflow_list)) {
1189 cqe = io_get_cqring(ctx);
1193 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1195 list_move(&req->list, &list);
1196 req->flags &= ~REQ_F_OVERFLOW;
1198 WRITE_ONCE(cqe->user_data, req->user_data);
1199 WRITE_ONCE(cqe->res, req->result);
1200 WRITE_ONCE(cqe->flags, req->cflags);
1202 WRITE_ONCE(ctx->rings->cq_overflow,
1203 atomic_inc_return(&ctx->cached_cq_overflow));
1207 io_commit_cqring(ctx);
1209 clear_bit(0, &ctx->sq_check_overflow);
1210 clear_bit(0, &ctx->cq_check_overflow);
1212 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1213 io_cqring_ev_posted(ctx);
1215 while (!list_empty(&list)) {
1216 req = list_first_entry(&list, struct io_kiocb, list);
1217 list_del(&req->list);
1224 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1226 struct io_ring_ctx *ctx = req->ctx;
1227 struct io_uring_cqe *cqe;
1229 trace_io_uring_complete(ctx, req->user_data, res);
1232 * If we can't get a cq entry, userspace overflowed the
1233 * submission (by quite a lot). Increment the overflow count in
1236 cqe = io_get_cqring(ctx);
1238 WRITE_ONCE(cqe->user_data, req->user_data);
1239 WRITE_ONCE(cqe->res, res);
1240 WRITE_ONCE(cqe->flags, cflags);
1241 } else if (ctx->cq_overflow_flushed) {
1242 WRITE_ONCE(ctx->rings->cq_overflow,
1243 atomic_inc_return(&ctx->cached_cq_overflow));
1245 if (list_empty(&ctx->cq_overflow_list)) {
1246 set_bit(0, &ctx->sq_check_overflow);
1247 set_bit(0, &ctx->cq_check_overflow);
1249 req->flags |= REQ_F_OVERFLOW;
1250 refcount_inc(&req->refs);
1252 req->cflags = cflags;
1253 list_add_tail(&req->list, &ctx->cq_overflow_list);
1257 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1259 __io_cqring_fill_event(req, res, 0);
1262 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1264 struct io_ring_ctx *ctx = req->ctx;
1265 unsigned long flags;
1267 spin_lock_irqsave(&ctx->completion_lock, flags);
1268 __io_cqring_fill_event(req, res, cflags);
1269 io_commit_cqring(ctx);
1270 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1272 io_cqring_ev_posted(ctx);
1275 static void io_cqring_add_event(struct io_kiocb *req, long res)
1277 __io_cqring_add_event(req, res, 0);
1280 static inline bool io_is_fallback_req(struct io_kiocb *req)
1282 return req == (struct io_kiocb *)
1283 ((unsigned long) req->ctx->fallback_req & ~1UL);
1286 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1288 struct io_kiocb *req;
1290 req = ctx->fallback_req;
1291 if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req))
1297 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1298 struct io_submit_state *state)
1300 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1301 struct io_kiocb *req;
1304 req = kmem_cache_alloc(req_cachep, gfp);
1307 } else if (!state->free_reqs) {
1311 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1312 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1315 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1316 * retry single alloc to be on the safe side.
1318 if (unlikely(ret <= 0)) {
1319 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1320 if (!state->reqs[0])
1324 state->free_reqs = ret - 1;
1325 req = state->reqs[ret - 1];
1328 req = state->reqs[state->free_reqs];
1333 return io_get_fallback_req(ctx);
1336 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1340 percpu_ref_put(req->fixed_file_refs);
1345 static void __io_req_aux_free(struct io_kiocb *req)
1347 if (req->flags & REQ_F_NEED_CLEANUP)
1348 io_cleanup_req(req);
1352 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1354 put_task_struct(req->task);
1356 io_req_work_drop_env(req);
1359 static void __io_free_req(struct io_kiocb *req)
1361 __io_req_aux_free(req);
1363 if (req->flags & REQ_F_INFLIGHT) {
1364 struct io_ring_ctx *ctx = req->ctx;
1365 unsigned long flags;
1367 spin_lock_irqsave(&ctx->inflight_lock, flags);
1368 list_del(&req->inflight_entry);
1369 if (waitqueue_active(&ctx->inflight_wait))
1370 wake_up(&ctx->inflight_wait);
1371 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1374 percpu_ref_put(&req->ctx->refs);
1375 if (likely(!io_is_fallback_req(req)))
1376 kmem_cache_free(req_cachep, req);
1378 clear_bit_unlock(0, (unsigned long *) &req->ctx->fallback_req);
1382 void *reqs[IO_IOPOLL_BATCH];
1387 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1391 if (rb->need_iter) {
1392 int i, inflight = 0;
1393 unsigned long flags;
1395 for (i = 0; i < rb->to_free; i++) {
1396 struct io_kiocb *req = rb->reqs[i];
1398 if (req->flags & REQ_F_INFLIGHT)
1400 __io_req_aux_free(req);
1405 spin_lock_irqsave(&ctx->inflight_lock, flags);
1406 for (i = 0; i < rb->to_free; i++) {
1407 struct io_kiocb *req = rb->reqs[i];
1409 if (req->flags & REQ_F_INFLIGHT) {
1410 list_del(&req->inflight_entry);
1415 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1417 if (waitqueue_active(&ctx->inflight_wait))
1418 wake_up(&ctx->inflight_wait);
1421 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1422 percpu_ref_put_many(&ctx->refs, rb->to_free);
1423 rb->to_free = rb->need_iter = 0;
1426 static bool io_link_cancel_timeout(struct io_kiocb *req)
1428 struct io_ring_ctx *ctx = req->ctx;
1431 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1433 io_cqring_fill_event(req, -ECANCELED);
1434 io_commit_cqring(ctx);
1435 req->flags &= ~REQ_F_LINK_HEAD;
1443 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1445 struct io_ring_ctx *ctx = req->ctx;
1446 bool wake_ev = false;
1448 /* Already got next link */
1449 if (req->flags & REQ_F_LINK_NEXT)
1453 * The list should never be empty when we are called here. But could
1454 * potentially happen if the chain is messed up, check to be on the
1457 while (!list_empty(&req->link_list)) {
1458 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1459 struct io_kiocb, link_list);
1461 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1462 (nxt->flags & REQ_F_TIMEOUT))) {
1463 list_del_init(&nxt->link_list);
1464 wake_ev |= io_link_cancel_timeout(nxt);
1465 req->flags &= ~REQ_F_LINK_TIMEOUT;
1469 list_del_init(&req->link_list);
1470 if (!list_empty(&nxt->link_list))
1471 nxt->flags |= REQ_F_LINK_HEAD;
1476 req->flags |= REQ_F_LINK_NEXT;
1478 io_cqring_ev_posted(ctx);
1482 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1484 static void io_fail_links(struct io_kiocb *req)
1486 struct io_ring_ctx *ctx = req->ctx;
1487 unsigned long flags;
1489 spin_lock_irqsave(&ctx->completion_lock, flags);
1491 while (!list_empty(&req->link_list)) {
1492 struct io_kiocb *link = list_first_entry(&req->link_list,
1493 struct io_kiocb, link_list);
1495 list_del_init(&link->link_list);
1496 trace_io_uring_fail_link(req, link);
1498 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1499 link->opcode == IORING_OP_LINK_TIMEOUT) {
1500 io_link_cancel_timeout(link);
1502 io_cqring_fill_event(link, -ECANCELED);
1503 __io_double_put_req(link);
1505 req->flags &= ~REQ_F_LINK_TIMEOUT;
1508 io_commit_cqring(ctx);
1509 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1510 io_cqring_ev_posted(ctx);
1513 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1515 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1519 * If LINK is set, we have dependent requests in this chain. If we
1520 * didn't fail this request, queue the first one up, moving any other
1521 * dependencies to the next request. In case of failure, fail the rest
1524 if (req->flags & REQ_F_FAIL_LINK) {
1526 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1527 REQ_F_LINK_TIMEOUT) {
1528 struct io_ring_ctx *ctx = req->ctx;
1529 unsigned long flags;
1532 * If this is a timeout link, we could be racing with the
1533 * timeout timer. Grab the completion lock for this case to
1534 * protect against that.
1536 spin_lock_irqsave(&ctx->completion_lock, flags);
1537 io_req_link_next(req, nxt);
1538 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1540 io_req_link_next(req, nxt);
1544 static void io_free_req(struct io_kiocb *req)
1546 struct io_kiocb *nxt = NULL;
1548 io_req_find_next(req, &nxt);
1552 io_queue_async_work(nxt);
1555 static void io_link_work_cb(struct io_wq_work **workptr)
1557 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1558 struct io_kiocb *link;
1560 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1561 io_queue_linked_timeout(link);
1562 io_wq_submit_work(workptr);
1565 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1567 struct io_kiocb *link;
1568 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1570 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1571 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1573 *workptr = &nxt->work;
1574 link = io_prep_linked_timeout(nxt);
1576 nxt->work.func = io_link_work_cb;
1580 * Drop reference to request, return next in chain (if there is one) if this
1581 * was the last reference to this request.
1583 __attribute__((nonnull))
1584 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1586 if (refcount_dec_and_test(&req->refs)) {
1587 io_req_find_next(req, nxtptr);
1592 static void io_put_req(struct io_kiocb *req)
1594 if (refcount_dec_and_test(&req->refs))
1598 static void io_steal_work(struct io_kiocb *req,
1599 struct io_wq_work **workptr)
1602 * It's in an io-wq worker, so there always should be at least
1603 * one reference, which will be dropped in io_put_work() just
1604 * after the current handler returns.
1606 * It also means, that if the counter dropped to 1, then there is
1607 * no asynchronous users left, so it's safe to steal the next work.
1609 if (refcount_read(&req->refs) == 1) {
1610 struct io_kiocb *nxt = NULL;
1612 io_req_find_next(req, &nxt);
1614 io_wq_assign_next(workptr, nxt);
1619 * Must only be used if we don't need to care about links, usually from
1620 * within the completion handling itself.
1622 static void __io_double_put_req(struct io_kiocb *req)
1624 /* drop both submit and complete references */
1625 if (refcount_sub_and_test(2, &req->refs))
1629 static void io_double_put_req(struct io_kiocb *req)
1631 /* drop both submit and complete references */
1632 if (refcount_sub_and_test(2, &req->refs))
1636 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1638 struct io_rings *rings = ctx->rings;
1640 if (test_bit(0, &ctx->cq_check_overflow)) {
1642 * noflush == true is from the waitqueue handler, just ensure
1643 * we wake up the task, and the next invocation will flush the
1644 * entries. We cannot safely to it from here.
1646 if (noflush && !list_empty(&ctx->cq_overflow_list))
1649 io_cqring_overflow_flush(ctx, false);
1652 /* See comment at the top of this file */
1654 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1657 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1659 struct io_rings *rings = ctx->rings;
1661 /* make sure SQ entry isn't read before tail */
1662 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1665 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1667 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1670 if (req->file || req->io)
1673 rb->reqs[rb->to_free++] = req;
1674 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1675 io_free_req_many(req->ctx, rb);
1679 static int io_put_kbuf(struct io_kiocb *req)
1681 struct io_buffer *kbuf;
1684 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1685 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1686 cflags |= IORING_CQE_F_BUFFER;
1693 * Find and free completed poll iocbs
1695 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1696 struct list_head *done)
1698 struct req_batch rb;
1699 struct io_kiocb *req;
1701 rb.to_free = rb.need_iter = 0;
1702 while (!list_empty(done)) {
1705 req = list_first_entry(done, struct io_kiocb, list);
1706 list_del(&req->list);
1708 if (req->flags & REQ_F_BUFFER_SELECTED)
1709 cflags = io_put_kbuf(req);
1711 __io_cqring_fill_event(req, req->result, cflags);
1714 if (refcount_dec_and_test(&req->refs) &&
1715 !io_req_multi_free(&rb, req))
1719 io_commit_cqring(ctx);
1720 if (ctx->flags & IORING_SETUP_SQPOLL)
1721 io_cqring_ev_posted(ctx);
1722 io_free_req_many(ctx, &rb);
1725 static void io_iopoll_queue(struct list_head *again)
1727 struct io_kiocb *req;
1730 req = list_first_entry(again, struct io_kiocb, list);
1731 list_del(&req->list);
1732 refcount_inc(&req->refs);
1733 io_queue_async_work(req);
1734 } while (!list_empty(again));
1737 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1740 struct io_kiocb *req, *tmp;
1747 * Only spin for completions if we don't have multiple devices hanging
1748 * off our complete list, and we're under the requested amount.
1750 spin = !ctx->poll_multi_file && *nr_events < min;
1753 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1754 struct kiocb *kiocb = &req->rw.kiocb;
1757 * Move completed and retryable entries to our local lists.
1758 * If we find a request that requires polling, break out
1759 * and complete those lists first, if we have entries there.
1761 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1762 list_move_tail(&req->list, &done);
1765 if (!list_empty(&done))
1768 if (req->result == -EAGAIN) {
1769 list_move_tail(&req->list, &again);
1772 if (!list_empty(&again))
1775 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1784 if (!list_empty(&done))
1785 io_iopoll_complete(ctx, nr_events, &done);
1787 if (!list_empty(&again))
1788 io_iopoll_queue(&again);
1794 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1795 * non-spinning poll check - we'll still enter the driver poll loop, but only
1796 * as a non-spinning completion check.
1798 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1801 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1804 ret = io_do_iopoll(ctx, nr_events, min);
1807 if (!min || *nr_events >= min)
1815 * We can't just wait for polled events to come to us, we have to actively
1816 * find and complete them.
1818 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1820 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1823 mutex_lock(&ctx->uring_lock);
1824 while (!list_empty(&ctx->poll_list)) {
1825 unsigned int nr_events = 0;
1827 io_iopoll_getevents(ctx, &nr_events, 1);
1830 * Ensure we allow local-to-the-cpu processing to take place,
1831 * in this case we need to ensure that we reap all events.
1835 mutex_unlock(&ctx->uring_lock);
1838 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1841 int iters = 0, ret = 0;
1844 * We disallow the app entering submit/complete with polling, but we
1845 * still need to lock the ring to prevent racing with polled issue
1846 * that got punted to a workqueue.
1848 mutex_lock(&ctx->uring_lock);
1853 * Don't enter poll loop if we already have events pending.
1854 * If we do, we can potentially be spinning for commands that
1855 * already triggered a CQE (eg in error).
1857 if (io_cqring_events(ctx, false))
1861 * If a submit got punted to a workqueue, we can have the
1862 * application entering polling for a command before it gets
1863 * issued. That app will hold the uring_lock for the duration
1864 * of the poll right here, so we need to take a breather every
1865 * now and then to ensure that the issue has a chance to add
1866 * the poll to the issued list. Otherwise we can spin here
1867 * forever, while the workqueue is stuck trying to acquire the
1870 if (!(++iters & 7)) {
1871 mutex_unlock(&ctx->uring_lock);
1872 mutex_lock(&ctx->uring_lock);
1875 if (*nr_events < min)
1876 tmin = min - *nr_events;
1878 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1882 } while (min && !*nr_events && !need_resched());
1884 mutex_unlock(&ctx->uring_lock);
1888 static void kiocb_end_write(struct io_kiocb *req)
1891 * Tell lockdep we inherited freeze protection from submission
1894 if (req->flags & REQ_F_ISREG) {
1895 struct inode *inode = file_inode(req->file);
1897 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1899 file_end_write(req->file);
1902 static inline void req_set_fail_links(struct io_kiocb *req)
1904 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1905 req->flags |= REQ_F_FAIL_LINK;
1908 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1910 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1913 if (kiocb->ki_flags & IOCB_WRITE)
1914 kiocb_end_write(req);
1916 if (res != req->result)
1917 req_set_fail_links(req);
1918 if (req->flags & REQ_F_BUFFER_SELECTED)
1919 cflags = io_put_kbuf(req);
1920 __io_cqring_add_event(req, res, cflags);
1923 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1925 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1927 io_complete_rw_common(kiocb, res);
1931 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1933 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1935 if (kiocb->ki_flags & IOCB_WRITE)
1936 kiocb_end_write(req);
1938 if (res != req->result)
1939 req_set_fail_links(req);
1942 req->flags |= REQ_F_IOPOLL_COMPLETED;
1946 * After the iocb has been issued, it's safe to be found on the poll list.
1947 * Adding the kiocb to the list AFTER submission ensures that we don't
1948 * find it from a io_iopoll_getevents() thread before the issuer is done
1949 * accessing the kiocb cookie.
1951 static void io_iopoll_req_issued(struct io_kiocb *req)
1953 struct io_ring_ctx *ctx = req->ctx;
1956 * Track whether we have multiple files in our lists. This will impact
1957 * how we do polling eventually, not spinning if we're on potentially
1958 * different devices.
1960 if (list_empty(&ctx->poll_list)) {
1961 ctx->poll_multi_file = false;
1962 } else if (!ctx->poll_multi_file) {
1963 struct io_kiocb *list_req;
1965 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1967 if (list_req->file != req->file)
1968 ctx->poll_multi_file = true;
1972 * For fast devices, IO may have already completed. If it has, add
1973 * it to the front so we find it first.
1975 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1976 list_add(&req->list, &ctx->poll_list);
1978 list_add_tail(&req->list, &ctx->poll_list);
1980 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1981 wq_has_sleeper(&ctx->sqo_wait))
1982 wake_up(&ctx->sqo_wait);
1985 static void io_file_put(struct io_submit_state *state)
1988 int diff = state->has_refs - state->used_refs;
1991 fput_many(state->file, diff);
1997 * Get as many references to a file as we have IOs left in this submission,
1998 * assuming most submissions are for one file, or at least that each file
1999 * has more than one submission.
