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;
365 struct sockaddr __user *addr;
366 int __user *addr_len;
368 unsigned long nofile;
392 /* NOTE: kiocb has the file as the first member, so don't do it here */
400 struct sockaddr __user *addr;
407 struct user_msghdr __user *msg;
413 struct io_buffer *kbuf;
422 struct filename *filename;
423 struct statx __user *buffer;
425 unsigned long nofile;
428 struct io_files_update {
454 struct epoll_event event;
458 struct file *file_out;
459 struct file *file_in;
466 struct io_provide_buf {
475 struct io_async_connect {
476 struct sockaddr_storage address;
479 struct io_async_msghdr {
480 struct iovec fast_iov[UIO_FASTIOV];
482 struct sockaddr __user *uaddr;
484 struct sockaddr_storage addr;
488 struct iovec fast_iov[UIO_FASTIOV];
494 struct io_async_ctx {
496 struct io_async_rw rw;
497 struct io_async_msghdr msg;
498 struct io_async_connect connect;
499 struct io_timeout_data timeout;
504 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
505 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
506 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
507 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
508 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
509 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
517 REQ_F_IOPOLL_COMPLETED_BIT,
518 REQ_F_LINK_TIMEOUT_BIT,
522 REQ_F_TIMEOUT_NOSEQ_BIT,
523 REQ_F_COMP_LOCKED_BIT,
524 REQ_F_NEED_CLEANUP_BIT,
527 REQ_F_BUFFER_SELECTED_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),
584 struct io_poll_iocb poll;
585 struct io_wq_work work;
589 * NOTE! Each of the iocb union members has the file pointer
590 * as the first entry in their struct definition. So you can
591 * access the file pointer through any of the sub-structs,
592 * or directly as just 'ki_filp' in this struct.
598 struct io_poll_iocb poll;
599 struct io_accept accept;
601 struct io_cancel cancel;
602 struct io_timeout timeout;
603 struct io_connect connect;
604 struct io_sr_msg sr_msg;
606 struct io_close close;
607 struct io_files_update files_update;
608 struct io_fadvise fadvise;
609 struct io_madvise madvise;
610 struct io_epoll epoll;
611 struct io_splice splice;
612 struct io_provide_buf pbuf;
615 struct io_async_ctx *io;
617 bool needs_fixed_file;
620 struct io_ring_ctx *ctx;
621 struct list_head list;
624 struct task_struct *task;
630 struct list_head link_list;
632 struct list_head inflight_entry;
634 struct percpu_ref *fixed_file_refs;
638 * Only commands that never go async can use the below fields,
639 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
640 * async armed poll handlers for regular commands. The latter
641 * restore the work, if needed.
644 struct callback_head task_work;
645 struct hlist_node hash_node;
646 struct async_poll *apoll;
648 struct io_wq_work work;
652 #define IO_PLUG_THRESHOLD 2
653 #define IO_IOPOLL_BATCH 8
655 struct io_submit_state {
656 struct blk_plug plug;
659 * io_kiocb alloc cache
661 void *reqs[IO_IOPOLL_BATCH];
662 unsigned int free_reqs;
665 * File reference cache
669 unsigned int has_refs;
670 unsigned int used_refs;
671 unsigned int ios_left;
675 /* needs req->io allocated for deferral/async */
676 unsigned async_ctx : 1;
677 /* needs current->mm setup, does mm access */
678 unsigned needs_mm : 1;
679 /* needs req->file assigned */
680 unsigned needs_file : 1;
681 /* needs req->file assigned IFF fd is >= 0 */
682 unsigned fd_non_neg : 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] = {
790 [IORING_OP_CLOSE] = {
794 [IORING_OP_FILES_UPDATE] = {
798 [IORING_OP_STATX] = {
807 .unbound_nonreg_file = 1,
811 [IORING_OP_WRITE] = {
814 .unbound_nonreg_file = 1,
817 [IORING_OP_FADVISE] = {
820 [IORING_OP_MADVISE] = {
826 .unbound_nonreg_file = 1,
832 .unbound_nonreg_file = 1,
836 [IORING_OP_OPENAT2] = {
842 [IORING_OP_EPOLL_CTL] = {
843 .unbound_nonreg_file = 1,
846 [IORING_OP_SPLICE] = {
849 .unbound_nonreg_file = 1,
851 [IORING_OP_PROVIDE_BUFFERS] = {},
852 [IORING_OP_REMOVE_BUFFERS] = {},
855 static void io_wq_submit_work(struct io_wq_work **workptr);
856 static void io_cqring_fill_event(struct io_kiocb *req, long res);
857 static void io_put_req(struct io_kiocb *req);
858 static void __io_double_put_req(struct io_kiocb *req);
859 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
860 static void io_queue_linked_timeout(struct io_kiocb *req);
861 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
862 struct io_uring_files_update *ip,
864 static int io_grab_files(struct io_kiocb *req);
865 static void io_cleanup_req(struct io_kiocb *req);
866 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
867 int fd, struct file **out_file, bool fixed);
868 static void __io_queue_sqe(struct io_kiocb *req,
869 const struct io_uring_sqe *sqe);
871 static struct kmem_cache *req_cachep;
873 static const struct file_operations io_uring_fops;
875 struct sock *io_uring_get_socket(struct file *file)
877 #if defined(CONFIG_UNIX)
878 if (file->f_op == &io_uring_fops) {
879 struct io_ring_ctx *ctx = file->private_data;
881 return ctx->ring_sock->sk;
886 EXPORT_SYMBOL(io_uring_get_socket);
888 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
890 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
892 complete(&ctx->completions[0]);
895 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
897 struct io_ring_ctx *ctx;
900 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
904 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
905 if (!ctx->fallback_req)
908 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
909 if (!ctx->completions)
913 * Use 5 bits less than the max cq entries, that should give us around
914 * 32 entries per hash list if totally full and uniformly spread.
916 hash_bits = ilog2(p->cq_entries);
920 ctx->cancel_hash_bits = hash_bits;
921 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
923 if (!ctx->cancel_hash)
925 __hash_init(ctx->cancel_hash, 1U << hash_bits);
927 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
928 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
931 ctx->flags = p->flags;
932 init_waitqueue_head(&ctx->cq_wait);
933 INIT_LIST_HEAD(&ctx->cq_overflow_list);
934 init_completion(&ctx->completions[0]);
935 init_completion(&ctx->completions[1]);
936 idr_init(&ctx->io_buffer_idr);
937 idr_init(&ctx->personality_idr);
938 mutex_init(&ctx->uring_lock);
939 init_waitqueue_head(&ctx->wait);
940 spin_lock_init(&ctx->completion_lock);
941 INIT_LIST_HEAD(&ctx->poll_list);
942 INIT_LIST_HEAD(&ctx->defer_list);
943 INIT_LIST_HEAD(&ctx->timeout_list);
944 init_waitqueue_head(&ctx->inflight_wait);
945 spin_lock_init(&ctx->inflight_lock);
946 INIT_LIST_HEAD(&ctx->inflight_list);
949 if (ctx->fallback_req)
950 kmem_cache_free(req_cachep, ctx->fallback_req);
951 kfree(ctx->completions);
952 kfree(ctx->cancel_hash);
957 static inline bool __req_need_defer(struct io_kiocb *req)
959 struct io_ring_ctx *ctx = req->ctx;
961 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
962 + atomic_read(&ctx->cached_cq_overflow);
965 static inline bool req_need_defer(struct io_kiocb *req)
967 if (unlikely(req->flags & REQ_F_IO_DRAIN))
968 return __req_need_defer(req);
973 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
975 struct io_kiocb *req;
977 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
978 if (req && !req_need_defer(req)) {
979 list_del_init(&req->list);
986 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
988 struct io_kiocb *req;
990 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
992 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
994 if (!__req_need_defer(req)) {
995 list_del_init(&req->list);
1003 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1005 struct io_rings *rings = ctx->rings;
1007 /* order cqe stores with ring update */
1008 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1010 if (wq_has_sleeper(&ctx->cq_wait)) {
1011 wake_up_interruptible(&ctx->cq_wait);
1012 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1016 static inline void io_req_work_grab_env(struct io_kiocb *req,
1017 const struct io_op_def *def)
1019 if (!req->work.mm && def->needs_mm) {
1020 mmgrab(current->mm);
1021 req->work.mm = current->mm;
1023 if (!req->work.creds)
1024 req->work.creds = get_current_cred();
1025 if (!req->work.fs && def->needs_fs) {
1026 spin_lock(¤t->fs->lock);
1027 if (!current->fs->in_exec) {
1028 req->work.fs = current->fs;
1029 req->work.fs->users++;
1031 req->work.flags |= IO_WQ_WORK_CANCEL;
1033 spin_unlock(¤t->fs->lock);
1035 if (!req->work.task_pid)
1036 req->work.task_pid = task_pid_vnr(current);
1039 static inline void io_req_work_drop_env(struct io_kiocb *req)
1042 mmdrop(req->work.mm);
1043 req->work.mm = NULL;
1045 if (req->work.creds) {
1046 put_cred(req->work.creds);
1047 req->work.creds = NULL;
1050 struct fs_struct *fs = req->work.fs;
1052 spin_lock(&req->work.fs->lock);
1055 spin_unlock(&req->work.fs->lock);
1061 static inline void io_prep_async_work(struct io_kiocb *req,
1062 struct io_kiocb **link)
1064 const struct io_op_def *def = &io_op_defs[req->opcode];
1066 if (req->flags & REQ_F_ISREG) {
1067 if (def->hash_reg_file)
1068 io_wq_hash_work(&req->work, file_inode(req->file));
1070 if (def->unbound_nonreg_file)
1071 req->work.flags |= IO_WQ_WORK_UNBOUND;
1074 io_req_work_grab_env(req, def);
1076 *link = io_prep_linked_timeout(req);
1079 static inline void io_queue_async_work(struct io_kiocb *req)
1081 struct io_ring_ctx *ctx = req->ctx;
1082 struct io_kiocb *link;
1084 io_prep_async_work(req, &link);
1086 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1087 &req->work, req->flags);
1088 io_wq_enqueue(ctx->io_wq, &req->work);
1091 io_queue_linked_timeout(link);
1094 static void io_kill_timeout(struct io_kiocb *req)
1098 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1100 atomic_inc(&req->ctx->cq_timeouts);
1101 list_del_init(&req->list);
1102 req->flags |= REQ_F_COMP_LOCKED;
1103 io_cqring_fill_event(req, 0);
1108 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1110 struct io_kiocb *req, *tmp;
1112 spin_lock_irq(&ctx->completion_lock);
1113 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1114 io_kill_timeout(req);
1115 spin_unlock_irq(&ctx->completion_lock);
1118 static void io_commit_cqring(struct io_ring_ctx *ctx)
1120 struct io_kiocb *req;
1122 while ((req = io_get_timeout_req(ctx)) != NULL)
1123 io_kill_timeout(req);
1125 __io_commit_cqring(ctx);
1127 while ((req = io_get_deferred_req(ctx)) != NULL)
1128 io_queue_async_work(req);
1131 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1133 struct io_rings *rings = ctx->rings;
1136 tail = ctx->cached_cq_tail;
1138 * writes to the cq entry need to come after reading head; the
1139 * control dependency is enough as we're using WRITE_ONCE to
1142 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1145 ctx->cached_cq_tail++;
1146 return &rings->cqes[tail & ctx->cq_mask];
1149 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1153 if (!ctx->eventfd_async)
1155 return io_wq_current_is_worker();
1158 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1160 if (waitqueue_active(&ctx->wait))
1161 wake_up(&ctx->wait);
1162 if (waitqueue_active(&ctx->sqo_wait))
1163 wake_up(&ctx->sqo_wait);
1164 if (io_should_trigger_evfd(ctx))
1165 eventfd_signal(ctx->cq_ev_fd, 1);
1168 /* Returns true if there are no backlogged entries after the flush */
1169 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1171 struct io_rings *rings = ctx->rings;
1172 struct io_uring_cqe *cqe;
1173 struct io_kiocb *req;
1174 unsigned long flags;
1178 if (list_empty_careful(&ctx->cq_overflow_list))
1180 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1181 rings->cq_ring_entries))
1185 spin_lock_irqsave(&ctx->completion_lock, flags);
1187 /* if force is set, the ring is going away. always drop after that */
1189 ctx->cq_overflow_flushed = 1;
1192 while (!list_empty(&ctx->cq_overflow_list)) {
1193 cqe = io_get_cqring(ctx);
1197 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1199 list_move(&req->list, &list);
1200 req->flags &= ~REQ_F_OVERFLOW;
1202 WRITE_ONCE(cqe->user_data, req->user_data);
1203 WRITE_ONCE(cqe->res, req->result);
1204 WRITE_ONCE(cqe->flags, req->cflags);
1206 WRITE_ONCE(ctx->rings->cq_overflow,
1207 atomic_inc_return(&ctx->cached_cq_overflow));
1211 io_commit_cqring(ctx);
1213 clear_bit(0, &ctx->sq_check_overflow);
1214 clear_bit(0, &ctx->cq_check_overflow);
1216 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1217 io_cqring_ev_posted(ctx);
1219 while (!list_empty(&list)) {
1220 req = list_first_entry(&list, struct io_kiocb, list);
1221 list_del(&req->list);
1228 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1230 struct io_ring_ctx *ctx = req->ctx;
1231 struct io_uring_cqe *cqe;
1233 trace_io_uring_complete(ctx, req->user_data, res);
1236 * If we can't get a cq entry, userspace overflowed the
1237 * submission (by quite a lot). Increment the overflow count in
1240 cqe = io_get_cqring(ctx);
1242 WRITE_ONCE(cqe->user_data, req->user_data);
1243 WRITE_ONCE(cqe->res, res);
1244 WRITE_ONCE(cqe->flags, cflags);
1245 } else if (ctx->cq_overflow_flushed) {
1246 WRITE_ONCE(ctx->rings->cq_overflow,
1247 atomic_inc_return(&ctx->cached_cq_overflow));
1249 if (list_empty(&ctx->cq_overflow_list)) {
1250 set_bit(0, &ctx->sq_check_overflow);
1251 set_bit(0, &ctx->cq_check_overflow);
1253 req->flags |= REQ_F_OVERFLOW;
1254 refcount_inc(&req->refs);
1256 req->cflags = cflags;
1257 list_add_tail(&req->list, &ctx->cq_overflow_list);
1261 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1263 __io_cqring_fill_event(req, res, 0);
1266 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1268 struct io_ring_ctx *ctx = req->ctx;
1269 unsigned long flags;
1271 spin_lock_irqsave(&ctx->completion_lock, flags);
1272 __io_cqring_fill_event(req, res, cflags);
1273 io_commit_cqring(ctx);
1274 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1276 io_cqring_ev_posted(ctx);
1279 static void io_cqring_add_event(struct io_kiocb *req, long res)
1281 __io_cqring_add_event(req, res, 0);
1284 static inline bool io_is_fallback_req(struct io_kiocb *req)
1286 return req == (struct io_kiocb *)
1287 ((unsigned long) req->ctx->fallback_req & ~1UL);
1290 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1292 struct io_kiocb *req;
1294 req = ctx->fallback_req;
1295 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1301 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1302 struct io_submit_state *state)
1304 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1305 struct io_kiocb *req;
1308 req = kmem_cache_alloc(req_cachep, gfp);
1311 } else if (!state->free_reqs) {
1315 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1316 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1319 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1320 * retry single alloc to be on the safe side.
1322 if (unlikely(ret <= 0)) {
1323 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1324 if (!state->reqs[0])
1328 state->free_reqs = ret - 1;
1329 req = state->reqs[ret - 1];
1332 req = state->reqs[state->free_reqs];
1337 return io_get_fallback_req(ctx);
1340 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1344 percpu_ref_put(req->fixed_file_refs);
1349 static void __io_req_aux_free(struct io_kiocb *req)
1351 if (req->flags & REQ_F_NEED_CLEANUP)
1352 io_cleanup_req(req);
1356 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1358 put_task_struct(req->task);
1360 io_req_work_drop_env(req);
1363 static void __io_free_req(struct io_kiocb *req)
1365 __io_req_aux_free(req);
1367 if (req->flags & REQ_F_INFLIGHT) {
1368 struct io_ring_ctx *ctx = req->ctx;
1369 unsigned long flags;
1371 spin_lock_irqsave(&ctx->inflight_lock, flags);
1372 list_del(&req->inflight_entry);
1373 if (waitqueue_active(&ctx->inflight_wait))
1374 wake_up(&ctx->inflight_wait);
1375 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1378 percpu_ref_put(&req->ctx->refs);
1379 if (likely(!io_is_fallback_req(req)))
1380 kmem_cache_free(req_cachep, req);
1382 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1386 void *reqs[IO_IOPOLL_BATCH];
1391 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1395 if (rb->need_iter) {
1396 int i, inflight = 0;
1397 unsigned long flags;
1399 for (i = 0; i < rb->to_free; i++) {
1400 struct io_kiocb *req = rb->reqs[i];
1402 if (req->flags & REQ_F_FIXED_FILE) {
1404 percpu_ref_put(req->fixed_file_refs);
1406 if (req->flags & REQ_F_INFLIGHT)
1408 __io_req_aux_free(req);
1413 spin_lock_irqsave(&ctx->inflight_lock, flags);
1414 for (i = 0; i < rb->to_free; i++) {
1415 struct io_kiocb *req = rb->reqs[i];
1417 if (req->flags & REQ_F_INFLIGHT) {
1418 list_del(&req->inflight_entry);
1423 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1425 if (waitqueue_active(&ctx->inflight_wait))
1426 wake_up(&ctx->inflight_wait);
1429 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1430 percpu_ref_put_many(&ctx->refs, rb->to_free);
1431 rb->to_free = rb->need_iter = 0;
1434 static bool io_link_cancel_timeout(struct io_kiocb *req)
1436 struct io_ring_ctx *ctx = req->ctx;
1439 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1441 io_cqring_fill_event(req, -ECANCELED);
1442 io_commit_cqring(ctx);
1443 req->flags &= ~REQ_F_LINK_HEAD;
1451 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1453 struct io_ring_ctx *ctx = req->ctx;
1454 bool wake_ev = false;
1456 /* Already got next link */
1457 if (req->flags & REQ_F_LINK_NEXT)
1461 * The list should never be empty when we are called here. But could
1462 * potentially happen if the chain is messed up, check to be on the
1465 while (!list_empty(&req->link_list)) {
1466 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1467 struct io_kiocb, link_list);
1469 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1470 (nxt->flags & REQ_F_TIMEOUT))) {
1471 list_del_init(&nxt->link_list);
1472 wake_ev |= io_link_cancel_timeout(nxt);
1473 req->flags &= ~REQ_F_LINK_TIMEOUT;
1477 list_del_init(&req->link_list);
1478 if (!list_empty(&nxt->link_list))
1479 nxt->flags |= REQ_F_LINK_HEAD;
1484 req->flags |= REQ_F_LINK_NEXT;
1486 io_cqring_ev_posted(ctx);
1490 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1492 static void io_fail_links(struct io_kiocb *req)
1494 struct io_ring_ctx *ctx = req->ctx;
1495 unsigned long flags;
1497 spin_lock_irqsave(&ctx->completion_lock, flags);
1499 while (!list_empty(&req->link_list)) {
1500 struct io_kiocb *link = list_first_entry(&req->link_list,
1501 struct io_kiocb, link_list);
1503 list_del_init(&link->link_list);
1504 trace_io_uring_fail_link(req, link);
1506 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1507 link->opcode == IORING_OP_LINK_TIMEOUT) {
1508 io_link_cancel_timeout(link);
1510 io_cqring_fill_event(link, -ECANCELED);
1511 __io_double_put_req(link);
1513 req->flags &= ~REQ_F_LINK_TIMEOUT;
1516 io_commit_cqring(ctx);
1517 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1518 io_cqring_ev_posted(ctx);
1521 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1523 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1527 * If LINK is set, we have dependent requests in this chain. If we
1528 * didn't fail this request, queue the first one up, moving any other
1529 * dependencies to the next request. In case of failure, fail the rest
1532 if (req->flags & REQ_F_FAIL_LINK) {
1534 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1535 REQ_F_LINK_TIMEOUT) {
1536 struct io_ring_ctx *ctx = req->ctx;
1537 unsigned long flags;
1540 * If this is a timeout link, we could be racing with the
1541 * timeout timer. Grab the completion lock for this case to
1542 * protect against that.
