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;
332 * First field must be the file pointer in all the
333 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
335 struct io_poll_iocb {
338 struct wait_queue_head *head;
344 struct wait_queue_entry wait;
349 struct file *put_file;
353 struct io_timeout_data {
354 struct io_kiocb *req;
355 struct hrtimer timer;
356 struct timespec64 ts;
357 enum hrtimer_mode mode;
363 struct sockaddr __user *addr;
364 int __user *addr_len;
366 unsigned long nofile;
390 /* NOTE: kiocb has the file as the first member, so don't do it here */
398 struct sockaddr __user *addr;
405 struct user_msghdr __user *msg;
411 struct io_buffer *kbuf;
420 struct filename *filename;
421 struct statx __user *buffer;
423 unsigned long nofile;
426 struct io_files_update {
452 struct epoll_event event;
456 struct file *file_out;
457 struct file *file_in;
464 struct io_provide_buf {
473 struct io_async_connect {
474 struct sockaddr_storage address;
477 struct io_async_msghdr {
478 struct iovec fast_iov[UIO_FASTIOV];
480 struct sockaddr __user *uaddr;
482 struct sockaddr_storage addr;
486 struct iovec fast_iov[UIO_FASTIOV];
492 struct io_async_ctx {
494 struct io_async_rw rw;
495 struct io_async_msghdr msg;
496 struct io_async_connect connect;
497 struct io_timeout_data timeout;
502 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
503 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
504 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
505 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
506 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
507 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
514 REQ_F_IOPOLL_COMPLETED_BIT,
515 REQ_F_LINK_TIMEOUT_BIT,
519 REQ_F_TIMEOUT_NOSEQ_BIT,
520 REQ_F_COMP_LOCKED_BIT,
521 REQ_F_NEED_CLEANUP_BIT,
524 REQ_F_BUFFER_SELECTED_BIT,
526 /* not a real bit, just to check we're not overflowing the space */
532 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
533 /* drain existing IO first */
534 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
536 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
537 /* doesn't sever on completion < 0 */
538 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
540 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
541 /* IOSQE_BUFFER_SELECT */
542 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
544 /* already grabbed next link */
545 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
546 /* fail rest of links */
547 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
548 /* on inflight list */
549 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
550 /* read/write uses file position */
551 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
552 /* must not punt to workers */
553 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
554 /* polled IO has completed */
555 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
556 /* has linked timeout */
557 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
558 /* timeout request */
559 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
561 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
562 /* must be punted even for NONBLOCK */
563 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
564 /* no timeout sequence */
565 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
566 /* completion under lock */
567 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
569 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
570 /* in overflow list */
571 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
572 /* already went through poll handler */
573 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
574 /* buffer already selected */
575 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
579 struct io_poll_iocb poll;
580 struct io_wq_work work;
584 * NOTE! Each of the iocb union members has the file pointer
585 * as the first entry in their struct definition. So you can
586 * access the file pointer through any of the sub-structs,
587 * or directly as just 'ki_filp' in this struct.
593 struct io_poll_iocb poll;
594 struct io_accept accept;
596 struct io_cancel cancel;
597 struct io_timeout timeout;
598 struct io_connect connect;
599 struct io_sr_msg sr_msg;
601 struct io_close close;
602 struct io_files_update files_update;
603 struct io_fadvise fadvise;
604 struct io_madvise madvise;
605 struct io_epoll epoll;
606 struct io_splice splice;
607 struct io_provide_buf pbuf;
610 struct io_async_ctx *io;
611 bool needs_fixed_file;
614 struct io_ring_ctx *ctx;
615 struct list_head list;
618 struct task_struct *task;
624 struct list_head link_list;
626 struct list_head inflight_entry;
628 struct percpu_ref *fixed_file_refs;
632 * Only commands that never go async can use the below fields,
633 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
634 * async armed poll handlers for regular commands. The latter
635 * restore the work, if needed.
638 struct callback_head task_work;
639 struct hlist_node hash_node;
640 struct async_poll *apoll;
643 struct io_wq_work work;
647 #define IO_PLUG_THRESHOLD 2
648 #define IO_IOPOLL_BATCH 8
650 struct io_submit_state {
651 struct blk_plug plug;
654 * io_kiocb alloc cache
656 void *reqs[IO_IOPOLL_BATCH];
657 unsigned int free_reqs;
660 * File reference cache
664 unsigned int has_refs;
665 unsigned int used_refs;
666 unsigned int ios_left;
670 /* needs req->io allocated for deferral/async */
671 unsigned async_ctx : 1;
672 /* needs current->mm setup, does mm access */
673 unsigned needs_mm : 1;
674 /* needs req->file assigned */
675 unsigned needs_file : 1;
676 /* needs req->file assigned IFF fd is >= 0 */
677 unsigned fd_non_neg : 1;
678 /* hash wq insertion if file is a regular file */
679 unsigned hash_reg_file : 1;
680 /* unbound wq insertion if file is a non-regular file */
681 unsigned unbound_nonreg_file : 1;
682 /* opcode is not supported by this kernel */
683 unsigned not_supported : 1;
684 /* needs file table */
685 unsigned file_table : 1;
687 unsigned needs_fs : 1;
688 /* set if opcode supports polled "wait" */
690 unsigned pollout : 1;
691 /* op supports buffer selection */
692 unsigned buffer_select : 1;
695 static const struct io_op_def io_op_defs[] = {
696 [IORING_OP_NOP] = {},
697 [IORING_OP_READV] = {
701 .unbound_nonreg_file = 1,
705 [IORING_OP_WRITEV] = {
710 .unbound_nonreg_file = 1,
713 [IORING_OP_FSYNC] = {
716 [IORING_OP_READ_FIXED] = {
718 .unbound_nonreg_file = 1,
721 [IORING_OP_WRITE_FIXED] = {
724 .unbound_nonreg_file = 1,
727 [IORING_OP_POLL_ADD] = {
729 .unbound_nonreg_file = 1,
731 [IORING_OP_POLL_REMOVE] = {},
732 [IORING_OP_SYNC_FILE_RANGE] = {
735 [IORING_OP_SENDMSG] = {
739 .unbound_nonreg_file = 1,
743 [IORING_OP_RECVMSG] = {
747 .unbound_nonreg_file = 1,
752 [IORING_OP_TIMEOUT] = {
756 [IORING_OP_TIMEOUT_REMOVE] = {},
757 [IORING_OP_ACCEPT] = {
760 .unbound_nonreg_file = 1,
764 [IORING_OP_ASYNC_CANCEL] = {},
765 [IORING_OP_LINK_TIMEOUT] = {
769 [IORING_OP_CONNECT] = {
773 .unbound_nonreg_file = 1,
776 [IORING_OP_FALLOCATE] = {
779 [IORING_OP_OPENAT] = {
785 [IORING_OP_CLOSE] = {
789 [IORING_OP_FILES_UPDATE] = {
793 [IORING_OP_STATX] = {
802 .unbound_nonreg_file = 1,
806 [IORING_OP_WRITE] = {
809 .unbound_nonreg_file = 1,
812 [IORING_OP_FADVISE] = {
815 [IORING_OP_MADVISE] = {
821 .unbound_nonreg_file = 1,
827 .unbound_nonreg_file = 1,
831 [IORING_OP_OPENAT2] = {
837 [IORING_OP_EPOLL_CTL] = {
838 .unbound_nonreg_file = 1,
841 [IORING_OP_SPLICE] = {
844 .unbound_nonreg_file = 1,
846 [IORING_OP_PROVIDE_BUFFERS] = {},
847 [IORING_OP_REMOVE_BUFFERS] = {},
850 static void io_wq_submit_work(struct io_wq_work **workptr);
851 static void io_cqring_fill_event(struct io_kiocb *req, long res);
852 static void io_put_req(struct io_kiocb *req);
853 static void __io_double_put_req(struct io_kiocb *req);
854 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
855 static void io_queue_linked_timeout(struct io_kiocb *req);
856 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
857 struct io_uring_files_update *ip,
859 static int io_grab_files(struct io_kiocb *req);
860 static void io_cleanup_req(struct io_kiocb *req);
861 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
862 int fd, struct file **out_file, bool fixed);
863 static void __io_queue_sqe(struct io_kiocb *req,
864 const struct io_uring_sqe *sqe);
866 static struct kmem_cache *req_cachep;
868 static const struct file_operations io_uring_fops;
870 struct sock *io_uring_get_socket(struct file *file)
872 #if defined(CONFIG_UNIX)
873 if (file->f_op == &io_uring_fops) {
874 struct io_ring_ctx *ctx = file->private_data;
876 return ctx->ring_sock->sk;
881 EXPORT_SYMBOL(io_uring_get_socket);
883 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
885 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
887 complete(&ctx->completions[0]);
890 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
892 struct io_ring_ctx *ctx;
895 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
899 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
900 if (!ctx->fallback_req)
903 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
904 if (!ctx->completions)
908 * Use 5 bits less than the max cq entries, that should give us around
909 * 32 entries per hash list if totally full and uniformly spread.
911 hash_bits = ilog2(p->cq_entries);
915 ctx->cancel_hash_bits = hash_bits;
916 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
918 if (!ctx->cancel_hash)
920 __hash_init(ctx->cancel_hash, 1U << hash_bits);
922 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
923 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
926 ctx->flags = p->flags;
927 init_waitqueue_head(&ctx->cq_wait);
928 INIT_LIST_HEAD(&ctx->cq_overflow_list);
929 init_completion(&ctx->completions[0]);
930 init_completion(&ctx->completions[1]);
931 idr_init(&ctx->io_buffer_idr);
932 idr_init(&ctx->personality_idr);
933 mutex_init(&ctx->uring_lock);
934 init_waitqueue_head(&ctx->wait);
935 spin_lock_init(&ctx->completion_lock);
936 INIT_LIST_HEAD(&ctx->poll_list);
937 INIT_LIST_HEAD(&ctx->defer_list);
938 INIT_LIST_HEAD(&ctx->timeout_list);
939 init_waitqueue_head(&ctx->inflight_wait);
940 spin_lock_init(&ctx->inflight_lock);
941 INIT_LIST_HEAD(&ctx->inflight_list);
944 if (ctx->fallback_req)
945 kmem_cache_free(req_cachep, ctx->fallback_req);
946 kfree(ctx->completions);
947 kfree(ctx->cancel_hash);
952 static inline bool __req_need_defer(struct io_kiocb *req)
954 struct io_ring_ctx *ctx = req->ctx;
956 return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped
957 + atomic_read(&ctx->cached_cq_overflow);
960 static inline bool req_need_defer(struct io_kiocb *req)
962 if (unlikely(req->flags & REQ_F_IO_DRAIN))
963 return __req_need_defer(req);
968 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
970 struct io_kiocb *req;
972 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
973 if (req && !req_need_defer(req)) {
974 list_del_init(&req->list);
981 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
983 struct io_kiocb *req;
985 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
987 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
989 if (!__req_need_defer(req)) {
990 list_del_init(&req->list);
998 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1000 struct io_rings *rings = ctx->rings;
1002 /* order cqe stores with ring update */
1003 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1005 if (wq_has_sleeper(&ctx->cq_wait)) {
1006 wake_up_interruptible(&ctx->cq_wait);
1007 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1011 static inline void io_req_work_grab_env(struct io_kiocb *req,
1012 const struct io_op_def *def)
1014 if (!req->work.mm && def->needs_mm) {
1015 mmgrab(current->mm);
1016 req->work.mm = current->mm;
1018 if (!req->work.creds)
1019 req->work.creds = get_current_cred();
1020 if (!req->work.fs && def->needs_fs) {
1021 spin_lock(¤t->fs->lock);
1022 if (!current->fs->in_exec) {
1023 req->work.fs = current->fs;
1024 req->work.fs->users++;
1026 req->work.flags |= IO_WQ_WORK_CANCEL;
1028 spin_unlock(¤t->fs->lock);
1030 if (!req->work.task_pid)
1031 req->work.task_pid = task_pid_vnr(current);
1034 static inline void io_req_work_drop_env(struct io_kiocb *req)
1037 mmdrop(req->work.mm);
1038 req->work.mm = NULL;
1040 if (req->work.creds) {
1041 put_cred(req->work.creds);
1042 req->work.creds = NULL;
1045 struct fs_struct *fs = req->work.fs;
1047 spin_lock(&req->work.fs->lock);
1050 spin_unlock(&req->work.fs->lock);
1056 static inline void io_prep_async_work(struct io_kiocb *req,
1057 struct io_kiocb **link)
1059 const struct io_op_def *def = &io_op_defs[req->opcode];
1061 if (req->flags & REQ_F_ISREG) {
1062 if (def->hash_reg_file)
1063 io_wq_hash_work(&req->work, file_inode(req->file));
1065 if (def->unbound_nonreg_file)
1066 req->work.flags |= IO_WQ_WORK_UNBOUND;
1069 io_req_work_grab_env(req, def);
1071 *link = io_prep_linked_timeout(req);
1074 static inline void io_queue_async_work(struct io_kiocb *req)
1076 struct io_ring_ctx *ctx = req->ctx;
1077 struct io_kiocb *link;
1079 io_prep_async_work(req, &link);
1081 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1082 &req->work, req->flags);
1083 io_wq_enqueue(ctx->io_wq, &req->work);
1086 io_queue_linked_timeout(link);
1089 static void io_kill_timeout(struct io_kiocb *req)
1093 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1095 atomic_inc(&req->ctx->cq_timeouts);
1096 list_del_init(&req->list);
1097 req->flags |= REQ_F_COMP_LOCKED;
1098 io_cqring_fill_event(req, 0);
1103 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1105 struct io_kiocb *req, *tmp;
1107 spin_lock_irq(&ctx->completion_lock);
1108 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1109 io_kill_timeout(req);
1110 spin_unlock_irq(&ctx->completion_lock);
1113 static void io_commit_cqring(struct io_ring_ctx *ctx)
1115 struct io_kiocb *req;
1117 while ((req = io_get_timeout_req(ctx)) != NULL)
1118 io_kill_timeout(req);
1120 __io_commit_cqring(ctx);
1122 while ((req = io_get_deferred_req(ctx)) != NULL)
1123 io_queue_async_work(req);
1126 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1128 struct io_rings *rings = ctx->rings;