2001 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2007 if (state->fd == fd) {
2014 state->file = fget_many(fd, state->ios_left);
2019 state->has_refs = state->ios_left;
2020 state->used_refs = 1;
2026 * If we tracked the file through the SCM inflight mechanism, we could support
2027 * any file. For now, just ensure that anything potentially problematic is done
2030 static bool io_file_supports_async(struct file *file, int rw)
2032 umode_t mode = file_inode(file)->i_mode;
2034 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2036 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2039 if (!(file->f_mode & FMODE_NOWAIT))
2043 return file->f_op->read_iter != NULL;
2045 return file->f_op->write_iter != NULL;
2048 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2049 bool force_nonblock)
2051 struct io_ring_ctx *ctx = req->ctx;
2052 struct kiocb *kiocb = &req->rw.kiocb;
2056 if (S_ISREG(file_inode(req->file)->i_mode))
2057 req->flags |= REQ_F_ISREG;
2059 kiocb->ki_pos = READ_ONCE(sqe->off);
2060 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2061 req->flags |= REQ_F_CUR_POS;
2062 kiocb->ki_pos = req->file->f_pos;
2064 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2065 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2066 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2070 ioprio = READ_ONCE(sqe->ioprio);
2072 ret = ioprio_check_cap(ioprio);
2076 kiocb->ki_ioprio = ioprio;
2078 kiocb->ki_ioprio = get_current_ioprio();
2080 /* don't allow async punt if RWF_NOWAIT was requested */
2081 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2082 (req->file->f_flags & O_NONBLOCK))
2083 req->flags |= REQ_F_NOWAIT;
2086 kiocb->ki_flags |= IOCB_NOWAIT;
2088 if (ctx->flags & IORING_SETUP_IOPOLL) {
2089 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2090 !kiocb->ki_filp->f_op->iopoll)
2093 kiocb->ki_flags |= IOCB_HIPRI;
2094 kiocb->ki_complete = io_complete_rw_iopoll;
2097 if (kiocb->ki_flags & IOCB_HIPRI)
2099 kiocb->ki_complete = io_complete_rw;
2102 req->rw.addr = READ_ONCE(sqe->addr);
2103 req->rw.len = READ_ONCE(sqe->len);
2104 /* we own ->private, reuse it for the buffer index / buffer ID */
2105 req->rw.kiocb.private = (void *) (unsigned long)
2106 READ_ONCE(sqe->buf_index);
2110 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2116 case -ERESTARTNOINTR:
2117 case -ERESTARTNOHAND:
2118 case -ERESTART_RESTARTBLOCK:
2120 * We can't just restart the syscall, since previously
2121 * submitted sqes may already be in progress. Just fail this
2127 kiocb->ki_complete(kiocb, ret, 0);
2131 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2133 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2135 if (req->flags & REQ_F_CUR_POS)
2136 req->file->f_pos = kiocb->ki_pos;
2137 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2138 io_complete_rw(kiocb, ret, 0);
2140 io_rw_done(kiocb, ret);
2143 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2144 struct iov_iter *iter)
2146 struct io_ring_ctx *ctx = req->ctx;
2147 size_t len = req->rw.len;
2148 struct io_mapped_ubuf *imu;
2149 unsigned index, buf_index;
2153 /* attempt to use fixed buffers without having provided iovecs */
2154 if (unlikely(!ctx->user_bufs))
2157 buf_index = (unsigned long) req->rw.kiocb.private;
2158 if (unlikely(buf_index >= ctx->nr_user_bufs))
2161 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2162 imu = &ctx->user_bufs[index];
2163 buf_addr = req->rw.addr;
2166 if (buf_addr + len < buf_addr)
2168 /* not inside the mapped region */
2169 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2173 * May not be a start of buffer, set size appropriately
2174 * and advance us to the beginning.
2176 offset = buf_addr - imu->ubuf;
2177 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2181 * Don't use iov_iter_advance() here, as it's really slow for
2182 * using the latter parts of a big fixed buffer - it iterates
2183 * over each segment manually. We can cheat a bit here, because
2186 * 1) it's a BVEC iter, we set it up
2187 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2188 * first and last bvec
2190 * So just find our index, and adjust the iterator afterwards.
2191 * If the offset is within the first bvec (or the whole first
2192 * bvec, just use iov_iter_advance(). This makes it easier
2193 * since we can just skip the first segment, which may not
2194 * be PAGE_SIZE aligned.
2196 const struct bio_vec *bvec = imu->bvec;
2198 if (offset <= bvec->bv_len) {
2199 iov_iter_advance(iter, offset);
2201 unsigned long seg_skip;
2203 /* skip first vec */
2204 offset -= bvec->bv_len;
2205 seg_skip = 1 + (offset >> PAGE_SHIFT);
2207 iter->bvec = bvec + seg_skip;
2208 iter->nr_segs -= seg_skip;
2209 iter->count -= bvec->bv_len + offset;
2210 iter->iov_offset = offset & ~PAGE_MASK;
2217 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2220 mutex_unlock(&ctx->uring_lock);
2223 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2226 * "Normal" inline submissions always hold the uring_lock, since we
2227 * grab it from the system call. Same is true for the SQPOLL offload.
2228 * The only exception is when we've detached the request and issue it
2229 * from an async worker thread, grab the lock for that case.
2232 mutex_lock(&ctx->uring_lock);
2235 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2236 int bgid, struct io_buffer *kbuf,
2239 struct io_buffer *head;
2241 if (req->flags & REQ_F_BUFFER_SELECTED)
2244 io_ring_submit_lock(req->ctx, needs_lock);
2246 lockdep_assert_held(&req->ctx->uring_lock);
2248 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2250 if (!list_empty(&head->list)) {
2251 kbuf = list_last_entry(&head->list, struct io_buffer,
2253 list_del(&kbuf->list);
2256 idr_remove(&req->ctx->io_buffer_idr, bgid);
2258 if (*len > kbuf->len)
2261 kbuf = ERR_PTR(-ENOBUFS);
2264 io_ring_submit_unlock(req->ctx, needs_lock);
2269 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2272 struct io_buffer *kbuf;
2275 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2276 bgid = (int) (unsigned long) req->rw.kiocb.private;
2277 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2280 req->rw.addr = (u64) (unsigned long) kbuf;
2281 req->flags |= REQ_F_BUFFER_SELECTED;
2282 return u64_to_user_ptr(kbuf->addr);
2285 #ifdef CONFIG_COMPAT
2286 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2289 struct compat_iovec __user *uiov;
2290 compat_ssize_t clen;
2294 uiov = u64_to_user_ptr(req->rw.addr);
2295 if (!access_ok(uiov, sizeof(*uiov)))
2297 if (__get_user(clen, &uiov->iov_len))
2303 buf = io_rw_buffer_select(req, &len, needs_lock);
2305 return PTR_ERR(buf);
2306 iov[0].iov_base = buf;
2307 iov[0].iov_len = (compat_size_t) len;
2312 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2315 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2319 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2322 len = iov[0].iov_len;
2325 buf = io_rw_buffer_select(req, &len, needs_lock);
2327 return PTR_ERR(buf);
2328 iov[0].iov_base = buf;
2329 iov[0].iov_len = len;
2333 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2336 if (req->flags & REQ_F_BUFFER_SELECTED)
2340 else if (req->rw.len > 1)
2343 #ifdef CONFIG_COMPAT
2344 if (req->ctx->compat)
2345 return io_compat_import(req, iov, needs_lock);
2348 return __io_iov_buffer_select(req, iov, needs_lock);
2351 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2352 struct iovec **iovec, struct iov_iter *iter,
2355 void __user *buf = u64_to_user_ptr(req->rw.addr);
2356 size_t sqe_len = req->rw.len;
2360 opcode = req->opcode;
2361 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2363 return io_import_fixed(req, rw, iter);
2366 /* buffer index only valid with fixed read/write, or buffer select */
2367 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2370 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2371 if (req->flags & REQ_F_BUFFER_SELECT) {
2372 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2375 return PTR_ERR(buf);
2377 req->rw.len = sqe_len;
2380 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2382 return ret < 0 ? ret : sqe_len;
2386 struct io_async_rw *iorw = &req->io->rw;
2389 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2390 if (iorw->iov == iorw->fast_iov)
2395 if (req->flags & REQ_F_BUFFER_SELECT) {
2396 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2398 ret = (*iovec)->iov_len;
2399 iov_iter_init(iter, rw, *iovec, 1, ret);
2405 #ifdef CONFIG_COMPAT
2406 if (req->ctx->compat)
2407 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2411 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2415 * For files that don't have ->read_iter() and ->write_iter(), handle them
2416 * by looping over ->read() or ->write() manually.
2418 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2419 struct iov_iter *iter)
2424 * Don't support polled IO through this interface, and we can't
2425 * support non-blocking either. For the latter, this just causes
2426 * the kiocb to be handled from an async context.
2428 if (kiocb->ki_flags & IOCB_HIPRI)
2430 if (kiocb->ki_flags & IOCB_NOWAIT)
2433 while (iov_iter_count(iter)) {
2437 if (!iov_iter_is_bvec(iter)) {
2438 iovec = iov_iter_iovec(iter);
2440 /* fixed buffers import bvec */
2441 iovec.iov_base = kmap(iter->bvec->bv_page)
2443 iovec.iov_len = min(iter->count,
2444 iter->bvec->bv_len - iter->iov_offset);
2448 nr = file->f_op->read(file, iovec.iov_base,
2449 iovec.iov_len, &kiocb->ki_pos);
2451 nr = file->f_op->write(file, iovec.iov_base,
2452 iovec.iov_len, &kiocb->ki_pos);
2455 if (iov_iter_is_bvec(iter))
2456 kunmap(iter->bvec->bv_page);
2464 if (nr != iovec.iov_len)
2466 iov_iter_advance(iter, nr);
2472 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2473 struct iovec *iovec, struct iovec *fast_iov,
2474 struct iov_iter *iter)
2476 req->io->rw.nr_segs = iter->nr_segs;
2477 req->io->rw.size = io_size;
2478 req->io->rw.iov = iovec;
2479 if (!req->io->rw.iov) {
2480 req->io->rw.iov = req->io->rw.fast_iov;
2481 if (req->io->rw.iov != fast_iov)
2482 memcpy(req->io->rw.iov, fast_iov,
2483 sizeof(struct iovec) * iter->nr_segs);
2485 req->flags |= REQ_F_NEED_CLEANUP;
2489 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2491 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2492 return req->io == NULL;
2495 static int io_alloc_async_ctx(struct io_kiocb *req)
2497 if (!io_op_defs[req->opcode].async_ctx)
2500 return __io_alloc_async_ctx(req);
2503 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2504 struct iovec *iovec, struct iovec *fast_iov,
2505 struct iov_iter *iter)
2507 if (!io_op_defs[req->opcode].async_ctx)
2510 if (__io_alloc_async_ctx(req))
2513 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2518 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2519 bool force_nonblock)
2521 struct io_async_ctx *io;
2522 struct iov_iter iter;
2525 ret = io_prep_rw(req, sqe, force_nonblock);
2529 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2532 /* either don't need iovec imported or already have it */
2533 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2537 io->rw.iov = io->rw.fast_iov;
2539 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2544 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2548 static int io_read(struct io_kiocb *req, bool force_nonblock)
2550 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2551 struct kiocb *kiocb = &req->rw.kiocb;
2552 struct iov_iter iter;
2554 ssize_t io_size, ret;
2556 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2560 /* Ensure we clear previously set non-block flag */
2561 if (!force_nonblock)
2562 kiocb->ki_flags &= ~IOCB_NOWAIT;
2566 if (req->flags & REQ_F_LINK_HEAD)
2567 req->result = io_size;
2570 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2571 * we know to async punt it even if it was opened O_NONBLOCK
2573 if (force_nonblock && !io_file_supports_async(req->file, READ))
2576 iov_count = iov_iter_count(&iter);
2577 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2581 if (req->file->f_op->read_iter)
2582 ret2 = call_read_iter(req->file, kiocb, &iter);
2584 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2586 /* Catch -EAGAIN return for forced non-blocking submission */
2587 if (!force_nonblock || ret2 != -EAGAIN) {
2588 kiocb_done(kiocb, ret2);
2591 ret = io_setup_async_rw(req, io_size, iovec,
2592 inline_vecs, &iter);
2595 /* any defer here is final, must blocking retry */
2596 if (!(req->flags & REQ_F_NOWAIT) &&
2597 !file_can_poll(req->file))
2598 req->flags |= REQ_F_MUST_PUNT;
2604 req->flags &= ~REQ_F_NEED_CLEANUP;
2608 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2609 bool force_nonblock)
2611 struct io_async_ctx *io;
2612 struct iov_iter iter;
2615 ret = io_prep_rw(req, sqe, force_nonblock);
2619 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2622 req->fsize = rlimit(RLIMIT_FSIZE);
2624 /* either don't need iovec imported or already have it */
2625 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2629 io->rw.iov = io->rw.fast_iov;
2631 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2636 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2640 static int io_write(struct io_kiocb *req, bool force_nonblock)
2642 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2643 struct kiocb *kiocb = &req->rw.kiocb;
2644 struct iov_iter iter;
2646 ssize_t ret, io_size;
2648 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2652 /* Ensure we clear previously set non-block flag */
2653 if (!force_nonblock)
2654 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2658 if (req->flags & REQ_F_LINK_HEAD)
2659 req->result = io_size;
2662 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2663 * we know to async punt it even if it was opened O_NONBLOCK
2665 if (force_nonblock && !io_file_supports_async(req->file, WRITE))
2668 /* file path doesn't support NOWAIT for non-direct_IO */
2669 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2670 (req->flags & REQ_F_ISREG))
2673 iov_count = iov_iter_count(&iter);
2674 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2679 * Open-code file_start_write here to grab freeze protection,
2680 * which will be released by another thread in
2681 * io_complete_rw(). Fool lockdep by telling it the lock got
2682 * released so that it doesn't complain about the held lock when
2683 * we return to userspace.
2685 if (req->flags & REQ_F_ISREG) {
2686 __sb_start_write(file_inode(req->file)->i_sb,
2687 SB_FREEZE_WRITE, true);
2688 __sb_writers_release(file_inode(req->file)->i_sb,
2691 kiocb->ki_flags |= IOCB_WRITE;
2693 if (!force_nonblock)
2694 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2696 if (req->file->f_op->write_iter)
2697 ret2 = call_write_iter(req->file, kiocb, &iter);
2699 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2701 if (!force_nonblock)
2702 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2705 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2706 * retry them without IOCB_NOWAIT.
2708 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2710 if (!force_nonblock || ret2 != -EAGAIN) {
2711 kiocb_done(kiocb, ret2);
2714 ret = io_setup_async_rw(req, io_size, iovec,
2715 inline_vecs, &iter);
2718 /* any defer here is final, must blocking retry */
2719 if (!file_can_poll(req->file))
2720 req->flags |= REQ_F_MUST_PUNT;
2725 req->flags &= ~REQ_F_NEED_CLEANUP;
2730 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2732 struct io_splice* sp = &req->splice;
2733 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2736 if (req->flags & REQ_F_NEED_CLEANUP)
2740 sp->off_in = READ_ONCE(sqe->splice_off_in);
2741 sp->off_out = READ_ONCE(sqe->off);
2742 sp->len = READ_ONCE(sqe->len);
2743 sp->flags = READ_ONCE(sqe->splice_flags);
2745 if (unlikely(sp->flags & ~valid_flags))
2748 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2749 (sp->flags & SPLICE_F_FD_IN_FIXED));
2752 req->flags |= REQ_F_NEED_CLEANUP;
2754 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2755 req->work.flags |= IO_WQ_WORK_UNBOUND;
2760 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2762 struct io_splice *sp = &req->splice;
2763 struct file *in = sp->file_in;
2764 struct file *out = sp->file_out;
2765 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2766 loff_t *poff_in, *poff_out;
2772 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2773 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2776 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2778 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2779 req->flags &= ~REQ_F_NEED_CLEANUP;
2781 io_cqring_add_event(req, ret);
2783 req_set_fail_links(req);
2789 * IORING_OP_NOP just posts a completion event, nothing else.