1544 spin_lock_irqsave(&ctx->completion_lock, flags);
1545 io_req_link_next(req, nxt);
1546 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1548 io_req_link_next(req, nxt);
1552 static void io_free_req(struct io_kiocb *req)
1554 struct io_kiocb *nxt = NULL;
1556 io_req_find_next(req, &nxt);
1560 io_queue_async_work(nxt);
1563 static void io_link_work_cb(struct io_wq_work **workptr)
1565 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1566 struct io_kiocb *link;
1568 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1569 io_queue_linked_timeout(link);
1570 io_wq_submit_work(workptr);
1573 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1575 struct io_kiocb *link;
1576 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1578 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1579 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1581 *workptr = &nxt->work;
1582 link = io_prep_linked_timeout(nxt);
1584 nxt->work.func = io_link_work_cb;
1588 * Drop reference to request, return next in chain (if there is one) if this
1589 * was the last reference to this request.
1591 __attribute__((nonnull))
1592 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1594 if (refcount_dec_and_test(&req->refs)) {
1595 io_req_find_next(req, nxtptr);
1600 static void io_put_req(struct io_kiocb *req)
1602 if (refcount_dec_and_test(&req->refs))
1606 static void io_steal_work(struct io_kiocb *req,
1607 struct io_wq_work **workptr)
1610 * It's in an io-wq worker, so there always should be at least
1611 * one reference, which will be dropped in io_put_work() just
1612 * after the current handler returns.
1614 * It also means, that if the counter dropped to 1, then there is
1615 * no asynchronous users left, so it's safe to steal the next work.
1617 if (refcount_read(&req->refs) == 1) {
1618 struct io_kiocb *nxt = NULL;
1620 io_req_find_next(req, &nxt);
1622 io_wq_assign_next(workptr, nxt);
1627 * Must only be used if we don't need to care about links, usually from
1628 * within the completion handling itself.
1630 static void __io_double_put_req(struct io_kiocb *req)
1632 /* drop both submit and complete references */
1633 if (refcount_sub_and_test(2, &req->refs))
1637 static void io_double_put_req(struct io_kiocb *req)
1639 /* drop both submit and complete references */
1640 if (refcount_sub_and_test(2, &req->refs))
1644 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1646 struct io_rings *rings = ctx->rings;
1648 if (test_bit(0, &ctx->cq_check_overflow)) {
1650 * noflush == true is from the waitqueue handler, just ensure
1651 * we wake up the task, and the next invocation will flush the
1652 * entries. We cannot safely to it from here.
1654 if (noflush && !list_empty(&ctx->cq_overflow_list))
1657 io_cqring_overflow_flush(ctx, false);
1660 /* See comment at the top of this file */
1662 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1665 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1667 struct io_rings *rings = ctx->rings;
1669 /* make sure SQ entry isn't read before tail */
1670 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1673 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1675 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1678 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1681 rb->reqs[rb->to_free++] = req;
1682 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1683 io_free_req_many(req->ctx, rb);
1687 static int io_put_kbuf(struct io_kiocb *req)
1689 struct io_buffer *kbuf;
1692 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1693 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1694 cflags |= IORING_CQE_F_BUFFER;
1701 * Find and free completed poll iocbs
1703 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1704 struct list_head *done)
1706 struct req_batch rb;
1707 struct io_kiocb *req;
1709 rb.to_free = rb.need_iter = 0;
1710 while (!list_empty(done)) {
1713 req = list_first_entry(done, struct io_kiocb, list);
1714 list_del(&req->list);
1716 if (req->flags & REQ_F_BUFFER_SELECTED)
1717 cflags = io_put_kbuf(req);
1719 __io_cqring_fill_event(req, req->result, cflags);
1722 if (refcount_dec_and_test(&req->refs) &&
1723 !io_req_multi_free(&rb, req))
1727 io_commit_cqring(ctx);
1728 if (ctx->flags & IORING_SETUP_SQPOLL)
1729 io_cqring_ev_posted(ctx);
1730 io_free_req_many(ctx, &rb);
1733 static void io_iopoll_queue(struct list_head *again)
1735 struct io_kiocb *req;
1738 req = list_first_entry(again, struct io_kiocb, list);
1739 list_del(&req->list);
1740 refcount_inc(&req->refs);
1741 io_queue_async_work(req);
1742 } while (!list_empty(again));
1745 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1748 struct io_kiocb *req, *tmp;
1755 * Only spin for completions if we don't have multiple devices hanging
1756 * off our complete list, and we're under the requested amount.
1758 spin = !ctx->poll_multi_file && *nr_events < min;
1761 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1762 struct kiocb *kiocb = &req->rw.kiocb;
1765 * Move completed and retryable entries to our local lists.
1766 * If we find a request that requires polling, break out
1767 * and complete those lists first, if we have entries there.
1769 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1770 list_move_tail(&req->list, &done);
1773 if (!list_empty(&done))
1776 if (req->result == -EAGAIN) {
1777 list_move_tail(&req->list, &again);
1780 if (!list_empty(&again))
1783 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1792 if (!list_empty(&done))
1793 io_iopoll_complete(ctx, nr_events, &done);
1795 if (!list_empty(&again))
1796 io_iopoll_queue(&again);
1802 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1803 * non-spinning poll check - we'll still enter the driver poll loop, but only
1804 * as a non-spinning completion check.
1806 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1809 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1812 ret = io_do_iopoll(ctx, nr_events, min);
1815 if (!min || *nr_events >= min)
1823 * We can't just wait for polled events to come to us, we have to actively
1824 * find and complete them.
1826 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1828 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1831 mutex_lock(&ctx->uring_lock);
1832 while (!list_empty(&ctx->poll_list)) {
1833 unsigned int nr_events = 0;
1835 io_iopoll_getevents(ctx, &nr_events, 1);
1838 * Ensure we allow local-to-the-cpu processing to take place,
1839 * in this case we need to ensure that we reap all events.
1843 mutex_unlock(&ctx->uring_lock);
1846 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1849 int iters = 0, ret = 0;
1852 * We disallow the app entering submit/complete with polling, but we
1853 * still need to lock the ring to prevent racing with polled issue
1854 * that got punted to a workqueue.
1856 mutex_lock(&ctx->uring_lock);
1861 * Don't enter poll loop if we already have events pending.
1862 * If we do, we can potentially be spinning for commands that
1863 * already triggered a CQE (eg in error).
1865 if (io_cqring_events(ctx, false))
1869 * If a submit got punted to a workqueue, we can have the
1870 * application entering polling for a command before it gets
1871 * issued. That app will hold the uring_lock for the duration
1872 * of the poll right here, so we need to take a breather every
1873 * now and then to ensure that the issue has a chance to add
1874 * the poll to the issued list. Otherwise we can spin here
1875 * forever, while the workqueue is stuck trying to acquire the
1878 if (!(++iters & 7)) {
1879 mutex_unlock(&ctx->uring_lock);
1880 mutex_lock(&ctx->uring_lock);
1883 if (*nr_events < min)
1884 tmin = min - *nr_events;
1886 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1890 } while (min && !*nr_events && !need_resched());
1892 mutex_unlock(&ctx->uring_lock);
1896 static void kiocb_end_write(struct io_kiocb *req)
1899 * Tell lockdep we inherited freeze protection from submission
1902 if (req->flags & REQ_F_ISREG) {
1903 struct inode *inode = file_inode(req->file);
1905 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1907 file_end_write(req->file);
1910 static inline void req_set_fail_links(struct io_kiocb *req)
1912 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1913 req->flags |= REQ_F_FAIL_LINK;
1916 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1918 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1921 if (kiocb->ki_flags & IOCB_WRITE)
1922 kiocb_end_write(req);
1924 if (res != req->result)
1925 req_set_fail_links(req);
1926 if (req->flags & REQ_F_BUFFER_SELECTED)
1927 cflags = io_put_kbuf(req);
1928 __io_cqring_add_event(req, res, cflags);
1931 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1933 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1935 io_complete_rw_common(kiocb, res);
1939 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1941 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1943 if (kiocb->ki_flags & IOCB_WRITE)
1944 kiocb_end_write(req);
1946 if (res != req->result)
1947 req_set_fail_links(req);
1950 req->flags |= REQ_F_IOPOLL_COMPLETED;
1954 * After the iocb has been issued, it's safe to be found on the poll list.
1955 * Adding the kiocb to the list AFTER submission ensures that we don't
1956 * find it from a io_iopoll_getevents() thread before the issuer is done
1957 * accessing the kiocb cookie.
1959 static void io_iopoll_req_issued(struct io_kiocb *req)
1961 struct io_ring_ctx *ctx = req->ctx;
1964 * Track whether we have multiple files in our lists. This will impact
1965 * how we do polling eventually, not spinning if we're on potentially
1966 * different devices.
1968 if (list_empty(&ctx->poll_list)) {
1969 ctx->poll_multi_file = false;
1970 } else if (!ctx->poll_multi_file) {
1971 struct io_kiocb *list_req;
1973 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1975 if (list_req->file != req->file)
1976 ctx->poll_multi_file = true;
1980 * For fast devices, IO may have already completed. If it has, add
1981 * it to the front so we find it first.
1983 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1984 list_add(&req->list, &ctx->poll_list);
1986 list_add_tail(&req->list, &ctx->poll_list);
1988 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1989 wq_has_sleeper(&ctx->sqo_wait))
1990 wake_up(&ctx->sqo_wait);
1993 static void io_file_put(struct io_submit_state *state)
1996 int diff = state->has_refs - state->used_refs;
1999 fput_many(state->file, diff);
2005 * Get as many references to a file as we have IOs left in this submission,
2006 * assuming most submissions are for one file, or at least that each file
2007 * has more than one submission.
2009 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2015 if (state->fd == fd) {
2022 state->file = fget_many(fd, state->ios_left);
2027 state->has_refs = state->ios_left;
2028 state->used_refs = 1;
2034 * If we tracked the file through the SCM inflight mechanism, we could support
2035 * any file. For now, just ensure that anything potentially problematic is done
2038 static bool io_file_supports_async(struct file *file)
2040 umode_t mode = file_inode(file)->i_mode;
2042 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2044 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2050 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2051 bool force_nonblock)
2053 struct io_ring_ctx *ctx = req->ctx;
2054 struct kiocb *kiocb = &req->rw.kiocb;
2058 if (S_ISREG(file_inode(req->file)->i_mode))
2059 req->flags |= REQ_F_ISREG;
2061 kiocb->ki_pos = READ_ONCE(sqe->off);
2062 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2063 req->flags |= REQ_F_CUR_POS;
2064 kiocb->ki_pos = req->file->f_pos;
2066 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2067 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2068 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2072 ioprio = READ_ONCE(sqe->ioprio);
2074 ret = ioprio_check_cap(ioprio);
2078 kiocb->ki_ioprio = ioprio;
2080 kiocb->ki_ioprio = get_current_ioprio();
2082 /* don't allow async punt if RWF_NOWAIT was requested */
2083 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2084 (req->file->f_flags & O_NONBLOCK))
2085 req->flags |= REQ_F_NOWAIT;
2088 kiocb->ki_flags |= IOCB_NOWAIT;
2090 if (ctx->flags & IORING_SETUP_IOPOLL) {
2091 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2092 !kiocb->ki_filp->f_op->iopoll)
2095 kiocb->ki_flags |= IOCB_HIPRI;
2096 kiocb->ki_complete = io_complete_rw_iopoll;
2099 if (kiocb->ki_flags & IOCB_HIPRI)
2101 kiocb->ki_complete = io_complete_rw;
2104 req->rw.addr = READ_ONCE(sqe->addr);
2105 req->rw.len = READ_ONCE(sqe->len);
2106 /* we own ->private, reuse it for the buffer index / buffer ID */
2107 req->rw.kiocb.private = (void *) (unsigned long)
2108 READ_ONCE(sqe->buf_index);
2112 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2118 case -ERESTARTNOINTR:
2119 case -ERESTARTNOHAND:
2120 case -ERESTART_RESTARTBLOCK:
2122 * We can't just restart the syscall, since previously
2123 * submitted sqes may already be in progress. Just fail this
2129 kiocb->ki_complete(kiocb, ret, 0);
2133 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2135 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2137 if (req->flags & REQ_F_CUR_POS)
2138 req->file->f_pos = kiocb->ki_pos;
2139 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2140 io_complete_rw(kiocb, ret, 0);
2142 io_rw_done(kiocb, ret);
2145 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2146 struct iov_iter *iter)
2148 struct io_ring_ctx *ctx = req->ctx;
2149 size_t len = req->rw.len;
2150 struct io_mapped_ubuf *imu;
2151 unsigned index, buf_index;
2155 /* attempt to use fixed buffers without having provided iovecs */
2156 if (unlikely(!ctx->user_bufs))
2159 buf_index = (unsigned long) req->rw.kiocb.private;
2160 if (unlikely(buf_index >= ctx->nr_user_bufs))
2163 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2164 imu = &ctx->user_bufs[index];
2165 buf_addr = req->rw.addr;
2168 if (buf_addr + len < buf_addr)
2170 /* not inside the mapped region */
2171 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2175 * May not be a start of buffer, set size appropriately
2176 * and advance us to the beginning.
2178 offset = buf_addr - imu->ubuf;
2179 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2183 * Don't use iov_iter_advance() here, as it's really slow for
2184 * using the latter parts of a big fixed buffer - it iterates
2185 * over each segment manually. We can cheat a bit here, because
2188 * 1) it's a BVEC iter, we set it up
2189 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2190 * first and last bvec
2192 * So just find our index, and adjust the iterator afterwards.
2193 * If the offset is within the first bvec (or the whole first
2194 * bvec, just use iov_iter_advance(). This makes it easier
2195 * since we can just skip the first segment, which may not
2196 * be PAGE_SIZE aligned.
2198 const struct bio_vec *bvec = imu->bvec;
2200 if (offset <= bvec->bv_len) {
2201 iov_iter_advance(iter, offset);
2203 unsigned long seg_skip;
2205 /* skip first vec */
2206 offset -= bvec->bv_len;
2207 seg_skip = 1 + (offset >> PAGE_SHIFT);
2209 iter->bvec = bvec + seg_skip;
2210 iter->nr_segs -= seg_skip;
2211 iter->count -= bvec->bv_len + offset;
2212 iter->iov_offset = offset & ~PAGE_MASK;
2219 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2222 mutex_unlock(&ctx->uring_lock);
2225 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2228 * "Normal" inline submissions always hold the uring_lock, since we
2229 * grab it from the system call. Same is true for the SQPOLL offload.
2230 * The only exception is when we've detached the request and issue it
2231 * from an async worker thread, grab the lock for that case.
2234 mutex_lock(&ctx->uring_lock);
2237 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2238 int bgid, struct io_buffer *kbuf,
2241 struct io_buffer *head;
2243 if (req->flags & REQ_F_BUFFER_SELECTED)
2246 io_ring_submit_lock(req->ctx, needs_lock);
2248 lockdep_assert_held(&req->ctx->uring_lock);
2250 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2252 if (!list_empty(&head->list)) {
2253 kbuf = list_last_entry(&head->list, struct io_buffer,
2255 list_del(&kbuf->list);
2258 idr_remove(&req->ctx->io_buffer_idr, bgid);
2260 if (*len > kbuf->len)
2263 kbuf = ERR_PTR(-ENOBUFS);
2266 io_ring_submit_unlock(req->ctx, needs_lock);
2271 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2274 struct io_buffer *kbuf;
2277 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2278 bgid = (int) (unsigned long) req->rw.kiocb.private;
2279 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2282 req->rw.addr = (u64) (unsigned long) kbuf;
2283 req->flags |= REQ_F_BUFFER_SELECTED;
2284 return u64_to_user_ptr(kbuf->addr);
2287 #ifdef CONFIG_COMPAT
2288 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2291 struct compat_iovec __user *uiov;
2292 compat_ssize_t clen;
2296 uiov = u64_to_user_ptr(req->rw.addr);
2297 if (!access_ok(uiov, sizeof(*uiov)))
2299 if (__get_user(clen, &uiov->iov_len))
2305 buf = io_rw_buffer_select(req, &len, needs_lock);
2307 return PTR_ERR(buf);
2308 iov[0].iov_base = buf;
2309 iov[0].iov_len = (compat_size_t) len;
2314 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2317 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2321 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2324 len = iov[0].iov_len;
2327 buf = io_rw_buffer_select(req, &len, needs_lock);
2329 return PTR_ERR(buf);
2330 iov[0].iov_base = buf;
2331 iov[0].iov_len = len;
2335 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2338 if (req->flags & REQ_F_BUFFER_SELECTED)
2342 else if (req->rw.len > 1)
2345 #ifdef CONFIG_COMPAT
2346 if (req->ctx->compat)
2347 return io_compat_import(req, iov, needs_lock);
2350 return __io_iov_buffer_select(req, iov, needs_lock);
2353 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2354 struct iovec **iovec, struct iov_iter *iter,
2357 void __user *buf = u64_to_user_ptr(req->rw.addr);
2358 size_t sqe_len = req->rw.len;
2362 opcode = req->opcode;
2363 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2365 return io_import_fixed(req, rw, iter);
2368 /* buffer index only valid with fixed read/write, or buffer select */
2369 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2372 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2373 if (req->flags & REQ_F_BUFFER_SELECT) {
2374 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2377 return PTR_ERR(buf);
2379 req->rw.len = sqe_len;
2382 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2384 return ret < 0 ? ret : sqe_len;
2388 struct io_async_rw *iorw = &req->io->rw;
2391 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2392 if (iorw->iov == iorw->fast_iov)
2397 if (req->flags & REQ_F_BUFFER_SELECT) {
2398 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2400 ret = (*iovec)->iov_len;
2401 iov_iter_init(iter, rw, *iovec, 1, ret);
2407 #ifdef CONFIG_COMPAT
2408 if (req->ctx->compat)
2409 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2413 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2417 * For files that don't have ->read_iter() and ->write_iter(), handle them
2418 * by looping over ->read() or ->write() manually.