1131 tail = ctx->cached_cq_tail;
1133 * writes to the cq entry need to come after reading head; the
1134 * control dependency is enough as we're using WRITE_ONCE to
1137 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1140 ctx->cached_cq_tail++;
1141 return &rings->cqes[tail & ctx->cq_mask];
1144 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1148 if (!ctx->eventfd_async)
1150 return io_wq_current_is_worker();
1153 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1155 if (waitqueue_active(&ctx->wait))
1156 wake_up(&ctx->wait);
1157 if (waitqueue_active(&ctx->sqo_wait))
1158 wake_up(&ctx->sqo_wait);
1159 if (io_should_trigger_evfd(ctx))
1160 eventfd_signal(ctx->cq_ev_fd, 1);
1163 /* Returns true if there are no backlogged entries after the flush */
1164 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1166 struct io_rings *rings = ctx->rings;
1167 struct io_uring_cqe *cqe;
1168 struct io_kiocb *req;
1169 unsigned long flags;
1173 if (list_empty_careful(&ctx->cq_overflow_list))
1175 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1176 rings->cq_ring_entries))
1180 spin_lock_irqsave(&ctx->completion_lock, flags);
1182 /* if force is set, the ring is going away. always drop after that */
1184 ctx->cq_overflow_flushed = 1;
1187 while (!list_empty(&ctx->cq_overflow_list)) {
1188 cqe = io_get_cqring(ctx);
1192 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1194 list_move(&req->list, &list);
1195 req->flags &= ~REQ_F_OVERFLOW;
1197 WRITE_ONCE(cqe->user_data, req->user_data);
1198 WRITE_ONCE(cqe->res, req->result);
1199 WRITE_ONCE(cqe->flags, req->cflags);
1201 WRITE_ONCE(ctx->rings->cq_overflow,
1202 atomic_inc_return(&ctx->cached_cq_overflow));
1206 io_commit_cqring(ctx);
1208 clear_bit(0, &ctx->sq_check_overflow);
1209 clear_bit(0, &ctx->cq_check_overflow);
1211 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1212 io_cqring_ev_posted(ctx);
1214 while (!list_empty(&list)) {
1215 req = list_first_entry(&list, struct io_kiocb, list);
1216 list_del(&req->list);
1223 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1225 struct io_ring_ctx *ctx = req->ctx;
1226 struct io_uring_cqe *cqe;
1228 trace_io_uring_complete(ctx, req->user_data, res);
1231 * If we can't get a cq entry, userspace overflowed the
1232 * submission (by quite a lot). Increment the overflow count in
1235 cqe = io_get_cqring(ctx);
1237 WRITE_ONCE(cqe->user_data, req->user_data);
1238 WRITE_ONCE(cqe->res, res);
1239 WRITE_ONCE(cqe->flags, cflags);
1240 } else if (ctx->cq_overflow_flushed) {
1241 WRITE_ONCE(ctx->rings->cq_overflow,
1242 atomic_inc_return(&ctx->cached_cq_overflow));
1244 if (list_empty(&ctx->cq_overflow_list)) {
1245 set_bit(0, &ctx->sq_check_overflow);
1246 set_bit(0, &ctx->cq_check_overflow);
1248 req->flags |= REQ_F_OVERFLOW;
1249 refcount_inc(&req->refs);
1251 req->cflags = cflags;
1252 list_add_tail(&req->list, &ctx->cq_overflow_list);
1256 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1258 __io_cqring_fill_event(req, res, 0);
1261 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1263 struct io_ring_ctx *ctx = req->ctx;
1264 unsigned long flags;
1266 spin_lock_irqsave(&ctx->completion_lock, flags);
1267 __io_cqring_fill_event(req, res, cflags);
1268 io_commit_cqring(ctx);
1269 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1271 io_cqring_ev_posted(ctx);
1274 static void io_cqring_add_event(struct io_kiocb *req, long res)
1276 __io_cqring_add_event(req, res, 0);
1279 static inline bool io_is_fallback_req(struct io_kiocb *req)
1281 return req == (struct io_kiocb *)
1282 ((unsigned long) req->ctx->fallback_req & ~1UL);
1285 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1287 struct io_kiocb *req;
1289 req = ctx->fallback_req;
1290 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1296 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
1297 struct io_submit_state *state)
1299 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1300 struct io_kiocb *req;
1303 req = kmem_cache_alloc(req_cachep, gfp);
1306 } else if (!state->free_reqs) {
1310 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1311 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1314 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1315 * retry single alloc to be on the safe side.
1317 if (unlikely(ret <= 0)) {
1318 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1319 if (!state->reqs[0])
1323 state->free_reqs = ret - 1;
1324 req = state->reqs[ret - 1];
1327 req = state->reqs[state->free_reqs];
1335 /* one is dropped after submission, the other at completion */
1336 refcount_set(&req->refs, 2);
1339 INIT_IO_WORK(&req->work, io_wq_submit_work);
1342 req = io_get_fallback_req(ctx);
1348 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1352 percpu_ref_put(req->fixed_file_refs);
1357 static void __io_req_do_free(struct io_kiocb *req)
1359 if (likely(!io_is_fallback_req(req)))
1360 kmem_cache_free(req_cachep, req);
1362 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1365 static void __io_req_aux_free(struct io_kiocb *req)
1367 if (req->flags & REQ_F_NEED_CLEANUP)
1368 io_cleanup_req(req);
1372 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1374 put_task_struct(req->task);
1376 io_req_work_drop_env(req);
1379 static void __io_free_req(struct io_kiocb *req)
1381 __io_req_aux_free(req);
1383 if (req->flags & REQ_F_INFLIGHT) {
1384 struct io_ring_ctx *ctx = req->ctx;
1385 unsigned long flags;
1387 spin_lock_irqsave(&ctx->inflight_lock, flags);
1388 list_del(&req->inflight_entry);
1389 if (waitqueue_active(&ctx->inflight_wait))
1390 wake_up(&ctx->inflight_wait);
1391 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1394 percpu_ref_put(&req->ctx->refs);
1395 __io_req_do_free(req);
1399 void *reqs[IO_IOPOLL_BATCH];
1404 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1408 if (rb->need_iter) {
1409 int i, inflight = 0;
1410 unsigned long flags;
1412 for (i = 0; i < rb->to_free; i++) {
1413 struct io_kiocb *req = rb->reqs[i];
1415 if (req->flags & REQ_F_FIXED_FILE) {
1417 percpu_ref_put(req->fixed_file_refs);
1419 if (req->flags & REQ_F_INFLIGHT)
1421 __io_req_aux_free(req);
1426 spin_lock_irqsave(&ctx->inflight_lock, flags);
1427 for (i = 0; i < rb->to_free; i++) {
1428 struct io_kiocb *req = rb->reqs[i];
1430 if (req->flags & REQ_F_INFLIGHT) {
1431 list_del(&req->inflight_entry);
1436 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1438 if (waitqueue_active(&ctx->inflight_wait))
1439 wake_up(&ctx->inflight_wait);
1442 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1443 percpu_ref_put_many(&ctx->refs, rb->to_free);
1444 rb->to_free = rb->need_iter = 0;
1447 static bool io_link_cancel_timeout(struct io_kiocb *req)
1449 struct io_ring_ctx *ctx = req->ctx;
1452 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1454 io_cqring_fill_event(req, -ECANCELED);
1455 io_commit_cqring(ctx);
1456 req->flags &= ~REQ_F_LINK;
1464 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1466 struct io_ring_ctx *ctx = req->ctx;
1467 bool wake_ev = false;
1469 /* Already got next link */
1470 if (req->flags & REQ_F_LINK_NEXT)
1474 * The list should never be empty when we are called here. But could
1475 * potentially happen if the chain is messed up, check to be on the
1478 while (!list_empty(&req->link_list)) {
1479 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1480 struct io_kiocb, link_list);
1482 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1483 (nxt->flags & REQ_F_TIMEOUT))) {
1484 list_del_init(&nxt->link_list);
1485 wake_ev |= io_link_cancel_timeout(nxt);
1486 req->flags &= ~REQ_F_LINK_TIMEOUT;
1490 list_del_init(&req->link_list);
1491 if (!list_empty(&nxt->link_list))
1492 nxt->flags |= REQ_F_LINK;
1497 req->flags |= REQ_F_LINK_NEXT;
1499 io_cqring_ev_posted(ctx);
1503 * Called if REQ_F_LINK is set, and we fail the head request
1505 static void io_fail_links(struct io_kiocb *req)
1507 struct io_ring_ctx *ctx = req->ctx;
1508 unsigned long flags;
1510 spin_lock_irqsave(&ctx->completion_lock, flags);
1512 while (!list_empty(&req->link_list)) {
1513 struct io_kiocb *link = list_first_entry(&req->link_list,
1514 struct io_kiocb, link_list);
1516 list_del_init(&link->link_list);
1517 trace_io_uring_fail_link(req, link);
1519 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1520 link->opcode == IORING_OP_LINK_TIMEOUT) {
1521 io_link_cancel_timeout(link);
1523 io_cqring_fill_event(link, -ECANCELED);
1524 __io_double_put_req(link);
1526 req->flags &= ~REQ_F_LINK_TIMEOUT;
1529 io_commit_cqring(ctx);
1530 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1531 io_cqring_ev_posted(ctx);
1534 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1536 if (likely(!(req->flags & REQ_F_LINK)))
1540 * If LINK is set, we have dependent requests in this chain. If we
1541 * didn't fail this request, queue the first one up, moving any other
1542 * dependencies to the next request. In case of failure, fail the rest
1545 if (req->flags & REQ_F_FAIL_LINK) {
1547 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1548 REQ_F_LINK_TIMEOUT) {
1549 struct io_ring_ctx *ctx = req->ctx;
1550 unsigned long flags;
1553 * If this is a timeout link, we could be racing with the
1554 * timeout timer. Grab the completion lock for this case to
1555 * protect against that.
1557 spin_lock_irqsave(&ctx->completion_lock, flags);
1558 io_req_link_next(req, nxt);
1559 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1561 io_req_link_next(req, nxt);
1565 static void io_free_req(struct io_kiocb *req)
1567 struct io_kiocb *nxt = NULL;
1569 io_req_find_next(req, &nxt);
1573 io_queue_async_work(nxt);
1576 static void io_link_work_cb(struct io_wq_work **workptr)
1578 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1579 struct io_kiocb *link;
1581 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1582 io_queue_linked_timeout(link);
1583 io_wq_submit_work(workptr);
1586 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1588 struct io_kiocb *link;
1589 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1591 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1592 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1594 *workptr = &nxt->work;
1595 link = io_prep_linked_timeout(nxt);
1597 nxt->work.func = io_link_work_cb;
1601 * Drop reference to request, return next in chain (if there is one) if this
1602 * was the last reference to this request.
1604 __attribute__((nonnull))
1605 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1607 if (refcount_dec_and_test(&req->refs)) {
1608 io_req_find_next(req, nxtptr);
1613 static void io_put_req(struct io_kiocb *req)
1615 if (refcount_dec_and_test(&req->refs))
1619 static void io_steal_work(struct io_kiocb *req,
1620 struct io_wq_work **workptr)
1623 * It's in an io-wq worker, so there always should be at least
1624 * one reference, which will be dropped in io_put_work() just
1625 * after the current handler returns.
1627 * It also means, that if the counter dropped to 1, then there is
1628 * no asynchronous users left, so it's safe to steal the next work.
1630 if (refcount_read(&req->refs) == 1) {
1631 struct io_kiocb *nxt = NULL;
1633 io_req_find_next(req, &nxt);
1635 io_wq_assign_next(workptr, nxt);
1640 * Must only be used if we don't need to care about links, usually from
1641 * within the completion handling itself.
1643 static void __io_double_put_req(struct io_kiocb *req)
1645 /* drop both submit and complete references */
1646 if (refcount_sub_and_test(2, &req->refs))
1650 static void io_double_put_req(struct io_kiocb *req)
1652 /* drop both submit and complete references */
1653 if (refcount_sub_and_test(2, &req->refs))
1657 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1659 struct io_rings *rings = ctx->rings;
1661 if (test_bit(0, &ctx->cq_check_overflow)) {
1663 * noflush == true is from the waitqueue handler, just ensure
1664 * we wake up the task, and the next invocation will flush the
1665 * entries. We cannot safely to it from here.
1667 if (noflush && !list_empty(&ctx->cq_overflow_list))
1670 io_cqring_overflow_flush(ctx, false);
1673 /* See comment at the top of this file */
1675 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1678 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1680 struct io_rings *rings = ctx->rings;
1682 /* make sure SQ entry isn't read before tail */
1683 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1686 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1688 if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req))
1691 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1694 rb->reqs[rb->to_free++] = req;
1695 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1696 io_free_req_many(req->ctx, rb);
1700 static int io_put_kbuf(struct io_kiocb *req)
1702 struct io_buffer *kbuf;
1705 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1706 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1707 cflags |= IORING_CQE_F_BUFFER;
1714 * Find and free completed poll iocbs
1716 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1717 struct list_head *done)
1719 struct req_batch rb;
1720 struct io_kiocb *req;
1722 rb.to_free = rb.need_iter = 0;
1723 while (!list_empty(done)) {
1726 req = list_first_entry(done, struct io_kiocb, list);
1727 list_del(&req->list);
1729 if (req->flags & REQ_F_BUFFER_SELECTED)
1730 cflags = io_put_kbuf(req);
1732 __io_cqring_fill_event(req, req->result, cflags);
1735 if (refcount_dec_and_test(&req->refs) &&
1736 !io_req_multi_free(&rb, req))
1740 io_commit_cqring(ctx);
1741 if (ctx->flags & IORING_SETUP_SQPOLL)
1742 io_cqring_ev_posted(ctx);
1743 io_free_req_many(ctx, &rb);
1746 static void io_iopoll_queue(struct list_head *again)
1748 struct io_kiocb *req;
1751 req = list_first_entry(again, struct io_kiocb, list);
1752 list_del(&req->list);
1753 refcount_inc(&req->refs);
1754 io_queue_async_work(req);
1755 } while (!list_empty(again));
1758 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1761 struct io_kiocb *req, *tmp;
1768 * Only spin for completions if we don't have multiple devices hanging
1769 * off our complete list, and we're under the requested amount.