2791 static int io_nop(struct io_kiocb *req)
2793 struct io_ring_ctx *ctx = req->ctx;
2795 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2798 io_cqring_add_event(req, 0);
2803 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2805 struct io_ring_ctx *ctx = req->ctx;
2810 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2812 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2815 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2816 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2819 req->sync.off = READ_ONCE(sqe->off);
2820 req->sync.len = READ_ONCE(sqe->len);
2824 static bool io_req_cancelled(struct io_kiocb *req)
2826 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2827 req_set_fail_links(req);
2828 io_cqring_add_event(req, -ECANCELED);
2836 static void __io_fsync(struct io_kiocb *req)
2838 loff_t end = req->sync.off + req->sync.len;
2841 ret = vfs_fsync_range(req->file, req->sync.off,
2842 end > 0 ? end : LLONG_MAX,
2843 req->sync.flags & IORING_FSYNC_DATASYNC);
2845 req_set_fail_links(req);
2846 io_cqring_add_event(req, ret);
2850 static void io_fsync_finish(struct io_wq_work **workptr)
2852 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2854 if (io_req_cancelled(req))
2857 io_steal_work(req, workptr);
2860 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2862 /* fsync always requires a blocking context */
2863 if (force_nonblock) {
2864 req->work.func = io_fsync_finish;
2871 static void __io_fallocate(struct io_kiocb *req)
2875 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2876 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2878 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2880 req_set_fail_links(req);
2881 io_cqring_add_event(req, ret);
2885 static void io_fallocate_finish(struct io_wq_work **workptr)
2887 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2889 if (io_req_cancelled(req))
2891 __io_fallocate(req);
2892 io_steal_work(req, workptr);
2895 static int io_fallocate_prep(struct io_kiocb *req,
2896 const struct io_uring_sqe *sqe)
2898 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2901 req->sync.off = READ_ONCE(sqe->off);
2902 req->sync.len = READ_ONCE(sqe->addr);
2903 req->sync.mode = READ_ONCE(sqe->len);
2904 req->fsize = rlimit(RLIMIT_FSIZE);
2908 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2910 /* fallocate always requiring blocking context */
2911 if (force_nonblock) {
2912 req->work.func = io_fallocate_finish;
2916 __io_fallocate(req);
2920 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2922 const char __user *fname;
2925 if (sqe->ioprio || sqe->buf_index)
2927 if (req->flags & REQ_F_FIXED_FILE)
2929 if (req->flags & REQ_F_NEED_CLEANUP)
2932 req->open.dfd = READ_ONCE(sqe->fd);
2933 req->open.how.mode = READ_ONCE(sqe->len);
2934 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2935 req->open.how.flags = READ_ONCE(sqe->open_flags);
2936 if (force_o_largefile())
2937 req->open.how.flags |= O_LARGEFILE;
2939 req->open.filename = getname(fname);
2940 if (IS_ERR(req->open.filename)) {
2941 ret = PTR_ERR(req->open.filename);
2942 req->open.filename = NULL;
2946 req->open.nofile = rlimit(RLIMIT_NOFILE);
2947 req->flags |= REQ_F_NEED_CLEANUP;
2951 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2953 struct open_how __user *how;
2954 const char __user *fname;
2958 if (sqe->ioprio || sqe->buf_index)
2960 if (req->flags & REQ_F_FIXED_FILE)
2962 if (req->flags & REQ_F_NEED_CLEANUP)
2965 req->open.dfd = READ_ONCE(sqe->fd);
2966 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2967 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2968 len = READ_ONCE(sqe->len);
2970 if (len < OPEN_HOW_SIZE_VER0)
2973 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2978 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2979 req->open.how.flags |= O_LARGEFILE;
2981 req->open.filename = getname(fname);
2982 if (IS_ERR(req->open.filename)) {
2983 ret = PTR_ERR(req->open.filename);
2984 req->open.filename = NULL;
2988 req->open.nofile = rlimit(RLIMIT_NOFILE);
2989 req->flags |= REQ_F_NEED_CLEANUP;
2993 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
2995 struct open_flags op;
3002 ret = build_open_flags(&req->open.how, &op);
3006 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3010 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3013 ret = PTR_ERR(file);
3015 fsnotify_open(file);
3016 fd_install(ret, file);
3019 putname(req->open.filename);
3020 req->flags &= ~REQ_F_NEED_CLEANUP;
3022 req_set_fail_links(req);
3023 io_cqring_add_event(req, ret);
3028 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3030 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3031 return io_openat2(req, force_nonblock);
3034 static int io_remove_buffers_prep(struct io_kiocb *req,
3035 const struct io_uring_sqe *sqe)
3037 struct io_provide_buf *p = &req->pbuf;
3040 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3043 tmp = READ_ONCE(sqe->fd);
3044 if (!tmp || tmp > USHRT_MAX)
3047 memset(p, 0, sizeof(*p));
3049 p->bgid = READ_ONCE(sqe->buf_group);
3053 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3054 int bgid, unsigned nbufs)
3058 /* shouldn't happen */
3062 /* the head kbuf is the list itself */
3063 while (!list_empty(&buf->list)) {
3064 struct io_buffer *nxt;
3066 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3067 list_del(&nxt->list);
3074 idr_remove(&ctx->io_buffer_idr, bgid);
3079 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3081 struct io_provide_buf *p = &req->pbuf;
3082 struct io_ring_ctx *ctx = req->ctx;
3083 struct io_buffer *head;
3086 io_ring_submit_lock(ctx, !force_nonblock);
3088 lockdep_assert_held(&ctx->uring_lock);
3091 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3093 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3095 io_ring_submit_lock(ctx, !force_nonblock);
3097 req_set_fail_links(req);
3098 io_cqring_add_event(req, ret);
3103 static int io_provide_buffers_prep(struct io_kiocb *req,
3104 const struct io_uring_sqe *sqe)
3106 struct io_provide_buf *p = &req->pbuf;
3109 if (sqe->ioprio || sqe->rw_flags)
3112 tmp = READ_ONCE(sqe->fd);
3113 if (!tmp || tmp > USHRT_MAX)
3116 p->addr = READ_ONCE(sqe->addr);
3117 p->len = READ_ONCE(sqe->len);
3119 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3122 p->bgid = READ_ONCE(sqe->buf_group);
3123 tmp = READ_ONCE(sqe->off);
3124 if (tmp > USHRT_MAX)
3130 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3132 struct io_buffer *buf;
3133 u64 addr = pbuf->addr;
3134 int i, bid = pbuf->bid;
3136 for (i = 0; i < pbuf->nbufs; i++) {
3137 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3142 buf->len = pbuf->len;
3147 INIT_LIST_HEAD(&buf->list);
3150 list_add_tail(&buf->list, &(*head)->list);
3154 return i ? i : -ENOMEM;
3157 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3159 struct io_provide_buf *p = &req->pbuf;
3160 struct io_ring_ctx *ctx = req->ctx;
3161 struct io_buffer *head, *list;
3164 io_ring_submit_lock(ctx, !force_nonblock);
3166 lockdep_assert_held(&ctx->uring_lock);
3168 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3170 ret = io_add_buffers(p, &head);
3175 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3178 __io_remove_buffers(ctx, head, p->bgid, -1U);
3183 io_ring_submit_unlock(ctx, !force_nonblock);
3185 req_set_fail_links(req);
3186 io_cqring_add_event(req, ret);
3191 static int io_epoll_ctl_prep(struct io_kiocb *req,
3192 const struct io_uring_sqe *sqe)
3194 #if defined(CONFIG_EPOLL)
3195 if (sqe->ioprio || sqe->buf_index)
3198 req->epoll.epfd = READ_ONCE(sqe->fd);
3199 req->epoll.op = READ_ONCE(sqe->len);
3200 req->epoll.fd = READ_ONCE(sqe->off);
3202 if (ep_op_has_event(req->epoll.op)) {
3203 struct epoll_event __user *ev;
3205 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3206 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3216 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3218 #if defined(CONFIG_EPOLL)
3219 struct io_epoll *ie = &req->epoll;
3222 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3223 if (force_nonblock && ret == -EAGAIN)
3227 req_set_fail_links(req);
3228 io_cqring_add_event(req, ret);
3236 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3238 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3239 if (sqe->ioprio || sqe->buf_index || sqe->off)
3242 req->madvise.addr = READ_ONCE(sqe->addr);
3243 req->madvise.len = READ_ONCE(sqe->len);
3244 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3251 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3253 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3254 struct io_madvise *ma = &req->madvise;
3260 ret = do_madvise(ma->addr, ma->len, ma->advice);
3262 req_set_fail_links(req);
3263 io_cqring_add_event(req, ret);
3271 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3273 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3276 req->fadvise.offset = READ_ONCE(sqe->off);
3277 req->fadvise.len = READ_ONCE(sqe->len);
3278 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3282 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3284 struct io_fadvise *fa = &req->fadvise;
3287 if (force_nonblock) {
3288 switch (fa->advice) {
3289 case POSIX_FADV_NORMAL:
3290 case POSIX_FADV_RANDOM:
3291 case POSIX_FADV_SEQUENTIAL:
3298 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3300 req_set_fail_links(req);
3301 io_cqring_add_event(req, ret);
3306 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3308 const char __user *fname;
3309 unsigned lookup_flags;
3312 if (sqe->ioprio || sqe->buf_index)
3314 if (req->flags & REQ_F_FIXED_FILE)
3316 if (req->flags & REQ_F_NEED_CLEANUP)
3319 req->open.dfd = READ_ONCE(sqe->fd);
3320 req->open.mask = READ_ONCE(sqe->len);
3321 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3322 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3323 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3325 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3328 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3329 if (IS_ERR(req->open.filename)) {
3330 ret = PTR_ERR(req->open.filename);
3331 req->open.filename = NULL;
3335 req->flags |= REQ_F_NEED_CLEANUP;
3339 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3341 struct io_open *ctx = &req->open;
3342 unsigned lookup_flags;
3347 if (force_nonblock) {
3348 /* only need file table for an actual valid fd */
3349 if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD)
3350 req->flags |= REQ_F_NO_FILE_TABLE;
3354 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3358 /* filename_lookup() drops it, keep a reference */
3359 ctx->filename->refcnt++;
3361 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3366 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3368 if (retry_estale(ret, lookup_flags)) {
3369 lookup_flags |= LOOKUP_REVAL;
3373 ret = cp_statx(&stat, ctx->buffer);
3375 putname(ctx->filename);
3376 req->flags &= ~REQ_F_NEED_CLEANUP;
3378 req_set_fail_links(req);
3379 io_cqring_add_event(req, ret);
3384 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3387 * If we queue this for async, it must not be cancellable. That would
3388 * leave the 'file' in an undeterminate state.
3390 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3392 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3393 sqe->rw_flags || sqe->buf_index)
3395 if (req->flags & REQ_F_FIXED_FILE)
3398 req->close.fd = READ_ONCE(sqe->fd);
3399 if (req->file->f_op == &io_uring_fops ||
3400 req->close.fd == req->ctx->ring_fd)
3406 /* only called when __close_fd_get_file() is done */
3407 static void __io_close_finish(struct io_kiocb *req)
3411 ret = filp_close(req->close.put_file, req->work.files);
3413 req_set_fail_links(req);
3414 io_cqring_add_event(req, ret);
3415 fput(req->close.put_file);
3419 static void io_close_finish(struct io_wq_work **workptr)
3421 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3423 /* not cancellable, don't do io_req_cancelled() */
3424 __io_close_finish(req);
3425 io_steal_work(req, workptr);
3428 static int io_close(struct io_kiocb *req, bool force_nonblock)
3432 req->close.put_file = NULL;
3433 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3437 /* if the file has a flush method, be safe and punt to async */
3438 if (req->close.put_file->f_op->flush && force_nonblock) {
3439 /* submission ref will be dropped, take it for async */
3440 refcount_inc(&req->refs);
3442 req->work.func = io_close_finish;
3444 * Do manual async queue here to avoid grabbing files - we don't
3445 * need the files, and it'll cause io_close_finish() to close
3446 * the file again and cause a double CQE entry for this request
3448 io_queue_async_work(req);
3453 * No ->flush(), safely close from here and just punt the
3454 * fput() to async context.
3456 __io_close_finish(req);
3460 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3462 struct io_ring_ctx *ctx = req->ctx;
3467 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3469 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3472 req->sync.off = READ_ONCE(sqe->off);
3473 req->sync.len = READ_ONCE(sqe->len);
3474 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3478 static void __io_sync_file_range(struct io_kiocb *req)
3482 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3485 req_set_fail_links(req);
3486 io_cqring_add_event(req, ret);
3491 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3493 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3495 if (io_req_cancelled(req))
3497 __io_sync_file_range(req);
3498 io_steal_work(req, workptr);
3501 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3503 /* sync_file_range always requires a blocking context */
3504 if (force_nonblock) {
3505 req->work.func = io_sync_file_range_finish;
3509 __io_sync_file_range(req);
3513 #if defined(CONFIG_NET)
3514 static int io_setup_async_msg(struct io_kiocb *req,
3515 struct io_async_msghdr *kmsg)
3519 if (io_alloc_async_ctx(req)) {
3520 if (kmsg->iov != kmsg->fast_iov)
3524 req->flags |= REQ_F_NEED_CLEANUP;
3525 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3529 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3531 struct io_sr_msg *sr = &req->sr_msg;
3532 struct io_async_ctx *io = req->io;
3535 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3536 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3537 sr->len = READ_ONCE(sqe->len);
3539 #ifdef CONFIG_COMPAT
3540 if (req->ctx->compat)
3541 sr->msg_flags |= MSG_CMSG_COMPAT;
3544 if (!io || req->opcode == IORING_OP_SEND)
3546 /* iovec is already imported */
3547 if (req->flags & REQ_F_NEED_CLEANUP)
3550 io->msg.iov = io->msg.fast_iov;
3551 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3554 req->flags |= REQ_F_NEED_CLEANUP;
3558 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3560 struct io_async_msghdr *kmsg = NULL;
3561 struct socket *sock;
3564 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3567 sock = sock_from_file(req->file, &ret);
3569 struct io_async_ctx io;
3573 kmsg = &req->io->msg;
3574 kmsg->msg.msg_name = &req->io->msg.addr;
3575 /* if iov is set, it's allocated already */
3577 kmsg->iov = kmsg->fast_iov;
3578 kmsg->msg.msg_iter.iov = kmsg->iov;
3580 struct io_sr_msg *sr = &req->sr_msg;
3583 kmsg->msg.msg_name = &io.msg.addr;
3585 io.msg.iov = io.msg.fast_iov;
3586 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3587 sr->msg_flags, &io.msg.iov);
3592 flags = req->sr_msg.msg_flags;
3593 if (flags & MSG_DONTWAIT)
3594 req->flags |= REQ_F_NOWAIT;
3595 else if (force_nonblock)
3596 flags |= MSG_DONTWAIT;
3598 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3599 if (force_nonblock && ret == -EAGAIN)
3600 return io_setup_async_msg(req, kmsg);
3601 if (ret == -ERESTARTSYS)
3605 if (kmsg && kmsg->iov != kmsg->fast_iov)
3607 req->flags &= ~REQ_F_NEED_CLEANUP;
3608 io_cqring_add_event(req, ret);
3610 req_set_fail_links(req);
3615 static int io_send(struct io_kiocb *req, bool force_nonblock)
3617 struct socket *sock;
3620 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3623 sock = sock_from_file(req->file, &ret);
3625 struct io_sr_msg *sr = &req->sr_msg;
3630 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3635 msg.msg_name = NULL;
3636 msg.msg_control = NULL;
3637 msg.msg_controllen = 0;
3638 msg.msg_namelen = 0;
3640 flags = req->sr_msg.msg_flags;
3641 if (flags & MSG_DONTWAIT)
3642 req->flags |= REQ_F_NOWAIT;
3643 else if (force_nonblock)
3644 flags |= MSG_DONTWAIT;
3646 msg.msg_flags = flags;
3647 ret = sock_sendmsg(sock, &msg);
3648 if (force_nonblock && ret == -EAGAIN)
3650 if (ret == -ERESTARTSYS)
3654 io_cqring_add_event(req, ret);
3656 req_set_fail_links(req);
3661 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3663 struct io_sr_msg *sr = &req->sr_msg;
3664 struct iovec __user *uiov;
3668 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3673 if (req->flags & REQ_F_BUFFER_SELECT) {
3676 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3678 sr->len = io->msg.iov[0].iov_len;
3679 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3683 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3684 &io->msg.iov, &io->msg.msg.msg_iter);
3692 #ifdef CONFIG_COMPAT
3693 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3694 struct io_async_ctx *io)
3696 struct compat_msghdr __user *msg_compat;
3697 struct io_sr_msg *sr = &req->sr_msg;
3698 struct compat_iovec __user *uiov;
3703 msg_compat = (struct compat_msghdr __user *) sr->msg;
3704 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3709 uiov = compat_ptr(ptr);
3710 if (req->flags & REQ_F_BUFFER_SELECT) {
3711 compat_ssize_t clen;
3715 if (!access_ok(uiov, sizeof(*uiov)))
3717 if (__get_user(clen, &uiov->iov_len))
3721 sr->len = io->msg.iov[0].iov_len;
3724 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3726 &io->msg.msg.msg_iter);
3735 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3737 io->msg.iov = io->msg.fast_iov;
3739 #ifdef CONFIG_COMPAT
3740 if (req->ctx->compat)
3741 return __io_compat_recvmsg_copy_hdr(req, io);
3744 return __io_recvmsg_copy_hdr(req, io);
3747 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3748 int *cflags, bool needs_lock)
3750 struct io_sr_msg *sr = &req->sr_msg;
3751 struct io_buffer *kbuf;
3753 if (!(req->flags & REQ_F_BUFFER_SELECT))
3756 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3761 req->flags |= REQ_F_BUFFER_SELECTED;
3763 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3764 *cflags |= IORING_CQE_F_BUFFER;
3768 static int io_recvmsg_prep(struct io_kiocb *req,
3769 const struct io_uring_sqe *sqe)
3771 struct io_sr_msg *sr = &req->sr_msg;
3772 struct io_async_ctx *io = req->io;
3775 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3776 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3777 sr->len = READ_ONCE(sqe->len);
3778 sr->bgid = READ_ONCE(sqe->buf_group);
3780 #ifdef CONFIG_COMPAT
3781 if (req->ctx->compat)
3782 sr->msg_flags |= MSG_CMSG_COMPAT;
3785 if (!io || req->opcode == IORING_OP_RECV)