2420 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2421 struct iov_iter *iter)
2426 * Don't support polled IO through this interface, and we can't
2427 * support non-blocking either. For the latter, this just causes
2428 * the kiocb to be handled from an async context.
2430 if (kiocb->ki_flags & IOCB_HIPRI)
2432 if (kiocb->ki_flags & IOCB_NOWAIT)
2435 while (iov_iter_count(iter)) {
2439 if (!iov_iter_is_bvec(iter)) {
2440 iovec = iov_iter_iovec(iter);
2442 /* fixed buffers import bvec */
2443 iovec.iov_base = kmap(iter->bvec->bv_page)
2445 iovec.iov_len = min(iter->count,
2446 iter->bvec->bv_len - iter->iov_offset);
2450 nr = file->f_op->read(file, iovec.iov_base,
2451 iovec.iov_len, &kiocb->ki_pos);
2453 nr = file->f_op->write(file, iovec.iov_base,
2454 iovec.iov_len, &kiocb->ki_pos);
2457 if (iov_iter_is_bvec(iter))
2458 kunmap(iter->bvec->bv_page);
2466 if (nr != iovec.iov_len)
2468 iov_iter_advance(iter, nr);
2474 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2475 struct iovec *iovec, struct iovec *fast_iov,
2476 struct iov_iter *iter)
2478 req->io->rw.nr_segs = iter->nr_segs;
2479 req->io->rw.size = io_size;
2480 req->io->rw.iov = iovec;
2481 if (!req->io->rw.iov) {
2482 req->io->rw.iov = req->io->rw.fast_iov;
2483 if (req->io->rw.iov != fast_iov)
2484 memcpy(req->io->rw.iov, fast_iov,
2485 sizeof(struct iovec) * iter->nr_segs);
2487 req->flags |= REQ_F_NEED_CLEANUP;
2491 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2493 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2494 return req->io == NULL;
2497 static int io_alloc_async_ctx(struct io_kiocb *req)
2499 if (!io_op_defs[req->opcode].async_ctx)
2502 return __io_alloc_async_ctx(req);
2505 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2506 struct iovec *iovec, struct iovec *fast_iov,
2507 struct iov_iter *iter)
2509 if (!io_op_defs[req->opcode].async_ctx)
2512 if (__io_alloc_async_ctx(req))
2515 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2520 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2521 bool force_nonblock)
2523 struct io_async_ctx *io;
2524 struct iov_iter iter;
2527 ret = io_prep_rw(req, sqe, force_nonblock);
2531 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2534 /* either don't need iovec imported or already have it */
2535 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2539 io->rw.iov = io->rw.fast_iov;
2541 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2546 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2550 static int io_read(struct io_kiocb *req, bool force_nonblock)
2552 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2553 struct kiocb *kiocb = &req->rw.kiocb;
2554 struct iov_iter iter;
2556 ssize_t io_size, ret;
2558 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2562 /* Ensure we clear previously set non-block flag */
2563 if (!force_nonblock)
2564 kiocb->ki_flags &= ~IOCB_NOWAIT;
2568 if (req->flags & REQ_F_LINK_HEAD)
2569 req->result = io_size;
2572 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2573 * we know to async punt it even if it was opened O_NONBLOCK
2575 if (force_nonblock && !io_file_supports_async(req->file))
2578 iov_count = iov_iter_count(&iter);
2579 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2583 if (req->file->f_op->read_iter)
2584 ret2 = call_read_iter(req->file, kiocb, &iter);
2586 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2588 /* Catch -EAGAIN return for forced non-blocking submission */
2589 if (!force_nonblock || ret2 != -EAGAIN) {
2590 kiocb_done(kiocb, ret2);
2593 ret = io_setup_async_rw(req, io_size, iovec,
2594 inline_vecs, &iter);
2597 /* any defer here is final, must blocking retry */
2598 if (!(req->flags & REQ_F_NOWAIT))
2599 req->flags |= REQ_F_MUST_PUNT;
2605 req->flags &= ~REQ_F_NEED_CLEANUP;
2609 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2610 bool force_nonblock)
2612 struct io_async_ctx *io;
2613 struct iov_iter iter;
2616 ret = io_prep_rw(req, sqe, force_nonblock);
2620 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2623 req->fsize = rlimit(RLIMIT_FSIZE);
2625 /* either don't need iovec imported or already have it */
2626 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2630 io->rw.iov = io->rw.fast_iov;
2632 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2637 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2641 static int io_write(struct io_kiocb *req, bool force_nonblock)
2643 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2644 struct kiocb *kiocb = &req->rw.kiocb;
2645 struct iov_iter iter;
2647 ssize_t ret, io_size;
2649 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2653 /* Ensure we clear previously set non-block flag */
2654 if (!force_nonblock)
2655 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2659 if (req->flags & REQ_F_LINK_HEAD)
2660 req->result = io_size;
2663 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2664 * we know to async punt it even if it was opened O_NONBLOCK
2666 if (force_nonblock && !io_file_supports_async(req->file))
2669 /* file path doesn't support NOWAIT for non-direct_IO */
2670 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2671 (req->flags & REQ_F_ISREG))
2674 iov_count = iov_iter_count(&iter);
2675 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2680 * Open-code file_start_write here to grab freeze protection,
2681 * which will be released by another thread in
2682 * io_complete_rw(). Fool lockdep by telling it the lock got
2683 * released so that it doesn't complain about the held lock when
2684 * we return to userspace.
2686 if (req->flags & REQ_F_ISREG) {
2687 __sb_start_write(file_inode(req->file)->i_sb,
2688 SB_FREEZE_WRITE, true);
2689 __sb_writers_release(file_inode(req->file)->i_sb,
2692 kiocb->ki_flags |= IOCB_WRITE;
2694 if (!force_nonblock)
2695 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2697 if (req->file->f_op->write_iter)
2698 ret2 = call_write_iter(req->file, kiocb, &iter);
2700 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2702 if (!force_nonblock)
2703 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2706 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2707 * retry them without IOCB_NOWAIT.
2709 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2711 if (!force_nonblock || ret2 != -EAGAIN) {
2712 kiocb_done(kiocb, ret2);
2715 ret = io_setup_async_rw(req, io_size, iovec,
2716 inline_vecs, &iter);
2719 /* any defer here is final, must blocking retry */
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 bool io_splice_punt(struct file *file)
2762 if (get_pipe_info(file))
2764 if (!io_file_supports_async(file))
2766 return !(file->f_mode & O_NONBLOCK);
2769 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2771 struct io_splice *sp = &req->splice;
2772 struct file *in = sp->file_in;
2773 struct file *out = sp->file_out;
2774 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2775 loff_t *poff_in, *poff_out;
2778 if (force_nonblock) {
2779 if (io_splice_punt(in) || io_splice_punt(out))
2781 flags |= SPLICE_F_NONBLOCK;
2784 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2785 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2786 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2787 if (force_nonblock && ret == -EAGAIN)
2790 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2791 req->flags &= ~REQ_F_NEED_CLEANUP;
2793 io_cqring_add_event(req, ret);
2795 req_set_fail_links(req);
2801 * IORING_OP_NOP just posts a completion event, nothing else.
2803 static int io_nop(struct io_kiocb *req)
2805 struct io_ring_ctx *ctx = req->ctx;
2807 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2810 io_cqring_add_event(req, 0);
2815 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2817 struct io_ring_ctx *ctx = req->ctx;
2822 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2824 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2827 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2828 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2831 req->sync.off = READ_ONCE(sqe->off);
2832 req->sync.len = READ_ONCE(sqe->len);
2836 static bool io_req_cancelled(struct io_kiocb *req)
2838 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2839 req_set_fail_links(req);
2840 io_cqring_add_event(req, -ECANCELED);
2848 static void __io_fsync(struct io_kiocb *req)
2850 loff_t end = req->sync.off + req->sync.len;
2853 ret = vfs_fsync_range(req->file, req->sync.off,
2854 end > 0 ? end : LLONG_MAX,
2855 req->sync.flags & IORING_FSYNC_DATASYNC);
2857 req_set_fail_links(req);
2858 io_cqring_add_event(req, ret);
2862 static void io_fsync_finish(struct io_wq_work **workptr)
2864 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2866 if (io_req_cancelled(req))
2869 io_steal_work(req, workptr);
2872 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2874 /* fsync always requires a blocking context */
2875 if (force_nonblock) {
2876 req->work.func = io_fsync_finish;
2883 static void __io_fallocate(struct io_kiocb *req)
2887 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2888 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2890 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2892 req_set_fail_links(req);
2893 io_cqring_add_event(req, ret);
2897 static void io_fallocate_finish(struct io_wq_work **workptr)
2899 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2901 if (io_req_cancelled(req))
2903 __io_fallocate(req);
2904 io_steal_work(req, workptr);
2907 static int io_fallocate_prep(struct io_kiocb *req,
2908 const struct io_uring_sqe *sqe)
2910 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2913 req->sync.off = READ_ONCE(sqe->off);
2914 req->sync.len = READ_ONCE(sqe->addr);
2915 req->sync.mode = READ_ONCE(sqe->len);
2916 req->fsize = rlimit(RLIMIT_FSIZE);
2920 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2922 /* fallocate always requiring blocking context */
2923 if (force_nonblock) {
2924 req->work.func = io_fallocate_finish;
2928 __io_fallocate(req);
2932 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2934 const char __user *fname;
2937 if (sqe->ioprio || sqe->buf_index)
2939 if (req->flags & REQ_F_FIXED_FILE)
2941 if (req->flags & REQ_F_NEED_CLEANUP)
2944 req->open.dfd = READ_ONCE(sqe->fd);
2945 req->open.how.mode = READ_ONCE(sqe->len);
2946 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2947 req->open.how.flags = READ_ONCE(sqe->open_flags);
2948 if (force_o_largefile())
2949 req->open.how.flags |= O_LARGEFILE;
2951 req->open.filename = getname(fname);
2952 if (IS_ERR(req->open.filename)) {
2953 ret = PTR_ERR(req->open.filename);
2954 req->open.filename = NULL;
2958 req->open.nofile = rlimit(RLIMIT_NOFILE);
2959 req->flags |= REQ_F_NEED_CLEANUP;
2963 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2965 struct open_how __user *how;
2966 const char __user *fname;
2970 if (sqe->ioprio || sqe->buf_index)
2972 if (req->flags & REQ_F_FIXED_FILE)
2974 if (req->flags & REQ_F_NEED_CLEANUP)
2977 req->open.dfd = READ_ONCE(sqe->fd);
2978 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2979 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2980 len = READ_ONCE(sqe->len);
2982 if (len < OPEN_HOW_SIZE_VER0)
2985 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2990 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2991 req->open.how.flags |= O_LARGEFILE;
2993 req->open.filename = getname(fname);
2994 if (IS_ERR(req->open.filename)) {
2995 ret = PTR_ERR(req->open.filename);
2996 req->open.filename = NULL;
3000 req->open.nofile = rlimit(RLIMIT_NOFILE);
3001 req->flags |= REQ_F_NEED_CLEANUP;
3005 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3007 struct open_flags op;
3014 ret = build_open_flags(&req->open.how, &op);
3018 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3022 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3025 ret = PTR_ERR(file);
3027 fsnotify_open(file);
3028 fd_install(ret, file);
3031 putname(req->open.filename);
3032 req->flags &= ~REQ_F_NEED_CLEANUP;
3034 req_set_fail_links(req);
3035 io_cqring_add_event(req, ret);
3040 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3042 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3043 return io_openat2(req, force_nonblock);
3046 static int io_remove_buffers_prep(struct io_kiocb *req,
3047 const struct io_uring_sqe *sqe)
3049 struct io_provide_buf *p = &req->pbuf;
3052 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3055 tmp = READ_ONCE(sqe->fd);
3056 if (!tmp || tmp > USHRT_MAX)
3059 memset(p, 0, sizeof(*p));
3061 p->bgid = READ_ONCE(sqe->buf_group);
3065 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3066 int bgid, unsigned nbufs)
3070 /* shouldn't happen */
3074 /* the head kbuf is the list itself */
3075 while (!list_empty(&buf->list)) {
3076 struct io_buffer *nxt;
3078 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3079 list_del(&nxt->list);
3086 idr_remove(&ctx->io_buffer_idr, bgid);
3091 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3093 struct io_provide_buf *p = &req->pbuf;
3094 struct io_ring_ctx *ctx = req->ctx;
3095 struct io_buffer *head;
3098 io_ring_submit_lock(ctx, !force_nonblock);
3100 lockdep_assert_held(&ctx->uring_lock);
3103 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3105 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3107 io_ring_submit_lock(ctx, !force_nonblock);
3109 req_set_fail_links(req);
3110 io_cqring_add_event(req, ret);
3115 static int io_provide_buffers_prep(struct io_kiocb *req,
3116 const struct io_uring_sqe *sqe)
3118 struct io_provide_buf *p = &req->pbuf;
3121 if (sqe->ioprio || sqe->rw_flags)
3124 tmp = READ_ONCE(sqe->fd);
3125 if (!tmp || tmp > USHRT_MAX)
3128 p->addr = READ_ONCE(sqe->addr);
3129 p->len = READ_ONCE(sqe->len);
3131 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3134 p->bgid = READ_ONCE(sqe->buf_group);
3135 tmp = READ_ONCE(sqe->off);
3136 if (tmp > USHRT_MAX)
3142 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3144 struct io_buffer *buf;
3145 u64 addr = pbuf->addr;
3146 int i, bid = pbuf->bid;
3148 for (i = 0; i < pbuf->nbufs; i++) {
3149 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3154 buf->len = pbuf->len;
3159 INIT_LIST_HEAD(&buf->list);
3162 list_add_tail(&buf->list, &(*head)->list);
3166 return i ? i : -ENOMEM;
3169 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3171 struct io_provide_buf *p = &req->pbuf;
3172 struct io_ring_ctx *ctx = req->ctx;
3173 struct io_buffer *head, *list;
3176 io_ring_submit_lock(ctx, !force_nonblock);
3178 lockdep_assert_held(&ctx->uring_lock);
3180 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3182 ret = io_add_buffers(p, &head);
3187 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3190 __io_remove_buffers(ctx, head, p->bgid, -1U);
3195 io_ring_submit_unlock(ctx, !force_nonblock);
3197 req_set_fail_links(req);
3198 io_cqring_add_event(req, ret);
3203 static int io_epoll_ctl_prep(struct io_kiocb *req,
3204 const struct io_uring_sqe *sqe)
3206 #if defined(CONFIG_EPOLL)
3207 if (sqe->ioprio || sqe->buf_index)
3210 req->epoll.epfd = READ_ONCE(sqe->fd);
3211 req->epoll.op = READ_ONCE(sqe->len);
3212 req->epoll.fd = READ_ONCE(sqe->off);
3214 if (ep_op_has_event(req->epoll.op)) {
3215 struct epoll_event __user *ev;
3217 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3218 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3228 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3230 #if defined(CONFIG_EPOLL)
3231 struct io_epoll *ie = &req->epoll;
3234 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3235 if (force_nonblock && ret == -EAGAIN)
3239 req_set_fail_links(req);
3240 io_cqring_add_event(req, ret);
3248 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3250 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3251 if (sqe->ioprio || sqe->buf_index || sqe->off)
3254 req->madvise.addr = READ_ONCE(sqe->addr);
3255 req->madvise.len = READ_ONCE(sqe->len);
3256 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3263 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3265 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3266 struct io_madvise *ma = &req->madvise;
3272 ret = do_madvise(ma->addr, ma->len, ma->advice);
3274 req_set_fail_links(req);
3275 io_cqring_add_event(req, ret);
3283 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3285 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3288 req->fadvise.offset = READ_ONCE(sqe->off);
3289 req->fadvise.len = READ_ONCE(sqe->len);
3290 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3294 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3296 struct io_fadvise *fa = &req->fadvise;
3299 if (force_nonblock) {
3300 switch (fa->advice) {
3301 case POSIX_FADV_NORMAL:
3302 case POSIX_FADV_RANDOM:
3303 case POSIX_FADV_SEQUENTIAL:
3310 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3312 req_set_fail_links(req);
3313 io_cqring_add_event(req, ret);
3318 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3320 const char __user *fname;
3321 unsigned lookup_flags;
3324 if (sqe->ioprio || sqe->buf_index)
3326 if (req->flags & REQ_F_FIXED_FILE)
3328 if (req->flags & REQ_F_NEED_CLEANUP)
3331 req->open.dfd = READ_ONCE(sqe->fd);
3332 req->open.mask = READ_ONCE(sqe->len);
3333 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3334 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3335 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3337 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3340 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3341 if (IS_ERR(req->open.filename)) {
3342 ret = PTR_ERR(req->open.filename);
3343 req->open.filename = NULL;
3347 req->flags |= REQ_F_NEED_CLEANUP;
3351 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3353 struct io_open *ctx = &req->open;
3354 unsigned lookup_flags;
3362 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3366 /* filename_lookup() drops it, keep a reference */
3367 ctx->filename->refcnt++;
3369 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3374 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3376 if (retry_estale(ret, lookup_flags)) {
3377 lookup_flags |= LOOKUP_REVAL;
3381 ret = cp_statx(&stat, ctx->buffer);
3383 putname(ctx->filename);
3384 req->flags &= ~REQ_F_NEED_CLEANUP;
3386 req_set_fail_links(req);
3387 io_cqring_add_event(req, ret);
3392 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3395 * If we queue this for async, it must not be cancellable. That would
3396 * leave the 'file' in an undeterminate state.
3398 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3400 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3401 sqe->rw_flags || sqe->buf_index)
3403 if (req->flags & REQ_F_FIXED_FILE)
3406 req->close.fd = READ_ONCE(sqe->fd);
3407 if (req->file->f_op == &io_uring_fops ||
3408 req->close.fd == req->ctx->ring_fd)
3414 /* only called when __close_fd_get_file() is done */
3415 static void __io_close_finish(struct io_kiocb *req)
3419 ret = filp_close(req->close.put_file, req->work.files);
3421 req_set_fail_links(req);
3422 io_cqring_add_event(req, ret);
3423 fput(req->close.put_file);
3427 static void io_close_finish(struct io_wq_work **workptr)
3429 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3431 /* not cancellable, don't do io_req_cancelled() */
3432 __io_close_finish(req);
3433 io_steal_work(req, workptr);
3436 static int io_close(struct io_kiocb *req, bool force_nonblock)
3440 req->close.put_file = NULL;
3441 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3445 /* if the file has a flush method, be safe and punt to async */
3446 if (req->close.put_file->f_op->flush && force_nonblock) {
3447 /* submission ref will be dropped, take it for async */
3448 refcount_inc(&req->refs);
3450 req->work.func = io_close_finish;
3452 * Do manual async queue here to avoid grabbing files - we don't
3453 * need the files, and it'll cause io_close_finish() to close
3454 * the file again and cause a double CQE entry for this request
3456 io_queue_async_work(req);
3461 * No ->flush(), safely close from here and just punt the
3462 * fput() to async context.