1771 spin = !ctx->poll_multi_file && *nr_events < min;
1774 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1775 struct kiocb *kiocb = &req->rw.kiocb;
1778 * Move completed and retryable entries to our local lists.
1779 * If we find a request that requires polling, break out
1780 * and complete those lists first, if we have entries there.
1782 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1783 list_move_tail(&req->list, &done);
1786 if (!list_empty(&done))
1789 if (req->result == -EAGAIN) {
1790 list_move_tail(&req->list, &again);
1793 if (!list_empty(&again))
1796 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1805 if (!list_empty(&done))
1806 io_iopoll_complete(ctx, nr_events, &done);
1808 if (!list_empty(&again))
1809 io_iopoll_queue(&again);
1815 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1816 * non-spinning poll check - we'll still enter the driver poll loop, but only
1817 * as a non-spinning completion check.
1819 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1822 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1825 ret = io_do_iopoll(ctx, nr_events, min);
1828 if (!min || *nr_events >= min)
1836 * We can't just wait for polled events to come to us, we have to actively
1837 * find and complete them.
1839 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1841 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1844 mutex_lock(&ctx->uring_lock);
1845 while (!list_empty(&ctx->poll_list)) {
1846 unsigned int nr_events = 0;
1848 io_iopoll_getevents(ctx, &nr_events, 1);
1851 * Ensure we allow local-to-the-cpu processing to take place,
1852 * in this case we need to ensure that we reap all events.
1856 mutex_unlock(&ctx->uring_lock);
1859 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1862 int iters = 0, ret = 0;
1865 * We disallow the app entering submit/complete with polling, but we
1866 * still need to lock the ring to prevent racing with polled issue
1867 * that got punted to a workqueue.
1869 mutex_lock(&ctx->uring_lock);
1874 * Don't enter poll loop if we already have events pending.
1875 * If we do, we can potentially be spinning for commands that
1876 * already triggered a CQE (eg in error).
1878 if (io_cqring_events(ctx, false))
1882 * If a submit got punted to a workqueue, we can have the
1883 * application entering polling for a command before it gets
1884 * issued. That app will hold the uring_lock for the duration
1885 * of the poll right here, so we need to take a breather every
1886 * now and then to ensure that the issue has a chance to add
1887 * the poll to the issued list. Otherwise we can spin here
1888 * forever, while the workqueue is stuck trying to acquire the
1891 if (!(++iters & 7)) {
1892 mutex_unlock(&ctx->uring_lock);
1893 mutex_lock(&ctx->uring_lock);
1896 if (*nr_events < min)
1897 tmin = min - *nr_events;
1899 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1903 } while (min && !*nr_events && !need_resched());
1905 mutex_unlock(&ctx->uring_lock);
1909 static void kiocb_end_write(struct io_kiocb *req)
1912 * Tell lockdep we inherited freeze protection from submission
1915 if (req->flags & REQ_F_ISREG) {
1916 struct inode *inode = file_inode(req->file);
1918 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1920 file_end_write(req->file);
1923 static inline void req_set_fail_links(struct io_kiocb *req)
1925 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1926 req->flags |= REQ_F_FAIL_LINK;
1929 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1931 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1934 if (kiocb->ki_flags & IOCB_WRITE)
1935 kiocb_end_write(req);
1937 if (res != req->result)
1938 req_set_fail_links(req);
1939 if (req->flags & REQ_F_BUFFER_SELECTED)
1940 cflags = io_put_kbuf(req);
1941 __io_cqring_add_event(req, res, cflags);
1944 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1946 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1948 io_complete_rw_common(kiocb, res);
1952 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1954 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1956 if (kiocb->ki_flags & IOCB_WRITE)
1957 kiocb_end_write(req);
1959 if (res != req->result)
1960 req_set_fail_links(req);
1963 req->flags |= REQ_F_IOPOLL_COMPLETED;
1967 * After the iocb has been issued, it's safe to be found on the poll list.
1968 * Adding the kiocb to the list AFTER submission ensures that we don't
1969 * find it from a io_iopoll_getevents() thread before the issuer is done
1970 * accessing the kiocb cookie.
1972 static void io_iopoll_req_issued(struct io_kiocb *req)
1974 struct io_ring_ctx *ctx = req->ctx;
1977 * Track whether we have multiple files in our lists. This will impact
1978 * how we do polling eventually, not spinning if we're on potentially
1979 * different devices.
1981 if (list_empty(&ctx->poll_list)) {
1982 ctx->poll_multi_file = false;
1983 } else if (!ctx->poll_multi_file) {
1984 struct io_kiocb *list_req;
1986 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1988 if (list_req->file != req->file)
1989 ctx->poll_multi_file = true;
1993 * For fast devices, IO may have already completed. If it has, add
1994 * it to the front so we find it first.
1996 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1997 list_add(&req->list, &ctx->poll_list);
1999 list_add_tail(&req->list, &ctx->poll_list);
2001 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
2002 wq_has_sleeper(&ctx->sqo_wait))
2003 wake_up(&ctx->sqo_wait);
2006 static void io_file_put(struct io_submit_state *state)
2009 int diff = state->has_refs - state->used_refs;
2012 fput_many(state->file, diff);
2018 * Get as many references to a file as we have IOs left in this submission,
2019 * assuming most submissions are for one file, or at least that each file
2020 * has more than one submission.
2022 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2028 if (state->fd == fd) {
2035 state->file = fget_many(fd, state->ios_left);
2040 state->has_refs = state->ios_left;
2041 state->used_refs = 1;
2047 * If we tracked the file through the SCM inflight mechanism, we could support
2048 * any file. For now, just ensure that anything potentially problematic is done
2051 static bool io_file_supports_async(struct file *file)
2053 umode_t mode = file_inode(file)->i_mode;
2055 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2057 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2063 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2064 bool force_nonblock)
2066 struct io_ring_ctx *ctx = req->ctx;
2067 struct kiocb *kiocb = &req->rw.kiocb;
2071 if (S_ISREG(file_inode(req->file)->i_mode))
2072 req->flags |= REQ_F_ISREG;
2074 kiocb->ki_pos = READ_ONCE(sqe->off);
2075 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2076 req->flags |= REQ_F_CUR_POS;
2077 kiocb->ki_pos = req->file->f_pos;
2079 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2080 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2081 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2085 ioprio = READ_ONCE(sqe->ioprio);
2087 ret = ioprio_check_cap(ioprio);
2091 kiocb->ki_ioprio = ioprio;
2093 kiocb->ki_ioprio = get_current_ioprio();
2095 /* don't allow async punt if RWF_NOWAIT was requested */
2096 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2097 (req->file->f_flags & O_NONBLOCK))
2098 req->flags |= REQ_F_NOWAIT;
2101 kiocb->ki_flags |= IOCB_NOWAIT;
2103 if (ctx->flags & IORING_SETUP_IOPOLL) {
2104 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2105 !kiocb->ki_filp->f_op->iopoll)
2108 kiocb->ki_flags |= IOCB_HIPRI;
2109 kiocb->ki_complete = io_complete_rw_iopoll;
2112 if (kiocb->ki_flags & IOCB_HIPRI)
2114 kiocb->ki_complete = io_complete_rw;
2117 req->rw.addr = READ_ONCE(sqe->addr);
2118 req->rw.len = READ_ONCE(sqe->len);
2119 /* we own ->private, reuse it for the buffer index / buffer ID */
2120 req->rw.kiocb.private = (void *) (unsigned long)
2121 READ_ONCE(sqe->buf_index);
2125 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2131 case -ERESTARTNOINTR:
2132 case -ERESTARTNOHAND:
2133 case -ERESTART_RESTARTBLOCK:
2135 * We can't just restart the syscall, since previously
2136 * submitted sqes may already be in progress. Just fail this
2142 kiocb->ki_complete(kiocb, ret, 0);
2146 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2148 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2150 if (req->flags & REQ_F_CUR_POS)
2151 req->file->f_pos = kiocb->ki_pos;
2152 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2153 io_complete_rw(kiocb, ret, 0);
2155 io_rw_done(kiocb, ret);
2158 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2159 struct iov_iter *iter)
2161 struct io_ring_ctx *ctx = req->ctx;
2162 size_t len = req->rw.len;
2163 struct io_mapped_ubuf *imu;
2164 unsigned index, buf_index;
2168 /* attempt to use fixed buffers without having provided iovecs */
2169 if (unlikely(!ctx->user_bufs))
2172 buf_index = (unsigned long) req->rw.kiocb.private;
2173 if (unlikely(buf_index >= ctx->nr_user_bufs))
2176 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2177 imu = &ctx->user_bufs[index];
2178 buf_addr = req->rw.addr;
2181 if (buf_addr + len < buf_addr)
2183 /* not inside the mapped region */
2184 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2188 * May not be a start of buffer, set size appropriately
2189 * and advance us to the beginning.
2191 offset = buf_addr - imu->ubuf;
2192 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2196 * Don't use iov_iter_advance() here, as it's really slow for
2197 * using the latter parts of a big fixed buffer - it iterates
2198 * over each segment manually. We can cheat a bit here, because
2201 * 1) it's a BVEC iter, we set it up
2202 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2203 * first and last bvec
2205 * So just find our index, and adjust the iterator afterwards.
2206 * If the offset is within the first bvec (or the whole first
2207 * bvec, just use iov_iter_advance(). This makes it easier
2208 * since we can just skip the first segment, which may not
2209 * be PAGE_SIZE aligned.
2211 const struct bio_vec *bvec = imu->bvec;
2213 if (offset <= bvec->bv_len) {
2214 iov_iter_advance(iter, offset);
2216 unsigned long seg_skip;
2218 /* skip first vec */
2219 offset -= bvec->bv_len;
2220 seg_skip = 1 + (offset >> PAGE_SHIFT);
2222 iter->bvec = bvec + seg_skip;
2223 iter->nr_segs -= seg_skip;
2224 iter->count -= bvec->bv_len + offset;
2225 iter->iov_offset = offset & ~PAGE_MASK;
2232 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2235 mutex_unlock(&ctx->uring_lock);
2238 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2241 * "Normal" inline submissions always hold the uring_lock, since we
2242 * grab it from the system call. Same is true for the SQPOLL offload.
2243 * The only exception is when we've detached the request and issue it
2244 * from an async worker thread, grab the lock for that case.
2247 mutex_lock(&ctx->uring_lock);
2250 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2251 int bgid, struct io_buffer *kbuf,
2254 struct io_buffer *head;
2256 if (req->flags & REQ_F_BUFFER_SELECTED)
2259 io_ring_submit_lock(req->ctx, needs_lock);
2261 lockdep_assert_held(&req->ctx->uring_lock);
2263 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2265 if (!list_empty(&head->list)) {
2266 kbuf = list_last_entry(&head->list, struct io_buffer,
2268 list_del(&kbuf->list);
2271 idr_remove(&req->ctx->io_buffer_idr, bgid);
2273 if (*len > kbuf->len)
2276 kbuf = ERR_PTR(-ENOBUFS);
2279 io_ring_submit_unlock(req->ctx, needs_lock);
2284 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2287 struct io_buffer *kbuf;
2290 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2291 bgid = (int) (unsigned long) req->rw.kiocb.private;
2292 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2295 req->rw.addr = (u64) (unsigned long) kbuf;
2296 req->flags |= REQ_F_BUFFER_SELECTED;
2297 return u64_to_user_ptr(kbuf->addr);
2300 #ifdef CONFIG_COMPAT
2301 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2304 struct compat_iovec __user *uiov;
2305 compat_ssize_t clen;
2309 uiov = u64_to_user_ptr(req->rw.addr);
2310 if (!access_ok(uiov, sizeof(*uiov)))
2312 if (__get_user(clen, &uiov->iov_len))
2318 buf = io_rw_buffer_select(req, &len, needs_lock);
2320 return PTR_ERR(buf);
2321 iov[0].iov_base = buf;
2322 iov[0].iov_len = (compat_size_t) len;
2327 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2330 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2334 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2337 len = iov[0].iov_len;
2340 buf = io_rw_buffer_select(req, &len, needs_lock);
2342 return PTR_ERR(buf);
2343 iov[0].iov_base = buf;
2344 iov[0].iov_len = len;
2348 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2351 if (req->flags & REQ_F_BUFFER_SELECTED)
2355 else if (req->rw.len > 1)
2358 #ifdef CONFIG_COMPAT
2359 if (req->ctx->compat)
2360 return io_compat_import(req, iov, needs_lock);
2363 return __io_iov_buffer_select(req, iov, needs_lock);
2366 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2367 struct iovec **iovec, struct iov_iter *iter,
2370 void __user *buf = u64_to_user_ptr(req->rw.addr);
2371 size_t sqe_len = req->rw.len;
2375 opcode = req->opcode;
2376 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2378 return io_import_fixed(req, rw, iter);
2381 /* buffer index only valid with fixed read/write, or buffer select */
2382 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2385 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2386 if (req->flags & REQ_F_BUFFER_SELECT) {
2387 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2390 return PTR_ERR(buf);
2392 req->rw.len = sqe_len;
2395 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2397 return ret < 0 ? ret : sqe_len;
2401 struct io_async_rw *iorw = &req->io->rw;
2404 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2405 if (iorw->iov == iorw->fast_iov)
2410 if (req->flags & REQ_F_BUFFER_SELECT) {
2411 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2413 ret = (*iovec)->iov_len;
2414 iov_iter_init(iter, rw, *iovec, 1, ret);
2420 #ifdef CONFIG_COMPAT
2421 if (req->ctx->compat)
2422 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2426 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2430 * For files that don't have ->read_iter() and ->write_iter(), handle them
2431 * by looping over ->read() or ->write() manually.