3787 /* iovec is already imported */
3788 if (req->flags & REQ_F_NEED_CLEANUP)
3791 ret = io_recvmsg_copy_hdr(req, io);
3793 req->flags |= REQ_F_NEED_CLEANUP;
3797 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3799 struct io_async_msghdr *kmsg = NULL;
3800 struct socket *sock;
3801 int ret, cflags = 0;
3803 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3806 sock = sock_from_file(req->file, &ret);
3808 struct io_buffer *kbuf;
3809 struct io_async_ctx io;
3813 kmsg = &req->io->msg;
3814 kmsg->msg.msg_name = &req->io->msg.addr;
3815 /* if iov is set, it's allocated already */
3817 kmsg->iov = kmsg->fast_iov;
3818 kmsg->msg.msg_iter.iov = kmsg->iov;
3821 kmsg->msg.msg_name = &io.msg.addr;
3823 ret = io_recvmsg_copy_hdr(req, &io);
3828 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3830 return PTR_ERR(kbuf);
3832 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3833 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3834 1, req->sr_msg.len);
3837 flags = req->sr_msg.msg_flags;
3838 if (flags & MSG_DONTWAIT)
3839 req->flags |= REQ_F_NOWAIT;
3840 else if (force_nonblock)
3841 flags |= MSG_DONTWAIT;
3843 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3844 kmsg->uaddr, flags);
3845 if (force_nonblock && ret == -EAGAIN)
3846 return io_setup_async_msg(req, kmsg);
3847 if (ret == -ERESTARTSYS)
3851 if (kmsg && kmsg->iov != kmsg->fast_iov)
3853 req->flags &= ~REQ_F_NEED_CLEANUP;
3854 __io_cqring_add_event(req, ret, cflags);
3856 req_set_fail_links(req);
3861 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3863 struct io_buffer *kbuf = NULL;
3864 struct socket *sock;
3865 int ret, cflags = 0;
3867 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3870 sock = sock_from_file(req->file, &ret);
3872 struct io_sr_msg *sr = &req->sr_msg;
3873 void __user *buf = sr->buf;
3878 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3880 return PTR_ERR(kbuf);
3882 buf = u64_to_user_ptr(kbuf->addr);
3884 ret = import_single_range(READ, buf, sr->len, &iov,
3891 req->flags |= REQ_F_NEED_CLEANUP;
3892 msg.msg_name = NULL;
3893 msg.msg_control = NULL;
3894 msg.msg_controllen = 0;
3895 msg.msg_namelen = 0;
3896 msg.msg_iocb = NULL;
3899 flags = req->sr_msg.msg_flags;
3900 if (flags & MSG_DONTWAIT)
3901 req->flags |= REQ_F_NOWAIT;
3902 else if (force_nonblock)
3903 flags |= MSG_DONTWAIT;
3905 ret = sock_recvmsg(sock, &msg, flags);
3906 if (force_nonblock && ret == -EAGAIN)
3908 if (ret == -ERESTARTSYS)
3913 req->flags &= ~REQ_F_NEED_CLEANUP;
3914 __io_cqring_add_event(req, ret, cflags);
3916 req_set_fail_links(req);
3921 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3923 struct io_accept *accept = &req->accept;
3925 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3927 if (sqe->ioprio || sqe->len || sqe->buf_index)
3930 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3931 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3932 accept->flags = READ_ONCE(sqe->accept_flags);
3933 accept->nofile = rlimit(RLIMIT_NOFILE);
3937 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3939 struct io_accept *accept = &req->accept;
3940 unsigned file_flags;
3943 file_flags = force_nonblock ? O_NONBLOCK : 0;
3944 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3945 accept->addr_len, accept->flags,
3947 if (ret == -EAGAIN && force_nonblock)
3949 if (ret == -ERESTARTSYS)
3952 req_set_fail_links(req);
3953 io_cqring_add_event(req, ret);
3958 static void io_accept_finish(struct io_wq_work **workptr)
3960 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3962 if (io_req_cancelled(req))
3964 __io_accept(req, false);
3965 io_steal_work(req, workptr);
3968 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3972 ret = __io_accept(req, force_nonblock);
3973 if (ret == -EAGAIN && force_nonblock) {
3974 req->work.func = io_accept_finish;
3980 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3982 struct io_connect *conn = &req->connect;
3983 struct io_async_ctx *io = req->io;
3985 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3987 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3990 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3991 conn->addr_len = READ_ONCE(sqe->addr2);
3996 return move_addr_to_kernel(conn->addr, conn->addr_len,
3997 &io->connect.address);
4000 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4002 struct io_async_ctx __io, *io;
4003 unsigned file_flags;
4009 ret = move_addr_to_kernel(req->connect.addr,
4010 req->connect.addr_len,
4011 &__io.connect.address);
4017 file_flags = force_nonblock ? O_NONBLOCK : 0;
4019 ret = __sys_connect_file(req->file, &io->connect.address,
4020 req->connect.addr_len, file_flags);
4021 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4024 if (io_alloc_async_ctx(req)) {
4028 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4031 if (ret == -ERESTARTSYS)
4035 req_set_fail_links(req);
4036 io_cqring_add_event(req, ret);
4040 #else /* !CONFIG_NET */
4041 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4046 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4051 static int io_send(struct io_kiocb *req, bool force_nonblock)
4056 static int io_recvmsg_prep(struct io_kiocb *req,
4057 const struct io_uring_sqe *sqe)
4062 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4067 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4072 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4077 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4082 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4087 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4091 #endif /* CONFIG_NET */
4093 struct io_poll_table {
4094 struct poll_table_struct pt;
4095 struct io_kiocb *req;
4099 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4100 struct wait_queue_head *head)
4102 if (unlikely(poll->head)) {
4103 pt->error = -EINVAL;
4109 add_wait_queue(head, &poll->wait);
4112 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4113 struct poll_table_struct *p)
4115 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4117 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4120 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4121 __poll_t mask, task_work_func_t func)
4123 struct task_struct *tsk;
4126 /* for instances that support it check for an event match first: */
4127 if (mask && !(mask & poll->events))
4130 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4132 list_del_init(&poll->wait.entry);
4136 init_task_work(&req->task_work, func);
4138 * If this fails, then the task is exiting. Punt to one of the io-wq
4139 * threads to ensure the work gets run, we can't always rely on exit
4140 * cancelation taking care of this.
4142 ret = task_work_add(tsk, &req->task_work, true);
4143 if (unlikely(ret)) {
4144 tsk = io_wq_get_task(req->ctx->io_wq);
4145 task_work_add(tsk, &req->task_work, true);
4147 wake_up_process(tsk);
4151 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4152 __acquires(&req->ctx->completion_lock)
4154 struct io_ring_ctx *ctx = req->ctx;
4156 if (!req->result && !READ_ONCE(poll->canceled)) {
4157 struct poll_table_struct pt = { ._key = poll->events };
4159 req->result = vfs_poll(req->file, &pt) & poll->events;
4162 spin_lock_irq(&ctx->completion_lock);
4163 if (!req->result && !READ_ONCE(poll->canceled)) {
4164 add_wait_queue(poll->head, &poll->wait);
4171 static void io_async_task_func(struct callback_head *cb)
4173 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4174 struct async_poll *apoll = req->apoll;
4175 struct io_ring_ctx *ctx = req->ctx;
4178 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4180 if (io_poll_rewait(req, &apoll->poll)) {
4181 spin_unlock_irq(&ctx->completion_lock);
4185 if (hash_hashed(&req->hash_node))
4186 hash_del(&req->hash_node);
4188 canceled = READ_ONCE(apoll->poll.canceled);
4190 io_cqring_fill_event(req, -ECANCELED);
4191 io_commit_cqring(ctx);
4194 spin_unlock_irq(&ctx->completion_lock);
4196 /* restore ->work in case we need to retry again */
4197 memcpy(&req->work, &apoll->work, sizeof(req->work));
4201 io_cqring_ev_posted(ctx);
4202 req_set_fail_links(req);
4203 io_double_put_req(req);
4207 __set_current_state(TASK_RUNNING);
4208 mutex_lock(&ctx->uring_lock);
4209 __io_queue_sqe(req, NULL);
4210 mutex_unlock(&ctx->uring_lock);
4215 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4218 struct io_kiocb *req = wait->private;
4219 struct io_poll_iocb *poll = &req->apoll->poll;
4221 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4224 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4227 static void io_poll_req_insert(struct io_kiocb *req)
4229 struct io_ring_ctx *ctx = req->ctx;
4230 struct hlist_head *list;
4232 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4233 hlist_add_head(&req->hash_node, list);
4236 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4237 struct io_poll_iocb *poll,
4238 struct io_poll_table *ipt, __poll_t mask,
4239 wait_queue_func_t wake_func)
4240 __acquires(&ctx->completion_lock)
4242 struct io_ring_ctx *ctx = req->ctx;
4243 bool cancel = false;
4245 poll->file = req->file;
4247 poll->done = poll->canceled = false;
4248 poll->events = mask;
4250 ipt->pt._key = mask;
4252 ipt->error = -EINVAL;
4254 INIT_LIST_HEAD(&poll->wait.entry);
4255 init_waitqueue_func_entry(&poll->wait, wake_func);
4256 poll->wait.private = req;
4258 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4260 spin_lock_irq(&ctx->completion_lock);
4261 if (likely(poll->head)) {
4262 spin_lock(&poll->head->lock);
4263 if (unlikely(list_empty(&poll->wait.entry))) {
4269 if (mask || ipt->error)
4270 list_del_init(&poll->wait.entry);
4272 WRITE_ONCE(poll->canceled, true);
4273 else if (!poll->done) /* actually waiting for an event */
4274 io_poll_req_insert(req);
4275 spin_unlock(&poll->head->lock);
4281 static bool io_arm_poll_handler(struct io_kiocb *req)
4283 const struct io_op_def *def = &io_op_defs[req->opcode];
4284 struct io_ring_ctx *ctx = req->ctx;
4285 struct async_poll *apoll;
4286 struct io_poll_table ipt;
4289 if (!req->file || !file_can_poll(req->file))
4291 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4293 if (!def->pollin && !def->pollout)
4296 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4297 if (unlikely(!apoll))
4300 req->flags |= REQ_F_POLLED;
4301 memcpy(&apoll->work, &req->work, sizeof(req->work));
4303 get_task_struct(current);
4304 req->task = current;
4306 INIT_HLIST_NODE(&req->hash_node);
4310 mask |= POLLIN | POLLRDNORM;
4312 mask |= POLLOUT | POLLWRNORM;
4313 mask |= POLLERR | POLLPRI;
4315 ipt.pt._qproc = io_async_queue_proc;
4317 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4321 apoll->poll.done = true;
4322 spin_unlock_irq(&ctx->completion_lock);
4323 memcpy(&req->work, &apoll->work, sizeof(req->work));
4327 spin_unlock_irq(&ctx->completion_lock);
4328 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4329 apoll->poll.events);
4333 static bool __io_poll_remove_one(struct io_kiocb *req,
4334 struct io_poll_iocb *poll)
4336 bool do_complete = false;
4338 spin_lock(&poll->head->lock);
4339 WRITE_ONCE(poll->canceled, true);
4340 if (!list_empty(&poll->wait.entry)) {
4341 list_del_init(&poll->wait.entry);
4344 spin_unlock(&poll->head->lock);
4348 static bool io_poll_remove_one(struct io_kiocb *req)
4350 struct async_poll *apoll = NULL;
4353 if (req->opcode == IORING_OP_POLL_ADD) {
4354 do_complete = __io_poll_remove_one(req, &req->poll);
4357 /* non-poll requests have submit ref still */
4358 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4363 hash_del(&req->hash_node);
4365 if (do_complete && apoll) {
4367 * restore ->work because we need to call io_req_work_drop_env.
4369 memcpy(&req->work, &apoll->work, sizeof(req->work));
4374 io_cqring_fill_event(req, -ECANCELED);
4375 io_commit_cqring(req->ctx);
4376 req->flags |= REQ_F_COMP_LOCKED;
4383 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4385 struct hlist_node *tmp;
4386 struct io_kiocb *req;
4389 spin_lock_irq(&ctx->completion_lock);
4390 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4391 struct hlist_head *list;
4393 list = &ctx->cancel_hash[i];
4394 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4395 posted += io_poll_remove_one(req);
4397 spin_unlock_irq(&ctx->completion_lock);
4400 io_cqring_ev_posted(ctx);
4403 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4405 struct hlist_head *list;
4406 struct io_kiocb *req;
4408 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4409 hlist_for_each_entry(req, list, hash_node) {
4410 if (sqe_addr != req->user_data)
4412 if (io_poll_remove_one(req))
4420 static int io_poll_remove_prep(struct io_kiocb *req,
4421 const struct io_uring_sqe *sqe)
4423 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4425 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4429 req->poll.addr = READ_ONCE(sqe->addr);
4434 * Find a running poll command that matches one specified in sqe->addr,
4435 * and remove it if found.
4437 static int io_poll_remove(struct io_kiocb *req)
4439 struct io_ring_ctx *ctx = req->ctx;
4443 addr = req->poll.addr;
4444 spin_lock_irq(&ctx->completion_lock);
4445 ret = io_poll_cancel(ctx, addr);
4446 spin_unlock_irq(&ctx->completion_lock);
4448 io_cqring_add_event(req, ret);
4450 req_set_fail_links(req);
4455 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4457 struct io_ring_ctx *ctx = req->ctx;
4459 req->poll.done = true;
4460 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4461 io_commit_cqring(ctx);
4464 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4466 struct io_ring_ctx *ctx = req->ctx;
4467 struct io_poll_iocb *poll = &req->poll;
4469 if (io_poll_rewait(req, poll)) {
4470 spin_unlock_irq(&ctx->completion_lock);
4474 hash_del(&req->hash_node);
4475 io_poll_complete(req, req->result, 0);
4476 req->flags |= REQ_F_COMP_LOCKED;
4477 io_put_req_find_next(req, nxt);
4478 spin_unlock_irq(&ctx->completion_lock);
4480 io_cqring_ev_posted(ctx);
4483 static void io_poll_task_func(struct callback_head *cb)
4485 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4486 struct io_kiocb *nxt = NULL;
4488 io_poll_task_handler(req, &nxt);
4490 struct io_ring_ctx *ctx = nxt->ctx;
4492 mutex_lock(&ctx->uring_lock);
4493 __io_queue_sqe(nxt, NULL);
4494 mutex_unlock(&ctx->uring_lock);
4498 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4501 struct io_kiocb *req = wait->private;
4502 struct io_poll_iocb *poll = &req->poll;
4504 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4507 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4508 struct poll_table_struct *p)
4510 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4512 __io_queue_proc(&pt->req->poll, pt, head);
4515 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4517 struct io_poll_iocb *poll = &req->poll;
4520 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4522 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4527 events = READ_ONCE(sqe->poll_events);
4528 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4530 get_task_struct(current);
4531 req->task = current;
4535 static int io_poll_add(struct io_kiocb *req)
4537 struct io_poll_iocb *poll = &req->poll;
4538 struct io_ring_ctx *ctx = req->ctx;
4539 struct io_poll_table ipt;
4542 INIT_HLIST_NODE(&req->hash_node);
4543 INIT_LIST_HEAD(&req->list);
4544 ipt.pt._qproc = io_poll_queue_proc;
4546 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4549 if (mask) { /* no async, we'd stolen it */
4551 io_poll_complete(req, mask, 0);
4553 spin_unlock_irq(&ctx->completion_lock);
4556 io_cqring_ev_posted(ctx);
4562 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4564 struct io_timeout_data *data = container_of(timer,
4565 struct io_timeout_data, timer);
4566 struct io_kiocb *req = data->req;
4567 struct io_ring_ctx *ctx = req->ctx;
4568 unsigned long flags;
4570 atomic_inc(&ctx->cq_timeouts);
4572 spin_lock_irqsave(&ctx->completion_lock, flags);
4574 * We could be racing with timeout deletion. If the list is empty,
4575 * then timeout lookup already found it and will be handling it.
4577 if (!list_empty(&req->list)) {
4578 struct io_kiocb *prev;
4581 * Adjust the reqs sequence before the current one because it
4582 * will consume a slot in the cq_ring and the cq_tail
4583 * pointer will be increased, otherwise other timeout reqs may
4584 * return in advance without waiting for enough wait_nr.
4587 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4589 list_del_init(&req->list);
4592 io_cqring_fill_event(req, -ETIME);
4593 io_commit_cqring(ctx);
4594 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4596 io_cqring_ev_posted(ctx);
4597 req_set_fail_links(req);
4599 return HRTIMER_NORESTART;
4602 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4604 struct io_kiocb *req;
4607 list_for_each_entry(req, &ctx->timeout_list, list) {
4608 if (user_data == req->user_data) {
4609 list_del_init(&req->list);
4618 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4622 req_set_fail_links(req);
4623 io_cqring_fill_event(req, -ECANCELED);
4628 static int io_timeout_remove_prep(struct io_kiocb *req,
4629 const struct io_uring_sqe *sqe)
4631 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4633 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4636 req->timeout.addr = READ_ONCE(sqe->addr);
4637 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4638 if (req->timeout.flags)
4645 * Remove or update an existing timeout command
4647 static int io_timeout_remove(struct io_kiocb *req)
4649 struct io_ring_ctx *ctx = req->ctx;
4652 spin_lock_irq(&ctx->completion_lock);
4653 ret = io_timeout_cancel(ctx, req->timeout.addr);
4655 io_cqring_fill_event(req, ret);
4656 io_commit_cqring(ctx);
4657 spin_unlock_irq(&ctx->completion_lock);
4658 io_cqring_ev_posted(ctx);
4660 req_set_fail_links(req);
4665 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4666 bool is_timeout_link)
4668 struct io_timeout_data *data;
4671 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4673 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4675 if (sqe->off && is_timeout_link)
4677 flags = READ_ONCE(sqe->timeout_flags);
4678 if (flags & ~IORING_TIMEOUT_ABS)
4681 req->timeout.count = READ_ONCE(sqe->off);
4683 if (!req->io && io_alloc_async_ctx(req))
4686 data = &req->io->timeout;
4688 req->flags |= REQ_F_TIMEOUT;
4690 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4693 if (flags & IORING_TIMEOUT_ABS)
4694 data->mode = HRTIMER_MODE_ABS;
4696 data->mode = HRTIMER_MODE_REL;
4698 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4702 static int io_timeout(struct io_kiocb *req)
4704 struct io_ring_ctx *ctx = req->ctx;
4705 struct io_timeout_data *data;
4706 struct list_head *entry;
4708 u32 count = req->timeout.count;
4709 u32 seq = req->sequence;
4711 data = &req->io->timeout;
4714 * sqe->off holds how many events that need to occur for this
4715 * timeout event to be satisfied. If it isn't set, then this is
4716 * a pure timeout request, sequence isn't used.