3464 __io_close_finish(req);
3468 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3470 struct io_ring_ctx *ctx = req->ctx;
3475 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3477 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3480 req->sync.off = READ_ONCE(sqe->off);
3481 req->sync.len = READ_ONCE(sqe->len);
3482 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3486 static void __io_sync_file_range(struct io_kiocb *req)
3490 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3493 req_set_fail_links(req);
3494 io_cqring_add_event(req, ret);
3499 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3501 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3503 if (io_req_cancelled(req))
3505 __io_sync_file_range(req);
3506 io_put_req(req); /* put submission ref */
3509 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3511 /* sync_file_range always requires a blocking context */
3512 if (force_nonblock) {
3513 req->work.func = io_sync_file_range_finish;
3517 __io_sync_file_range(req);
3521 #if defined(CONFIG_NET)
3522 static int io_setup_async_msg(struct io_kiocb *req,
3523 struct io_async_msghdr *kmsg)
3527 if (io_alloc_async_ctx(req)) {
3528 if (kmsg->iov != kmsg->fast_iov)
3532 req->flags |= REQ_F_NEED_CLEANUP;
3533 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3537 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3539 struct io_sr_msg *sr = &req->sr_msg;
3540 struct io_async_ctx *io = req->io;
3543 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3544 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3545 sr->len = READ_ONCE(sqe->len);
3547 #ifdef CONFIG_COMPAT
3548 if (req->ctx->compat)
3549 sr->msg_flags |= MSG_CMSG_COMPAT;
3552 if (!io || req->opcode == IORING_OP_SEND)
3554 /* iovec is already imported */
3555 if (req->flags & REQ_F_NEED_CLEANUP)
3558 io->msg.iov = io->msg.fast_iov;
3559 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3562 req->flags |= REQ_F_NEED_CLEANUP;
3566 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3568 struct io_async_msghdr *kmsg = NULL;
3569 struct socket *sock;
3572 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3575 sock = sock_from_file(req->file, &ret);
3577 struct io_async_ctx io;
3581 kmsg = &req->io->msg;
3582 kmsg->msg.msg_name = &req->io->msg.addr;
3583 /* if iov is set, it's allocated already */
3585 kmsg->iov = kmsg->fast_iov;
3586 kmsg->msg.msg_iter.iov = kmsg->iov;
3588 struct io_sr_msg *sr = &req->sr_msg;
3591 kmsg->msg.msg_name = &io.msg.addr;
3593 io.msg.iov = io.msg.fast_iov;
3594 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3595 sr->msg_flags, &io.msg.iov);
3600 flags = req->sr_msg.msg_flags;
3601 if (flags & MSG_DONTWAIT)
3602 req->flags |= REQ_F_NOWAIT;
3603 else if (force_nonblock)
3604 flags |= MSG_DONTWAIT;
3606 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3607 if (force_nonblock && ret == -EAGAIN)
3608 return io_setup_async_msg(req, kmsg);
3609 if (ret == -ERESTARTSYS)
3613 if (kmsg && kmsg->iov != kmsg->fast_iov)
3615 req->flags &= ~REQ_F_NEED_CLEANUP;
3616 io_cqring_add_event(req, ret);
3618 req_set_fail_links(req);
3623 static int io_send(struct io_kiocb *req, bool force_nonblock)
3625 struct socket *sock;
3628 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3631 sock = sock_from_file(req->file, &ret);
3633 struct io_sr_msg *sr = &req->sr_msg;
3638 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3643 msg.msg_name = NULL;
3644 msg.msg_control = NULL;
3645 msg.msg_controllen = 0;
3646 msg.msg_namelen = 0;
3648 flags = req->sr_msg.msg_flags;
3649 if (flags & MSG_DONTWAIT)
3650 req->flags |= REQ_F_NOWAIT;
3651 else if (force_nonblock)
3652 flags |= MSG_DONTWAIT;
3654 msg.msg_flags = flags;
3655 ret = sock_sendmsg(sock, &msg);
3656 if (force_nonblock && ret == -EAGAIN)
3658 if (ret == -ERESTARTSYS)
3662 io_cqring_add_event(req, ret);
3664 req_set_fail_links(req);
3669 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3671 struct io_sr_msg *sr = &req->sr_msg;
3672 struct iovec __user *uiov;
3676 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3681 if (req->flags & REQ_F_BUFFER_SELECT) {
3684 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3686 sr->len = io->msg.iov[0].iov_len;
3687 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3691 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3692 &io->msg.iov, &io->msg.msg.msg_iter);
3700 #ifdef CONFIG_COMPAT
3701 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3702 struct io_async_ctx *io)
3704 struct compat_msghdr __user *msg_compat;
3705 struct io_sr_msg *sr = &req->sr_msg;
3706 struct compat_iovec __user *uiov;
3711 msg_compat = (struct compat_msghdr __user *) sr->msg;
3712 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3717 uiov = compat_ptr(ptr);
3718 if (req->flags & REQ_F_BUFFER_SELECT) {
3719 compat_ssize_t clen;
3723 if (!access_ok(uiov, sizeof(*uiov)))
3725 if (__get_user(clen, &uiov->iov_len))
3729 sr->len = io->msg.iov[0].iov_len;
3732 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3734 &io->msg.msg.msg_iter);
3743 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3745 io->msg.iov = io->msg.fast_iov;
3747 #ifdef CONFIG_COMPAT
3748 if (req->ctx->compat)
3749 return __io_compat_recvmsg_copy_hdr(req, io);
3752 return __io_recvmsg_copy_hdr(req, io);
3755 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3756 int *cflags, bool needs_lock)
3758 struct io_sr_msg *sr = &req->sr_msg;
3759 struct io_buffer *kbuf;
3761 if (!(req->flags & REQ_F_BUFFER_SELECT))
3764 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3769 req->flags |= REQ_F_BUFFER_SELECTED;
3771 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3772 *cflags |= IORING_CQE_F_BUFFER;
3776 static int io_recvmsg_prep(struct io_kiocb *req,
3777 const struct io_uring_sqe *sqe)
3779 struct io_sr_msg *sr = &req->sr_msg;
3780 struct io_async_ctx *io = req->io;
3783 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3784 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3785 sr->len = READ_ONCE(sqe->len);
3786 sr->bgid = READ_ONCE(sqe->buf_group);
3788 #ifdef CONFIG_COMPAT
3789 if (req->ctx->compat)
3790 sr->msg_flags |= MSG_CMSG_COMPAT;
3793 if (!io || req->opcode == IORING_OP_RECV)
3795 /* iovec is already imported */
3796 if (req->flags & REQ_F_NEED_CLEANUP)
3799 ret = io_recvmsg_copy_hdr(req, io);
3801 req->flags |= REQ_F_NEED_CLEANUP;
3805 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3807 struct io_async_msghdr *kmsg = NULL;
3808 struct socket *sock;
3809 int ret, cflags = 0;
3811 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3814 sock = sock_from_file(req->file, &ret);
3816 struct io_buffer *kbuf;
3817 struct io_async_ctx io;
3821 kmsg = &req->io->msg;
3822 kmsg->msg.msg_name = &req->io->msg.addr;
3823 /* if iov is set, it's allocated already */
3825 kmsg->iov = kmsg->fast_iov;
3826 kmsg->msg.msg_iter.iov = kmsg->iov;
3829 kmsg->msg.msg_name = &io.msg.addr;
3831 ret = io_recvmsg_copy_hdr(req, &io);
3836 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3838 return PTR_ERR(kbuf);
3840 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3841 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3842 1, req->sr_msg.len);
3845 flags = req->sr_msg.msg_flags;
3846 if (flags & MSG_DONTWAIT)
3847 req->flags |= REQ_F_NOWAIT;
3848 else if (force_nonblock)
3849 flags |= MSG_DONTWAIT;
3851 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3852 kmsg->uaddr, flags);
3853 if (force_nonblock && ret == -EAGAIN)
3854 return io_setup_async_msg(req, kmsg);
3855 if (ret == -ERESTARTSYS)
3859 if (kmsg && kmsg->iov != kmsg->fast_iov)
3861 req->flags &= ~REQ_F_NEED_CLEANUP;
3862 __io_cqring_add_event(req, ret, cflags);
3864 req_set_fail_links(req);
3869 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3871 struct io_buffer *kbuf = NULL;
3872 struct socket *sock;
3873 int ret, cflags = 0;
3875 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3878 sock = sock_from_file(req->file, &ret);
3880 struct io_sr_msg *sr = &req->sr_msg;
3881 void __user *buf = sr->buf;
3886 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3888 return PTR_ERR(kbuf);
3890 buf = u64_to_user_ptr(kbuf->addr);
3892 ret = import_single_range(READ, buf, sr->len, &iov,
3899 req->flags |= REQ_F_NEED_CLEANUP;
3900 msg.msg_name = NULL;
3901 msg.msg_control = NULL;
3902 msg.msg_controllen = 0;
3903 msg.msg_namelen = 0;
3904 msg.msg_iocb = NULL;
3907 flags = req->sr_msg.msg_flags;
3908 if (flags & MSG_DONTWAIT)
3909 req->flags |= REQ_F_NOWAIT;
3910 else if (force_nonblock)
3911 flags |= MSG_DONTWAIT;
3913 ret = sock_recvmsg(sock, &msg, flags);
3914 if (force_nonblock && ret == -EAGAIN)
3916 if (ret == -ERESTARTSYS)
3921 req->flags &= ~REQ_F_NEED_CLEANUP;
3922 __io_cqring_add_event(req, ret, cflags);
3924 req_set_fail_links(req);
3929 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3931 struct io_accept *accept = &req->accept;
3933 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3935 if (sqe->ioprio || sqe->len || sqe->buf_index)
3938 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3939 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3940 accept->flags = READ_ONCE(sqe->accept_flags);
3941 accept->nofile = rlimit(RLIMIT_NOFILE);
3945 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3947 struct io_accept *accept = &req->accept;
3948 unsigned file_flags;
3951 file_flags = force_nonblock ? O_NONBLOCK : 0;
3952 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3953 accept->addr_len, accept->flags,
3955 if (ret == -EAGAIN && force_nonblock)
3957 if (ret == -ERESTARTSYS)
3960 req_set_fail_links(req);
3961 io_cqring_add_event(req, ret);
3966 static void io_accept_finish(struct io_wq_work **workptr)
3968 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3970 if (io_req_cancelled(req))
3972 __io_accept(req, false);
3973 io_steal_work(req, workptr);
3976 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3980 ret = __io_accept(req, force_nonblock);
3981 if (ret == -EAGAIN && force_nonblock) {
3982 req->work.func = io_accept_finish;
3988 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3990 struct io_connect *conn = &req->connect;
3991 struct io_async_ctx *io = req->io;
3993 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3995 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3998 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3999 conn->addr_len = READ_ONCE(sqe->addr2);
4004 return move_addr_to_kernel(conn->addr, conn->addr_len,
4005 &io->connect.address);
4008 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4010 struct io_async_ctx __io, *io;
4011 unsigned file_flags;
4017 ret = move_addr_to_kernel(req->connect.addr,
4018 req->connect.addr_len,
4019 &__io.connect.address);
4025 file_flags = force_nonblock ? O_NONBLOCK : 0;
4027 ret = __sys_connect_file(req->file, &io->connect.address,
4028 req->connect.addr_len, file_flags);
4029 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4032 if (io_alloc_async_ctx(req)) {
4036 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4039 if (ret == -ERESTARTSYS)
4043 req_set_fail_links(req);
4044 io_cqring_add_event(req, ret);
4048 #else /* !CONFIG_NET */
4049 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4054 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4059 static int io_send(struct io_kiocb *req, bool force_nonblock)
4064 static int io_recvmsg_prep(struct io_kiocb *req,
4065 const struct io_uring_sqe *sqe)
4070 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4075 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4080 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4085 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4090 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4095 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4099 #endif /* CONFIG_NET */
4101 struct io_poll_table {
4102 struct poll_table_struct pt;
4103 struct io_kiocb *req;
4107 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4108 struct wait_queue_head *head)
4110 if (unlikely(poll->head)) {
4111 pt->error = -EINVAL;
4117 add_wait_queue(head, &poll->wait);
4120 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4121 struct poll_table_struct *p)
4123 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4125 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4128 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4129 __poll_t mask, task_work_func_t func)
4131 struct task_struct *tsk;
4134 /* for instances that support it check for an event match first: */
4135 if (mask && !(mask & poll->events))
4138 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4140 list_del_init(&poll->wait.entry);
4144 init_task_work(&req->task_work, func);
4146 * If this fails, then the task is exiting. Punt to one of the io-wq
4147 * threads to ensure the work gets run, we can't always rely on exit
4148 * cancelation taking care of this.
4150 ret = task_work_add(tsk, &req->task_work, true);
4151 if (unlikely(ret)) {
4152 tsk = io_wq_get_task(req->ctx->io_wq);
4153 task_work_add(tsk, &req->task_work, true);
4155 wake_up_process(tsk);
4159 static void io_async_task_func(struct callback_head *cb)
4161 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4162 struct async_poll *apoll = req->apoll;
4163 struct io_ring_ctx *ctx = req->ctx;
4165 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4167 WARN_ON_ONCE(!list_empty(&req->apoll->poll.wait.entry));
4169 if (hash_hashed(&req->hash_node)) {
4170 spin_lock_irq(&ctx->completion_lock);
4171 hash_del(&req->hash_node);
4172 spin_unlock_irq(&ctx->completion_lock);
4175 /* restore ->work in case we need to retry again */
4176 memcpy(&req->work, &apoll->work, sizeof(req->work));
4178 __set_current_state(TASK_RUNNING);
4179 mutex_lock(&ctx->uring_lock);
4180 __io_queue_sqe(req, NULL);
4181 mutex_unlock(&ctx->uring_lock);
4186 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4189 struct io_kiocb *req = wait->private;
4190 struct io_poll_iocb *poll = &req->apoll->poll;
4192 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4195 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4198 static void io_poll_req_insert(struct io_kiocb *req)
4200 struct io_ring_ctx *ctx = req->ctx;
4201 struct hlist_head *list;
4203 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4204 hlist_add_head(&req->hash_node, list);
4207 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4208 struct io_poll_iocb *poll,
4209 struct io_poll_table *ipt, __poll_t mask,
4210 wait_queue_func_t wake_func)
4211 __acquires(&ctx->completion_lock)
4213 struct io_ring_ctx *ctx = req->ctx;
4214 bool cancel = false;
4216 poll->file = req->file;
4218 poll->done = poll->canceled = false;
4219 poll->events = mask;
4221 ipt->pt._key = mask;
4223 ipt->error = -EINVAL;
4225 INIT_LIST_HEAD(&poll->wait.entry);
4226 init_waitqueue_func_entry(&poll->wait, wake_func);
4227 poll->wait.private = req;
4229 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4231 spin_lock_irq(&ctx->completion_lock);
4232 if (likely(poll->head)) {
4233 spin_lock(&poll->head->lock);
4234 if (unlikely(list_empty(&poll->wait.entry))) {
4240 if (mask || ipt->error)
4241 list_del_init(&poll->wait.entry);
4243 WRITE_ONCE(poll->canceled, true);
4244 else if (!poll->done) /* actually waiting for an event */
4245 io_poll_req_insert(req);
4246 spin_unlock(&poll->head->lock);
4252 static bool io_arm_poll_handler(struct io_kiocb *req)
4254 const struct io_op_def *def = &io_op_defs[req->opcode];
4255 struct io_ring_ctx *ctx = req->ctx;
4256 struct async_poll *apoll;
4257 struct io_poll_table ipt;
4260 if (!req->file || !file_can_poll(req->file))
4262 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4264 if (!def->pollin && !def->pollout)
4267 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4268 if (unlikely(!apoll))
4271 req->flags |= REQ_F_POLLED;
4272 memcpy(&apoll->work, &req->work, sizeof(req->work));
4274 get_task_struct(current);
4275 req->task = current;
4277 INIT_HLIST_NODE(&req->hash_node);
4281 mask |= POLLIN | POLLRDNORM;
4283 mask |= POLLOUT | POLLWRNORM;
4284 mask |= POLLERR | POLLPRI;
4286 ipt.pt._qproc = io_async_queue_proc;
4288 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4292 apoll->poll.done = true;
4293 spin_unlock_irq(&ctx->completion_lock);
4294 memcpy(&req->work, &apoll->work, sizeof(req->work));
4298 spin_unlock_irq(&ctx->completion_lock);
4299 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4300 apoll->poll.events);
4304 static bool __io_poll_remove_one(struct io_kiocb *req,
4305 struct io_poll_iocb *poll)
4307 bool do_complete = false;
4309 spin_lock(&poll->head->lock);
4310 WRITE_ONCE(poll->canceled, true);
4311 if (!list_empty(&poll->wait.entry)) {
4312 list_del_init(&poll->wait.entry);
4315 spin_unlock(&poll->head->lock);
4319 static bool io_poll_remove_one(struct io_kiocb *req)
4323 if (req->opcode == IORING_OP_POLL_ADD) {
4324 do_complete = __io_poll_remove_one(req, &req->poll);
4326 /* non-poll requests have submit ref still */
4327 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4332 hash_del(&req->hash_node);
4335 io_cqring_fill_event(req, -ECANCELED);
4336 io_commit_cqring(req->ctx);
4337 req->flags |= REQ_F_COMP_LOCKED;
4344 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4346 struct hlist_node *tmp;
4347 struct io_kiocb *req;
4350 spin_lock_irq(&ctx->completion_lock);
4351 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4352 struct hlist_head *list;
4354 list = &ctx->cancel_hash[i];
4355 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4356 io_poll_remove_one(req);
4358 spin_unlock_irq(&ctx->completion_lock);
4360 io_cqring_ev_posted(ctx);
4363 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4365 struct hlist_head *list;
4366 struct io_kiocb *req;
4368 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4369 hlist_for_each_entry(req, list, hash_node) {
4370 if (sqe_addr != req->user_data)
4372 if (io_poll_remove_one(req))
4380 static int io_poll_remove_prep(struct io_kiocb *req,
4381 const struct io_uring_sqe *sqe)
4383 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4385 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4389 req->poll.addr = READ_ONCE(sqe->addr);
4394 * Find a running poll command that matches one specified in sqe->addr,
4395 * and remove it if found.