2433 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2434 struct iov_iter *iter)
2439 * Don't support polled IO through this interface, and we can't
2440 * support non-blocking either. For the latter, this just causes
2441 * the kiocb to be handled from an async context.
2443 if (kiocb->ki_flags & IOCB_HIPRI)
2445 if (kiocb->ki_flags & IOCB_NOWAIT)
2448 while (iov_iter_count(iter)) {
2452 if (!iov_iter_is_bvec(iter)) {
2453 iovec = iov_iter_iovec(iter);
2455 /* fixed buffers import bvec */
2456 iovec.iov_base = kmap(iter->bvec->bv_page)
2458 iovec.iov_len = min(iter->count,
2459 iter->bvec->bv_len - iter->iov_offset);
2463 nr = file->f_op->read(file, iovec.iov_base,
2464 iovec.iov_len, &kiocb->ki_pos);
2466 nr = file->f_op->write(file, iovec.iov_base,
2467 iovec.iov_len, &kiocb->ki_pos);
2470 if (iov_iter_is_bvec(iter))
2471 kunmap(iter->bvec->bv_page);
2479 if (nr != iovec.iov_len)
2481 iov_iter_advance(iter, nr);
2487 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2488 struct iovec *iovec, struct iovec *fast_iov,
2489 struct iov_iter *iter)
2491 req->io->rw.nr_segs = iter->nr_segs;
2492 req->io->rw.size = io_size;
2493 req->io->rw.iov = iovec;
2494 if (!req->io->rw.iov) {
2495 req->io->rw.iov = req->io->rw.fast_iov;
2496 memcpy(req->io->rw.iov, fast_iov,
2497 sizeof(struct iovec) * iter->nr_segs);
2499 req->flags |= REQ_F_NEED_CLEANUP;
2503 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2505 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2506 return req->io == NULL;
2509 static int io_alloc_async_ctx(struct io_kiocb *req)
2511 if (!io_op_defs[req->opcode].async_ctx)
2514 return __io_alloc_async_ctx(req);
2517 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2518 struct iovec *iovec, struct iovec *fast_iov,
2519 struct iov_iter *iter)
2521 if (!io_op_defs[req->opcode].async_ctx)
2524 if (__io_alloc_async_ctx(req))
2527 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2532 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2533 bool force_nonblock)
2535 struct io_async_ctx *io;
2536 struct iov_iter iter;
2539 ret = io_prep_rw(req, sqe, force_nonblock);
2543 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2546 /* either don't need iovec imported or already have it */
2547 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2551 io->rw.iov = io->rw.fast_iov;
2553 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2558 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2562 static int io_read(struct io_kiocb *req, bool force_nonblock)
2564 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2565 struct kiocb *kiocb = &req->rw.kiocb;
2566 struct iov_iter iter;
2568 ssize_t io_size, ret;
2570 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2574 /* Ensure we clear previously set non-block flag */
2575 if (!force_nonblock)
2576 kiocb->ki_flags &= ~IOCB_NOWAIT;
2580 if (req->flags & REQ_F_LINK)
2581 req->result = io_size;
2584 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2585 * we know to async punt it even if it was opened O_NONBLOCK
2587 if (force_nonblock && !io_file_supports_async(req->file))
2590 iov_count = iov_iter_count(&iter);
2591 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2595 if (req->file->f_op->read_iter)
2596 ret2 = call_read_iter(req->file, kiocb, &iter);
2598 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2600 /* Catch -EAGAIN return for forced non-blocking submission */
2601 if (!force_nonblock || ret2 != -EAGAIN) {
2602 kiocb_done(kiocb, ret2);
2605 ret = io_setup_async_rw(req, io_size, iovec,
2606 inline_vecs, &iter);
2609 /* any defer here is final, must blocking retry */
2610 if (!(req->flags & REQ_F_NOWAIT))
2611 req->flags |= REQ_F_MUST_PUNT;
2617 req->flags &= ~REQ_F_NEED_CLEANUP;
2621 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2622 bool force_nonblock)
2624 struct io_async_ctx *io;
2625 struct iov_iter iter;
2628 ret = io_prep_rw(req, sqe, force_nonblock);
2632 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2635 req->fsize = rlimit(RLIMIT_FSIZE);
2637 /* either don't need iovec imported or already have it */
2638 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2642 io->rw.iov = io->rw.fast_iov;
2644 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2649 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2653 static int io_write(struct io_kiocb *req, bool force_nonblock)
2655 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2656 struct kiocb *kiocb = &req->rw.kiocb;
2657 struct iov_iter iter;
2659 ssize_t ret, io_size;
2661 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2665 /* Ensure we clear previously set non-block flag */
2666 if (!force_nonblock)
2667 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2671 if (req->flags & REQ_F_LINK)
2672 req->result = io_size;
2675 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2676 * we know to async punt it even if it was opened O_NONBLOCK
2678 if (force_nonblock && !io_file_supports_async(req->file))
2681 /* file path doesn't support NOWAIT for non-direct_IO */
2682 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2683 (req->flags & REQ_F_ISREG))
2686 iov_count = iov_iter_count(&iter);
2687 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2692 * Open-code file_start_write here to grab freeze protection,
2693 * which will be released by another thread in
2694 * io_complete_rw(). Fool lockdep by telling it the lock got
2695 * released so that it doesn't complain about the held lock when
2696 * we return to userspace.
2698 if (req->flags & REQ_F_ISREG) {
2699 __sb_start_write(file_inode(req->file)->i_sb,
2700 SB_FREEZE_WRITE, true);
2701 __sb_writers_release(file_inode(req->file)->i_sb,
2704 kiocb->ki_flags |= IOCB_WRITE;
2706 if (!force_nonblock)
2707 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2709 if (req->file->f_op->write_iter)
2710 ret2 = call_write_iter(req->file, kiocb, &iter);
2712 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2714 if (!force_nonblock)
2715 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2718 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2719 * retry them without IOCB_NOWAIT.
2721 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2723 if (!force_nonblock || ret2 != -EAGAIN) {
2724 kiocb_done(kiocb, ret2);
2727 ret = io_setup_async_rw(req, io_size, iovec,
2728 inline_vecs, &iter);
2731 /* any defer here is final, must blocking retry */
2732 req->flags |= REQ_F_MUST_PUNT;
2737 req->flags &= ~REQ_F_NEED_CLEANUP;
2742 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2744 struct io_splice* sp = &req->splice;
2745 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2748 if (req->flags & REQ_F_NEED_CLEANUP)
2752 sp->off_in = READ_ONCE(sqe->splice_off_in);
2753 sp->off_out = READ_ONCE(sqe->off);
2754 sp->len = READ_ONCE(sqe->len);
2755 sp->flags = READ_ONCE(sqe->splice_flags);
2757 if (unlikely(sp->flags & ~valid_flags))
2760 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2761 (sp->flags & SPLICE_F_FD_IN_FIXED));
2764 req->flags |= REQ_F_NEED_CLEANUP;
2766 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2767 req->work.flags |= IO_WQ_WORK_UNBOUND;
2772 static bool io_splice_punt(struct file *file)
2774 if (get_pipe_info(file))
2776 if (!io_file_supports_async(file))
2778 return !(file->f_mode & O_NONBLOCK);
2781 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2783 struct io_splice *sp = &req->splice;
2784 struct file *in = sp->file_in;
2785 struct file *out = sp->file_out;
2786 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2787 loff_t *poff_in, *poff_out;
2790 if (force_nonblock) {
2791 if (io_splice_punt(in) || io_splice_punt(out))
2793 flags |= SPLICE_F_NONBLOCK;
2796 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2797 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2798 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2799 if (force_nonblock && ret == -EAGAIN)
2802 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2803 req->flags &= ~REQ_F_NEED_CLEANUP;
2805 io_cqring_add_event(req, ret);
2807 req_set_fail_links(req);
2813 * IORING_OP_NOP just posts a completion event, nothing else.
2815 static int io_nop(struct io_kiocb *req)
2817 struct io_ring_ctx *ctx = req->ctx;
2819 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2822 io_cqring_add_event(req, 0);
2827 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2829 struct io_ring_ctx *ctx = req->ctx;
2834 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2836 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2839 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2840 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2843 req->sync.off = READ_ONCE(sqe->off);
2844 req->sync.len = READ_ONCE(sqe->len);
2848 static bool io_req_cancelled(struct io_kiocb *req)
2850 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2851 req_set_fail_links(req);
2852 io_cqring_add_event(req, -ECANCELED);
2860 static void __io_fsync(struct io_kiocb *req)
2862 loff_t end = req->sync.off + req->sync.len;
2865 ret = vfs_fsync_range(req->file, req->sync.off,
2866 end > 0 ? end : LLONG_MAX,
2867 req->sync.flags & IORING_FSYNC_DATASYNC);
2869 req_set_fail_links(req);
2870 io_cqring_add_event(req, ret);
2874 static void io_fsync_finish(struct io_wq_work **workptr)
2876 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2878 if (io_req_cancelled(req))
2881 io_steal_work(req, workptr);
2884 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2886 /* fsync always requires a blocking context */
2887 if (force_nonblock) {
2888 req->work.func = io_fsync_finish;
2895 static void __io_fallocate(struct io_kiocb *req)
2899 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2900 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2902 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2904 req_set_fail_links(req);
2905 io_cqring_add_event(req, ret);
2909 static void io_fallocate_finish(struct io_wq_work **workptr)
2911 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2913 if (io_req_cancelled(req))
2915 __io_fallocate(req);
2916 io_steal_work(req, workptr);
2919 static int io_fallocate_prep(struct io_kiocb *req,
2920 const struct io_uring_sqe *sqe)
2922 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2925 req->sync.off = READ_ONCE(sqe->off);
2926 req->sync.len = READ_ONCE(sqe->addr);
2927 req->sync.mode = READ_ONCE(sqe->len);
2928 req->fsize = rlimit(RLIMIT_FSIZE);
2932 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2934 /* fallocate always requiring blocking context */
2935 if (force_nonblock) {
2936 req->work.func = io_fallocate_finish;
2940 __io_fallocate(req);
2944 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2946 const char __user *fname;
2949 if (sqe->ioprio || sqe->buf_index)
2951 if (sqe->flags & IOSQE_FIXED_FILE)
2953 if (req->flags & REQ_F_NEED_CLEANUP)
2956 req->open.dfd = READ_ONCE(sqe->fd);
2957 req->open.how.mode = READ_ONCE(sqe->len);
2958 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2959 req->open.how.flags = READ_ONCE(sqe->open_flags);
2961 req->open.filename = getname(fname);
2962 if (IS_ERR(req->open.filename)) {
2963 ret = PTR_ERR(req->open.filename);
2964 req->open.filename = NULL;
2968 req->open.nofile = rlimit(RLIMIT_NOFILE);
2969 req->flags |= REQ_F_NEED_CLEANUP;
2973 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2975 struct open_how __user *how;
2976 const char __user *fname;
2980 if (sqe->ioprio || sqe->buf_index)
2982 if (sqe->flags & IOSQE_FIXED_FILE)
2984 if (req->flags & REQ_F_NEED_CLEANUP)
2987 req->open.dfd = READ_ONCE(sqe->fd);
2988 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2989 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2990 len = READ_ONCE(sqe->len);
2992 if (len < OPEN_HOW_SIZE_VER0)
2995 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
3000 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
3001 req->open.how.flags |= O_LARGEFILE;
3003 req->open.filename = getname(fname);
3004 if (IS_ERR(req->open.filename)) {
3005 ret = PTR_ERR(req->open.filename);
3006 req->open.filename = NULL;
3010 req->open.nofile = rlimit(RLIMIT_NOFILE);
3011 req->flags |= REQ_F_NEED_CLEANUP;
3015 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3017 struct open_flags op;
3024 ret = build_open_flags(&req->open.how, &op);
3028 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3032 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3035 ret = PTR_ERR(file);
3037 fsnotify_open(file);
3038 fd_install(ret, file);
3041 putname(req->open.filename);
3042 req->flags &= ~REQ_F_NEED_CLEANUP;
3044 req_set_fail_links(req);
3045 io_cqring_add_event(req, ret);
3050 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3052 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3053 return io_openat2(req, force_nonblock);
3056 static int io_remove_buffers_prep(struct io_kiocb *req,
3057 const struct io_uring_sqe *sqe)
3059 struct io_provide_buf *p = &req->pbuf;
3062 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3065 tmp = READ_ONCE(sqe->fd);
3066 if (!tmp || tmp > USHRT_MAX)
3069 memset(p, 0, sizeof(*p));
3071 p->bgid = READ_ONCE(sqe->buf_group);
3075 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3076 int bgid, unsigned nbufs)
3080 /* shouldn't happen */
3084 /* the head kbuf is the list itself */
3085 while (!list_empty(&buf->list)) {
3086 struct io_buffer *nxt;
3088 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3089 list_del(&nxt->list);
3096 idr_remove(&ctx->io_buffer_idr, bgid);
3101 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3103 struct io_provide_buf *p = &req->pbuf;
3104 struct io_ring_ctx *ctx = req->ctx;
3105 struct io_buffer *head;
3108 io_ring_submit_lock(ctx, !force_nonblock);
3110 lockdep_assert_held(&ctx->uring_lock);
3113 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3115 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3117 io_ring_submit_lock(ctx, !force_nonblock);
3119 req_set_fail_links(req);
3120 io_cqring_add_event(req, ret);
3125 static int io_provide_buffers_prep(struct io_kiocb *req,
3126 const struct io_uring_sqe *sqe)
3128 struct io_provide_buf *p = &req->pbuf;