4719 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4720 spin_lock_irq(&ctx->completion_lock);
4721 entry = ctx->timeout_list.prev;
4725 req->sequence = seq + count;
4728 * Insertion sort, ensuring the first entry in the list is always
4729 * the one we need first.
4731 spin_lock_irq(&ctx->completion_lock);
4732 list_for_each_prev(entry, &ctx->timeout_list) {
4733 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4735 long long tmp, tmp_nxt;
4736 u32 nxt_offset = nxt->timeout.count;
4738 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4742 * Since seq + count can overflow, use type long
4745 tmp = (long long)seq + count;
4746 nxt_seq = nxt->sequence - nxt_offset;
4747 tmp_nxt = (long long)nxt_seq + nxt_offset;
4750 * cached_sq_head may overflow, and it will never overflow twice
4751 * once there is some timeout req still be valid.
4760 * Sequence of reqs after the insert one and itself should
4761 * be adjusted because each timeout req consumes a slot.
4766 req->sequence -= span;
4768 list_add(&req->list, entry);
4769 data->timer.function = io_timeout_fn;
4770 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4771 spin_unlock_irq(&ctx->completion_lock);
4775 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4777 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4779 return req->user_data == (unsigned long) data;
4782 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4784 enum io_wq_cancel cancel_ret;
4787 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4788 switch (cancel_ret) {
4789 case IO_WQ_CANCEL_OK:
4792 case IO_WQ_CANCEL_RUNNING:
4795 case IO_WQ_CANCEL_NOTFOUND:
4803 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4804 struct io_kiocb *req, __u64 sqe_addr,
4807 unsigned long flags;
4810 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4811 if (ret != -ENOENT) {
4812 spin_lock_irqsave(&ctx->completion_lock, flags);
4816 spin_lock_irqsave(&ctx->completion_lock, flags);
4817 ret = io_timeout_cancel(ctx, sqe_addr);
4820 ret = io_poll_cancel(ctx, sqe_addr);
4824 io_cqring_fill_event(req, ret);
4825 io_commit_cqring(ctx);
4826 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4827 io_cqring_ev_posted(ctx);
4830 req_set_fail_links(req);
4834 static int io_async_cancel_prep(struct io_kiocb *req,
4835 const struct io_uring_sqe *sqe)
4837 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4839 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4843 req->cancel.addr = READ_ONCE(sqe->addr);
4847 static int io_async_cancel(struct io_kiocb *req)
4849 struct io_ring_ctx *ctx = req->ctx;
4851 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4855 static int io_files_update_prep(struct io_kiocb *req,
4856 const struct io_uring_sqe *sqe)
4858 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4861 req->files_update.offset = READ_ONCE(sqe->off);
4862 req->files_update.nr_args = READ_ONCE(sqe->len);
4863 if (!req->files_update.nr_args)
4865 req->files_update.arg = READ_ONCE(sqe->addr);
4869 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4871 struct io_ring_ctx *ctx = req->ctx;
4872 struct io_uring_files_update up;
4878 up.offset = req->files_update.offset;
4879 up.fds = req->files_update.arg;
4881 mutex_lock(&ctx->uring_lock);
4882 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4883 mutex_unlock(&ctx->uring_lock);
4886 req_set_fail_links(req);
4887 io_cqring_add_event(req, ret);
4892 static int io_req_defer_prep(struct io_kiocb *req,
4893 const struct io_uring_sqe *sqe)
4900 if (io_op_defs[req->opcode].file_table) {
4901 ret = io_grab_files(req);
4906 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4908 switch (req->opcode) {
4911 case IORING_OP_READV:
4912 case IORING_OP_READ_FIXED:
4913 case IORING_OP_READ:
4914 ret = io_read_prep(req, sqe, true);
4916 case IORING_OP_WRITEV:
4917 case IORING_OP_WRITE_FIXED:
4918 case IORING_OP_WRITE:
4919 ret = io_write_prep(req, sqe, true);
4921 case IORING_OP_POLL_ADD:
4922 ret = io_poll_add_prep(req, sqe);
4924 case IORING_OP_POLL_REMOVE:
4925 ret = io_poll_remove_prep(req, sqe);
4927 case IORING_OP_FSYNC:
4928 ret = io_prep_fsync(req, sqe);
4930 case IORING_OP_SYNC_FILE_RANGE:
4931 ret = io_prep_sfr(req, sqe);
4933 case IORING_OP_SENDMSG:
4934 case IORING_OP_SEND:
4935 ret = io_sendmsg_prep(req, sqe);
4937 case IORING_OP_RECVMSG:
4938 case IORING_OP_RECV:
4939 ret = io_recvmsg_prep(req, sqe);
4941 case IORING_OP_CONNECT:
4942 ret = io_connect_prep(req, sqe);
4944 case IORING_OP_TIMEOUT:
4945 ret = io_timeout_prep(req, sqe, false);
4947 case IORING_OP_TIMEOUT_REMOVE:
4948 ret = io_timeout_remove_prep(req, sqe);
4950 case IORING_OP_ASYNC_CANCEL:
4951 ret = io_async_cancel_prep(req, sqe);
4953 case IORING_OP_LINK_TIMEOUT:
4954 ret = io_timeout_prep(req, sqe, true);
4956 case IORING_OP_ACCEPT:
4957 ret = io_accept_prep(req, sqe);
4959 case IORING_OP_FALLOCATE:
4960 ret = io_fallocate_prep(req, sqe);
4962 case IORING_OP_OPENAT:
4963 ret = io_openat_prep(req, sqe);
4965 case IORING_OP_CLOSE:
4966 ret = io_close_prep(req, sqe);
4968 case IORING_OP_FILES_UPDATE:
4969 ret = io_files_update_prep(req, sqe);
4971 case IORING_OP_STATX:
4972 ret = io_statx_prep(req, sqe);
4974 case IORING_OP_FADVISE:
4975 ret = io_fadvise_prep(req, sqe);
4977 case IORING_OP_MADVISE:
4978 ret = io_madvise_prep(req, sqe);
4980 case IORING_OP_OPENAT2:
4981 ret = io_openat2_prep(req, sqe);
4983 case IORING_OP_EPOLL_CTL:
4984 ret = io_epoll_ctl_prep(req, sqe);
4986 case IORING_OP_SPLICE:
4987 ret = io_splice_prep(req, sqe);
4989 case IORING_OP_PROVIDE_BUFFERS:
4990 ret = io_provide_buffers_prep(req, sqe);
4992 case IORING_OP_REMOVE_BUFFERS:
4993 ret = io_remove_buffers_prep(req, sqe);
4996 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5005 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5007 struct io_ring_ctx *ctx = req->ctx;
5010 /* Still need defer if there is pending req in defer list. */
5011 if (!req_need_defer(req) && list_empty_careful(&ctx->defer_list))
5015 if (io_alloc_async_ctx(req))
5017 ret = io_req_defer_prep(req, sqe);
5022 spin_lock_irq(&ctx->completion_lock);
5023 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5024 spin_unlock_irq(&ctx->completion_lock);
5028 trace_io_uring_defer(ctx, req, req->user_data);
5029 list_add_tail(&req->list, &ctx->defer_list);
5030 spin_unlock_irq(&ctx->completion_lock);
5031 return -EIOCBQUEUED;
5034 static void io_cleanup_req(struct io_kiocb *req)
5036 struct io_async_ctx *io = req->io;
5038 switch (req->opcode) {
5039 case IORING_OP_READV:
5040 case IORING_OP_READ_FIXED:
5041 case IORING_OP_READ:
5042 if (req->flags & REQ_F_BUFFER_SELECTED)
5043 kfree((void *)(unsigned long)req->rw.addr);
5045 case IORING_OP_WRITEV:
5046 case IORING_OP_WRITE_FIXED:
5047 case IORING_OP_WRITE:
5048 if (io->rw.iov != io->rw.fast_iov)
5051 case IORING_OP_RECVMSG:
5052 if (req->flags & REQ_F_BUFFER_SELECTED)
5053 kfree(req->sr_msg.kbuf);
5055 case IORING_OP_SENDMSG:
5056 if (io->msg.iov != io->msg.fast_iov)
5059 case IORING_OP_RECV:
5060 if (req->flags & REQ_F_BUFFER_SELECTED)
5061 kfree(req->sr_msg.kbuf);
5063 case IORING_OP_OPENAT:
5064 case IORING_OP_OPENAT2:
5065 case IORING_OP_STATX:
5066 putname(req->open.filename);
5068 case IORING_OP_SPLICE:
5069 io_put_file(req, req->splice.file_in,
5070 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5074 req->flags &= ~REQ_F_NEED_CLEANUP;
5077 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5078 bool force_nonblock)
5080 struct io_ring_ctx *ctx = req->ctx;
5083 switch (req->opcode) {
5087 case IORING_OP_READV:
5088 case IORING_OP_READ_FIXED:
5089 case IORING_OP_READ:
5091 ret = io_read_prep(req, sqe, force_nonblock);
5095 ret = io_read(req, force_nonblock);
5097 case IORING_OP_WRITEV:
5098 case IORING_OP_WRITE_FIXED:
5099 case IORING_OP_WRITE:
5101 ret = io_write_prep(req, sqe, force_nonblock);
5105 ret = io_write(req, force_nonblock);
5107 case IORING_OP_FSYNC:
5109 ret = io_prep_fsync(req, sqe);
5113 ret = io_fsync(req, force_nonblock);
5115 case IORING_OP_POLL_ADD:
5117 ret = io_poll_add_prep(req, sqe);
5121 ret = io_poll_add(req);
5123 case IORING_OP_POLL_REMOVE:
5125 ret = io_poll_remove_prep(req, sqe);
5129 ret = io_poll_remove(req);
5131 case IORING_OP_SYNC_FILE_RANGE:
5133 ret = io_prep_sfr(req, sqe);
5137 ret = io_sync_file_range(req, force_nonblock);
5139 case IORING_OP_SENDMSG:
5140 case IORING_OP_SEND:
5142 ret = io_sendmsg_prep(req, sqe);
5146 if (req->opcode == IORING_OP_SENDMSG)
5147 ret = io_sendmsg(req, force_nonblock);
5149 ret = io_send(req, force_nonblock);
5151 case IORING_OP_RECVMSG:
5152 case IORING_OP_RECV:
5154 ret = io_recvmsg_prep(req, sqe);
5158 if (req->opcode == IORING_OP_RECVMSG)
5159 ret = io_recvmsg(req, force_nonblock);
5161 ret = io_recv(req, force_nonblock);
5163 case IORING_OP_TIMEOUT:
5165 ret = io_timeout_prep(req, sqe, false);
5169 ret = io_timeout(req);
5171 case IORING_OP_TIMEOUT_REMOVE:
5173 ret = io_timeout_remove_prep(req, sqe);
5177 ret = io_timeout_remove(req);
5179 case IORING_OP_ACCEPT:
5181 ret = io_accept_prep(req, sqe);
5185 ret = io_accept(req, force_nonblock);
5187 case IORING_OP_CONNECT:
5189 ret = io_connect_prep(req, sqe);
5193 ret = io_connect(req, force_nonblock);
5195 case IORING_OP_ASYNC_CANCEL:
5197 ret = io_async_cancel_prep(req, sqe);
5201 ret = io_async_cancel(req);
5203 case IORING_OP_FALLOCATE:
5205 ret = io_fallocate_prep(req, sqe);
5209 ret = io_fallocate(req, force_nonblock);
5211 case IORING_OP_OPENAT:
5213 ret = io_openat_prep(req, sqe);
5217 ret = io_openat(req, force_nonblock);
5219 case IORING_OP_CLOSE:
5221 ret = io_close_prep(req, sqe);
5225 ret = io_close(req, force_nonblock);
5227 case IORING_OP_FILES_UPDATE:
5229 ret = io_files_update_prep(req, sqe);
5233 ret = io_files_update(req, force_nonblock);
5235 case IORING_OP_STATX:
5237 ret = io_statx_prep(req, sqe);
5241 ret = io_statx(req, force_nonblock);
5243 case IORING_OP_FADVISE:
5245 ret = io_fadvise_prep(req, sqe);
5249 ret = io_fadvise(req, force_nonblock);
5251 case IORING_OP_MADVISE:
5253 ret = io_madvise_prep(req, sqe);
5257 ret = io_madvise(req, force_nonblock);
5259 case IORING_OP_OPENAT2:
5261 ret = io_openat2_prep(req, sqe);
5265 ret = io_openat2(req, force_nonblock);
5267 case IORING_OP_EPOLL_CTL:
5269 ret = io_epoll_ctl_prep(req, sqe);
5273 ret = io_epoll_ctl(req, force_nonblock);
5275 case IORING_OP_SPLICE:
5277 ret = io_splice_prep(req, sqe);
5281 ret = io_splice(req, force_nonblock);
5283 case IORING_OP_PROVIDE_BUFFERS:
5285 ret = io_provide_buffers_prep(req, sqe);
5289 ret = io_provide_buffers(req, force_nonblock);
5291 case IORING_OP_REMOVE_BUFFERS:
5293 ret = io_remove_buffers_prep(req, sqe);
5297 ret = io_remove_buffers(req, force_nonblock);
5307 if (ctx->flags & IORING_SETUP_IOPOLL) {
5308 const bool in_async = io_wq_current_is_worker();
5310 if (req->result == -EAGAIN)
5313 /* workqueue context doesn't hold uring_lock, grab it now */
5315 mutex_lock(&ctx->uring_lock);
5317 io_iopoll_req_issued(req);
5320 mutex_unlock(&ctx->uring_lock);
5326 static void io_wq_submit_work(struct io_wq_work **workptr)
5328 struct io_wq_work *work = *workptr;
5329 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5332 /* if NO_CANCEL is set, we must still run the work */
5333 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5334 IO_WQ_WORK_CANCEL) {
5340 ret = io_issue_sqe(req, NULL, false);
5342 * We can get EAGAIN for polled IO even though we're
5343 * forcing a sync submission from here, since we can't
5344 * wait for request slots on the block side.
5353 req_set_fail_links(req);
5354 io_cqring_add_event(req, ret);
5358 io_steal_work(req, workptr);
5361 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5364 struct fixed_file_table *table;
5366 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5367 return table->files[index & IORING_FILE_TABLE_MASK];;
5370 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5371 int fd, struct file **out_file, bool fixed)
5373 struct io_ring_ctx *ctx = req->ctx;
5377 if (unlikely(!ctx->file_data ||
5378 (unsigned) fd >= ctx->nr_user_files))
5380 fd = array_index_nospec(fd, ctx->nr_user_files);
5381 file = io_file_from_index(ctx, fd);
5384 req->fixed_file_refs = ctx->file_data->cur_refs;
5385 percpu_ref_get(req->fixed_file_refs);
5387 trace_io_uring_file_get(ctx, fd);
5388 file = __io_file_get(state, fd);
5389 if (unlikely(!file))
5397 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5402 fixed = (req->flags & REQ_F_FIXED_FILE) != 0;
5403 if (unlikely(!fixed && req->needs_fixed_file))
5406 return io_file_get(state, req, fd, &req->file, fixed);
5409 static int io_grab_files(struct io_kiocb *req)
5412 struct io_ring_ctx *ctx = req->ctx;
5414 if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE))
5416 if (!ctx->ring_file)
5420 spin_lock_irq(&ctx->inflight_lock);
5422 * We use the f_ops->flush() handler to ensure that we can flush
5423 * out work accessing these files if the fd is closed. Check if
5424 * the fd has changed since we started down this path, and disallow
5425 * this operation if it has.
5427 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5428 list_add(&req->inflight_entry, &ctx->inflight_list);
5429 req->flags |= REQ_F_INFLIGHT;
5430 req->work.files = current->files;
5433 spin_unlock_irq(&ctx->inflight_lock);
5439 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5441 struct io_timeout_data *data = container_of(timer,
5442 struct io_timeout_data, timer);
5443 struct io_kiocb *req = data->req;
5444 struct io_ring_ctx *ctx = req->ctx;
5445 struct io_kiocb *prev = NULL;
5446 unsigned long flags;
5448 spin_lock_irqsave(&ctx->completion_lock, flags);
5451 * We don't expect the list to be empty, that will only happen if we
5452 * race with the completion of the linked work.
5454 if (!list_empty(&req->link_list)) {
5455 prev = list_entry(req->link_list.prev, struct io_kiocb,
5457 if (refcount_inc_not_zero(&prev->refs)) {
5458 list_del_init(&req->link_list);
5459 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5464 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5467 req_set_fail_links(prev);
5468 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5471 io_cqring_add_event(req, -ETIME);
5474 return HRTIMER_NORESTART;
5477 static void io_queue_linked_timeout(struct io_kiocb *req)
5479 struct io_ring_ctx *ctx = req->ctx;
5482 * If the list is now empty, then our linked request finished before
5483 * we got a chance to setup the timer
5485 spin_lock_irq(&ctx->completion_lock);
5486 if (!list_empty(&req->link_list)) {
5487 struct io_timeout_data *data = &req->io->timeout;
5489 data->timer.function = io_link_timeout_fn;
5490 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5493 spin_unlock_irq(&ctx->completion_lock);
5495 /* drop submission reference */
5499 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5501 struct io_kiocb *nxt;
5503 if (!(req->flags & REQ_F_LINK_HEAD))
5505 /* for polled retry, if flag is set, we already went through here */
5506 if (req->flags & REQ_F_POLLED)
5509 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5511 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5514 req->flags |= REQ_F_LINK_TIMEOUT;
5518 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5520 struct io_kiocb *linked_timeout;
5521 struct io_kiocb *nxt;
5522 const struct cred *old_creds = NULL;
5526 linked_timeout = io_prep_linked_timeout(req);
5528 if (req->work.creds && req->work.creds != current_cred()) {
5530 revert_creds(old_creds);
5531 if (old_creds == req->work.creds)
5532 old_creds = NULL; /* restored original creds */
5534 old_creds = override_creds(req->work.creds);
5537 ret = io_issue_sqe(req, sqe, true);
5540 * We async punt it if the file wasn't marked NOWAIT, or if the file
5541 * doesn't support non-blocking read/write attempts
5543 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5544 (req->flags & REQ_F_MUST_PUNT))) {
5545 if (io_arm_poll_handler(req)) {
5547 io_queue_linked_timeout(linked_timeout);
5551 if (io_op_defs[req->opcode].file_table) {
5552 ret = io_grab_files(req);
5558 * Queued up for async execution, worker will release
5559 * submit reference when the iocb is actually submitted.