4397 static int io_poll_remove(struct io_kiocb *req)
4399 struct io_ring_ctx *ctx = req->ctx;
4403 addr = req->poll.addr;
4404 spin_lock_irq(&ctx->completion_lock);
4405 ret = io_poll_cancel(ctx, addr);
4406 spin_unlock_irq(&ctx->completion_lock);
4408 io_cqring_add_event(req, ret);
4410 req_set_fail_links(req);
4415 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4417 struct io_ring_ctx *ctx = req->ctx;
4419 req->poll.done = true;
4420 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4421 io_commit_cqring(ctx);
4424 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4426 struct io_ring_ctx *ctx = req->ctx;
4427 struct io_poll_iocb *poll = &req->poll;
4429 if (!req->result && !READ_ONCE(poll->canceled)) {
4430 struct poll_table_struct pt = { ._key = poll->events };
4432 req->result = vfs_poll(req->file, &pt) & poll->events;
4435 spin_lock_irq(&ctx->completion_lock);
4436 if (!req->result && !READ_ONCE(poll->canceled)) {
4437 add_wait_queue(poll->head, &poll->wait);
4438 spin_unlock_irq(&ctx->completion_lock);
4441 hash_del(&req->hash_node);
4442 io_poll_complete(req, req->result, 0);
4443 req->flags |= REQ_F_COMP_LOCKED;
4444 io_put_req_find_next(req, nxt);
4445 spin_unlock_irq(&ctx->completion_lock);
4447 io_cqring_ev_posted(ctx);
4450 static void io_poll_task_func(struct callback_head *cb)
4452 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4453 struct io_kiocb *nxt = NULL;
4455 io_poll_task_handler(req, &nxt);
4457 struct io_ring_ctx *ctx = nxt->ctx;
4459 mutex_lock(&ctx->uring_lock);
4460 __io_queue_sqe(nxt, NULL);
4461 mutex_unlock(&ctx->uring_lock);
4465 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4468 struct io_kiocb *req = wait->private;
4469 struct io_poll_iocb *poll = &req->poll;
4471 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4474 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4475 struct poll_table_struct *p)
4477 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4479 __io_queue_proc(&pt->req->poll, pt, head);
4482 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4484 struct io_poll_iocb *poll = &req->poll;
4487 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4489 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4494 events = READ_ONCE(sqe->poll_events);
4495 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4497 get_task_struct(current);
4498 req->task = current;
4502 static int io_poll_add(struct io_kiocb *req)
4504 struct io_poll_iocb *poll = &req->poll;
4505 struct io_ring_ctx *ctx = req->ctx;
4506 struct io_poll_table ipt;
4509 INIT_HLIST_NODE(&req->hash_node);
4510 INIT_LIST_HEAD(&req->list);
4511 ipt.pt._qproc = io_poll_queue_proc;
4513 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4516 if (mask) { /* no async, we'd stolen it */
4518 io_poll_complete(req, mask, 0);
4520 spin_unlock_irq(&ctx->completion_lock);
4523 io_cqring_ev_posted(ctx);
4529 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4531 struct io_timeout_data *data = container_of(timer,
4532 struct io_timeout_data, timer);
4533 struct io_kiocb *req = data->req;
4534 struct io_ring_ctx *ctx = req->ctx;
4535 unsigned long flags;
4537 atomic_inc(&ctx->cq_timeouts);
4539 spin_lock_irqsave(&ctx->completion_lock, flags);
4541 * We could be racing with timeout deletion. If the list is empty,
4542 * then timeout lookup already found it and will be handling it.
4544 if (!list_empty(&req->list)) {
4545 struct io_kiocb *prev;
4548 * Adjust the reqs sequence before the current one because it
4549 * will consume a slot in the cq_ring and the cq_tail
4550 * pointer will be increased, otherwise other timeout reqs may
4551 * return in advance without waiting for enough wait_nr.
4554 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4556 list_del_init(&req->list);
4559 io_cqring_fill_event(req, -ETIME);
4560 io_commit_cqring(ctx);
4561 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4563 io_cqring_ev_posted(ctx);
4564 req_set_fail_links(req);
4566 return HRTIMER_NORESTART;
4569 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4571 struct io_kiocb *req;
4574 list_for_each_entry(req, &ctx->timeout_list, list) {
4575 if (user_data == req->user_data) {
4576 list_del_init(&req->list);
4585 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4589 req_set_fail_links(req);
4590 io_cqring_fill_event(req, -ECANCELED);
4595 static int io_timeout_remove_prep(struct io_kiocb *req,
4596 const struct io_uring_sqe *sqe)
4598 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4600 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4603 req->timeout.addr = READ_ONCE(sqe->addr);
4604 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4605 if (req->timeout.flags)
4612 * Remove or update an existing timeout command
4614 static int io_timeout_remove(struct io_kiocb *req)
4616 struct io_ring_ctx *ctx = req->ctx;
4619 spin_lock_irq(&ctx->completion_lock);
4620 ret = io_timeout_cancel(ctx, req->timeout.addr);
4622 io_cqring_fill_event(req, ret);
4623 io_commit_cqring(ctx);
4624 spin_unlock_irq(&ctx->completion_lock);
4625 io_cqring_ev_posted(ctx);
4627 req_set_fail_links(req);
4632 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4633 bool is_timeout_link)
4635 struct io_timeout_data *data;
4638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4640 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4642 if (sqe->off && is_timeout_link)
4644 flags = READ_ONCE(sqe->timeout_flags);
4645 if (flags & ~IORING_TIMEOUT_ABS)
4648 req->timeout.count = READ_ONCE(sqe->off);
4650 if (!req->io && io_alloc_async_ctx(req))
4653 data = &req->io->timeout;
4655 req->flags |= REQ_F_TIMEOUT;
4657 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4660 if (flags & IORING_TIMEOUT_ABS)
4661 data->mode = HRTIMER_MODE_ABS;
4663 data->mode = HRTIMER_MODE_REL;
4665 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4669 static int io_timeout(struct io_kiocb *req)
4672 struct io_ring_ctx *ctx = req->ctx;
4673 struct io_timeout_data *data;
4674 struct list_head *entry;
4677 data = &req->io->timeout;
4680 * sqe->off holds how many events that need to occur for this
4681 * timeout event to be satisfied. If it isn't set, then this is
4682 * a pure timeout request, sequence isn't used.
4684 count = req->timeout.count;
4686 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4687 spin_lock_irq(&ctx->completion_lock);
4688 entry = ctx->timeout_list.prev;
4692 req->sequence = ctx->cached_sq_head + count - 1;
4693 data->seq_offset = count;
4696 * Insertion sort, ensuring the first entry in the list is always
4697 * the one we need first.
4699 spin_lock_irq(&ctx->completion_lock);
4700 list_for_each_prev(entry, &ctx->timeout_list) {
4701 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4702 unsigned nxt_sq_head;
4703 long long tmp, tmp_nxt;
4704 u32 nxt_offset = nxt->io->timeout.seq_offset;
4706 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4710 * Since cached_sq_head + count - 1 can overflow, use type long
4713 tmp = (long long)ctx->cached_sq_head + count - 1;
4714 nxt_sq_head = nxt->sequence - nxt_offset + 1;
4715 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
4718 * cached_sq_head may overflow, and it will never overflow twice
4719 * once there is some timeout req still be valid.
4721 if (ctx->cached_sq_head < nxt_sq_head)
4728 * Sequence of reqs after the insert one and itself should
4729 * be adjusted because each timeout req consumes a slot.
4734 req->sequence -= span;
4736 list_add(&req->list, entry);
4737 data->timer.function = io_timeout_fn;
4738 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4739 spin_unlock_irq(&ctx->completion_lock);
4743 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4745 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4747 return req->user_data == (unsigned long) data;
4750 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4752 enum io_wq_cancel cancel_ret;
4755 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4756 switch (cancel_ret) {
4757 case IO_WQ_CANCEL_OK:
4760 case IO_WQ_CANCEL_RUNNING:
4763 case IO_WQ_CANCEL_NOTFOUND:
4771 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4772 struct io_kiocb *req, __u64 sqe_addr,
4775 unsigned long flags;
4778 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4779 if (ret != -ENOENT) {
4780 spin_lock_irqsave(&ctx->completion_lock, flags);
4784 spin_lock_irqsave(&ctx->completion_lock, flags);
4785 ret = io_timeout_cancel(ctx, sqe_addr);
4788 ret = io_poll_cancel(ctx, sqe_addr);
4792 io_cqring_fill_event(req, ret);
4793 io_commit_cqring(ctx);
4794 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4795 io_cqring_ev_posted(ctx);
4798 req_set_fail_links(req);
4802 static int io_async_cancel_prep(struct io_kiocb *req,
4803 const struct io_uring_sqe *sqe)
4805 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4807 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4811 req->cancel.addr = READ_ONCE(sqe->addr);
4815 static int io_async_cancel(struct io_kiocb *req)
4817 struct io_ring_ctx *ctx = req->ctx;
4819 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4823 static int io_files_update_prep(struct io_kiocb *req,
4824 const struct io_uring_sqe *sqe)
4826 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4829 req->files_update.offset = READ_ONCE(sqe->off);
4830 req->files_update.nr_args = READ_ONCE(sqe->len);
4831 if (!req->files_update.nr_args)
4833 req->files_update.arg = READ_ONCE(sqe->addr);
4837 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4839 struct io_ring_ctx *ctx = req->ctx;
4840 struct io_uring_files_update up;
4846 up.offset = req->files_update.offset;
4847 up.fds = req->files_update.arg;
4849 mutex_lock(&ctx->uring_lock);
4850 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4851 mutex_unlock(&ctx->uring_lock);
4854 req_set_fail_links(req);
4855 io_cqring_add_event(req, ret);
4860 static int io_req_defer_prep(struct io_kiocb *req,
4861 const struct io_uring_sqe *sqe)
4868 if (io_op_defs[req->opcode].file_table) {
4869 ret = io_grab_files(req);
4874 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4876 switch (req->opcode) {
4879 case IORING_OP_READV:
4880 case IORING_OP_READ_FIXED:
4881 case IORING_OP_READ:
4882 ret = io_read_prep(req, sqe, true);
4884 case IORING_OP_WRITEV:
4885 case IORING_OP_WRITE_FIXED:
4886 case IORING_OP_WRITE:
4887 ret = io_write_prep(req, sqe, true);
4889 case IORING_OP_POLL_ADD:
4890 ret = io_poll_add_prep(req, sqe);
4892 case IORING_OP_POLL_REMOVE:
4893 ret = io_poll_remove_prep(req, sqe);
4895 case IORING_OP_FSYNC:
4896 ret = io_prep_fsync(req, sqe);
4898 case IORING_OP_SYNC_FILE_RANGE:
4899 ret = io_prep_sfr(req, sqe);
4901 case IORING_OP_SENDMSG:
4902 case IORING_OP_SEND:
4903 ret = io_sendmsg_prep(req, sqe);
4905 case IORING_OP_RECVMSG:
4906 case IORING_OP_RECV:
4907 ret = io_recvmsg_prep(req, sqe);
4909 case IORING_OP_CONNECT:
4910 ret = io_connect_prep(req, sqe);
4912 case IORING_OP_TIMEOUT:
4913 ret = io_timeout_prep(req, sqe, false);
4915 case IORING_OP_TIMEOUT_REMOVE:
4916 ret = io_timeout_remove_prep(req, sqe);
4918 case IORING_OP_ASYNC_CANCEL:
4919 ret = io_async_cancel_prep(req, sqe);
4921 case IORING_OP_LINK_TIMEOUT:
4922 ret = io_timeout_prep(req, sqe, true);
4924 case IORING_OP_ACCEPT:
4925 ret = io_accept_prep(req, sqe);
4927 case IORING_OP_FALLOCATE:
4928 ret = io_fallocate_prep(req, sqe);
4930 case IORING_OP_OPENAT:
4931 ret = io_openat_prep(req, sqe);
4933 case IORING_OP_CLOSE:
4934 ret = io_close_prep(req, sqe);
4936 case IORING_OP_FILES_UPDATE:
4937 ret = io_files_update_prep(req, sqe);
4939 case IORING_OP_STATX:
4940 ret = io_statx_prep(req, sqe);
4942 case IORING_OP_FADVISE:
4943 ret = io_fadvise_prep(req, sqe);
4945 case IORING_OP_MADVISE:
4946 ret = io_madvise_prep(req, sqe);
4948 case IORING_OP_OPENAT2:
4949 ret = io_openat2_prep(req, sqe);
4951 case IORING_OP_EPOLL_CTL:
4952 ret = io_epoll_ctl_prep(req, sqe);
4954 case IORING_OP_SPLICE:
4955 ret = io_splice_prep(req, sqe);
4957 case IORING_OP_PROVIDE_BUFFERS:
4958 ret = io_provide_buffers_prep(req, sqe);
4960 case IORING_OP_REMOVE_BUFFERS:
4961 ret = io_remove_buffers_prep(req, sqe);
4964 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4973 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4975 struct io_ring_ctx *ctx = req->ctx;
4978 /* Still need defer if there is pending req in defer list. */
4979 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4982 if (!req->io && io_alloc_async_ctx(req))
4985 ret = io_req_defer_prep(req, sqe);
4989 spin_lock_irq(&ctx->completion_lock);
4990 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
4991 spin_unlock_irq(&ctx->completion_lock);
4995 trace_io_uring_defer(ctx, req, req->user_data);
4996 list_add_tail(&req->list, &ctx->defer_list);
4997 spin_unlock_irq(&ctx->completion_lock);
4998 return -EIOCBQUEUED;
5001 static void io_cleanup_req(struct io_kiocb *req)
5003 struct io_async_ctx *io = req->io;
5005 switch (req->opcode) {
5006 case IORING_OP_READV:
5007 case IORING_OP_READ_FIXED:
5008 case IORING_OP_READ:
5009 if (req->flags & REQ_F_BUFFER_SELECTED)
5010 kfree((void *)(unsigned long)req->rw.addr);
5012 case IORING_OP_WRITEV:
5013 case IORING_OP_WRITE_FIXED:
5014 case IORING_OP_WRITE:
5015 if (io->rw.iov != io->rw.fast_iov)
5018 case IORING_OP_RECVMSG:
5019 if (req->flags & REQ_F_BUFFER_SELECTED)
5020 kfree(req->sr_msg.kbuf);
5022 case IORING_OP_SENDMSG:
5023 if (io->msg.iov != io->msg.fast_iov)
5026 case IORING_OP_RECV:
5027 if (req->flags & REQ_F_BUFFER_SELECTED)
5028 kfree(req->sr_msg.kbuf);
5030 case IORING_OP_OPENAT:
5031 case IORING_OP_OPENAT2:
5032 case IORING_OP_STATX:
5033 putname(req->open.filename);
5035 case IORING_OP_SPLICE:
5036 io_put_file(req, req->splice.file_in,
5037 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5041 req->flags &= ~REQ_F_NEED_CLEANUP;
5044 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5045 bool force_nonblock)
5047 struct io_ring_ctx *ctx = req->ctx;
5050 switch (req->opcode) {
5054 case IORING_OP_READV:
5055 case IORING_OP_READ_FIXED:
5056 case IORING_OP_READ:
5058 ret = io_read_prep(req, sqe, force_nonblock);
5062 ret = io_read(req, force_nonblock);
5064 case IORING_OP_WRITEV:
5065 case IORING_OP_WRITE_FIXED:
5066 case IORING_OP_WRITE:
5068 ret = io_write_prep(req, sqe, force_nonblock);
5072 ret = io_write(req, force_nonblock);
5074 case IORING_OP_FSYNC:
5076 ret = io_prep_fsync(req, sqe);
5080 ret = io_fsync(req, force_nonblock);
5082 case IORING_OP_POLL_ADD:
5084 ret = io_poll_add_prep(req, sqe);
5088 ret = io_poll_add(req);
5090 case IORING_OP_POLL_REMOVE:
5092 ret = io_poll_remove_prep(req, sqe);
5096 ret = io_poll_remove(req);
5098 case IORING_OP_SYNC_FILE_RANGE:
5100 ret = io_prep_sfr(req, sqe);
5104 ret = io_sync_file_range(req, force_nonblock);
5106 case IORING_OP_SENDMSG:
5107 case IORING_OP_SEND:
5109 ret = io_sendmsg_prep(req, sqe);
5113 if (req->opcode == IORING_OP_SENDMSG)
5114 ret = io_sendmsg(req, force_nonblock);
5116 ret = io_send(req, force_nonblock);
5118 case IORING_OP_RECVMSG:
5119 case IORING_OP_RECV:
5121 ret = io_recvmsg_prep(req, sqe);
5125 if (req->opcode == IORING_OP_RECVMSG)
5126 ret = io_recvmsg(req, force_nonblock);
5128 ret = io_recv(req, force_nonblock);
5130 case IORING_OP_TIMEOUT:
5132 ret = io_timeout_prep(req, sqe, false);
5136 ret = io_timeout(req);
5138 case IORING_OP_TIMEOUT_REMOVE:
5140 ret = io_timeout_remove_prep(req, sqe);
5144 ret = io_timeout_remove(req);
5146 case IORING_OP_ACCEPT:
5148 ret = io_accept_prep(req, sqe);
5152 ret = io_accept(req, force_nonblock);
5154 case IORING_OP_CONNECT:
5156 ret = io_connect_prep(req, sqe);
5160 ret = io_connect(req, force_nonblock);
5162 case IORING_OP_ASYNC_CANCEL:
5164 ret = io_async_cancel_prep(req, sqe);
5168 ret = io_async_cancel(req);
5170 case IORING_OP_FALLOCATE:
5172 ret = io_fallocate_prep(req, sqe);
5176 ret = io_fallocate(req, force_nonblock);
5178 case IORING_OP_OPENAT:
5180 ret = io_openat_prep(req, sqe);
5184 ret = io_openat(req, force_nonblock);
5186 case IORING_OP_CLOSE:
5188 ret = io_close_prep(req, sqe);
5192 ret = io_close(req, force_nonblock);
5194 case IORING_OP_FILES_UPDATE:
5196 ret = io_files_update_prep(req, sqe);
5200 ret = io_files_update(req, force_nonblock);
5202 case IORING_OP_STATX:
5204 ret = io_statx_prep(req, sqe);
5208 ret = io_statx(req, force_nonblock);
5210 case IORING_OP_FADVISE:
5212 ret = io_fadvise_prep(req, sqe);
5216 ret = io_fadvise(req, force_nonblock);
5218 case IORING_OP_MADVISE:
5220 ret = io_madvise_prep(req, sqe);
5224 ret = io_madvise(req, force_nonblock);
5226 case IORING_OP_OPENAT2:
5228 ret = io_openat2_prep(req, sqe);
5232 ret = io_openat2(req, force_nonblock);
5234 case IORING_OP_EPOLL_CTL:
5236 ret = io_epoll_ctl_prep(req, sqe);
5240 ret = io_epoll_ctl(req, force_nonblock);
5242 case IORING_OP_SPLICE:
5244 ret = io_splice_prep(req, sqe);
5248 ret = io_splice(req, force_nonblock);
5250 case IORING_OP_PROVIDE_BUFFERS:
5252 ret = io_provide_buffers_prep(req, sqe);
5256 ret = io_provide_buffers(req, force_nonblock);
5258 case IORING_OP_REMOVE_BUFFERS:
5260 ret = io_remove_buffers_prep(req, sqe);
5264 ret = io_remove_buffers(req, force_nonblock);
5274 if (ctx->flags & IORING_SETUP_IOPOLL) {
5275 const bool in_async = io_wq_current_is_worker();
5277 if (req->result == -EAGAIN)
5280 /* workqueue context doesn't hold uring_lock, grab it now */
5282 mutex_lock(&ctx->uring_lock);
5284 io_iopoll_req_issued(req);
5287 mutex_unlock(&ctx->uring_lock);
5293 static void io_wq_submit_work(struct io_wq_work **workptr)
5295 struct io_wq_work *work = *workptr;
5296 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5299 /* if NO_CANCEL is set, we must still run the work */
5300 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5301 IO_WQ_WORK_CANCEL) {
5307 ret = io_issue_sqe(req, NULL, false);
5309 * We can get EAGAIN for polled IO even though we're
5310 * forcing a sync submission from here, since we can't
5311 * wait for request slots on the block side.