3131 if (sqe->ioprio || sqe->rw_flags)
3134 tmp = READ_ONCE(sqe->fd);
3135 if (!tmp || tmp > USHRT_MAX)
3138 p->addr = READ_ONCE(sqe->addr);
3139 p->len = READ_ONCE(sqe->len);
3141 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3144 p->bgid = READ_ONCE(sqe->buf_group);
3145 tmp = READ_ONCE(sqe->off);
3146 if (tmp > USHRT_MAX)
3152 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3154 struct io_buffer *buf;
3155 u64 addr = pbuf->addr;
3156 int i, bid = pbuf->bid;
3158 for (i = 0; i < pbuf->nbufs; i++) {
3159 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3164 buf->len = pbuf->len;
3169 INIT_LIST_HEAD(&buf->list);
3172 list_add_tail(&buf->list, &(*head)->list);
3176 return i ? i : -ENOMEM;
3179 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3181 struct io_provide_buf *p = &req->pbuf;
3182 struct io_ring_ctx *ctx = req->ctx;
3183 struct io_buffer *head, *list;
3186 io_ring_submit_lock(ctx, !force_nonblock);
3188 lockdep_assert_held(&ctx->uring_lock);
3190 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3192 ret = io_add_buffers(p, &head);
3197 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3200 __io_remove_buffers(ctx, head, p->bgid, -1U);
3205 io_ring_submit_unlock(ctx, !force_nonblock);
3207 req_set_fail_links(req);
3208 io_cqring_add_event(req, ret);
3213 static int io_epoll_ctl_prep(struct io_kiocb *req,
3214 const struct io_uring_sqe *sqe)
3216 #if defined(CONFIG_EPOLL)
3217 if (sqe->ioprio || sqe->buf_index)
3220 req->epoll.epfd = READ_ONCE(sqe->fd);
3221 req->epoll.op = READ_ONCE(sqe->len);
3222 req->epoll.fd = READ_ONCE(sqe->off);
3224 if (ep_op_has_event(req->epoll.op)) {
3225 struct epoll_event __user *ev;
3227 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3228 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3238 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3240 #if defined(CONFIG_EPOLL)
3241 struct io_epoll *ie = &req->epoll;
3244 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3245 if (force_nonblock && ret == -EAGAIN)
3249 req_set_fail_links(req);
3250 io_cqring_add_event(req, ret);
3258 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3260 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3261 if (sqe->ioprio || sqe->buf_index || sqe->off)
3264 req->madvise.addr = READ_ONCE(sqe->addr);
3265 req->madvise.len = READ_ONCE(sqe->len);
3266 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3273 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3275 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3276 struct io_madvise *ma = &req->madvise;
3282 ret = do_madvise(ma->addr, ma->len, ma->advice);
3284 req_set_fail_links(req);
3285 io_cqring_add_event(req, ret);
3293 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3295 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3298 req->fadvise.offset = READ_ONCE(sqe->off);
3299 req->fadvise.len = READ_ONCE(sqe->len);
3300 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3304 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3306 struct io_fadvise *fa = &req->fadvise;
3309 if (force_nonblock) {
3310 switch (fa->advice) {
3311 case POSIX_FADV_NORMAL:
3312 case POSIX_FADV_RANDOM:
3313 case POSIX_FADV_SEQUENTIAL:
3320 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3322 req_set_fail_links(req);
3323 io_cqring_add_event(req, ret);
3328 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3330 const char __user *fname;
3331 unsigned lookup_flags;
3334 if (sqe->ioprio || sqe->buf_index)
3336 if (sqe->flags & IOSQE_FIXED_FILE)
3338 if (req->flags & REQ_F_NEED_CLEANUP)
3341 req->open.dfd = READ_ONCE(sqe->fd);
3342 req->open.mask = READ_ONCE(sqe->len);
3343 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3344 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3345 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3347 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3350 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3351 if (IS_ERR(req->open.filename)) {
3352 ret = PTR_ERR(req->open.filename);
3353 req->open.filename = NULL;
3357 req->flags |= REQ_F_NEED_CLEANUP;
3361 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3363 struct io_open *ctx = &req->open;
3364 unsigned lookup_flags;
3372 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3376 /* filename_lookup() drops it, keep a reference */
3377 ctx->filename->refcnt++;
3379 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3384 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3386 if (retry_estale(ret, lookup_flags)) {
3387 lookup_flags |= LOOKUP_REVAL;
3391 ret = cp_statx(&stat, ctx->buffer);
3393 putname(ctx->filename);
3394 req->flags &= ~REQ_F_NEED_CLEANUP;
3396 req_set_fail_links(req);
3397 io_cqring_add_event(req, ret);
3402 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3405 * If we queue this for async, it must not be cancellable. That would
3406 * leave the 'file' in an undeterminate state.
3408 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3410 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3411 sqe->rw_flags || sqe->buf_index)
3413 if (sqe->flags & IOSQE_FIXED_FILE)
3416 req->close.fd = READ_ONCE(sqe->fd);
3417 if (req->file->f_op == &io_uring_fops ||
3418 req->close.fd == req->ctx->ring_fd)
3424 /* only called when __close_fd_get_file() is done */
3425 static void __io_close_finish(struct io_kiocb *req)
3429 ret = filp_close(req->close.put_file, req->work.files);
3431 req_set_fail_links(req);
3432 io_cqring_add_event(req, ret);
3433 fput(req->close.put_file);
3437 static void io_close_finish(struct io_wq_work **workptr)
3439 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3441 /* not cancellable, don't do io_req_cancelled() */
3442 __io_close_finish(req);
3443 io_steal_work(req, workptr);
3446 static int io_close(struct io_kiocb *req, bool force_nonblock)
3450 req->close.put_file = NULL;
3451 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3455 /* if the file has a flush method, be safe and punt to async */
3456 if (req->close.put_file->f_op->flush && force_nonblock) {
3457 /* submission ref will be dropped, take it for async */
3458 refcount_inc(&req->refs);
3460 req->work.func = io_close_finish;
3462 * Do manual async queue here to avoid grabbing files - we don't
3463 * need the files, and it'll cause io_close_finish() to close
3464 * the file again and cause a double CQE entry for this request
3466 io_queue_async_work(req);
3471 * No ->flush(), safely close from here and just punt the
3472 * fput() to async context.
3474 __io_close_finish(req);
3478 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3480 struct io_ring_ctx *ctx = req->ctx;
3485 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3487 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3490 req->sync.off = READ_ONCE(sqe->off);
3491 req->sync.len = READ_ONCE(sqe->len);
3492 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3496 static void __io_sync_file_range(struct io_kiocb *req)
3500 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3503 req_set_fail_links(req);
3504 io_cqring_add_event(req, ret);
3509 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3511 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3512 struct io_kiocb *nxt = NULL;
3514 if (io_req_cancelled(req))
3516 __io_sync_file_range(req);
3517 io_put_req(req); /* put submission ref */
3519 io_wq_assign_next(workptr, nxt);
3522 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3524 /* sync_file_range always requires a blocking context */
3525 if (force_nonblock) {
3526 req->work.func = io_sync_file_range_finish;
3530 __io_sync_file_range(req);
3534 #if defined(CONFIG_NET)
3535 static int io_setup_async_msg(struct io_kiocb *req,
3536 struct io_async_msghdr *kmsg)
3540 if (io_alloc_async_ctx(req)) {
3541 if (kmsg->iov != kmsg->fast_iov)
3545 req->flags |= REQ_F_NEED_CLEANUP;
3546 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3550 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3552 struct io_sr_msg *sr = &req->sr_msg;
3553 struct io_async_ctx *io = req->io;
3556 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3557 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3558 sr->len = READ_ONCE(sqe->len);
3560 #ifdef CONFIG_COMPAT
3561 if (req->ctx->compat)
3562 sr->msg_flags |= MSG_CMSG_COMPAT;
3565 if (!io || req->opcode == IORING_OP_SEND)
3567 /* iovec is already imported */
3568 if (req->flags & REQ_F_NEED_CLEANUP)
3571 io->msg.iov = io->msg.fast_iov;
3572 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3575 req->flags |= REQ_F_NEED_CLEANUP;
3579 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3581 struct io_async_msghdr *kmsg = NULL;
3582 struct socket *sock;
3585 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3588 sock = sock_from_file(req->file, &ret);
3590 struct io_async_ctx io;
3594 kmsg = &req->io->msg;
3595 kmsg->msg.msg_name = &req->io->msg.addr;
3596 /* if iov is set, it's allocated already */
3598 kmsg->iov = kmsg->fast_iov;
3599 kmsg->msg.msg_iter.iov = kmsg->iov;
3601 struct io_sr_msg *sr = &req->sr_msg;
3604 kmsg->msg.msg_name = &io.msg.addr;
3606 io.msg.iov = io.msg.fast_iov;
3607 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3608 sr->msg_flags, &io.msg.iov);
3613 flags = req->sr_msg.msg_flags;
3614 if (flags & MSG_DONTWAIT)
3615 req->flags |= REQ_F_NOWAIT;
3616 else if (force_nonblock)
3617 flags |= MSG_DONTWAIT;
3619 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3620 if (force_nonblock && ret == -EAGAIN)
3621 return io_setup_async_msg(req, kmsg);
3622 if (ret == -ERESTARTSYS)
3626 if (kmsg && kmsg->iov != kmsg->fast_iov)
3628 req->flags &= ~REQ_F_NEED_CLEANUP;
3629 io_cqring_add_event(req, ret);
3631 req_set_fail_links(req);
3636 static int io_send(struct io_kiocb *req, bool force_nonblock)
3638 struct socket *sock;
3641 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3644 sock = sock_from_file(req->file, &ret);
3646 struct io_sr_msg *sr = &req->sr_msg;
3651 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3656 msg.msg_name = NULL;
3657 msg.msg_control = NULL;
3658 msg.msg_controllen = 0;
3659 msg.msg_namelen = 0;
3661 flags = req->sr_msg.msg_flags;
3662 if (flags & MSG_DONTWAIT)
3663 req->flags |= REQ_F_NOWAIT;
3664 else if (force_nonblock)
3665 flags |= MSG_DONTWAIT;
3667 msg.msg_flags = flags;
3668 ret = sock_sendmsg(sock, &msg);
3669 if (force_nonblock && ret == -EAGAIN)
3671 if (ret == -ERESTARTSYS)
3675 io_cqring_add_event(req, ret);
3677 req_set_fail_links(req);
3682 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3684 struct io_sr_msg *sr = &req->sr_msg;
3685 struct iovec __user *uiov;
3689 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3694 if (req->flags & REQ_F_BUFFER_SELECT) {
3697 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3699 sr->len = io->msg.iov[0].iov_len;
3700 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3704 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3705 &io->msg.iov, &io->msg.msg.msg_iter);
3713 #ifdef CONFIG_COMPAT
3714 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3715 struct io_async_ctx *io)
3717 struct compat_msghdr __user *msg_compat;
3718 struct io_sr_msg *sr = &req->sr_msg;
3719 struct compat_iovec __user *uiov;
3724 msg_compat = (struct compat_msghdr __user *) sr->msg;
3725 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3730 uiov = compat_ptr(ptr);
3731 if (req->flags & REQ_F_BUFFER_SELECT) {
3732 compat_ssize_t clen;
3736 if (!access_ok(uiov, sizeof(*uiov)))
3738 if (__get_user(clen, &uiov->iov_len))
3742 sr->len = io->msg.iov[0].iov_len;
3745 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3747 &io->msg.msg.msg_iter);
3756 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3758 io->msg.iov = io->msg.fast_iov;
3760 #ifdef CONFIG_COMPAT
3761 if (req->ctx->compat)
3762 return __io_compat_recvmsg_copy_hdr(req, io);
3765 return __io_recvmsg_copy_hdr(req, io);
3768 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3769 int *cflags, bool needs_lock)
3771 struct io_sr_msg *sr = &req->sr_msg;
3772 struct io_buffer *kbuf;
3774 if (!(req->flags & REQ_F_BUFFER_SELECT))
3777 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3782 req->flags |= REQ_F_BUFFER_SELECTED;
3784 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3785 *cflags |= IORING_CQE_F_BUFFER;
3789 static int io_recvmsg_prep(struct io_kiocb *req,
3790 const struct io_uring_sqe *sqe)
3792 struct io_sr_msg *sr = &req->sr_msg;
3793 struct io_async_ctx *io = req->io;
3796 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3797 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3798 sr->len = READ_ONCE(sqe->len);
3799 sr->bgid = READ_ONCE(sqe->buf_group);
3801 #ifdef CONFIG_COMPAT
3802 if (req->ctx->compat)
3803 sr->msg_flags |= MSG_CMSG_COMPAT;
3806 if (!io || req->opcode == IORING_OP_RECV)
3808 /* iovec is already imported */
3809 if (req->flags & REQ_F_NEED_CLEANUP)
3812 ret = io_recvmsg_copy_hdr(req, io);
3814 req->flags |= REQ_F_NEED_CLEANUP;
3818 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3820 struct io_async_msghdr *kmsg = NULL;
3821 struct socket *sock;
3822 int ret, cflags = 0;
3824 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3827 sock = sock_from_file(req->file, &ret);
3829 struct io_buffer *kbuf;
3830 struct io_async_ctx io;
3834 kmsg = &req->io->msg;
3835 kmsg->msg.msg_name = &req->io->msg.addr;
3836 /* if iov is set, it's allocated already */
3838 kmsg->iov = kmsg->fast_iov;
3839 kmsg->msg.msg_iter.iov = kmsg->iov;
3842 kmsg->msg.msg_name = &io.msg.addr;
3844 ret = io_recvmsg_copy_hdr(req, &io);
3849 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3851 return PTR_ERR(kbuf);
3853 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3854 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3855 1, req->sr_msg.len);
3858 flags = req->sr_msg.msg_flags;
3859 if (flags & MSG_DONTWAIT)
3860 req->flags |= REQ_F_NOWAIT;
3861 else if (force_nonblock)
3862 flags |= MSG_DONTWAIT;
3864 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3865 kmsg->uaddr, flags);
3866 if (force_nonblock && ret == -EAGAIN)
3867 return io_setup_async_msg(req, kmsg);
3868 if (ret == -ERESTARTSYS)