5561 io_queue_async_work(req);
5567 /* drop submission reference */
5568 io_put_req_find_next(req, &nxt);
5570 if (linked_timeout) {
5572 io_queue_linked_timeout(linked_timeout);
5574 io_put_req(linked_timeout);
5577 /* and drop final reference, if we failed */
5579 io_cqring_add_event(req, ret);
5580 req_set_fail_links(req);
5586 if (req->flags & REQ_F_FORCE_ASYNC)
5592 revert_creds(old_creds);
5595 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5599 ret = io_req_defer(req, sqe);
5601 if (ret != -EIOCBQUEUED) {
5603 io_cqring_add_event(req, ret);
5604 req_set_fail_links(req);
5605 io_double_put_req(req);
5607 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5610 if (io_alloc_async_ctx(req))
5612 ret = io_req_defer_prep(req, sqe);
5613 if (unlikely(ret < 0))
5618 * Never try inline submit of IOSQE_ASYNC is set, go straight
5619 * to async execution.
5621 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5622 io_queue_async_work(req);
5624 __io_queue_sqe(req, sqe);
5628 static inline void io_queue_link_head(struct io_kiocb *req)
5630 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5631 io_cqring_add_event(req, -ECANCELED);
5632 io_double_put_req(req);
5634 io_queue_sqe(req, NULL);
5637 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5638 struct io_submit_state *state, struct io_kiocb **link)
5640 struct io_ring_ctx *ctx = req->ctx;
5644 * If we already have a head request, queue this one for async
5645 * submittal once the head completes. If we don't have a head but
5646 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5647 * submitted sync once the chain is complete. If none of those
5648 * conditions are true (normal request), then just queue it.
5651 struct io_kiocb *head = *link;
5654 * Taking sequential execution of a link, draining both sides
5655 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5656 * requests in the link. So, it drains the head and the
5657 * next after the link request. The last one is done via
5658 * drain_next flag to persist the effect across calls.
5660 if (req->flags & REQ_F_IO_DRAIN) {
5661 head->flags |= REQ_F_IO_DRAIN;
5662 ctx->drain_next = 1;
5664 if (io_alloc_async_ctx(req))
5667 ret = io_req_defer_prep(req, sqe);
5669 /* fail even hard links since we don't submit */
5670 head->flags |= REQ_F_FAIL_LINK;
5673 trace_io_uring_link(ctx, req, head);
5674 list_add_tail(&req->link_list, &head->link_list);
5676 /* last request of a link, enqueue the link */
5677 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5678 io_queue_link_head(head);
5682 if (unlikely(ctx->drain_next)) {
5683 req->flags |= REQ_F_IO_DRAIN;
5684 ctx->drain_next = 0;
5686 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5687 req->flags |= REQ_F_LINK_HEAD;
5688 INIT_LIST_HEAD(&req->link_list);
5690 if (io_alloc_async_ctx(req))
5693 ret = io_req_defer_prep(req, sqe);
5695 req->flags |= REQ_F_FAIL_LINK;
5698 io_queue_sqe(req, sqe);
5706 * Batched submission is done, ensure local IO is flushed out.
5708 static void io_submit_state_end(struct io_submit_state *state)
5710 blk_finish_plug(&state->plug);
5712 if (state->free_reqs)
5713 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5717 * Start submission side cache.
5719 static void io_submit_state_start(struct io_submit_state *state,
5720 unsigned int max_ios)
5722 blk_start_plug(&state->plug);
5723 state->free_reqs = 0;
5725 state->ios_left = max_ios;
5728 static void io_commit_sqring(struct io_ring_ctx *ctx)
5730 struct io_rings *rings = ctx->rings;
5733 * Ensure any loads from the SQEs are done at this point,
5734 * since once we write the new head, the application could
5735 * write new data to them.
5737 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5741 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5742 * that is mapped by userspace. This means that care needs to be taken to
5743 * ensure that reads are stable, as we cannot rely on userspace always
5744 * being a good citizen. If members of the sqe are validated and then later
5745 * used, it's important that those reads are done through READ_ONCE() to
5746 * prevent a re-load down the line.
5748 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5750 u32 *sq_array = ctx->sq_array;
5754 * The cached sq head (or cq tail) serves two purposes:
5756 * 1) allows us to batch the cost of updating the user visible
5758 * 2) allows the kernel side to track the head on its own, even
5759 * though the application is the one updating it.
5761 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5762 if (likely(head < ctx->sq_entries))
5763 return &ctx->sq_sqes[head];
5765 /* drop invalid entries */
5766 ctx->cached_sq_dropped++;
5767 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5771 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5773 ctx->cached_sq_head++;
5776 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5777 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5778 IOSQE_BUFFER_SELECT)
5780 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5781 const struct io_uring_sqe *sqe,
5782 struct io_submit_state *state, bool async)
5784 unsigned int sqe_flags;
5788 * All io need record the previous position, if LINK vs DARIN,
5789 * it can be used to mark the position of the first IO in the
5792 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5793 req->opcode = READ_ONCE(sqe->opcode);
5794 req->user_data = READ_ONCE(sqe->user_data);
5799 /* one is dropped after submission, the other at completion */
5800 refcount_set(&req->refs, 2);
5803 req->needs_fixed_file = async;
5804 INIT_IO_WORK(&req->work, io_wq_submit_work);
5806 if (unlikely(req->opcode >= IORING_OP_LAST))
5809 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5810 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5812 use_mm(ctx->sqo_mm);
5815 sqe_flags = READ_ONCE(sqe->flags);
5816 /* enforce forwards compatibility on users */
5817 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5820 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5821 !io_op_defs[req->opcode].buffer_select)
5824 id = READ_ONCE(sqe->personality);
5826 req->work.creds = idr_find(&ctx->personality_idr, id);
5827 if (unlikely(!req->work.creds))
5829 get_cred(req->work.creds);
5832 /* same numerical values with corresponding REQ_F_*, safe to copy */
5833 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5834 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5835 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5837 if (!io_op_defs[req->opcode].needs_file)
5840 return io_req_set_file(state, req, READ_ONCE(sqe->fd));
5843 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5844 struct file *ring_file, int ring_fd, bool async)
5846 struct io_submit_state state, *statep = NULL;
5847 struct io_kiocb *link = NULL;
5848 int i, submitted = 0;
5850 /* if we have a backlog and couldn't flush it all, return BUSY */
5851 if (test_bit(0, &ctx->sq_check_overflow)) {
5852 if (!list_empty(&ctx->cq_overflow_list) &&
5853 !io_cqring_overflow_flush(ctx, false))
5857 /* make sure SQ entry isn't read before tail */
5858 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5860 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5863 if (nr > IO_PLUG_THRESHOLD) {
5864 io_submit_state_start(&state, nr);
5868 ctx->ring_fd = ring_fd;
5869 ctx->ring_file = ring_file;
5871 for (i = 0; i < nr; i++) {
5872 const struct io_uring_sqe *sqe;
5873 struct io_kiocb *req;
5876 sqe = io_get_sqe(ctx);
5877 if (unlikely(!sqe)) {
5878 io_consume_sqe(ctx);
5881 req = io_alloc_req(ctx, statep);
5882 if (unlikely(!req)) {
5884 submitted = -EAGAIN;
5888 err = io_init_req(ctx, req, sqe, statep, async);
5889 io_consume_sqe(ctx);
5890 /* will complete beyond this point, count as submitted */
5893 if (unlikely(err)) {
5895 io_cqring_add_event(req, err);
5896 io_double_put_req(req);
5900 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5902 err = io_submit_sqe(req, sqe, statep, &link);
5907 if (unlikely(submitted != nr)) {
5908 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5910 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5913 io_queue_link_head(link);
5915 io_submit_state_end(&state);
5917 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5918 io_commit_sqring(ctx);
5923 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5925 struct mm_struct *mm = current->mm;
5933 static int io_sq_thread(void *data)
5935 struct io_ring_ctx *ctx = data;
5936 const struct cred *old_cred;
5937 mm_segment_t old_fs;
5939 unsigned long timeout;
5942 complete(&ctx->completions[1]);
5946 old_cred = override_creds(ctx->creds);
5948 timeout = jiffies + ctx->sq_thread_idle;
5949 while (!kthread_should_park()) {
5950 unsigned int to_submit;
5952 if (!list_empty(&ctx->poll_list)) {
5953 unsigned nr_events = 0;
5955 mutex_lock(&ctx->uring_lock);
5956 if (!list_empty(&ctx->poll_list))
5957 io_iopoll_getevents(ctx, &nr_events, 0);
5959 timeout = jiffies + ctx->sq_thread_idle;
5960 mutex_unlock(&ctx->uring_lock);
5963 to_submit = io_sqring_entries(ctx);
5966 * If submit got -EBUSY, flag us as needing the application
5967 * to enter the kernel to reap and flush events.
5969 if (!to_submit || ret == -EBUSY) {
5971 * Drop cur_mm before scheduling, we can't hold it for
5972 * long periods (or over schedule()). Do this before
5973 * adding ourselves to the waitqueue, as the unuse/drop
5976 io_sq_thread_drop_mm(ctx);
5979 * We're polling. If we're within the defined idle
5980 * period, then let us spin without work before going
5981 * to sleep. The exception is if we got EBUSY doing
5982 * more IO, we should wait for the application to
5983 * reap events and wake us up.
5985 if (!list_empty(&ctx->poll_list) ||
5986 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5987 !percpu_ref_is_dying(&ctx->refs))) {
5988 if (current->task_works)
5994 prepare_to_wait(&ctx->sqo_wait, &wait,
5995 TASK_INTERRUPTIBLE);
5998 * While doing polled IO, before going to sleep, we need
5999 * to check if there are new reqs added to poll_list, it
6000 * is because reqs may have been punted to io worker and
6001 * will be added to poll_list later, hence check the
6004 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6005 !list_empty_careful(&ctx->poll_list)) {
6006 finish_wait(&ctx->sqo_wait, &wait);
6010 /* Tell userspace we may need a wakeup call */
6011 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6012 /* make sure to read SQ tail after writing flags */
6015 to_submit = io_sqring_entries(ctx);
6016 if (!to_submit || ret == -EBUSY) {
6017 if (kthread_should_park()) {
6018 finish_wait(&ctx->sqo_wait, &wait);
6021 if (current->task_works) {
6023 finish_wait(&ctx->sqo_wait, &wait);
6026 if (signal_pending(current))
6027 flush_signals(current);
6029 finish_wait(&ctx->sqo_wait, &wait);
6031 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6034 finish_wait(&ctx->sqo_wait, &wait);
6036 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6039 mutex_lock(&ctx->uring_lock);
6040 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6041 mutex_unlock(&ctx->uring_lock);
6042 timeout = jiffies + ctx->sq_thread_idle;
6045 if (current->task_works)
6049 io_sq_thread_drop_mm(ctx);
6050 revert_creds(old_cred);
6057 struct io_wait_queue {
6058 struct wait_queue_entry wq;
6059 struct io_ring_ctx *ctx;
6061 unsigned nr_timeouts;
6064 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6066 struct io_ring_ctx *ctx = iowq->ctx;
6069 * Wake up if we have enough events, or if a timeout occurred since we
6070 * started waiting. For timeouts, we always want to return to userspace,
6071 * regardless of event count.
6073 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6074 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6077 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6078 int wake_flags, void *key)
6080 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6083 /* use noflush == true, as we can't safely rely on locking context */
6084 if (!io_should_wake(iowq, true))
6087 return autoremove_wake_function(curr, mode, wake_flags, key);
6091 * Wait until events become available, if we don't already have some. The
6092 * application must reap them itself, as they reside on the shared cq ring.
6094 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6095 const sigset_t __user *sig, size_t sigsz)
6097 struct io_wait_queue iowq = {
6100 .func = io_wake_function,
6101 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6104 .to_wait = min_events,
6106 struct io_rings *rings = ctx->rings;
6110 if (io_cqring_events(ctx, false) >= min_events)
6112 if (!current->task_works)
6118 #ifdef CONFIG_COMPAT
6119 if (in_compat_syscall())
6120 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6124 ret = set_user_sigmask(sig, sigsz);
6130 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6131 trace_io_uring_cqring_wait(ctx, min_events);
6133 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6134 TASK_INTERRUPTIBLE);
6135 if (current->task_works)
6137 if (io_should_wake(&iowq, false))
6140 if (signal_pending(current)) {
6145 finish_wait(&ctx->wait, &iowq.wq);
6147 restore_saved_sigmask_unless(ret == -EINTR);
6149 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6152 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6154 #if defined(CONFIG_UNIX)
6155 if (ctx->ring_sock) {
6156 struct sock *sock = ctx->ring_sock->sk;
6157 struct sk_buff *skb;
6159 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6165 for (i = 0; i < ctx->nr_user_files; i++) {
6168 file = io_file_from_index(ctx, i);
6175 static void io_file_ref_kill(struct percpu_ref *ref)
6177 struct fixed_file_data *data;
6179 data = container_of(ref, struct fixed_file_data, refs);
6180 complete(&data->done);
6183 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6185 struct fixed_file_data *data = ctx->file_data;
6186 struct fixed_file_ref_node *ref_node = NULL;
6187 unsigned nr_tables, i;
6188 unsigned long flags;
6193 spin_lock_irqsave(&data->lock, flags);
6194 if (!list_empty(&data->ref_list))
6195 ref_node = list_first_entry(&data->ref_list,
6196 struct fixed_file_ref_node, node);
6197 spin_unlock_irqrestore(&data->lock, flags);
6199 percpu_ref_kill(&ref_node->refs);
6201 percpu_ref_kill(&data->refs);
6203 /* wait for all refs nodes to complete */
6204 wait_for_completion(&data->done);
6206 __io_sqe_files_unregister(ctx);
6207 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6208 for (i = 0; i < nr_tables; i++)
6209 kfree(data->table[i].files);
6211 percpu_ref_exit(&data->refs);
6213 ctx->file_data = NULL;
6214 ctx->nr_user_files = 0;
6218 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6220 if (ctx->sqo_thread) {
6221 wait_for_completion(&ctx->completions[1]);
6223 * The park is a bit of a work-around, without it we get
6224 * warning spews on shutdown with SQPOLL set and affinity
6225 * set to a single CPU.
6227 kthread_park(ctx->sqo_thread);
6228 kthread_stop(ctx->sqo_thread);
6229 ctx->sqo_thread = NULL;
6233 static void io_finish_async(struct io_ring_ctx *ctx)
6235 io_sq_thread_stop(ctx);
6238 io_wq_destroy(ctx->io_wq);
6243 #if defined(CONFIG_UNIX)
6245 * Ensure the UNIX gc is aware of our file set, so we are certain that
6246 * the io_uring can be safely unregistered on process exit, even if we have
6247 * loops in the file referencing.
6249 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6251 struct sock *sk = ctx->ring_sock->sk;
6252 struct scm_fp_list *fpl;
6253 struct sk_buff *skb;
6256 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6260 skb = alloc_skb(0, GFP_KERNEL);
6269 fpl->user = get_uid(ctx->user);
6270 for (i = 0; i < nr; i++) {
6271 struct file *file = io_file_from_index(ctx, i + offset);
6275 fpl->fp[nr_files] = get_file(file);
6276 unix_inflight(fpl->user, fpl->fp[nr_files]);
6281 fpl->max = SCM_MAX_FD;
6282 fpl->count = nr_files;
6283 UNIXCB(skb).fp = fpl;
6284 skb->destructor = unix_destruct_scm;
6285 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6286 skb_queue_head(&sk->sk_receive_queue, skb);
6288 for (i = 0; i < nr_files; i++)
6299 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6300 * causes regular reference counting to break down. We rely on the UNIX
6301 * garbage collection to take care of this problem for us.
6303 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6305 unsigned left, total;
6309 left = ctx->nr_user_files;
6311 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6313 ret = __io_sqe_files_scm(ctx, this_files, total);
6317 total += this_files;
6323 while (total < ctx->nr_user_files) {
6324 struct file *file = io_file_from_index(ctx, total);
6334 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6340 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6345 for (i = 0; i < nr_tables; i++) {
6346 struct fixed_file_table *table = &ctx->file_data->table[i];
6347 unsigned this_files;
6349 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6350 table->files = kcalloc(this_files, sizeof(struct file *),
6354 nr_files -= this_files;
6360 for (i = 0; i < nr_tables; i++) {
6361 struct fixed_file_table *table = &ctx->file_data->table[i];
6362 kfree(table->files);
6367 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6369 #if defined(CONFIG_UNIX)
6370 struct sock *sock = ctx->ring_sock->sk;
6371 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6372 struct sk_buff *skb;
6375 __skb_queue_head_init(&list);
6378 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6379 * remove this entry and rearrange the file array.