5320 req_set_fail_links(req);
5321 io_cqring_add_event(req, ret);
5325 io_steal_work(req, workptr);
5328 static int io_req_needs_file(struct io_kiocb *req, int fd)
5330 if (!io_op_defs[req->opcode].needs_file)
5332 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5337 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5340 struct fixed_file_table *table;
5342 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5343 return table->files[index & IORING_FILE_TABLE_MASK];;
5346 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5347 int fd, struct file **out_file, bool fixed)
5349 struct io_ring_ctx *ctx = req->ctx;
5353 if (unlikely(!ctx->file_data ||
5354 (unsigned) fd >= ctx->nr_user_files))
5356 fd = array_index_nospec(fd, ctx->nr_user_files);
5357 file = io_file_from_index(ctx, fd);
5360 req->fixed_file_refs = ctx->file_data->cur_refs;
5361 percpu_ref_get(req->fixed_file_refs);
5363 trace_io_uring_file_get(ctx, fd);
5364 file = __io_file_get(state, fd);
5365 if (unlikely(!file))
5373 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5374 int fd, unsigned int flags)
5378 if (!io_req_needs_file(req, fd))
5381 fixed = (flags & IOSQE_FIXED_FILE);
5382 if (unlikely(!fixed && req->needs_fixed_file))
5385 return io_file_get(state, req, fd, &req->file, fixed);
5388 static int io_grab_files(struct io_kiocb *req)
5391 struct io_ring_ctx *ctx = req->ctx;
5393 if (req->work.files)
5395 if (!ctx->ring_file)
5399 spin_lock_irq(&ctx->inflight_lock);
5401 * We use the f_ops->flush() handler to ensure that we can flush
5402 * out work accessing these files if the fd is closed. Check if
5403 * the fd has changed since we started down this path, and disallow
5404 * this operation if it has.
5406 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5407 list_add(&req->inflight_entry, &ctx->inflight_list);
5408 req->flags |= REQ_F_INFLIGHT;
5409 req->work.files = current->files;
5412 spin_unlock_irq(&ctx->inflight_lock);
5418 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5420 struct io_timeout_data *data = container_of(timer,
5421 struct io_timeout_data, timer);
5422 struct io_kiocb *req = data->req;
5423 struct io_ring_ctx *ctx = req->ctx;
5424 struct io_kiocb *prev = NULL;
5425 unsigned long flags;
5427 spin_lock_irqsave(&ctx->completion_lock, flags);
5430 * We don't expect the list to be empty, that will only happen if we
5431 * race with the completion of the linked work.
5433 if (!list_empty(&req->link_list)) {
5434 prev = list_entry(req->link_list.prev, struct io_kiocb,
5436 if (refcount_inc_not_zero(&prev->refs)) {
5437 list_del_init(&req->link_list);
5438 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5443 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5446 req_set_fail_links(prev);
5447 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5450 io_cqring_add_event(req, -ETIME);
5453 return HRTIMER_NORESTART;
5456 static void io_queue_linked_timeout(struct io_kiocb *req)
5458 struct io_ring_ctx *ctx = req->ctx;
5461 * If the list is now empty, then our linked request finished before
5462 * we got a chance to setup the timer
5464 spin_lock_irq(&ctx->completion_lock);
5465 if (!list_empty(&req->link_list)) {
5466 struct io_timeout_data *data = &req->io->timeout;
5468 data->timer.function = io_link_timeout_fn;
5469 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5472 spin_unlock_irq(&ctx->completion_lock);
5474 /* drop submission reference */
5478 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5480 struct io_kiocb *nxt;
5482 if (!(req->flags & REQ_F_LINK_HEAD))
5484 /* for polled retry, if flag is set, we already went through here */
5485 if (req->flags & REQ_F_POLLED)
5488 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5490 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5493 req->flags |= REQ_F_LINK_TIMEOUT;
5497 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5499 struct io_kiocb *linked_timeout;
5500 struct io_kiocb *nxt;
5501 const struct cred *old_creds = NULL;
5505 linked_timeout = io_prep_linked_timeout(req);
5507 if (req->work.creds && req->work.creds != current_cred()) {
5509 revert_creds(old_creds);
5510 if (old_creds == req->work.creds)
5511 old_creds = NULL; /* restored original creds */
5513 old_creds = override_creds(req->work.creds);
5516 ret = io_issue_sqe(req, sqe, true);
5519 * We async punt it if the file wasn't marked NOWAIT, or if the file
5520 * doesn't support non-blocking read/write attempts
5522 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5523 (req->flags & REQ_F_MUST_PUNT))) {
5524 if (io_arm_poll_handler(req)) {
5526 io_queue_linked_timeout(linked_timeout);
5530 if (io_op_defs[req->opcode].file_table) {
5531 ret = io_grab_files(req);
5537 * Queued up for async execution, worker will release
5538 * submit reference when the iocb is actually submitted.
5540 io_queue_async_work(req);
5546 /* drop submission reference */
5547 io_put_req_find_next(req, &nxt);
5549 if (linked_timeout) {
5551 io_queue_linked_timeout(linked_timeout);
5553 io_put_req(linked_timeout);
5556 /* and drop final reference, if we failed */
5558 io_cqring_add_event(req, ret);
5559 req_set_fail_links(req);
5565 if (req->flags & REQ_F_FORCE_ASYNC)
5571 revert_creds(old_creds);
5574 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5578 ret = io_req_defer(req, sqe);
5580 if (ret != -EIOCBQUEUED) {
5582 io_cqring_add_event(req, ret);
5583 req_set_fail_links(req);
5584 io_double_put_req(req);
5586 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5587 ret = io_req_defer_prep(req, sqe);
5588 if (unlikely(ret < 0))
5591 * Never try inline submit of IOSQE_ASYNC is set, go straight
5592 * to async execution.
5594 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5595 io_queue_async_work(req);
5597 __io_queue_sqe(req, sqe);
5601 static inline void io_queue_link_head(struct io_kiocb *req)
5603 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5604 io_cqring_add_event(req, -ECANCELED);
5605 io_double_put_req(req);
5607 io_queue_sqe(req, NULL);
5610 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5611 struct io_submit_state *state, struct io_kiocb **link)
5613 struct io_ring_ctx *ctx = req->ctx;
5617 * If we already have a head request, queue this one for async
5618 * submittal once the head completes. If we don't have a head but
5619 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5620 * submitted sync once the chain is complete. If none of those
5621 * conditions are true (normal request), then just queue it.
5624 struct io_kiocb *head = *link;
5627 * Taking sequential execution of a link, draining both sides
5628 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5629 * requests in the link. So, it drains the head and the
5630 * next after the link request. The last one is done via
5631 * drain_next flag to persist the effect across calls.
5633 if (req->flags & REQ_F_IO_DRAIN) {
5634 head->flags |= REQ_F_IO_DRAIN;
5635 ctx->drain_next = 1;
5637 if (io_alloc_async_ctx(req))
5640 ret = io_req_defer_prep(req, sqe);
5642 /* fail even hard links since we don't submit */
5643 head->flags |= REQ_F_FAIL_LINK;
5646 trace_io_uring_link(ctx, req, head);
5647 list_add_tail(&req->link_list, &head->link_list);
5649 /* last request of a link, enqueue the link */
5650 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5651 io_queue_link_head(head);
5655 if (unlikely(ctx->drain_next)) {
5656 req->flags |= REQ_F_IO_DRAIN;
5657 ctx->drain_next = 0;
5659 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5660 req->flags |= REQ_F_LINK_HEAD;
5661 INIT_LIST_HEAD(&req->link_list);
5663 if (io_alloc_async_ctx(req))
5666 ret = io_req_defer_prep(req, sqe);
5668 req->flags |= REQ_F_FAIL_LINK;
5671 io_queue_sqe(req, sqe);
5679 * Batched submission is done, ensure local IO is flushed out.
5681 static void io_submit_state_end(struct io_submit_state *state)
5683 blk_finish_plug(&state->plug);
5685 if (state->free_reqs)
5686 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5690 * Start submission side cache.
5692 static void io_submit_state_start(struct io_submit_state *state,
5693 unsigned int max_ios)
5695 blk_start_plug(&state->plug);
5696 state->free_reqs = 0;
5698 state->ios_left = max_ios;
5701 static void io_commit_sqring(struct io_ring_ctx *ctx)
5703 struct io_rings *rings = ctx->rings;
5706 * Ensure any loads from the SQEs are done at this point,
5707 * since once we write the new head, the application could
5708 * write new data to them.
5710 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5714 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5715 * that is mapped by userspace. This means that care needs to be taken to
5716 * ensure that reads are stable, as we cannot rely on userspace always
5717 * being a good citizen. If members of the sqe are validated and then later
5718 * used, it's important that those reads are done through READ_ONCE() to
5719 * prevent a re-load down the line.
5721 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5723 u32 *sq_array = ctx->sq_array;
5727 * The cached sq head (or cq tail) serves two purposes:
5729 * 1) allows us to batch the cost of updating the user visible
5731 * 2) allows the kernel side to track the head on its own, even
5732 * though the application is the one updating it.
5734 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5735 if (likely(head < ctx->sq_entries))
5736 return &ctx->sq_sqes[head];
5738 /* drop invalid entries */
5739 ctx->cached_sq_dropped++;
5740 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5744 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5746 ctx->cached_sq_head++;
5749 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5750 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5751 IOSQE_BUFFER_SELECT)
5753 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5754 const struct io_uring_sqe *sqe,
5755 struct io_submit_state *state, bool async)
5757 unsigned int sqe_flags;
5761 * All io need record the previous position, if LINK vs DARIN,
5762 * it can be used to mark the position of the first IO in the
5765 req->sequence = ctx->cached_sq_head;
5766 req->opcode = READ_ONCE(sqe->opcode);
5767 req->user_data = READ_ONCE(sqe->user_data);
5772 /* one is dropped after submission, the other at completion */
5773 refcount_set(&req->refs, 2);
5776 req->needs_fixed_file = async;
5777 INIT_IO_WORK(&req->work, io_wq_submit_work);
5779 if (unlikely(req->opcode >= IORING_OP_LAST))
5782 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5783 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5785 use_mm(ctx->sqo_mm);
5788 sqe_flags = READ_ONCE(sqe->flags);
5789 /* enforce forwards compatibility on users */
5790 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5793 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5794 !io_op_defs[req->opcode].buffer_select)
5797 id = READ_ONCE(sqe->personality);
5799 req->work.creds = idr_find(&ctx->personality_idr, id);
5800 if (unlikely(!req->work.creds))
5802 get_cred(req->work.creds);
5805 /* same numerical values with corresponding REQ_F_*, safe to copy */
5806 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5807 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5808 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5810 fd = READ_ONCE(sqe->fd);
5811 return io_req_set_file(state, req, fd, sqe_flags);
5814 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5815 struct file *ring_file, int ring_fd, bool async)
5817 struct io_submit_state state, *statep = NULL;
5818 struct io_kiocb *link = NULL;
5819 int i, submitted = 0;
5821 /* if we have a backlog and couldn't flush it all, return BUSY */
5822 if (test_bit(0, &ctx->sq_check_overflow)) {
5823 if (!list_empty(&ctx->cq_overflow_list) &&
5824 !io_cqring_overflow_flush(ctx, false))
5828 /* make sure SQ entry isn't read before tail */
5829 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5831 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5834 if (nr > IO_PLUG_THRESHOLD) {
5835 io_submit_state_start(&state, nr);
5839 ctx->ring_fd = ring_fd;
5840 ctx->ring_file = ring_file;
5842 for (i = 0; i < nr; i++) {
5843 const struct io_uring_sqe *sqe;
5844 struct io_kiocb *req;
5847 sqe = io_get_sqe(ctx);
5848 if (unlikely(!sqe)) {
5849 io_consume_sqe(ctx);
5852 req = io_alloc_req(ctx, statep);
5853 if (unlikely(!req)) {
5855 submitted = -EAGAIN;
5859 err = io_init_req(ctx, req, sqe, statep, async);
5860 io_consume_sqe(ctx);
5861 /* will complete beyond this point, count as submitted */
5864 if (unlikely(err)) {
5866 io_cqring_add_event(req, err);
5867 io_double_put_req(req);
5871 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5873 err = io_submit_sqe(req, sqe, statep, &link);
5878 if (unlikely(submitted != nr)) {
5879 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5881 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5884 io_queue_link_head(link);
5886 io_submit_state_end(&state);
5888 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5889 io_commit_sqring(ctx);
5894 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5896 struct mm_struct *mm = current->mm;
5904 static int io_sq_thread(void *data)
5906 struct io_ring_ctx *ctx = data;
5907 const struct cred *old_cred;
5908 mm_segment_t old_fs;
5910 unsigned long timeout;
5913 complete(&ctx->completions[1]);
5917 old_cred = override_creds(ctx->creds);
5919 timeout = jiffies + ctx->sq_thread_idle;
5920 while (!kthread_should_park()) {
5921 unsigned int to_submit;
5923 if (!list_empty(&ctx->poll_list)) {
5924 unsigned nr_events = 0;
5926 mutex_lock(&ctx->uring_lock);
5927 if (!list_empty(&ctx->poll_list))
5928 io_iopoll_getevents(ctx, &nr_events, 0);
5930 timeout = jiffies + ctx->sq_thread_idle;
5931 mutex_unlock(&ctx->uring_lock);
5934 to_submit = io_sqring_entries(ctx);
5937 * If submit got -EBUSY, flag us as needing the application
5938 * to enter the kernel to reap and flush events.
5940 if (!to_submit || ret == -EBUSY) {
5942 * Drop cur_mm before scheduling, we can't hold it for
5943 * long periods (or over schedule()). Do this before
5944 * adding ourselves to the waitqueue, as the unuse/drop
5947 io_sq_thread_drop_mm(ctx);
5950 * We're polling. If we're within the defined idle
5951 * period, then let us spin without work before going
5952 * to sleep. The exception is if we got EBUSY doing
5953 * more IO, we should wait for the application to
5954 * reap events and wake us up.
5956 if (!list_empty(&ctx->poll_list) ||
5957 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5958 !percpu_ref_is_dying(&ctx->refs))) {
5959 if (current->task_works)
5965 prepare_to_wait(&ctx->sqo_wait, &wait,
5966 TASK_INTERRUPTIBLE);
5969 * While doing polled IO, before going to sleep, we need
5970 * to check if there are new reqs added to poll_list, it
5971 * is because reqs may have been punted to io worker and
5972 * will be added to poll_list later, hence check the
5975 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5976 !list_empty_careful(&ctx->poll_list)) {
5977 finish_wait(&ctx->sqo_wait, &wait);
5981 /* Tell userspace we may need a wakeup call */
5982 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5983 /* make sure to read SQ tail after writing flags */
5986 to_submit = io_sqring_entries(ctx);
5987 if (!to_submit || ret == -EBUSY) {
5988 if (kthread_should_park()) {
5989 finish_wait(&ctx->sqo_wait, &wait);
5992 if (current->task_works) {
5994 finish_wait(&ctx->sqo_wait, &wait);
5997 if (signal_pending(current))
5998 flush_signals(current);
6000 finish_wait(&ctx->sqo_wait, &wait);
6002 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6005 finish_wait(&ctx->sqo_wait, &wait);
6007 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6010 mutex_lock(&ctx->uring_lock);
6011 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6012 mutex_unlock(&ctx->uring_lock);
6013 timeout = jiffies + ctx->sq_thread_idle;
6016 if (current->task_works)
6020 io_sq_thread_drop_mm(ctx);
6021 revert_creds(old_cred);
6028 struct io_wait_queue {
6029 struct wait_queue_entry wq;
6030 struct io_ring_ctx *ctx;
6032 unsigned nr_timeouts;
6035 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6037 struct io_ring_ctx *ctx = iowq->ctx;
6040 * Wake up if we have enough events, or if a timeout occurred since we
6041 * started waiting. For timeouts, we always want to return to userspace,
6042 * regardless of event count.
6044 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6045 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6048 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6049 int wake_flags, void *key)
6051 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6054 /* use noflush == true, as we can't safely rely on locking context */
6055 if (!io_should_wake(iowq, true))
6058 return autoremove_wake_function(curr, mode, wake_flags, key);
6062 * Wait until events become available, if we don't already have some. The
6063 * application must reap them itself, as they reside on the shared cq ring.
6065 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6066 const sigset_t __user *sig, size_t sigsz)
6068 struct io_wait_queue iowq = {
6071 .func = io_wake_function,
6072 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6075 .to_wait = min_events,
6077 struct io_rings *rings = ctx->rings;
6081 if (io_cqring_events(ctx, false) >= min_events)
6083 if (!current->task_works)
6089 #ifdef CONFIG_COMPAT
6090 if (in_compat_syscall())
6091 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6095 ret = set_user_sigmask(sig, sigsz);
6101 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6102 trace_io_uring_cqring_wait(ctx, min_events);
6104 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6105 TASK_INTERRUPTIBLE);
6106 if (current->task_works)
6108 if (io_should_wake(&iowq, false))
6111 if (signal_pending(current)) {
6116 finish_wait(&ctx->wait, &iowq.wq);
6118 restore_saved_sigmask_unless(ret == -EINTR);
6120 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6123 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6125 #if defined(CONFIG_UNIX)
6126 if (ctx->ring_sock) {
6127 struct sock *sock = ctx->ring_sock->sk;
6128 struct sk_buff *skb;
6130 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6136 for (i = 0; i < ctx->nr_user_files; i++) {
6139 file = io_file_from_index(ctx, i);
6146 static void io_file_ref_kill(struct percpu_ref *ref)
6148 struct fixed_file_data *data;
6150 data = container_of(ref, struct fixed_file_data, refs);
6151 complete(&data->done);
6154 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6156 struct fixed_file_data *data = ctx->file_data;
6157 struct fixed_file_ref_node *ref_node = NULL;
6158 unsigned nr_tables, i;
6159 unsigned long flags;
6164 spin_lock_irqsave(&data->lock, flags);
6165 if (!list_empty(&data->ref_list))
6166 ref_node = list_first_entry(&data->ref_list,
6167 struct fixed_file_ref_node, node);
6168 spin_unlock_irqrestore(&data->lock, flags);
6170 percpu_ref_kill(&ref_node->refs);
6172 percpu_ref_kill(&data->refs);
6174 /* wait for all refs nodes to complete */
6175 wait_for_completion(&data->done);
6177 __io_sqe_files_unregister(ctx);
6178 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6179 for (i = 0; i < nr_tables; i++)
6180 kfree(data->table[i].files);
6182 percpu_ref_exit(&data->refs);
6184 ctx->file_data = NULL;
6185 ctx->nr_user_files = 0;
6189 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6191 if (ctx->sqo_thread) {
6192 wait_for_completion(&ctx->completions[1]);
6194 * The park is a bit of a work-around, without it we get
6195 * warning spews on shutdown with SQPOLL set and affinity
6196 * set to a single CPU.