3872 if (kmsg && kmsg->iov != kmsg->fast_iov)
3874 req->flags &= ~REQ_F_NEED_CLEANUP;
3875 __io_cqring_add_event(req, ret, cflags);
3877 req_set_fail_links(req);
3882 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3884 struct io_buffer *kbuf = NULL;
3885 struct socket *sock;
3886 int ret, cflags = 0;
3888 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3891 sock = sock_from_file(req->file, &ret);
3893 struct io_sr_msg *sr = &req->sr_msg;
3894 void __user *buf = sr->buf;
3899 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3901 return PTR_ERR(kbuf);
3903 buf = u64_to_user_ptr(kbuf->addr);
3905 ret = import_single_range(READ, buf, sr->len, &iov,
3912 req->flags |= REQ_F_NEED_CLEANUP;
3913 msg.msg_name = NULL;
3914 msg.msg_control = NULL;
3915 msg.msg_controllen = 0;
3916 msg.msg_namelen = 0;
3917 msg.msg_iocb = NULL;
3920 flags = req->sr_msg.msg_flags;
3921 if (flags & MSG_DONTWAIT)
3922 req->flags |= REQ_F_NOWAIT;
3923 else if (force_nonblock)
3924 flags |= MSG_DONTWAIT;
3926 ret = sock_recvmsg(sock, &msg, flags);
3927 if (force_nonblock && ret == -EAGAIN)
3929 if (ret == -ERESTARTSYS)
3934 req->flags &= ~REQ_F_NEED_CLEANUP;
3935 __io_cqring_add_event(req, ret, cflags);
3937 req_set_fail_links(req);
3942 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3944 struct io_accept *accept = &req->accept;
3946 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3948 if (sqe->ioprio || sqe->len || sqe->buf_index)
3951 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3952 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3953 accept->flags = READ_ONCE(sqe->accept_flags);
3954 accept->nofile = rlimit(RLIMIT_NOFILE);
3958 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3960 struct io_accept *accept = &req->accept;
3961 unsigned file_flags;
3964 file_flags = force_nonblock ? O_NONBLOCK : 0;
3965 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3966 accept->addr_len, accept->flags,
3968 if (ret == -EAGAIN && force_nonblock)
3970 if (ret == -ERESTARTSYS)
3973 req_set_fail_links(req);
3974 io_cqring_add_event(req, ret);
3979 static void io_accept_finish(struct io_wq_work **workptr)
3981 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3983 if (io_req_cancelled(req))
3985 __io_accept(req, false);
3986 io_steal_work(req, workptr);
3989 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3993 ret = __io_accept(req, force_nonblock);
3994 if (ret == -EAGAIN && force_nonblock) {
3995 req->work.func = io_accept_finish;
4001 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4003 struct io_connect *conn = &req->connect;
4004 struct io_async_ctx *io = req->io;
4006 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
4008 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
4011 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
4012 conn->addr_len = READ_ONCE(sqe->addr2);
4017 return move_addr_to_kernel(conn->addr, conn->addr_len,
4018 &io->connect.address);
4021 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4023 struct io_async_ctx __io, *io;
4024 unsigned file_flags;
4030 ret = move_addr_to_kernel(req->connect.addr,
4031 req->connect.addr_len,
4032 &__io.connect.address);
4038 file_flags = force_nonblock ? O_NONBLOCK : 0;
4040 ret = __sys_connect_file(req->file, &io->connect.address,
4041 req->connect.addr_len, file_flags);
4042 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4045 if (io_alloc_async_ctx(req)) {
4049 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4052 if (ret == -ERESTARTSYS)
4056 req_set_fail_links(req);
4057 io_cqring_add_event(req, ret);
4061 #else /* !CONFIG_NET */
4062 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4067 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4072 static int io_send(struct io_kiocb *req, bool force_nonblock)
4077 static int io_recvmsg_prep(struct io_kiocb *req,
4078 const struct io_uring_sqe *sqe)
4083 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4088 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4093 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4098 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4103 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4108 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4112 #endif /* CONFIG_NET */
4114 struct io_poll_table {
4115 struct poll_table_struct pt;
4116 struct io_kiocb *req;
4120 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4121 struct wait_queue_head *head)
4123 if (unlikely(poll->head)) {
4124 pt->error = -EINVAL;
4130 add_wait_queue(head, &poll->wait);
4133 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4134 struct poll_table_struct *p)
4136 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4138 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4141 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4142 __poll_t mask, task_work_func_t func)
4144 struct task_struct *tsk;
4147 /* for instances that support it check for an event match first: */
4148 if (mask && !(mask & poll->events))
4151 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4153 list_del_init(&poll->wait.entry);
4157 init_task_work(&req->task_work, func);
4159 * If this fails, then the task is exiting. Punt to one of the io-wq
4160 * threads to ensure the work gets run, we can't always rely on exit
4161 * cancelation taking care of this.
4163 ret = task_work_add(tsk, &req->task_work, true);
4164 if (unlikely(ret)) {
4165 tsk = io_wq_get_task(req->ctx->io_wq);
4166 task_work_add(tsk, &req->task_work, true);
4168 wake_up_process(tsk);
4172 static void io_async_task_func(struct callback_head *cb)
4174 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4175 struct async_poll *apoll = req->apoll;
4176 struct io_ring_ctx *ctx = req->ctx;
4178 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4180 WARN_ON_ONCE(!list_empty(&req->apoll->poll.wait.entry));
4182 if (hash_hashed(&req->hash_node)) {
4183 spin_lock_irq(&ctx->completion_lock);
4184 hash_del(&req->hash_node);
4185 spin_unlock_irq(&ctx->completion_lock);
4188 /* restore ->work in case we need to retry again */
4189 memcpy(&req->work, &apoll->work, sizeof(req->work));
4191 __set_current_state(TASK_RUNNING);
4192 mutex_lock(&ctx->uring_lock);
4193 __io_queue_sqe(req, NULL);
4194 mutex_unlock(&ctx->uring_lock);
4199 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4202 struct io_kiocb *req = wait->private;
4203 struct io_poll_iocb *poll = &req->apoll->poll;
4205 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4208 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4211 static void io_poll_req_insert(struct io_kiocb *req)
4213 struct io_ring_ctx *ctx = req->ctx;
4214 struct hlist_head *list;
4216 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4217 hlist_add_head(&req->hash_node, list);
4220 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4221 struct io_poll_iocb *poll,
4222 struct io_poll_table *ipt, __poll_t mask,
4223 wait_queue_func_t wake_func)
4224 __acquires(&ctx->completion_lock)
4226 struct io_ring_ctx *ctx = req->ctx;
4227 bool cancel = false;
4229 poll->file = req->file;
4231 poll->done = poll->canceled = false;
4232 poll->events = mask;
4234 ipt->pt._key = mask;
4236 ipt->error = -EINVAL;
4238 INIT_LIST_HEAD(&poll->wait.entry);
4239 init_waitqueue_func_entry(&poll->wait, wake_func);
4240 poll->wait.private = req;
4242 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4244 spin_lock_irq(&ctx->completion_lock);
4245 if (likely(poll->head)) {
4246 spin_lock(&poll->head->lock);
4247 if (unlikely(list_empty(&poll->wait.entry))) {
4253 if (mask || ipt->error)
4254 list_del_init(&poll->wait.entry);
4256 WRITE_ONCE(poll->canceled, true);
4257 else if (!poll->done) /* actually waiting for an event */
4258 io_poll_req_insert(req);
4259 spin_unlock(&poll->head->lock);
4265 static bool io_arm_poll_handler(struct io_kiocb *req)
4267 const struct io_op_def *def = &io_op_defs[req->opcode];
4268 struct io_ring_ctx *ctx = req->ctx;
4269 struct async_poll *apoll;
4270 struct io_poll_table ipt;
4273 if (!req->file || !file_can_poll(req->file))
4275 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4277 if (!def->pollin && !def->pollout)
4280 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4281 if (unlikely(!apoll))
4284 req->flags |= REQ_F_POLLED;
4285 memcpy(&apoll->work, &req->work, sizeof(req->work));
4287 get_task_struct(current);
4288 req->task = current;
4290 INIT_HLIST_NODE(&req->hash_node);
4294 mask |= POLLIN | POLLRDNORM;
4296 mask |= POLLOUT | POLLWRNORM;
4297 mask |= POLLERR | POLLPRI;
4299 ipt.pt._qproc = io_async_queue_proc;
4301 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4305 apoll->poll.done = true;
4306 spin_unlock_irq(&ctx->completion_lock);
4307 memcpy(&req->work, &apoll->work, sizeof(req->work));
4311 spin_unlock_irq(&ctx->completion_lock);
4312 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4313 apoll->poll.events);
4317 static bool __io_poll_remove_one(struct io_kiocb *req,
4318 struct io_poll_iocb *poll)
4320 bool do_complete = false;
4322 spin_lock(&poll->head->lock);
4323 WRITE_ONCE(poll->canceled, true);
4324 if (!list_empty(&poll->wait.entry)) {
4325 list_del_init(&poll->wait.entry);
4328 spin_unlock(&poll->head->lock);
4332 static bool io_poll_remove_one(struct io_kiocb *req)
4336 if (req->opcode == IORING_OP_POLL_ADD) {
4337 do_complete = __io_poll_remove_one(req, &req->poll);
4339 /* non-poll requests have submit ref still */
4340 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4345 hash_del(&req->hash_node);
4348 io_cqring_fill_event(req, -ECANCELED);
4349 io_commit_cqring(req->ctx);
4350 req->flags |= REQ_F_COMP_LOCKED;
4357 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4359 struct hlist_node *tmp;
4360 struct io_kiocb *req;
4363 spin_lock_irq(&ctx->completion_lock);
4364 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4365 struct hlist_head *list;
4367 list = &ctx->cancel_hash[i];
4368 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4369 io_poll_remove_one(req);
4371 spin_unlock_irq(&ctx->completion_lock);
4373 io_cqring_ev_posted(ctx);
4376 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4378 struct hlist_head *list;
4379 struct io_kiocb *req;
4381 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4382 hlist_for_each_entry(req, list, hash_node) {
4383 if (sqe_addr != req->user_data)
4385 if (io_poll_remove_one(req))
4393 static int io_poll_remove_prep(struct io_kiocb *req,
4394 const struct io_uring_sqe *sqe)
4396 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4398 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4402 req->poll.addr = READ_ONCE(sqe->addr);
4407 * Find a running poll command that matches one specified in sqe->addr,
4408 * and remove it if found.
4410 static int io_poll_remove(struct io_kiocb *req)
4412 struct io_ring_ctx *ctx = req->ctx;
4416 addr = req->poll.addr;
4417 spin_lock_irq(&ctx->completion_lock);
4418 ret = io_poll_cancel(ctx, addr);
4419 spin_unlock_irq(&ctx->completion_lock);
4421 io_cqring_add_event(req, ret);
4423 req_set_fail_links(req);
4428 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4430 struct io_ring_ctx *ctx = req->ctx;
4432 req->poll.done = true;
4433 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4434 io_commit_cqring(ctx);
4437 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4439 struct io_ring_ctx *ctx = req->ctx;
4440 struct io_poll_iocb *poll = &req->poll;
4442 if (!req->result && !READ_ONCE(poll->canceled)) {
4443 struct poll_table_struct pt = { ._key = poll->events };
4445 req->result = vfs_poll(req->file, &pt) & poll->events;
4448 spin_lock_irq(&ctx->completion_lock);
4449 if (!req->result && !READ_ONCE(poll->canceled)) {
4450 add_wait_queue(poll->head, &poll->wait);
4451 spin_unlock_irq(&ctx->completion_lock);
4454 hash_del(&req->hash_node);
4455 io_poll_complete(req, req->result, 0);
4456 req->flags |= REQ_F_COMP_LOCKED;
4457 io_put_req_find_next(req, nxt);
4458 spin_unlock_irq(&ctx->completion_lock);
4460 io_cqring_ev_posted(ctx);
4463 static void io_poll_task_func(struct callback_head *cb)
4465 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4466 struct io_kiocb *nxt = NULL;
4468 io_poll_task_handler(req, &nxt);
4470 struct io_ring_ctx *ctx = nxt->ctx;
4472 mutex_lock(&ctx->uring_lock);
4473 __io_queue_sqe(nxt, NULL);
4474 mutex_unlock(&ctx->uring_lock);
4478 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4481 struct io_kiocb *req = wait->private;
4482 struct io_poll_iocb *poll = &req->poll;
4484 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4487 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4488 struct poll_table_struct *p)
4490 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4492 __io_queue_proc(&pt->req->poll, pt, head);
4495 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4497 struct io_poll_iocb *poll = &req->poll;
4500 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4502 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4507 events = READ_ONCE(sqe->poll_events);
4508 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4510 get_task_struct(current);
4511 req->task = current;
4515 static int io_poll_add(struct io_kiocb *req)
4517 struct io_poll_iocb *poll = &req->poll;
4518 struct io_ring_ctx *ctx = req->ctx;
4519 struct io_poll_table ipt;
4522 INIT_HLIST_NODE(&req->hash_node);
4523 INIT_LIST_HEAD(&req->list);
4524 ipt.pt._qproc = io_poll_queue_proc;
4526 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4529 if (mask) { /* no async, we'd stolen it */
4531 io_poll_complete(req, mask, 0);
4533 spin_unlock_irq(&ctx->completion_lock);
4536 io_cqring_ev_posted(ctx);
4542 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4544 struct io_timeout_data *data = container_of(timer,
4545 struct io_timeout_data, timer);
4546 struct io_kiocb *req = data->req;
4547 struct io_ring_ctx *ctx = req->ctx;
4548 unsigned long flags;
4550 atomic_inc(&ctx->cq_timeouts);
4552 spin_lock_irqsave(&ctx->completion_lock, flags);
4554 * We could be racing with timeout deletion. If the list is empty,
4555 * then timeout lookup already found it and will be handling it.