6381 skb = skb_dequeue(head);
6383 struct scm_fp_list *fp;
6385 fp = UNIXCB(skb).fp;
6386 for (i = 0; i < fp->count; i++) {
6389 if (fp->fp[i] != file)
6392 unix_notinflight(fp->user, fp->fp[i]);
6393 left = fp->count - 1 - i;
6395 memmove(&fp->fp[i], &fp->fp[i + 1],
6396 left * sizeof(struct file *));
6403 __skb_queue_tail(&list, skb);
6413 __skb_queue_tail(&list, skb);
6415 skb = skb_dequeue(head);
6418 if (skb_peek(&list)) {
6419 spin_lock_irq(&head->lock);
6420 while ((skb = __skb_dequeue(&list)) != NULL)
6421 __skb_queue_tail(head, skb);
6422 spin_unlock_irq(&head->lock);
6429 struct io_file_put {
6430 struct list_head list;
6434 static void io_file_put_work(struct work_struct *work)
6436 struct fixed_file_ref_node *ref_node;
6437 struct fixed_file_data *file_data;
6438 struct io_ring_ctx *ctx;
6439 struct io_file_put *pfile, *tmp;
6440 unsigned long flags;
6442 ref_node = container_of(work, struct fixed_file_ref_node, work);
6443 file_data = ref_node->file_data;
6444 ctx = file_data->ctx;
6446 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6447 list_del_init(&pfile->list);
6448 io_ring_file_put(ctx, pfile->file);
6452 spin_lock_irqsave(&file_data->lock, flags);
6453 list_del_init(&ref_node->node);
6454 spin_unlock_irqrestore(&file_data->lock, flags);
6456 percpu_ref_exit(&ref_node->refs);
6458 percpu_ref_put(&file_data->refs);
6461 static void io_file_data_ref_zero(struct percpu_ref *ref)
6463 struct fixed_file_ref_node *ref_node;
6465 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6467 queue_work(system_wq, &ref_node->work);
6470 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6471 struct io_ring_ctx *ctx)
6473 struct fixed_file_ref_node *ref_node;
6475 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6477 return ERR_PTR(-ENOMEM);
6479 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6482 return ERR_PTR(-ENOMEM);
6484 INIT_LIST_HEAD(&ref_node->node);
6485 INIT_LIST_HEAD(&ref_node->file_list);
6486 INIT_WORK(&ref_node->work, io_file_put_work);
6487 ref_node->file_data = ctx->file_data;
6492 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6494 percpu_ref_exit(&ref_node->refs);
6498 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6501 __s32 __user *fds = (__s32 __user *) arg;
6506 struct fixed_file_ref_node *ref_node;
6507 unsigned long flags;
6513 if (nr_args > IORING_MAX_FIXED_FILES)
6516 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6517 if (!ctx->file_data)
6519 ctx->file_data->ctx = ctx;
6520 init_completion(&ctx->file_data->done);
6521 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6522 spin_lock_init(&ctx->file_data->lock);
6524 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6525 ctx->file_data->table = kcalloc(nr_tables,
6526 sizeof(struct fixed_file_table),
6528 if (!ctx->file_data->table) {
6529 kfree(ctx->file_data);
6530 ctx->file_data = NULL;
6534 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6535 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6536 kfree(ctx->file_data->table);
6537 kfree(ctx->file_data);
6538 ctx->file_data = NULL;
6542 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6543 percpu_ref_exit(&ctx->file_data->refs);
6544 kfree(ctx->file_data->table);
6545 kfree(ctx->file_data);
6546 ctx->file_data = NULL;
6550 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6551 struct fixed_file_table *table;
6555 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6557 /* allow sparse sets */
6563 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6564 index = i & IORING_FILE_TABLE_MASK;
6572 * Don't allow io_uring instances to be registered. If UNIX
6573 * isn't enabled, then this causes a reference cycle and this
6574 * instance can never get freed. If UNIX is enabled we'll
6575 * handle it just fine, but there's still no point in allowing
6576 * a ring fd as it doesn't support regular read/write anyway.
6578 if (file->f_op == &io_uring_fops) {
6583 table->files[index] = file;
6587 for (i = 0; i < ctx->nr_user_files; i++) {
6588 file = io_file_from_index(ctx, i);
6592 for (i = 0; i < nr_tables; i++)
6593 kfree(ctx->file_data->table[i].files);
6595 kfree(ctx->file_data->table);
6596 kfree(ctx->file_data);
6597 ctx->file_data = NULL;
6598 ctx->nr_user_files = 0;
6602 ret = io_sqe_files_scm(ctx);
6604 io_sqe_files_unregister(ctx);
6608 ref_node = alloc_fixed_file_ref_node(ctx);
6609 if (IS_ERR(ref_node)) {
6610 io_sqe_files_unregister(ctx);
6611 return PTR_ERR(ref_node);
6614 ctx->file_data->cur_refs = &ref_node->refs;
6615 spin_lock_irqsave(&ctx->file_data->lock, flags);
6616 list_add(&ref_node->node, &ctx->file_data->ref_list);
6617 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6618 percpu_ref_get(&ctx->file_data->refs);
6622 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6625 #if defined(CONFIG_UNIX)
6626 struct sock *sock = ctx->ring_sock->sk;
6627 struct sk_buff_head *head = &sock->sk_receive_queue;
6628 struct sk_buff *skb;
6631 * See if we can merge this file into an existing skb SCM_RIGHTS
6632 * file set. If there's no room, fall back to allocating a new skb
6633 * and filling it in.
6635 spin_lock_irq(&head->lock);
6636 skb = skb_peek(head);
6638 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6640 if (fpl->count < SCM_MAX_FD) {
6641 __skb_unlink(skb, head);
6642 spin_unlock_irq(&head->lock);
6643 fpl->fp[fpl->count] = get_file(file);
6644 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6646 spin_lock_irq(&head->lock);
6647 __skb_queue_head(head, skb);
6652 spin_unlock_irq(&head->lock);
6659 return __io_sqe_files_scm(ctx, 1, index);
6665 static int io_queue_file_removal(struct fixed_file_data *data,
6668 struct io_file_put *pfile;
6669 struct percpu_ref *refs = data->cur_refs;
6670 struct fixed_file_ref_node *ref_node;
6672 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6676 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6678 list_add(&pfile->list, &ref_node->file_list);
6683 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6684 struct io_uring_files_update *up,
6687 struct fixed_file_data *data = ctx->file_data;
6688 struct fixed_file_ref_node *ref_node;
6693 unsigned long flags;
6694 bool needs_switch = false;
6696 if (check_add_overflow(up->offset, nr_args, &done))
6698 if (done > ctx->nr_user_files)
6701 ref_node = alloc_fixed_file_ref_node(ctx);
6702 if (IS_ERR(ref_node))
6703 return PTR_ERR(ref_node);
6706 fds = u64_to_user_ptr(up->fds);
6708 struct fixed_file_table *table;
6712 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6716 i = array_index_nospec(up->offset, ctx->nr_user_files);
6717 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6718 index = i & IORING_FILE_TABLE_MASK;
6719 if (table->files[index]) {
6720 file = io_file_from_index(ctx, index);
6721 err = io_queue_file_removal(data, file);
6724 table->files[index] = NULL;
6725 needs_switch = true;
6734 * Don't allow io_uring instances to be registered. If
6735 * UNIX isn't enabled, then this causes a reference
6736 * cycle and this instance can never get freed. If UNIX
6737 * is enabled we'll handle it just fine, but there's
6738 * still no point in allowing a ring fd as it doesn't
6739 * support regular read/write anyway.
6741 if (file->f_op == &io_uring_fops) {
6746 table->files[index] = file;
6747 err = io_sqe_file_register(ctx, file, i);
6757 percpu_ref_kill(data->cur_refs);
6758 spin_lock_irqsave(&data->lock, flags);
6759 list_add(&ref_node->node, &data->ref_list);
6760 data->cur_refs = &ref_node->refs;
6761 spin_unlock_irqrestore(&data->lock, flags);
6762 percpu_ref_get(&ctx->file_data->refs);
6764 destroy_fixed_file_ref_node(ref_node);
6766 return done ? done : err;
6769 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6772 struct io_uring_files_update up;
6774 if (!ctx->file_data)
6778 if (copy_from_user(&up, arg, sizeof(up)))
6783 return __io_sqe_files_update(ctx, &up, nr_args);
6786 static void io_free_work(struct io_wq_work *work)
6788 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6790 /* Consider that io_steal_work() relies on this ref */
6794 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6795 struct io_uring_params *p)
6797 struct io_wq_data data;
6799 struct io_ring_ctx *ctx_attach;
6800 unsigned int concurrency;
6803 data.user = ctx->user;
6804 data.free_work = io_free_work;
6806 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6807 /* Do QD, or 4 * CPUS, whatever is smallest */
6808 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6810 ctx->io_wq = io_wq_create(concurrency, &data);
6811 if (IS_ERR(ctx->io_wq)) {
6812 ret = PTR_ERR(ctx->io_wq);
6818 f = fdget(p->wq_fd);
6822 if (f.file->f_op != &io_uring_fops) {
6827 ctx_attach = f.file->private_data;
6828 /* @io_wq is protected by holding the fd */
6829 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6834 ctx->io_wq = ctx_attach->io_wq;
6840 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6841 struct io_uring_params *p)
6845 mmgrab(current->mm);
6846 ctx->sqo_mm = current->mm;
6848 if (ctx->flags & IORING_SETUP_SQPOLL) {
6850 if (!capable(CAP_SYS_ADMIN))
6853 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6854 if (!ctx->sq_thread_idle)
6855 ctx->sq_thread_idle = HZ;
6857 if (p->flags & IORING_SETUP_SQ_AFF) {
6858 int cpu = p->sq_thread_cpu;
6861 if (cpu >= nr_cpu_ids)
6863 if (!cpu_online(cpu))
6866 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6870 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6873 if (IS_ERR(ctx->sqo_thread)) {
6874 ret = PTR_ERR(ctx->sqo_thread);
6875 ctx->sqo_thread = NULL;
6878 wake_up_process(ctx->sqo_thread);
6879 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6880 /* Can't have SQ_AFF without SQPOLL */
6885 ret = io_init_wq_offload(ctx, p);
6891 io_finish_async(ctx);
6892 mmdrop(ctx->sqo_mm);
6897 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6899 atomic_long_sub(nr_pages, &user->locked_vm);
6902 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6904 unsigned long page_limit, cur_pages, new_pages;
6906 /* Don't allow more pages than we can safely lock */
6907 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6910 cur_pages = atomic_long_read(&user->locked_vm);
6911 new_pages = cur_pages + nr_pages;
6912 if (new_pages > page_limit)
6914 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6915 new_pages) != cur_pages);
6920 static void io_mem_free(void *ptr)
6927 page = virt_to_head_page(ptr);
6928 if (put_page_testzero(page))
6929 free_compound_page(page);
6932 static void *io_mem_alloc(size_t size)
6934 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6937 return (void *) __get_free_pages(gfp_flags, get_order(size));
6940 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6943 struct io_rings *rings;
6944 size_t off, sq_array_size;
6946 off = struct_size(rings, cqes, cq_entries);
6947 if (off == SIZE_MAX)
6951 off = ALIGN(off, SMP_CACHE_BYTES);
6956 sq_array_size = array_size(sizeof(u32), sq_entries);
6957 if (sq_array_size == SIZE_MAX)
6960 if (check_add_overflow(off, sq_array_size, &off))
6969 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6973 pages = (size_t)1 << get_order(
6974 rings_size(sq_entries, cq_entries, NULL));
6975 pages += (size_t)1 << get_order(
6976 array_size(sizeof(struct io_uring_sqe), sq_entries));
6981 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6985 if (!ctx->user_bufs)
6988 for (i = 0; i < ctx->nr_user_bufs; i++) {
6989 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6991 for (j = 0; j < imu->nr_bvecs; j++)
6992 unpin_user_page(imu->bvec[j].bv_page);
6994 if (ctx->account_mem)
6995 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7000 kfree(ctx->user_bufs);
7001 ctx->user_bufs = NULL;
7002 ctx->nr_user_bufs = 0;
7006 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7007 void __user *arg, unsigned index)
7009 struct iovec __user *src;
7011 #ifdef CONFIG_COMPAT
7013 struct compat_iovec __user *ciovs;
7014 struct compat_iovec ciov;
7016 ciovs = (struct compat_iovec __user *) arg;
7017 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7020 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7021 dst->iov_len = ciov.iov_len;
7025 src = (struct iovec __user *) arg;
7026 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7031 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7034 struct vm_area_struct **vmas = NULL;
7035 struct page **pages = NULL;
7036 int i, j, got_pages = 0;
7041 if (!nr_args || nr_args > UIO_MAXIOV)
7044 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7046 if (!ctx->user_bufs)
7049 for (i = 0; i < nr_args; i++) {
7050 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7051 unsigned long off, start, end, ubuf;
7056 ret = io_copy_iov(ctx, &iov, arg, i);
7061 * Don't impose further limits on the size and buffer
7062 * constraints here, we'll -EINVAL later when IO is
7063 * submitted if they are wrong.
7066 if (!iov.iov_base || !iov.iov_len)
7069 /* arbitrary limit, but we need something */
7070 if (iov.iov_len > SZ_1G)
7073 ubuf = (unsigned long) iov.iov_base;
7074 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7075 start = ubuf >> PAGE_SHIFT;
7076 nr_pages = end - start;
7078 if (ctx->account_mem) {
7079 ret = io_account_mem(ctx->user, nr_pages);
7085 if (!pages || nr_pages > got_pages) {
7088 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7090 vmas = kvmalloc_array(nr_pages,
7091 sizeof(struct vm_area_struct *),
7093 if (!pages || !vmas) {
7095 if (ctx->account_mem)
7096 io_unaccount_mem(ctx->user, nr_pages);
7099 got_pages = nr_pages;
7102 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7106 if (ctx->account_mem)
7107 io_unaccount_mem(ctx->user, nr_pages);
7112 down_read(¤t->mm->mmap_sem);
7113 pret = pin_user_pages(ubuf, nr_pages,
7114 FOLL_WRITE | FOLL_LONGTERM,
7116 if (pret == nr_pages) {
7117 /* don't support file backed memory */
7118 for (j = 0; j < nr_pages; j++) {
7119 struct vm_area_struct *vma = vmas[j];
7122 !is_file_hugepages(vma->vm_file)) {
7128 ret = pret < 0 ? pret : -EFAULT;
7130 up_read(¤t->mm->mmap_sem);
7133 * if we did partial map, or found file backed vmas,
7134 * release any pages we did get
7137 unpin_user_pages(pages, pret);
7138 if (ctx->account_mem)
7139 io_unaccount_mem(ctx->user, nr_pages);
7144 off = ubuf & ~PAGE_MASK;
7146 for (j = 0; j < nr_pages; j++) {
7149 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7150 imu->bvec[j].bv_page = pages[j];
7151 imu->bvec[j].bv_len = vec_len;
7152 imu->bvec[j].bv_offset = off;
7156 /* store original address for later verification */
7158 imu->len = iov.iov_len;
7159 imu->nr_bvecs = nr_pages;
7161 ctx->nr_user_bufs++;
7169 io_sqe_buffer_unregister(ctx);
7173 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7175 __s32 __user *fds = arg;
7181 if (copy_from_user(&fd, fds, sizeof(*fds)))
7184 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7185 if (IS_ERR(ctx->cq_ev_fd)) {
7186 int ret = PTR_ERR(ctx->cq_ev_fd);
7187 ctx->cq_ev_fd = NULL;
7194 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7196 if (ctx->cq_ev_fd) {
7197 eventfd_ctx_put(ctx->cq_ev_fd);
7198 ctx->cq_ev_fd = NULL;
7205 static int __io_destroy_buffers(int id, void *p, void *data)
7207 struct io_ring_ctx *ctx = data;
7208 struct io_buffer *buf = p;
7210 __io_remove_buffers(ctx, buf, id, -1U);
7214 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7216 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7217 idr_destroy(&ctx->io_buffer_idr);
7220 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7222 io_finish_async(ctx);
7224 mmdrop(ctx->sqo_mm);
7226 io_iopoll_reap_events(ctx);
7227 io_sqe_buffer_unregister(ctx);
7228 io_sqe_files_unregister(ctx);
7229 io_eventfd_unregister(ctx);
7230 io_destroy_buffers(ctx);
7231 idr_destroy(&ctx->personality_idr);
7233 #if defined(CONFIG_UNIX)
7234 if (ctx->ring_sock) {
7235 ctx->ring_sock->file = NULL; /* so that iput() is called */
7236 sock_release(ctx->ring_sock);
7240 io_mem_free(ctx->rings);
7241 io_mem_free(ctx->sq_sqes);
7243 percpu_ref_exit(&ctx->refs);
7244 if (ctx->account_mem)
7245 io_unaccount_mem(ctx->user,
7246 ring_pages(ctx->sq_entries, ctx->cq_entries));
7247 free_uid(ctx->user);
7248 put_cred(ctx->creds);
7249 kfree(ctx->completions);
7250 kfree(ctx->cancel_hash);
7251 kmem_cache_free(req_cachep, ctx->fallback_req);
7255 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7257 struct io_ring_ctx *ctx = file->private_data;
7260 poll_wait(file, &ctx->cq_wait, wait);
7262 * synchronizes with barrier from wq_has_sleeper call in
7266 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7267 ctx->rings->sq_ring_entries)
7268 mask |= EPOLLOUT | EPOLLWRNORM;
7269 if (io_cqring_events(ctx, false))
7270 mask |= EPOLLIN | EPOLLRDNORM;
7275 static int io_uring_fasync(int fd, struct file *file, int on)
7277 struct io_ring_ctx *ctx = file->private_data;
7279 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7282 static int io_remove_personalities(int id, void *p, void *data)
7284 struct io_ring_ctx *ctx = data;
7285 const struct cred *cred;
7287 cred = idr_remove(&ctx->personality_idr, id);
7293 static void io_ring_exit_work(struct work_struct *work)
7295 struct io_ring_ctx *ctx;
7297 ctx = container_of(work, struct io_ring_ctx, exit_work);
7299 io_cqring_overflow_flush(ctx, true);
7301 wait_for_completion(&ctx->completions[0]);
7302 io_ring_ctx_free(ctx);
7305 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7307 mutex_lock(&ctx->uring_lock);
7308 percpu_ref_kill(&ctx->refs);
7309 mutex_unlock(&ctx->uring_lock);
7312 * Wait for sq thread to idle, if we have one. It won't spin on new
7313 * work after we've killed the ctx ref above. This is important to do
7314 * before we cancel existing commands, as the thread could otherwise
7315 * be queueing new work post that. If that's work we need to cancel,
7316 * it could cause shutdown to hang.