6198 kthread_park(ctx->sqo_thread);
6199 kthread_stop(ctx->sqo_thread);
6200 ctx->sqo_thread = NULL;
6204 static void io_finish_async(struct io_ring_ctx *ctx)
6206 io_sq_thread_stop(ctx);
6209 io_wq_destroy(ctx->io_wq);
6214 #if defined(CONFIG_UNIX)
6216 * Ensure the UNIX gc is aware of our file set, so we are certain that
6217 * the io_uring can be safely unregistered on process exit, even if we have
6218 * loops in the file referencing.
6220 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6222 struct sock *sk = ctx->ring_sock->sk;
6223 struct scm_fp_list *fpl;
6224 struct sk_buff *skb;
6227 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6231 skb = alloc_skb(0, GFP_KERNEL);
6240 fpl->user = get_uid(ctx->user);
6241 for (i = 0; i < nr; i++) {
6242 struct file *file = io_file_from_index(ctx, i + offset);
6246 fpl->fp[nr_files] = get_file(file);
6247 unix_inflight(fpl->user, fpl->fp[nr_files]);
6252 fpl->max = SCM_MAX_FD;
6253 fpl->count = nr_files;
6254 UNIXCB(skb).fp = fpl;
6255 skb->destructor = unix_destruct_scm;
6256 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6257 skb_queue_head(&sk->sk_receive_queue, skb);
6259 for (i = 0; i < nr_files; i++)
6270 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6271 * causes regular reference counting to break down. We rely on the UNIX
6272 * garbage collection to take care of this problem for us.
6274 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6276 unsigned left, total;
6280 left = ctx->nr_user_files;
6282 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6284 ret = __io_sqe_files_scm(ctx, this_files, total);
6288 total += this_files;
6294 while (total < ctx->nr_user_files) {
6295 struct file *file = io_file_from_index(ctx, total);
6305 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6311 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6316 for (i = 0; i < nr_tables; i++) {
6317 struct fixed_file_table *table = &ctx->file_data->table[i];
6318 unsigned this_files;
6320 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6321 table->files = kcalloc(this_files, sizeof(struct file *),
6325 nr_files -= this_files;
6331 for (i = 0; i < nr_tables; i++) {
6332 struct fixed_file_table *table = &ctx->file_data->table[i];
6333 kfree(table->files);
6338 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6340 #if defined(CONFIG_UNIX)
6341 struct sock *sock = ctx->ring_sock->sk;
6342 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6343 struct sk_buff *skb;
6346 __skb_queue_head_init(&list);
6349 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6350 * remove this entry and rearrange the file array.
6352 skb = skb_dequeue(head);
6354 struct scm_fp_list *fp;
6356 fp = UNIXCB(skb).fp;
6357 for (i = 0; i < fp->count; i++) {
6360 if (fp->fp[i] != file)
6363 unix_notinflight(fp->user, fp->fp[i]);
6364 left = fp->count - 1 - i;
6366 memmove(&fp->fp[i], &fp->fp[i + 1],
6367 left * sizeof(struct file *));
6374 __skb_queue_tail(&list, skb);
6384 __skb_queue_tail(&list, skb);
6386 skb = skb_dequeue(head);
6389 if (skb_peek(&list)) {
6390 spin_lock_irq(&head->lock);
6391 while ((skb = __skb_dequeue(&list)) != NULL)
6392 __skb_queue_tail(head, skb);
6393 spin_unlock_irq(&head->lock);
6400 struct io_file_put {
6401 struct list_head list;
6405 static void io_file_put_work(struct work_struct *work)
6407 struct fixed_file_ref_node *ref_node;
6408 struct fixed_file_data *file_data;
6409 struct io_ring_ctx *ctx;
6410 struct io_file_put *pfile, *tmp;
6411 unsigned long flags;
6413 ref_node = container_of(work, struct fixed_file_ref_node, work);
6414 file_data = ref_node->file_data;
6415 ctx = file_data->ctx;
6417 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6418 list_del_init(&pfile->list);
6419 io_ring_file_put(ctx, pfile->file);
6423 spin_lock_irqsave(&file_data->lock, flags);
6424 list_del_init(&ref_node->node);
6425 spin_unlock_irqrestore(&file_data->lock, flags);
6427 percpu_ref_exit(&ref_node->refs);
6429 percpu_ref_put(&file_data->refs);
6432 static void io_file_data_ref_zero(struct percpu_ref *ref)
6434 struct fixed_file_ref_node *ref_node;
6436 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6438 queue_work(system_wq, &ref_node->work);
6441 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6442 struct io_ring_ctx *ctx)
6444 struct fixed_file_ref_node *ref_node;
6446 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6448 return ERR_PTR(-ENOMEM);
6450 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6453 return ERR_PTR(-ENOMEM);
6455 INIT_LIST_HEAD(&ref_node->node);
6456 INIT_LIST_HEAD(&ref_node->file_list);
6457 INIT_WORK(&ref_node->work, io_file_put_work);
6458 ref_node->file_data = ctx->file_data;
6463 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6465 percpu_ref_exit(&ref_node->refs);
6469 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6472 __s32 __user *fds = (__s32 __user *) arg;
6477 struct fixed_file_ref_node *ref_node;
6478 unsigned long flags;
6484 if (nr_args > IORING_MAX_FIXED_FILES)
6487 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6488 if (!ctx->file_data)
6490 ctx->file_data->ctx = ctx;
6491 init_completion(&ctx->file_data->done);
6492 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6493 spin_lock_init(&ctx->file_data->lock);
6495 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6496 ctx->file_data->table = kcalloc(nr_tables,
6497 sizeof(struct fixed_file_table),
6499 if (!ctx->file_data->table) {
6500 kfree(ctx->file_data);
6501 ctx->file_data = NULL;
6505 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6506 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6507 kfree(ctx->file_data->table);
6508 kfree(ctx->file_data);
6509 ctx->file_data = NULL;
6513 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6514 percpu_ref_exit(&ctx->file_data->refs);
6515 kfree(ctx->file_data->table);
6516 kfree(ctx->file_data);
6517 ctx->file_data = NULL;
6521 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6522 struct fixed_file_table *table;
6526 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6528 /* allow sparse sets */
6534 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6535 index = i & IORING_FILE_TABLE_MASK;
6543 * Don't allow io_uring instances to be registered. If UNIX
6544 * isn't enabled, then this causes a reference cycle and this
6545 * instance can never get freed. If UNIX is enabled we'll
6546 * handle it just fine, but there's still no point in allowing
6547 * a ring fd as it doesn't support regular read/write anyway.
6549 if (file->f_op == &io_uring_fops) {
6554 table->files[index] = file;
6558 for (i = 0; i < ctx->nr_user_files; i++) {
6559 file = io_file_from_index(ctx, i);
6563 for (i = 0; i < nr_tables; i++)
6564 kfree(ctx->file_data->table[i].files);
6566 kfree(ctx->file_data->table);
6567 kfree(ctx->file_data);
6568 ctx->file_data = NULL;
6569 ctx->nr_user_files = 0;
6573 ret = io_sqe_files_scm(ctx);
6575 io_sqe_files_unregister(ctx);
6579 ref_node = alloc_fixed_file_ref_node(ctx);
6580 if (IS_ERR(ref_node)) {
6581 io_sqe_files_unregister(ctx);
6582 return PTR_ERR(ref_node);
6585 ctx->file_data->cur_refs = &ref_node->refs;
6586 spin_lock_irqsave(&ctx->file_data->lock, flags);
6587 list_add(&ref_node->node, &ctx->file_data->ref_list);
6588 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6589 percpu_ref_get(&ctx->file_data->refs);
6593 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6596 #if defined(CONFIG_UNIX)
6597 struct sock *sock = ctx->ring_sock->sk;
6598 struct sk_buff_head *head = &sock->sk_receive_queue;
6599 struct sk_buff *skb;
6602 * See if we can merge this file into an existing skb SCM_RIGHTS
6603 * file set. If there's no room, fall back to allocating a new skb
6604 * and filling it in.
6606 spin_lock_irq(&head->lock);
6607 skb = skb_peek(head);
6609 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6611 if (fpl->count < SCM_MAX_FD) {
6612 __skb_unlink(skb, head);
6613 spin_unlock_irq(&head->lock);
6614 fpl->fp[fpl->count] = get_file(file);
6615 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6617 spin_lock_irq(&head->lock);
6618 __skb_queue_head(head, skb);
6623 spin_unlock_irq(&head->lock);
6630 return __io_sqe_files_scm(ctx, 1, index);
6636 static int io_queue_file_removal(struct fixed_file_data *data,
6639 struct io_file_put *pfile;
6640 struct percpu_ref *refs = data->cur_refs;
6641 struct fixed_file_ref_node *ref_node;
6643 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6647 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6649 list_add(&pfile->list, &ref_node->file_list);
6654 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6655 struct io_uring_files_update *up,
6658 struct fixed_file_data *data = ctx->file_data;
6659 struct fixed_file_ref_node *ref_node;
6664 unsigned long flags;
6665 bool needs_switch = false;
6667 if (check_add_overflow(up->offset, nr_args, &done))
6669 if (done > ctx->nr_user_files)
6672 ref_node = alloc_fixed_file_ref_node(ctx);
6673 if (IS_ERR(ref_node))
6674 return PTR_ERR(ref_node);
6677 fds = u64_to_user_ptr(up->fds);
6679 struct fixed_file_table *table;
6683 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6687 i = array_index_nospec(up->offset, ctx->nr_user_files);
6688 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6689 index = i & IORING_FILE_TABLE_MASK;
6690 if (table->files[index]) {
6691 file = io_file_from_index(ctx, index);
6692 err = io_queue_file_removal(data, file);
6695 table->files[index] = NULL;
6696 needs_switch = true;
6705 * Don't allow io_uring instances to be registered. If
6706 * UNIX isn't enabled, then this causes a reference
6707 * cycle and this instance can never get freed. If UNIX
6708 * is enabled we'll handle it just fine, but there's
6709 * still no point in allowing a ring fd as it doesn't
6710 * support regular read/write anyway.
6712 if (file->f_op == &io_uring_fops) {
6717 table->files[index] = file;
6718 err = io_sqe_file_register(ctx, file, i);
6728 percpu_ref_kill(data->cur_refs);
6729 spin_lock_irqsave(&data->lock, flags);
6730 list_add(&ref_node->node, &data->ref_list);
6731 data->cur_refs = &ref_node->refs;
6732 spin_unlock_irqrestore(&data->lock, flags);
6733 percpu_ref_get(&ctx->file_data->refs);
6735 destroy_fixed_file_ref_node(ref_node);
6737 return done ? done : err;
6740 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6743 struct io_uring_files_update up;
6745 if (!ctx->file_data)
6749 if (copy_from_user(&up, arg, sizeof(up)))
6754 return __io_sqe_files_update(ctx, &up, nr_args);
6757 static void io_free_work(struct io_wq_work *work)
6759 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6761 /* Consider that io_steal_work() relies on this ref */
6765 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6766 struct io_uring_params *p)
6768 struct io_wq_data data;
6770 struct io_ring_ctx *ctx_attach;
6771 unsigned int concurrency;
6774 data.user = ctx->user;
6775 data.free_work = io_free_work;
6777 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6778 /* Do QD, or 4 * CPUS, whatever is smallest */
6779 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6781 ctx->io_wq = io_wq_create(concurrency, &data);
6782 if (IS_ERR(ctx->io_wq)) {
6783 ret = PTR_ERR(ctx->io_wq);
6789 f = fdget(p->wq_fd);
6793 if (f.file->f_op != &io_uring_fops) {
6798 ctx_attach = f.file->private_data;
6799 /* @io_wq is protected by holding the fd */
6800 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6805 ctx->io_wq = ctx_attach->io_wq;
6811 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6812 struct io_uring_params *p)
6816 init_waitqueue_head(&ctx->sqo_wait);
6817 mmgrab(current->mm);
6818 ctx->sqo_mm = current->mm;
6820 if (ctx->flags & IORING_SETUP_SQPOLL) {
6822 if (!capable(CAP_SYS_ADMIN))
6825 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6826 if (!ctx->sq_thread_idle)
6827 ctx->sq_thread_idle = HZ;
6829 if (p->flags & IORING_SETUP_SQ_AFF) {
6830 int cpu = p->sq_thread_cpu;
6833 if (cpu >= nr_cpu_ids)
6835 if (!cpu_online(cpu))
6838 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6842 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6845 if (IS_ERR(ctx->sqo_thread)) {
6846 ret = PTR_ERR(ctx->sqo_thread);
6847 ctx->sqo_thread = NULL;
6850 wake_up_process(ctx->sqo_thread);
6851 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6852 /* Can't have SQ_AFF without SQPOLL */
6857 ret = io_init_wq_offload(ctx, p);
6863 io_finish_async(ctx);
6864 mmdrop(ctx->sqo_mm);
6869 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6871 atomic_long_sub(nr_pages, &user->locked_vm);
6874 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6876 unsigned long page_limit, cur_pages, new_pages;
6878 /* Don't allow more pages than we can safely lock */
6879 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6882 cur_pages = atomic_long_read(&user->locked_vm);
6883 new_pages = cur_pages + nr_pages;
6884 if (new_pages > page_limit)
6886 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6887 new_pages) != cur_pages);
6892 static void io_mem_free(void *ptr)
6899 page = virt_to_head_page(ptr);
6900 if (put_page_testzero(page))
6901 free_compound_page(page);
6904 static void *io_mem_alloc(size_t size)
6906 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6909 return (void *) __get_free_pages(gfp_flags, get_order(size));
6912 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6915 struct io_rings *rings;
6916 size_t off, sq_array_size;
6918 off = struct_size(rings, cqes, cq_entries);
6919 if (off == SIZE_MAX)
6923 off = ALIGN(off, SMP_CACHE_BYTES);
6928 sq_array_size = array_size(sizeof(u32), sq_entries);
6929 if (sq_array_size == SIZE_MAX)
6932 if (check_add_overflow(off, sq_array_size, &off))
6941 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6945 pages = (size_t)1 << get_order(
6946 rings_size(sq_entries, cq_entries, NULL));
6947 pages += (size_t)1 << get_order(
6948 array_size(sizeof(struct io_uring_sqe), sq_entries));
6953 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6957 if (!ctx->user_bufs)
6960 for (i = 0; i < ctx->nr_user_bufs; i++) {
6961 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6963 for (j = 0; j < imu->nr_bvecs; j++)
6964 unpin_user_page(imu->bvec[j].bv_page);
6966 if (ctx->account_mem)
6967 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6972 kfree(ctx->user_bufs);
6973 ctx->user_bufs = NULL;
6974 ctx->nr_user_bufs = 0;
6978 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6979 void __user *arg, unsigned index)
6981 struct iovec __user *src;
6983 #ifdef CONFIG_COMPAT
6985 struct compat_iovec __user *ciovs;
6986 struct compat_iovec ciov;
6988 ciovs = (struct compat_iovec __user *) arg;
6989 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
6992 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
6993 dst->iov_len = ciov.iov_len;
6997 src = (struct iovec __user *) arg;
6998 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7003 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7006 struct vm_area_struct **vmas = NULL;
7007 struct page **pages = NULL;
7008 int i, j, got_pages = 0;
7013 if (!nr_args || nr_args > UIO_MAXIOV)
7016 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7018 if (!ctx->user_bufs)
7021 for (i = 0; i < nr_args; i++) {
7022 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7023 unsigned long off, start, end, ubuf;
7028 ret = io_copy_iov(ctx, &iov, arg, i);
7033 * Don't impose further limits on the size and buffer
7034 * constraints here, we'll -EINVAL later when IO is
7035 * submitted if they are wrong.
7038 if (!iov.iov_base || !iov.iov_len)
7041 /* arbitrary limit, but we need something */
7042 if (iov.iov_len > SZ_1G)
7045 ubuf = (unsigned long) iov.iov_base;
7046 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7047 start = ubuf >> PAGE_SHIFT;
7048 nr_pages = end - start;
7050 if (ctx->account_mem) {
7051 ret = io_account_mem(ctx->user, nr_pages);
7057 if (!pages || nr_pages > got_pages) {
7060 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7062 vmas = kvmalloc_array(nr_pages,
7063 sizeof(struct vm_area_struct *),
7065 if (!pages || !vmas) {
7067 if (ctx->account_mem)
7068 io_unaccount_mem(ctx->user, nr_pages);
7071 got_pages = nr_pages;
7074 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7078 if (ctx->account_mem)
7079 io_unaccount_mem(ctx->user, nr_pages);
7084 down_read(¤t->mm->mmap_sem);
7085 pret = pin_user_pages(ubuf, nr_pages,
7086 FOLL_WRITE | FOLL_LONGTERM,
7088 if (pret == nr_pages) {
7089 /* don't support file backed memory */
7090 for (j = 0; j < nr_pages; j++) {
7091 struct vm_area_struct *vma = vmas[j];
7094 !is_file_hugepages(vma->vm_file)) {
7100 ret = pret < 0 ? pret : -EFAULT;
7102 up_read(¤t->mm->mmap_sem);
7105 * if we did partial map, or found file backed vmas,
7106 * release any pages we did get
7109 unpin_user_pages(pages, pret);
7110 if (ctx->account_mem)
7111 io_unaccount_mem(ctx->user, nr_pages);
7116 off = ubuf & ~PAGE_MASK;
7118 for (j = 0; j < nr_pages; j++) {
7121 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7122 imu->bvec[j].bv_page = pages[j];
7123 imu->bvec[j].bv_len = vec_len;
7124 imu->bvec[j].bv_offset = off;
7128 /* store original address for later verification */
7130 imu->len = iov.iov_len;
7131 imu->nr_bvecs = nr_pages;
7133 ctx->nr_user_bufs++;
7141 io_sqe_buffer_unregister(ctx);
7145 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7147 __s32 __user *fds = arg;
7153 if (copy_from_user(&fd, fds, sizeof(*fds)))
7156 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7157 if (IS_ERR(ctx->cq_ev_fd)) {
7158 int ret = PTR_ERR(ctx->cq_ev_fd);
7159 ctx->cq_ev_fd = NULL;
7166 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7168 if (ctx->cq_ev_fd) {
7169 eventfd_ctx_put(ctx->cq_ev_fd);
7170 ctx->cq_ev_fd = NULL;
7177 static int __io_destroy_buffers(int id, void *p, void *data)
7179 struct io_ring_ctx *ctx = data;
7180 struct io_buffer *buf = p;
7182 __io_remove_buffers(ctx, buf, id, -1U);
7186 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7188 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7189 idr_destroy(&ctx->io_buffer_idr);
7192 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7194 io_finish_async(ctx);
7196 mmdrop(ctx->sqo_mm);
7198 io_iopoll_reap_events(ctx);
7199 io_sqe_buffer_unregister(ctx);
7200 io_sqe_files_unregister(ctx);
7201 io_eventfd_unregister(ctx);
7202 io_destroy_buffers(ctx);
7203 idr_destroy(&ctx->personality_idr);
7205 #if defined(CONFIG_UNIX)
7206 if (ctx->ring_sock) {
7207 ctx->ring_sock->file = NULL; /* so that iput() is called */
7208 sock_release(ctx->ring_sock);
7212 io_mem_free(ctx->rings);
7213 io_mem_free(ctx->sq_sqes);
7215 percpu_ref_exit(&ctx->refs);
7216 if (ctx->account_mem)
7217 io_unaccount_mem(ctx->user,
7218 ring_pages(ctx->sq_entries, ctx->cq_entries));
7219 free_uid(ctx->user);
7220 put_cred(ctx->creds);
7221 kfree(ctx->completions);
7222 kfree(ctx->cancel_hash);
7223 kmem_cache_free(req_cachep, ctx->fallback_req);
7227 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7229 struct io_ring_ctx *ctx = file->private_data;
7232 poll_wait(file, &ctx->cq_wait, wait);
7234 * synchronizes with barrier from wq_has_sleeper call in
7238 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7239 ctx->rings->sq_ring_entries)
7240 mask |= EPOLLOUT | EPOLLWRNORM;
7241 if (io_cqring_events(ctx, false))
7242 mask |= EPOLLIN | EPOLLRDNORM;
7247 static int io_uring_fasync(int fd, struct file *file, int on)
7249 struct io_ring_ctx *ctx = file->private_data;
7251 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7254 static int io_remove_personalities(int id, void *p, void *data)
7256 struct io_ring_ctx *ctx = data;
7257 const struct cred *cred;
7259 cred = idr_remove(&ctx->personality_idr, id);
7265 static void io_ring_exit_work(struct work_struct *work)
7267 struct io_ring_ctx *ctx;
7269 ctx = container_of(work, struct io_ring_ctx, exit_work);
7271 io_cqring_overflow_flush(ctx, true);
7273 wait_for_completion(&ctx->completions[0]);
7274 io_ring_ctx_free(ctx);
7277 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7279 mutex_lock(&ctx->uring_lock);
7280 percpu_ref_kill(&ctx->refs);
7281 mutex_unlock(&ctx->uring_lock);
7284 * Wait for sq thread to idle, if we have one. It won't spin on new
7285 * work after we've killed the ctx ref above. This is important to do
7286 * before we cancel existing commands, as the thread could otherwise
7287 * be queueing new work post that. If that's work we need to cancel,
7288 * it could cause shutdown to hang.