4557 if (!list_empty(&req->list)) {
4558 struct io_kiocb *prev;
4561 * Adjust the reqs sequence before the current one because it
4562 * will consume a slot in the cq_ring and the cq_tail
4563 * pointer will be increased, otherwise other timeout reqs may
4564 * return in advance without waiting for enough wait_nr.
4567 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4569 list_del_init(&req->list);
4572 io_cqring_fill_event(req, -ETIME);
4573 io_commit_cqring(ctx);
4574 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4576 io_cqring_ev_posted(ctx);
4577 req_set_fail_links(req);
4579 return HRTIMER_NORESTART;
4582 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4584 struct io_kiocb *req;
4587 list_for_each_entry(req, &ctx->timeout_list, list) {
4588 if (user_data == req->user_data) {
4589 list_del_init(&req->list);
4598 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4602 req_set_fail_links(req);
4603 io_cqring_fill_event(req, -ECANCELED);
4608 static int io_timeout_remove_prep(struct io_kiocb *req,
4609 const struct io_uring_sqe *sqe)
4611 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4613 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4616 req->timeout.addr = READ_ONCE(sqe->addr);
4617 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4618 if (req->timeout.flags)
4625 * Remove or update an existing timeout command
4627 static int io_timeout_remove(struct io_kiocb *req)
4629 struct io_ring_ctx *ctx = req->ctx;
4632 spin_lock_irq(&ctx->completion_lock);
4633 ret = io_timeout_cancel(ctx, req->timeout.addr);
4635 io_cqring_fill_event(req, ret);
4636 io_commit_cqring(ctx);
4637 spin_unlock_irq(&ctx->completion_lock);
4638 io_cqring_ev_posted(ctx);
4640 req_set_fail_links(req);
4645 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4646 bool is_timeout_link)
4648 struct io_timeout_data *data;
4651 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4653 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4655 if (sqe->off && is_timeout_link)
4657 flags = READ_ONCE(sqe->timeout_flags);
4658 if (flags & ~IORING_TIMEOUT_ABS)
4661 req->timeout.count = READ_ONCE(sqe->off);
4663 if (!req->io && io_alloc_async_ctx(req))
4666 data = &req->io->timeout;
4668 req->flags |= REQ_F_TIMEOUT;
4670 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4673 if (flags & IORING_TIMEOUT_ABS)
4674 data->mode = HRTIMER_MODE_ABS;
4676 data->mode = HRTIMER_MODE_REL;
4678 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4682 static int io_timeout(struct io_kiocb *req)
4685 struct io_ring_ctx *ctx = req->ctx;
4686 struct io_timeout_data *data;
4687 struct list_head *entry;
4690 data = &req->io->timeout;
4693 * sqe->off holds how many events that need to occur for this
4694 * timeout event to be satisfied. If it isn't set, then this is
4695 * a pure timeout request, sequence isn't used.
4697 count = req->timeout.count;
4699 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4700 spin_lock_irq(&ctx->completion_lock);
4701 entry = ctx->timeout_list.prev;
4705 req->sequence = ctx->cached_sq_head + count - 1;
4706 data->seq_offset = count;
4709 * Insertion sort, ensuring the first entry in the list is always
4710 * the one we need first.
4712 spin_lock_irq(&ctx->completion_lock);
4713 list_for_each_prev(entry, &ctx->timeout_list) {
4714 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4715 unsigned nxt_sq_head;
4716 long long tmp, tmp_nxt;
4717 u32 nxt_offset = nxt->io->timeout.seq_offset;
4719 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4723 * Since cached_sq_head + count - 1 can overflow, use type long
4726 tmp = (long long)ctx->cached_sq_head + count - 1;
4727 nxt_sq_head = nxt->sequence - nxt_offset + 1;
4728 tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1;
4731 * cached_sq_head may overflow, and it will never overflow twice
4732 * once there is some timeout req still be valid.
4734 if (ctx->cached_sq_head < nxt_sq_head)
4741 * Sequence of reqs after the insert one and itself should
4742 * be adjusted because each timeout req consumes a slot.
4747 req->sequence -= span;
4749 list_add(&req->list, entry);
4750 data->timer.function = io_timeout_fn;
4751 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4752 spin_unlock_irq(&ctx->completion_lock);
4756 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4758 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4760 return req->user_data == (unsigned long) data;
4763 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4765 enum io_wq_cancel cancel_ret;
4768 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4769 switch (cancel_ret) {
4770 case IO_WQ_CANCEL_OK:
4773 case IO_WQ_CANCEL_RUNNING:
4776 case IO_WQ_CANCEL_NOTFOUND:
4784 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4785 struct io_kiocb *req, __u64 sqe_addr,
4788 unsigned long flags;
4791 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4792 if (ret != -ENOENT) {
4793 spin_lock_irqsave(&ctx->completion_lock, flags);
4797 spin_lock_irqsave(&ctx->completion_lock, flags);
4798 ret = io_timeout_cancel(ctx, sqe_addr);
4801 ret = io_poll_cancel(ctx, sqe_addr);
4805 io_cqring_fill_event(req, ret);
4806 io_commit_cqring(ctx);
4807 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4808 io_cqring_ev_posted(ctx);
4811 req_set_fail_links(req);
4815 static int io_async_cancel_prep(struct io_kiocb *req,
4816 const struct io_uring_sqe *sqe)
4818 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4820 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4824 req->cancel.addr = READ_ONCE(sqe->addr);
4828 static int io_async_cancel(struct io_kiocb *req)
4830 struct io_ring_ctx *ctx = req->ctx;
4832 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4836 static int io_files_update_prep(struct io_kiocb *req,
4837 const struct io_uring_sqe *sqe)
4839 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4842 req->files_update.offset = READ_ONCE(sqe->off);
4843 req->files_update.nr_args = READ_ONCE(sqe->len);
4844 if (!req->files_update.nr_args)
4846 req->files_update.arg = READ_ONCE(sqe->addr);
4850 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4852 struct io_ring_ctx *ctx = req->ctx;
4853 struct io_uring_files_update up;
4859 up.offset = req->files_update.offset;
4860 up.fds = req->files_update.arg;
4862 mutex_lock(&ctx->uring_lock);
4863 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4864 mutex_unlock(&ctx->uring_lock);
4867 req_set_fail_links(req);
4868 io_cqring_add_event(req, ret);
4873 static int io_req_defer_prep(struct io_kiocb *req,
4874 const struct io_uring_sqe *sqe)
4881 if (io_op_defs[req->opcode].file_table) {
4882 ret = io_grab_files(req);
4887 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4889 switch (req->opcode) {
4892 case IORING_OP_READV:
4893 case IORING_OP_READ_FIXED:
4894 case IORING_OP_READ:
4895 ret = io_read_prep(req, sqe, true);
4897 case IORING_OP_WRITEV:
4898 case IORING_OP_WRITE_FIXED:
4899 case IORING_OP_WRITE:
4900 ret = io_write_prep(req, sqe, true);
4902 case IORING_OP_POLL_ADD:
4903 ret = io_poll_add_prep(req, sqe);
4905 case IORING_OP_POLL_REMOVE:
4906 ret = io_poll_remove_prep(req, sqe);
4908 case IORING_OP_FSYNC:
4909 ret = io_prep_fsync(req, sqe);
4911 case IORING_OP_SYNC_FILE_RANGE:
4912 ret = io_prep_sfr(req, sqe);
4914 case IORING_OP_SENDMSG:
4915 case IORING_OP_SEND:
4916 ret = io_sendmsg_prep(req, sqe);
4918 case IORING_OP_RECVMSG:
4919 case IORING_OP_RECV:
4920 ret = io_recvmsg_prep(req, sqe);
4922 case IORING_OP_CONNECT:
4923 ret = io_connect_prep(req, sqe);
4925 case IORING_OP_TIMEOUT:
4926 ret = io_timeout_prep(req, sqe, false);
4928 case IORING_OP_TIMEOUT_REMOVE:
4929 ret = io_timeout_remove_prep(req, sqe);
4931 case IORING_OP_ASYNC_CANCEL:
4932 ret = io_async_cancel_prep(req, sqe);
4934 case IORING_OP_LINK_TIMEOUT:
4935 ret = io_timeout_prep(req, sqe, true);
4937 case IORING_OP_ACCEPT:
4938 ret = io_accept_prep(req, sqe);
4940 case IORING_OP_FALLOCATE:
4941 ret = io_fallocate_prep(req, sqe);
4943 case IORING_OP_OPENAT:
4944 ret = io_openat_prep(req, sqe);
4946 case IORING_OP_CLOSE:
4947 ret = io_close_prep(req, sqe);
4949 case IORING_OP_FILES_UPDATE:
4950 ret = io_files_update_prep(req, sqe);
4952 case IORING_OP_STATX:
4953 ret = io_statx_prep(req, sqe);
4955 case IORING_OP_FADVISE:
4956 ret = io_fadvise_prep(req, sqe);
4958 case IORING_OP_MADVISE:
4959 ret = io_madvise_prep(req, sqe);
4961 case IORING_OP_OPENAT2:
4962 ret = io_openat2_prep(req, sqe);
4964 case IORING_OP_EPOLL_CTL:
4965 ret = io_epoll_ctl_prep(req, sqe);
4967 case IORING_OP_SPLICE:
4968 ret = io_splice_prep(req, sqe);
4970 case IORING_OP_PROVIDE_BUFFERS:
4971 ret = io_provide_buffers_prep(req, sqe);
4973 case IORING_OP_REMOVE_BUFFERS:
4974 ret = io_remove_buffers_prep(req, sqe);
4977 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
4986 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4988 struct io_ring_ctx *ctx = req->ctx;
4991 /* Still need defer if there is pending req in defer list. */
4992 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
4995 if (!req->io && io_alloc_async_ctx(req))
4998 ret = io_req_defer_prep(req, sqe);
5002 spin_lock_irq(&ctx->completion_lock);
5003 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5004 spin_unlock_irq(&ctx->completion_lock);
5008 trace_io_uring_defer(ctx, req, req->user_data);
5009 list_add_tail(&req->list, &ctx->defer_list);
5010 spin_unlock_irq(&ctx->completion_lock);
5011 return -EIOCBQUEUED;
5014 static void io_cleanup_req(struct io_kiocb *req)
5016 struct io_async_ctx *io = req->io;
5018 switch (req->opcode) {
5019 case IORING_OP_READV:
5020 case IORING_OP_READ_FIXED:
5021 case IORING_OP_READ:
5022 if (req->flags & REQ_F_BUFFER_SELECTED)
5023 kfree((void *)(unsigned long)req->rw.addr);
5025 case IORING_OP_WRITEV:
5026 case IORING_OP_WRITE_FIXED:
5027 case IORING_OP_WRITE:
5028 if (io->rw.iov != io->rw.fast_iov)
5031 case IORING_OP_RECVMSG:
5032 if (req->flags & REQ_F_BUFFER_SELECTED)
5033 kfree(req->sr_msg.kbuf);
5035 case IORING_OP_SENDMSG:
5036 if (io->msg.iov != io->msg.fast_iov)
5039 case IORING_OP_RECV:
5040 if (req->flags & REQ_F_BUFFER_SELECTED)
5041 kfree(req->sr_msg.kbuf);
5043 case IORING_OP_OPENAT:
5044 case IORING_OP_OPENAT2:
5045 case IORING_OP_STATX:
5046 putname(req->open.filename);
5048 case IORING_OP_SPLICE:
5049 io_put_file(req, req->splice.file_in,
5050 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5054 req->flags &= ~REQ_F_NEED_CLEANUP;
5057 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5058 bool force_nonblock)
5060 struct io_ring_ctx *ctx = req->ctx;
5063 switch (req->opcode) {
5067 case IORING_OP_READV:
5068 case IORING_OP_READ_FIXED:
5069 case IORING_OP_READ:
5071 ret = io_read_prep(req, sqe, force_nonblock);
5075 ret = io_read(req, force_nonblock);
5077 case IORING_OP_WRITEV:
5078 case IORING_OP_WRITE_FIXED:
5079 case IORING_OP_WRITE:
5081 ret = io_write_prep(req, sqe, force_nonblock);
5085 ret = io_write(req, force_nonblock);
5087 case IORING_OP_FSYNC:
5089 ret = io_prep_fsync(req, sqe);
5093 ret = io_fsync(req, force_nonblock);
5095 case IORING_OP_POLL_ADD:
5097 ret = io_poll_add_prep(req, sqe);
5101 ret = io_poll_add(req);
5103 case IORING_OP_POLL_REMOVE:
5105 ret = io_poll_remove_prep(req, sqe);
5109 ret = io_poll_remove(req);
5111 case IORING_OP_SYNC_FILE_RANGE:
5113 ret = io_prep_sfr(req, sqe);
5117 ret = io_sync_file_range(req, force_nonblock);
5119 case IORING_OP_SENDMSG:
5120 case IORING_OP_SEND:
5122 ret = io_sendmsg_prep(req, sqe);
5126 if (req->opcode == IORING_OP_SENDMSG)
5127 ret = io_sendmsg(req, force_nonblock);
5129 ret = io_send(req, force_nonblock);
5131 case IORING_OP_RECVMSG:
5132 case IORING_OP_RECV:
5134 ret = io_recvmsg_prep(req, sqe);
5138 if (req->opcode == IORING_OP_RECVMSG)
5139 ret = io_recvmsg(req, force_nonblock);
5141 ret = io_recv(req, force_nonblock);
5143 case IORING_OP_TIMEOUT:
5145 ret = io_timeout_prep(req, sqe, false);
5149 ret = io_timeout(req);
5151 case IORING_OP_TIMEOUT_REMOVE:
5153 ret = io_timeout_remove_prep(req, sqe);
5157 ret = io_timeout_remove(req);
5159 case IORING_OP_ACCEPT:
5161 ret = io_accept_prep(req, sqe);
5165 ret = io_accept(req, force_nonblock);
5167 case IORING_OP_CONNECT:
5169 ret = io_connect_prep(req, sqe);
5173 ret = io_connect(req, force_nonblock);
5175 case IORING_OP_ASYNC_CANCEL:
5177 ret = io_async_cancel_prep(req, sqe);
5181 ret = io_async_cancel(req);
5183 case IORING_OP_FALLOCATE:
5185 ret = io_fallocate_prep(req, sqe);
5189 ret = io_fallocate(req, force_nonblock);
5191 case IORING_OP_OPENAT:
5193 ret = io_openat_prep(req, sqe);
5197 ret = io_openat(req, force_nonblock);
5199 case IORING_OP_CLOSE:
5201 ret = io_close_prep(req, sqe);
5205 ret = io_close(req, force_nonblock);
5207 case IORING_OP_FILES_UPDATE:
5209 ret = io_files_update_prep(req, sqe);
5213 ret = io_files_update(req, force_nonblock);
5215 case IORING_OP_STATX:
5217 ret = io_statx_prep(req, sqe);
5221 ret = io_statx(req, force_nonblock);
5223 case IORING_OP_FADVISE:
5225 ret = io_fadvise_prep(req, sqe);
5229 ret = io_fadvise(req, force_nonblock);
5231 case IORING_OP_MADVISE:
5233 ret = io_madvise_prep(req, sqe);
5237 ret = io_madvise(req, force_nonblock);
5239 case IORING_OP_OPENAT2:
5241 ret = io_openat2_prep(req, sqe);
5245 ret = io_openat2(req, force_nonblock);
5247 case IORING_OP_EPOLL_CTL:
5249 ret = io_epoll_ctl_prep(req, sqe);
5253 ret = io_epoll_ctl(req, force_nonblock);
5255 case IORING_OP_SPLICE:
5257 ret = io_splice_prep(req, sqe);
5261 ret = io_splice(req, force_nonblock);
5263 case IORING_OP_PROVIDE_BUFFERS:
5265 ret = io_provide_buffers_prep(req, sqe);
5269 ret = io_provide_buffers(req, force_nonblock);
5271 case IORING_OP_REMOVE_BUFFERS:
5273 ret = io_remove_buffers_prep(req, sqe);
5277 ret = io_remove_buffers(req, force_nonblock);
5287 if (ctx->flags & IORING_SETUP_IOPOLL) {
5288 const bool in_async = io_wq_current_is_worker();
5290 if (req->result == -EAGAIN)
5293 /* workqueue context doesn't hold uring_lock, grab it now */
5295 mutex_lock(&ctx->uring_lock);
5297 io_iopoll_req_issued(req);
5300 mutex_unlock(&ctx->uring_lock);
5306 static void io_wq_submit_work(struct io_wq_work **workptr)
5308 struct io_wq_work *work = *workptr;
5309 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5312 /* if NO_CANCEL is set, we must still run the work */
5313 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5314 IO_WQ_WORK_CANCEL) {
5320 ret = io_issue_sqe(req, NULL, false);
5322 * We can get EAGAIN for polled IO even though we're
5323 * forcing a sync submission from here, since we can't
5324 * wait for request slots on the block side.