7318 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7321 io_kill_timeouts(ctx);
7322 io_poll_remove_all(ctx);
7325 io_wq_cancel_all(ctx->io_wq);
7327 io_iopoll_reap_events(ctx);
7328 /* if we failed setting up the ctx, we might not have any rings */
7330 io_cqring_overflow_flush(ctx, true);
7331 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7332 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7333 queue_work(system_wq, &ctx->exit_work);
7336 static int io_uring_release(struct inode *inode, struct file *file)
7338 struct io_ring_ctx *ctx = file->private_data;
7340 file->private_data = NULL;
7341 io_ring_ctx_wait_and_kill(ctx);
7345 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7346 struct files_struct *files)
7348 while (!list_empty_careful(&ctx->inflight_list)) {
7349 struct io_kiocb *cancel_req = NULL, *req;
7352 spin_lock_irq(&ctx->inflight_lock);
7353 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7354 if (req->work.files != files)
7356 /* req is being completed, ignore */
7357 if (!refcount_inc_not_zero(&req->refs))
7363 prepare_to_wait(&ctx->inflight_wait, &wait,
7364 TASK_UNINTERRUPTIBLE);
7365 spin_unlock_irq(&ctx->inflight_lock);
7367 /* We need to keep going until we don't find a matching req */
7371 if (cancel_req->flags & REQ_F_OVERFLOW) {
7372 spin_lock_irq(&ctx->completion_lock);
7373 list_del(&cancel_req->list);
7374 cancel_req->flags &= ~REQ_F_OVERFLOW;
7375 if (list_empty(&ctx->cq_overflow_list)) {
7376 clear_bit(0, &ctx->sq_check_overflow);
7377 clear_bit(0, &ctx->cq_check_overflow);
7379 spin_unlock_irq(&ctx->completion_lock);
7381 WRITE_ONCE(ctx->rings->cq_overflow,
7382 atomic_inc_return(&ctx->cached_cq_overflow));
7385 * Put inflight ref and overflow ref. If that's
7386 * all we had, then we're done with this request.
7388 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7389 io_put_req(cancel_req);
7390 finish_wait(&ctx->inflight_wait, &wait);
7395 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7396 io_put_req(cancel_req);
7398 finish_wait(&ctx->inflight_wait, &wait);
7402 static int io_uring_flush(struct file *file, void *data)
7404 struct io_ring_ctx *ctx = file->private_data;
7406 io_uring_cancel_files(ctx, data);
7409 * If the task is going away, cancel work it may have pending
7411 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7412 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7417 static void *io_uring_validate_mmap_request(struct file *file,
7418 loff_t pgoff, size_t sz)
7420 struct io_ring_ctx *ctx = file->private_data;
7421 loff_t offset = pgoff << PAGE_SHIFT;
7426 case IORING_OFF_SQ_RING:
7427 case IORING_OFF_CQ_RING:
7430 case IORING_OFF_SQES:
7434 return ERR_PTR(-EINVAL);
7437 page = virt_to_head_page(ptr);
7438 if (sz > page_size(page))
7439 return ERR_PTR(-EINVAL);
7446 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7448 size_t sz = vma->vm_end - vma->vm_start;
7452 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7454 return PTR_ERR(ptr);
7456 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7457 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7460 #else /* !CONFIG_MMU */
7462 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7464 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7467 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7469 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7472 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7473 unsigned long addr, unsigned long len,
7474 unsigned long pgoff, unsigned long flags)
7478 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7480 return PTR_ERR(ptr);
7482 return (unsigned long) ptr;
7485 #endif /* !CONFIG_MMU */
7487 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7488 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7491 struct io_ring_ctx *ctx;
7496 if (current->task_works)
7499 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7507 if (f.file->f_op != &io_uring_fops)
7511 ctx = f.file->private_data;
7512 if (!percpu_ref_tryget(&ctx->refs))
7516 * For SQ polling, the thread will do all submissions and completions.
7517 * Just return the requested submit count, and wake the thread if
7521 if (ctx->flags & IORING_SETUP_SQPOLL) {
7522 if (!list_empty_careful(&ctx->cq_overflow_list))
7523 io_cqring_overflow_flush(ctx, false);
7524 if (flags & IORING_ENTER_SQ_WAKEUP)
7525 wake_up(&ctx->sqo_wait);
7526 submitted = to_submit;
7527 } else if (to_submit) {
7528 mutex_lock(&ctx->uring_lock);
7529 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7530 mutex_unlock(&ctx->uring_lock);
7532 if (submitted != to_submit)
7535 if (flags & IORING_ENTER_GETEVENTS) {
7536 unsigned nr_events = 0;
7538 min_complete = min(min_complete, ctx->cq_entries);
7541 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7542 * space applications don't need to do io completion events
7543 * polling again, they can rely on io_sq_thread to do polling
7544 * work, which can reduce cpu usage and uring_lock contention.
7546 if (ctx->flags & IORING_SETUP_IOPOLL &&
7547 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7548 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7550 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7555 percpu_ref_put(&ctx->refs);
7558 return submitted ? submitted : ret;
7561 #ifdef CONFIG_PROC_FS
7562 static int io_uring_show_cred(int id, void *p, void *data)
7564 const struct cred *cred = p;
7565 struct seq_file *m = data;
7566 struct user_namespace *uns = seq_user_ns(m);
7567 struct group_info *gi;
7572 seq_printf(m, "%5d\n", id);
7573 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7574 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7575 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7576 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7577 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7578 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7579 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7580 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7581 seq_puts(m, "\n\tGroups:\t");
7582 gi = cred->group_info;
7583 for (g = 0; g < gi->ngroups; g++) {
7584 seq_put_decimal_ull(m, g ? " " : "",
7585 from_kgid_munged(uns, gi->gid[g]));
7587 seq_puts(m, "\n\tCapEff:\t");
7588 cap = cred->cap_effective;
7589 CAP_FOR_EACH_U32(__capi)
7590 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7595 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7599 mutex_lock(&ctx->uring_lock);
7600 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7601 for (i = 0; i < ctx->nr_user_files; i++) {
7602 struct fixed_file_table *table;
7605 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7606 f = table->files[i & IORING_FILE_TABLE_MASK];
7608 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7610 seq_printf(m, "%5u: <none>\n", i);
7612 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7613 for (i = 0; i < ctx->nr_user_bufs; i++) {
7614 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7616 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7617 (unsigned int) buf->len);
7619 if (!idr_is_empty(&ctx->personality_idr)) {
7620 seq_printf(m, "Personalities:\n");
7621 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7623 seq_printf(m, "PollList:\n");
7624 spin_lock_irq(&ctx->completion_lock);
7625 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7626 struct hlist_head *list = &ctx->cancel_hash[i];
7627 struct io_kiocb *req;
7629 hlist_for_each_entry(req, list, hash_node)
7630 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7631 req->task->task_works != NULL);
7633 spin_unlock_irq(&ctx->completion_lock);
7634 mutex_unlock(&ctx->uring_lock);
7637 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7639 struct io_ring_ctx *ctx = f->private_data;
7641 if (percpu_ref_tryget(&ctx->refs)) {
7642 __io_uring_show_fdinfo(ctx, m);
7643 percpu_ref_put(&ctx->refs);
7648 static const struct file_operations io_uring_fops = {
7649 .release = io_uring_release,
7650 .flush = io_uring_flush,
7651 .mmap = io_uring_mmap,
7653 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7654 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7656 .poll = io_uring_poll,
7657 .fasync = io_uring_fasync,
7658 #ifdef CONFIG_PROC_FS
7659 .show_fdinfo = io_uring_show_fdinfo,
7663 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7664 struct io_uring_params *p)
7666 struct io_rings *rings;
7667 size_t size, sq_array_offset;
7669 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7670 if (size == SIZE_MAX)
7673 rings = io_mem_alloc(size);
7678 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7679 rings->sq_ring_mask = p->sq_entries - 1;
7680 rings->cq_ring_mask = p->cq_entries - 1;
7681 rings->sq_ring_entries = p->sq_entries;
7682 rings->cq_ring_entries = p->cq_entries;
7683 ctx->sq_mask = rings->sq_ring_mask;
7684 ctx->cq_mask = rings->cq_ring_mask;
7685 ctx->sq_entries = rings->sq_ring_entries;
7686 ctx->cq_entries = rings->cq_ring_entries;
7688 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7689 if (size == SIZE_MAX) {
7690 io_mem_free(ctx->rings);
7695 ctx->sq_sqes = io_mem_alloc(size);
7696 if (!ctx->sq_sqes) {
7697 io_mem_free(ctx->rings);
7706 * Allocate an anonymous fd, this is what constitutes the application
7707 * visible backing of an io_uring instance. The application mmaps this
7708 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7709 * we have to tie this fd to a socket for file garbage collection purposes.
7711 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7716 #if defined(CONFIG_UNIX)
7717 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7723 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7727 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7728 O_RDWR | O_CLOEXEC);
7731 ret = PTR_ERR(file);
7735 #if defined(CONFIG_UNIX)
7736 ctx->ring_sock->file = file;
7738 fd_install(ret, file);
7741 #if defined(CONFIG_UNIX)
7742 sock_release(ctx->ring_sock);
7743 ctx->ring_sock = NULL;
7748 static int io_uring_create(unsigned entries, struct io_uring_params *p,
7749 struct io_uring_params __user *params)
7751 struct user_struct *user = NULL;
7752 struct io_ring_ctx *ctx;
7758 if (entries > IORING_MAX_ENTRIES) {
7759 if (!(p->flags & IORING_SETUP_CLAMP))
7761 entries = IORING_MAX_ENTRIES;
7765 * Use twice as many entries for the CQ ring. It's possible for the
7766 * application to drive a higher depth than the size of the SQ ring,
7767 * since the sqes are only used at submission time. This allows for
7768 * some flexibility in overcommitting a bit. If the application has
7769 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7770 * of CQ ring entries manually.
7772 p->sq_entries = roundup_pow_of_two(entries);
7773 if (p->flags & IORING_SETUP_CQSIZE) {
7775 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7776 * to a power-of-two, if it isn't already. We do NOT impose
7777 * any cq vs sq ring sizing.
7779 if (p->cq_entries < p->sq_entries)
7781 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7782 if (!(p->flags & IORING_SETUP_CLAMP))
7784 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7786 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7788 p->cq_entries = 2 * p->sq_entries;
7791 user = get_uid(current_user());
7792 account_mem = !capable(CAP_IPC_LOCK);
7795 ret = io_account_mem(user,
7796 ring_pages(p->sq_entries, p->cq_entries));
7803 ctx = io_ring_ctx_alloc(p);
7806 io_unaccount_mem(user, ring_pages(p->sq_entries,
7811 ctx->compat = in_compat_syscall();
7812 ctx->account_mem = account_mem;
7814 ctx->creds = get_current_cred();
7816 ret = io_allocate_scq_urings(ctx, p);
7820 ret = io_sq_offload_start(ctx, p);
7824 memset(&p->sq_off, 0, sizeof(p->sq_off));
7825 p->sq_off.head = offsetof(struct io_rings, sq.head);
7826 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7827 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7828 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7829 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7830 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7831 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7833 memset(&p->cq_off, 0, sizeof(p->cq_off));
7834 p->cq_off.head = offsetof(struct io_rings, cq.head);
7835 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7836 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7837 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7838 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7839 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7841 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7842 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7843 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7845 if (copy_to_user(params, p, sizeof(*p))) {
7850 * Install ring fd as the very last thing, so we don't risk someone
7851 * having closed it before we finish setup
7853 ret = io_uring_get_fd(ctx);
7857 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7860 io_ring_ctx_wait_and_kill(ctx);
7865 * Sets up an aio uring context, and returns the fd. Applications asks for a
7866 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7867 * params structure passed in.
7869 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7871 struct io_uring_params p;
7874 if (copy_from_user(&p, params, sizeof(p)))
7876 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7881 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7882 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7883 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7886 return io_uring_create(entries, &p, params);
7889 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7890 struct io_uring_params __user *, params)
7892 return io_uring_setup(entries, params);
7895 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7897 struct io_uring_probe *p;
7901 size = struct_size(p, ops, nr_args);
7902 if (size == SIZE_MAX)
7904 p = kzalloc(size, GFP_KERNEL);
7909 if (copy_from_user(p, arg, size))
7912 if (memchr_inv(p, 0, size))
7915 p->last_op = IORING_OP_LAST - 1;
7916 if (nr_args > IORING_OP_LAST)
7917 nr_args = IORING_OP_LAST;
7919 for (i = 0; i < nr_args; i++) {
7921 if (!io_op_defs[i].not_supported)
7922 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7927 if (copy_to_user(arg, p, size))
7934 static int io_register_personality(struct io_ring_ctx *ctx)
7936 const struct cred *creds = get_current_cred();
7939 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7940 USHRT_MAX, GFP_KERNEL);
7946 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7948 const struct cred *old_creds;
7950 old_creds = idr_remove(&ctx->personality_idr, id);
7952 put_cred(old_creds);
7959 static bool io_register_op_must_quiesce(int op)
7962 case IORING_UNREGISTER_FILES:
7963 case IORING_REGISTER_FILES_UPDATE:
7964 case IORING_REGISTER_PROBE:
7965 case IORING_REGISTER_PERSONALITY:
7966 case IORING_UNREGISTER_PERSONALITY:
7973 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7974 void __user *arg, unsigned nr_args)
7975 __releases(ctx->uring_lock)
7976 __acquires(ctx->uring_lock)
7981 * We're inside the ring mutex, if the ref is already dying, then
7982 * someone else killed the ctx or is already going through
7983 * io_uring_register().
7985 if (percpu_ref_is_dying(&ctx->refs))
7988 if (io_register_op_must_quiesce(opcode)) {
7989 percpu_ref_kill(&ctx->refs);
7992 * Drop uring mutex before waiting for references to exit. If
7993 * another thread is currently inside io_uring_enter() it might
7994 * need to grab the uring_lock to make progress. If we hold it
7995 * here across the drain wait, then we can deadlock. It's safe
7996 * to drop the mutex here, since no new references will come in
7997 * after we've killed the percpu ref.
7999 mutex_unlock(&ctx->uring_lock);
8000 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8001 mutex_lock(&ctx->uring_lock);
8003 percpu_ref_resurrect(&ctx->refs);
8010 case IORING_REGISTER_BUFFERS:
8011 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8013 case IORING_UNREGISTER_BUFFERS:
8017 ret = io_sqe_buffer_unregister(ctx);
8019 case IORING_REGISTER_FILES:
8020 ret = io_sqe_files_register(ctx, arg, nr_args);
8022 case IORING_UNREGISTER_FILES:
8026 ret = io_sqe_files_unregister(ctx);
8028 case IORING_REGISTER_FILES_UPDATE:
8029 ret = io_sqe_files_update(ctx, arg, nr_args);
8031 case IORING_REGISTER_EVENTFD:
8032 case IORING_REGISTER_EVENTFD_ASYNC:
8036 ret = io_eventfd_register(ctx, arg);
8039 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8040 ctx->eventfd_async = 1;
8042 ctx->eventfd_async = 0;
8044 case IORING_UNREGISTER_EVENTFD:
8048 ret = io_eventfd_unregister(ctx);
8050 case IORING_REGISTER_PROBE:
8052 if (!arg || nr_args > 256)
8054 ret = io_probe(ctx, arg, nr_args);
8056 case IORING_REGISTER_PERSONALITY:
8060 ret = io_register_personality(ctx);
8062 case IORING_UNREGISTER_PERSONALITY:
8066 ret = io_unregister_personality(ctx, nr_args);
8073 if (io_register_op_must_quiesce(opcode)) {
8074 /* bring the ctx back to life */
8075 percpu_ref_reinit(&ctx->refs);
8077 reinit_completion(&ctx->completions[0]);
8082 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8083 void __user *, arg, unsigned int, nr_args)
8085 struct io_ring_ctx *ctx;
8094 if (f.file->f_op != &io_uring_fops)
8097 ctx = f.file->private_data;
8099 mutex_lock(&ctx->uring_lock);
8100 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8101 mutex_unlock(&ctx->uring_lock);
8102 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8103 ctx->cq_ev_fd != NULL, ret);
8109 static int __init io_uring_init(void)
8111 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8112 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8113 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8116 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8117 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8118 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8119 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8120 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8121 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8122 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8123 BUILD_BUG_SQE_ELEM(8, __u64, off);
8124 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8125 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8126 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8127 BUILD_BUG_SQE_ELEM(24, __u32, len);
8128 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8129 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8130 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8131 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8132 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8133 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8134 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8135 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8136 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8137 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8138 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8139 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8140 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8141 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8142 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8143 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8144 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8145 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8147 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8148 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8149 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8152 __initcall(io_uring_init);