7290 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7293 io_kill_timeouts(ctx);
7294 io_poll_remove_all(ctx);
7297 io_wq_cancel_all(ctx->io_wq);
7299 io_iopoll_reap_events(ctx);
7300 /* if we failed setting up the ctx, we might not have any rings */
7302 io_cqring_overflow_flush(ctx, true);
7303 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7304 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7305 queue_work(system_wq, &ctx->exit_work);
7308 static int io_uring_release(struct inode *inode, struct file *file)
7310 struct io_ring_ctx *ctx = file->private_data;
7312 file->private_data = NULL;
7313 io_ring_ctx_wait_and_kill(ctx);
7317 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7318 struct files_struct *files)
7320 struct io_kiocb *req;
7323 while (!list_empty_careful(&ctx->inflight_list)) {
7324 struct io_kiocb *cancel_req = NULL;
7326 spin_lock_irq(&ctx->inflight_lock);
7327 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7328 if (req->work.files != files)
7330 /* req is being completed, ignore */
7331 if (!refcount_inc_not_zero(&req->refs))
7337 prepare_to_wait(&ctx->inflight_wait, &wait,
7338 TASK_UNINTERRUPTIBLE);
7339 spin_unlock_irq(&ctx->inflight_lock);
7341 /* We need to keep going until we don't find a matching req */
7345 if (cancel_req->flags & REQ_F_OVERFLOW) {
7346 spin_lock_irq(&ctx->completion_lock);
7347 list_del(&cancel_req->list);
7348 cancel_req->flags &= ~REQ_F_OVERFLOW;
7349 if (list_empty(&ctx->cq_overflow_list)) {
7350 clear_bit(0, &ctx->sq_check_overflow);
7351 clear_bit(0, &ctx->cq_check_overflow);
7353 spin_unlock_irq(&ctx->completion_lock);
7355 WRITE_ONCE(ctx->rings->cq_overflow,
7356 atomic_inc_return(&ctx->cached_cq_overflow));
7359 * Put inflight ref and overflow ref. If that's
7360 * all we had, then we're done with this request.
7362 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7363 io_put_req(cancel_req);
7368 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7369 io_put_req(cancel_req);
7372 finish_wait(&ctx->inflight_wait, &wait);
7375 static int io_uring_flush(struct file *file, void *data)
7377 struct io_ring_ctx *ctx = file->private_data;
7379 io_uring_cancel_files(ctx, data);
7382 * If the task is going away, cancel work it may have pending
7384 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7385 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7390 static void *io_uring_validate_mmap_request(struct file *file,
7391 loff_t pgoff, size_t sz)
7393 struct io_ring_ctx *ctx = file->private_data;
7394 loff_t offset = pgoff << PAGE_SHIFT;
7399 case IORING_OFF_SQ_RING:
7400 case IORING_OFF_CQ_RING:
7403 case IORING_OFF_SQES:
7407 return ERR_PTR(-EINVAL);
7410 page = virt_to_head_page(ptr);
7411 if (sz > page_size(page))
7412 return ERR_PTR(-EINVAL);
7419 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7421 size_t sz = vma->vm_end - vma->vm_start;
7425 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7427 return PTR_ERR(ptr);
7429 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7430 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7433 #else /* !CONFIG_MMU */
7435 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7437 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7440 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7442 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7445 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7446 unsigned long addr, unsigned long len,
7447 unsigned long pgoff, unsigned long flags)
7451 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7453 return PTR_ERR(ptr);
7455 return (unsigned long) ptr;
7458 #endif /* !CONFIG_MMU */
7460 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7461 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7464 struct io_ring_ctx *ctx;
7469 if (current->task_works)
7472 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7480 if (f.file->f_op != &io_uring_fops)
7484 ctx = f.file->private_data;
7485 if (!percpu_ref_tryget(&ctx->refs))
7489 * For SQ polling, the thread will do all submissions and completions.
7490 * Just return the requested submit count, and wake the thread if
7494 if (ctx->flags & IORING_SETUP_SQPOLL) {
7495 if (!list_empty_careful(&ctx->cq_overflow_list))
7496 io_cqring_overflow_flush(ctx, false);
7497 if (flags & IORING_ENTER_SQ_WAKEUP)
7498 wake_up(&ctx->sqo_wait);
7499 submitted = to_submit;
7500 } else if (to_submit) {
7501 mutex_lock(&ctx->uring_lock);
7502 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7503 mutex_unlock(&ctx->uring_lock);
7505 if (submitted != to_submit)
7508 if (flags & IORING_ENTER_GETEVENTS) {
7509 unsigned nr_events = 0;
7511 min_complete = min(min_complete, ctx->cq_entries);
7514 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7515 * space applications don't need to do io completion events
7516 * polling again, they can rely on io_sq_thread to do polling
7517 * work, which can reduce cpu usage and uring_lock contention.
7519 if (ctx->flags & IORING_SETUP_IOPOLL &&
7520 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7521 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7523 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7528 percpu_ref_put(&ctx->refs);
7531 return submitted ? submitted : ret;
7534 #ifdef CONFIG_PROC_FS
7535 static int io_uring_show_cred(int id, void *p, void *data)
7537 const struct cred *cred = p;
7538 struct seq_file *m = data;
7539 struct user_namespace *uns = seq_user_ns(m);
7540 struct group_info *gi;
7545 seq_printf(m, "%5d\n", id);
7546 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7547 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7548 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7549 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7550 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7551 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7552 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7553 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7554 seq_puts(m, "\n\tGroups:\t");
7555 gi = cred->group_info;
7556 for (g = 0; g < gi->ngroups; g++) {
7557 seq_put_decimal_ull(m, g ? " " : "",
7558 from_kgid_munged(uns, gi->gid[g]));
7560 seq_puts(m, "\n\tCapEff:\t");
7561 cap = cred->cap_effective;
7562 CAP_FOR_EACH_U32(__capi)
7563 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7568 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7572 mutex_lock(&ctx->uring_lock);
7573 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7574 for (i = 0; i < ctx->nr_user_files; i++) {
7575 struct fixed_file_table *table;
7578 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7579 f = table->files[i & IORING_FILE_TABLE_MASK];
7581 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7583 seq_printf(m, "%5u: <none>\n", i);
7585 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7586 for (i = 0; i < ctx->nr_user_bufs; i++) {
7587 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7589 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7590 (unsigned int) buf->len);
7592 if (!idr_is_empty(&ctx->personality_idr)) {
7593 seq_printf(m, "Personalities:\n");
7594 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7596 seq_printf(m, "PollList:\n");
7597 spin_lock_irq(&ctx->completion_lock);
7598 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7599 struct hlist_head *list = &ctx->cancel_hash[i];
7600 struct io_kiocb *req;
7602 hlist_for_each_entry(req, list, hash_node)
7603 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7604 req->task->task_works != NULL);
7606 spin_unlock_irq(&ctx->completion_lock);
7607 mutex_unlock(&ctx->uring_lock);
7610 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7612 struct io_ring_ctx *ctx = f->private_data;
7614 if (percpu_ref_tryget(&ctx->refs)) {
7615 __io_uring_show_fdinfo(ctx, m);
7616 percpu_ref_put(&ctx->refs);
7621 static const struct file_operations io_uring_fops = {
7622 .release = io_uring_release,
7623 .flush = io_uring_flush,
7624 .mmap = io_uring_mmap,
7626 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7627 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7629 .poll = io_uring_poll,
7630 .fasync = io_uring_fasync,
7631 #ifdef CONFIG_PROC_FS
7632 .show_fdinfo = io_uring_show_fdinfo,
7636 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7637 struct io_uring_params *p)
7639 struct io_rings *rings;
7640 size_t size, sq_array_offset;
7642 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7643 if (size == SIZE_MAX)
7646 rings = io_mem_alloc(size);
7651 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7652 rings->sq_ring_mask = p->sq_entries - 1;
7653 rings->cq_ring_mask = p->cq_entries - 1;
7654 rings->sq_ring_entries = p->sq_entries;
7655 rings->cq_ring_entries = p->cq_entries;
7656 ctx->sq_mask = rings->sq_ring_mask;
7657 ctx->cq_mask = rings->cq_ring_mask;
7658 ctx->sq_entries = rings->sq_ring_entries;
7659 ctx->cq_entries = rings->cq_ring_entries;
7661 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7662 if (size == SIZE_MAX) {
7663 io_mem_free(ctx->rings);
7668 ctx->sq_sqes = io_mem_alloc(size);
7669 if (!ctx->sq_sqes) {
7670 io_mem_free(ctx->rings);
7679 * Allocate an anonymous fd, this is what constitutes the application
7680 * visible backing of an io_uring instance. The application mmaps this
7681 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7682 * we have to tie this fd to a socket for file garbage collection purposes.
7684 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7689 #if defined(CONFIG_UNIX)
7690 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7696 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7700 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7701 O_RDWR | O_CLOEXEC);
7704 ret = PTR_ERR(file);
7708 #if defined(CONFIG_UNIX)
7709 ctx->ring_sock->file = file;
7711 fd_install(ret, file);
7714 #if defined(CONFIG_UNIX)
7715 sock_release(ctx->ring_sock);
7716 ctx->ring_sock = NULL;
7721 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7723 struct user_struct *user = NULL;
7724 struct io_ring_ctx *ctx;
7730 if (entries > IORING_MAX_ENTRIES) {
7731 if (!(p->flags & IORING_SETUP_CLAMP))
7733 entries = IORING_MAX_ENTRIES;
7737 * Use twice as many entries for the CQ ring. It's possible for the
7738 * application to drive a higher depth than the size of the SQ ring,
7739 * since the sqes are only used at submission time. This allows for
7740 * some flexibility in overcommitting a bit. If the application has
7741 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7742 * of CQ ring entries manually.
7744 p->sq_entries = roundup_pow_of_two(entries);
7745 if (p->flags & IORING_SETUP_CQSIZE) {
7747 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7748 * to a power-of-two, if it isn't already. We do NOT impose
7749 * any cq vs sq ring sizing.
7751 if (p->cq_entries < p->sq_entries)
7753 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7754 if (!(p->flags & IORING_SETUP_CLAMP))
7756 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7758 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7760 p->cq_entries = 2 * p->sq_entries;
7763 user = get_uid(current_user());
7764 account_mem = !capable(CAP_IPC_LOCK);
7767 ret = io_account_mem(user,
7768 ring_pages(p->sq_entries, p->cq_entries));
7775 ctx = io_ring_ctx_alloc(p);
7778 io_unaccount_mem(user, ring_pages(p->sq_entries,
7783 ctx->compat = in_compat_syscall();
7784 ctx->account_mem = account_mem;
7786 ctx->creds = get_current_cred();
7788 ret = io_allocate_scq_urings(ctx, p);
7792 ret = io_sq_offload_start(ctx, p);
7796 memset(&p->sq_off, 0, sizeof(p->sq_off));
7797 p->sq_off.head = offsetof(struct io_rings, sq.head);
7798 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7799 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7800 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7801 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7802 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7803 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7805 memset(&p->cq_off, 0, sizeof(p->cq_off));
7806 p->cq_off.head = offsetof(struct io_rings, cq.head);
7807 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7808 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7809 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7810 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7811 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7814 * Install ring fd as the very last thing, so we don't risk someone
7815 * having closed it before we finish setup
7817 ret = io_uring_get_fd(ctx);
7821 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7822 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7823 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7824 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7827 io_ring_ctx_wait_and_kill(ctx);
7832 * Sets up an aio uring context, and returns the fd. Applications asks for a
7833 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7834 * params structure passed in.
7836 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7838 struct io_uring_params p;
7842 if (copy_from_user(&p, params, sizeof(p)))
7844 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7849 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7850 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7851 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7854 ret = io_uring_create(entries, &p);
7858 if (copy_to_user(params, &p, sizeof(p)))
7864 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7865 struct io_uring_params __user *, params)
7867 return io_uring_setup(entries, params);
7870 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7872 struct io_uring_probe *p;
7876 size = struct_size(p, ops, nr_args);
7877 if (size == SIZE_MAX)
7879 p = kzalloc(size, GFP_KERNEL);
7884 if (copy_from_user(p, arg, size))
7887 if (memchr_inv(p, 0, size))
7890 p->last_op = IORING_OP_LAST - 1;
7891 if (nr_args > IORING_OP_LAST)
7892 nr_args = IORING_OP_LAST;
7894 for (i = 0; i < nr_args; i++) {
7896 if (!io_op_defs[i].not_supported)
7897 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7902 if (copy_to_user(arg, p, size))
7909 static int io_register_personality(struct io_ring_ctx *ctx)
7911 const struct cred *creds = get_current_cred();
7914 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7915 USHRT_MAX, GFP_KERNEL);
7921 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7923 const struct cred *old_creds;
7925 old_creds = idr_remove(&ctx->personality_idr, id);
7927 put_cred(old_creds);
7934 static bool io_register_op_must_quiesce(int op)
7937 case IORING_UNREGISTER_FILES:
7938 case IORING_REGISTER_FILES_UPDATE:
7939 case IORING_REGISTER_PROBE:
7940 case IORING_REGISTER_PERSONALITY:
7941 case IORING_UNREGISTER_PERSONALITY:
7948 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7949 void __user *arg, unsigned nr_args)
7950 __releases(ctx->uring_lock)
7951 __acquires(ctx->uring_lock)
7956 * We're inside the ring mutex, if the ref is already dying, then
7957 * someone else killed the ctx or is already going through
7958 * io_uring_register().
7960 if (percpu_ref_is_dying(&ctx->refs))
7963 if (io_register_op_must_quiesce(opcode)) {
7964 percpu_ref_kill(&ctx->refs);
7967 * Drop uring mutex before waiting for references to exit. If
7968 * another thread is currently inside io_uring_enter() it might
7969 * need to grab the uring_lock to make progress. If we hold it
7970 * here across the drain wait, then we can deadlock. It's safe
7971 * to drop the mutex here, since no new references will come in
7972 * after we've killed the percpu ref.
7974 mutex_unlock(&ctx->uring_lock);
7975 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7976 mutex_lock(&ctx->uring_lock);
7978 percpu_ref_resurrect(&ctx->refs);
7985 case IORING_REGISTER_BUFFERS:
7986 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7988 case IORING_UNREGISTER_BUFFERS:
7992 ret = io_sqe_buffer_unregister(ctx);
7994 case IORING_REGISTER_FILES:
7995 ret = io_sqe_files_register(ctx, arg, nr_args);
7997 case IORING_UNREGISTER_FILES:
8001 ret = io_sqe_files_unregister(ctx);
8003 case IORING_REGISTER_FILES_UPDATE:
8004 ret = io_sqe_files_update(ctx, arg, nr_args);
8006 case IORING_REGISTER_EVENTFD:
8007 case IORING_REGISTER_EVENTFD_ASYNC:
8011 ret = io_eventfd_register(ctx, arg);
8014 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8015 ctx->eventfd_async = 1;
8017 ctx->eventfd_async = 0;
8019 case IORING_UNREGISTER_EVENTFD:
8023 ret = io_eventfd_unregister(ctx);
8025 case IORING_REGISTER_PROBE:
8027 if (!arg || nr_args > 256)
8029 ret = io_probe(ctx, arg, nr_args);
8031 case IORING_REGISTER_PERSONALITY:
8035 ret = io_register_personality(ctx);
8037 case IORING_UNREGISTER_PERSONALITY:
8041 ret = io_unregister_personality(ctx, nr_args);
8048 if (io_register_op_must_quiesce(opcode)) {
8049 /* bring the ctx back to life */
8050 percpu_ref_reinit(&ctx->refs);
8052 reinit_completion(&ctx->completions[0]);
8057 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8058 void __user *, arg, unsigned int, nr_args)
8060 struct io_ring_ctx *ctx;
8069 if (f.file->f_op != &io_uring_fops)
8072 ctx = f.file->private_data;
8074 mutex_lock(&ctx->uring_lock);
8075 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8076 mutex_unlock(&ctx->uring_lock);
8077 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8078 ctx->cq_ev_fd != NULL, ret);
8084 static int __init io_uring_init(void)
8086 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8087 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8088 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8091 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8092 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8093 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8094 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8095 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8096 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8097 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8098 BUILD_BUG_SQE_ELEM(8, __u64, off);
8099 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8100 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8101 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8102 BUILD_BUG_SQE_ELEM(24, __u32, len);
8103 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8104 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8105 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8106 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8107 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8108 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8109 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8110 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8111 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8112 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8113 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8114 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8115 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8116 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8117 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8118 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8119 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8120 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8122 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8123 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8124 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8127 __initcall(io_uring_init);