5333 req_set_fail_links(req);
5334 io_cqring_add_event(req, ret);
5338 io_steal_work(req, workptr);
5341 static int io_req_needs_file(struct io_kiocb *req, int fd)
5343 if (!io_op_defs[req->opcode].needs_file)
5345 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5350 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5353 struct fixed_file_table *table;
5355 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5356 return table->files[index & IORING_FILE_TABLE_MASK];;
5359 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5360 int fd, struct file **out_file, bool fixed)
5362 struct io_ring_ctx *ctx = req->ctx;
5366 if (unlikely(!ctx->file_data ||
5367 (unsigned) fd >= ctx->nr_user_files))
5369 fd = array_index_nospec(fd, ctx->nr_user_files);
5370 file = io_file_from_index(ctx, fd);
5373 req->fixed_file_refs = ctx->file_data->cur_refs;
5374 percpu_ref_get(req->fixed_file_refs);
5376 trace_io_uring_file_get(ctx, fd);
5377 file = __io_file_get(state, fd);
5378 if (unlikely(!file))
5386 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5387 const struct io_uring_sqe *sqe)
5393 flags = READ_ONCE(sqe->flags);
5394 fd = READ_ONCE(sqe->fd);
5396 if (!io_req_needs_file(req, fd))
5399 fixed = (flags & IOSQE_FIXED_FILE);
5400 if (unlikely(!fixed && req->needs_fixed_file))
5403 return io_file_get(state, req, fd, &req->file, fixed);
5406 static int io_grab_files(struct io_kiocb *req)
5409 struct io_ring_ctx *ctx = req->ctx;
5411 if (req->work.files)
5413 if (!ctx->ring_file)
5417 spin_lock_irq(&ctx->inflight_lock);
5419 * We use the f_ops->flush() handler to ensure that we can flush
5420 * out work accessing these files if the fd is closed. Check if
5421 * the fd has changed since we started down this path, and disallow
5422 * this operation if it has.
5424 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5425 list_add(&req->inflight_entry, &ctx->inflight_list);
5426 req->flags |= REQ_F_INFLIGHT;
5427 req->work.files = current->files;
5430 spin_unlock_irq(&ctx->inflight_lock);
5436 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5438 struct io_timeout_data *data = container_of(timer,
5439 struct io_timeout_data, timer);
5440 struct io_kiocb *req = data->req;
5441 struct io_ring_ctx *ctx = req->ctx;
5442 struct io_kiocb *prev = NULL;
5443 unsigned long flags;
5445 spin_lock_irqsave(&ctx->completion_lock, flags);
5448 * We don't expect the list to be empty, that will only happen if we
5449 * race with the completion of the linked work.
5451 if (!list_empty(&req->link_list)) {
5452 prev = list_entry(req->link_list.prev, struct io_kiocb,
5454 if (refcount_inc_not_zero(&prev->refs)) {
5455 list_del_init(&req->link_list);
5456 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5461 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5464 req_set_fail_links(prev);
5465 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5468 io_cqring_add_event(req, -ETIME);
5471 return HRTIMER_NORESTART;
5474 static void io_queue_linked_timeout(struct io_kiocb *req)
5476 struct io_ring_ctx *ctx = req->ctx;
5479 * If the list is now empty, then our linked request finished before
5480 * we got a chance to setup the timer
5482 spin_lock_irq(&ctx->completion_lock);
5483 if (!list_empty(&req->link_list)) {
5484 struct io_timeout_data *data = &req->io->timeout;
5486 data->timer.function = io_link_timeout_fn;
5487 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5490 spin_unlock_irq(&ctx->completion_lock);
5492 /* drop submission reference */
5496 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5498 struct io_kiocb *nxt;
5500 if (!(req->flags & REQ_F_LINK))
5502 /* for polled retry, if flag is set, we already went through here */
5503 if (req->flags & REQ_F_POLLED)
5506 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5508 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5511 req->flags |= REQ_F_LINK_TIMEOUT;
5515 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5517 struct io_kiocb *linked_timeout;
5518 struct io_kiocb *nxt;
5519 const struct cred *old_creds = NULL;
5523 linked_timeout = io_prep_linked_timeout(req);
5525 if (req->work.creds && req->work.creds != current_cred()) {
5527 revert_creds(old_creds);
5528 if (old_creds == req->work.creds)
5529 old_creds = NULL; /* restored original creds */
5531 old_creds = override_creds(req->work.creds);
5534 ret = io_issue_sqe(req, sqe, true);
5537 * We async punt it if the file wasn't marked NOWAIT, or if the file
5538 * doesn't support non-blocking read/write attempts
5540 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5541 (req->flags & REQ_F_MUST_PUNT))) {
5542 if (io_arm_poll_handler(req)) {
5544 io_queue_linked_timeout(linked_timeout);
5548 if (io_op_defs[req->opcode].file_table) {
5549 ret = io_grab_files(req);
5555 * Queued up for async execution, worker will release
5556 * submit reference when the iocb is actually submitted.
5558 io_queue_async_work(req);
5564 /* drop submission reference */
5565 io_put_req_find_next(req, &nxt);
5567 if (linked_timeout) {
5569 io_queue_linked_timeout(linked_timeout);
5571 io_put_req(linked_timeout);
5574 /* and drop final reference, if we failed */
5576 io_cqring_add_event(req, ret);
5577 req_set_fail_links(req);
5583 if (req->flags & REQ_F_FORCE_ASYNC)
5589 revert_creds(old_creds);
5592 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5596 ret = io_req_defer(req, sqe);
5598 if (ret != -EIOCBQUEUED) {
5600 io_cqring_add_event(req, ret);
5601 req_set_fail_links(req);
5602 io_double_put_req(req);
5604 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5605 ret = io_req_defer_prep(req, sqe);
5606 if (unlikely(ret < 0))
5609 * Never try inline submit of IOSQE_ASYNC is set, go straight
5610 * to async execution.
5612 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5613 io_queue_async_work(req);
5615 __io_queue_sqe(req, sqe);
5619 static inline void io_queue_link_head(struct io_kiocb *req)
5621 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5622 io_cqring_add_event(req, -ECANCELED);
5623 io_double_put_req(req);
5625 io_queue_sqe(req, NULL);
5628 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5629 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5630 IOSQE_BUFFER_SELECT)
5632 static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5633 struct io_submit_state *state, struct io_kiocb **link)
5635 struct io_ring_ctx *ctx = req->ctx;
5636 unsigned int sqe_flags;
5639 sqe_flags = READ_ONCE(sqe->flags);
5641 /* enforce forwards compatibility on users */
5642 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
5647 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5648 !io_op_defs[req->opcode].buffer_select) {
5653 id = READ_ONCE(sqe->personality);
5655 req->work.creds = idr_find(&ctx->personality_idr, id);
5656 if (unlikely(!req->work.creds)) {
5660 get_cred(req->work.creds);
5663 /* same numerical values with corresponding REQ_F_*, safe to copy */
5664 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5665 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5666 IOSQE_BUFFER_SELECT);
5668 ret = io_req_set_file(state, req, sqe);
5669 if (unlikely(ret)) {
5671 io_cqring_add_event(req, ret);
5672 io_double_put_req(req);
5677 * If we already have a head request, queue this one for async
5678 * submittal once the head completes. If we don't have a head but
5679 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5680 * submitted sync once the chain is complete. If none of those
5681 * conditions are true (normal request), then just queue it.
5684 struct io_kiocb *head = *link;
5687 * Taking sequential execution of a link, draining both sides
5688 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5689 * requests in the link. So, it drains the head and the
5690 * next after the link request. The last one is done via
5691 * drain_next flag to persist the effect across calls.
5693 if (sqe_flags & IOSQE_IO_DRAIN) {
5694 head->flags |= REQ_F_IO_DRAIN;
5695 ctx->drain_next = 1;
5697 if (io_alloc_async_ctx(req)) {
5702 ret = io_req_defer_prep(req, sqe);
5704 /* fail even hard links since we don't submit */
5705 head->flags |= REQ_F_FAIL_LINK;
5708 trace_io_uring_link(ctx, req, head);
5709 list_add_tail(&req->link_list, &head->link_list);
5711 /* last request of a link, enqueue the link */
5712 if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) {
5713 io_queue_link_head(head);
5717 if (unlikely(ctx->drain_next)) {
5718 req->flags |= REQ_F_IO_DRAIN;
5719 req->ctx->drain_next = 0;
5721 if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
5722 req->flags |= REQ_F_LINK;
5723 INIT_LIST_HEAD(&req->link_list);
5725 if (io_alloc_async_ctx(req)) {
5729 ret = io_req_defer_prep(req, sqe);
5731 req->flags |= REQ_F_FAIL_LINK;
5734 io_queue_sqe(req, sqe);
5742 * Batched submission is done, ensure local IO is flushed out.
5744 static void io_submit_state_end(struct io_submit_state *state)
5746 blk_finish_plug(&state->plug);
5748 if (state->free_reqs)
5749 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5753 * Start submission side cache.
5755 static void io_submit_state_start(struct io_submit_state *state,
5756 unsigned int max_ios)
5758 blk_start_plug(&state->plug);
5759 state->free_reqs = 0;
5761 state->ios_left = max_ios;
5764 static void io_commit_sqring(struct io_ring_ctx *ctx)
5766 struct io_rings *rings = ctx->rings;
5769 * Ensure any loads from the SQEs are done at this point,
5770 * since once we write the new head, the application could
5771 * write new data to them.
5773 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5777 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5778 * that is mapped by userspace. This means that care needs to be taken to
5779 * ensure that reads are stable, as we cannot rely on userspace always
5780 * being a good citizen. If members of the sqe are validated and then later
5781 * used, it's important that those reads are done through READ_ONCE() to
5782 * prevent a re-load down the line.
5784 static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
5785 const struct io_uring_sqe **sqe_ptr)
5787 u32 *sq_array = ctx->sq_array;
5791 * The cached sq head (or cq tail) serves two purposes:
5793 * 1) allows us to batch the cost of updating the user visible
5795 * 2) allows the kernel side to track the head on its own, even
5796 * though the application is the one updating it.
5798 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5799 if (likely(head < ctx->sq_entries)) {
5801 * All io need record the previous position, if LINK vs DARIN,
5802 * it can be used to mark the position of the first IO in the
5805 req->sequence = ctx->cached_sq_head;
5806 *sqe_ptr = &ctx->sq_sqes[head];
5807 req->opcode = READ_ONCE((*sqe_ptr)->opcode);
5808 req->user_data = READ_ONCE((*sqe_ptr)->user_data);
5809 ctx->cached_sq_head++;
5813 /* drop invalid entries */
5814 ctx->cached_sq_head++;
5815 ctx->cached_sq_dropped++;
5816 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5820 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5821 struct file *ring_file, int ring_fd,
5822 struct mm_struct **mm, bool async)
5824 struct io_submit_state state, *statep = NULL;
5825 struct io_kiocb *link = NULL;
5826 int i, submitted = 0;
5827 bool mm_fault = false;
5829 /* if we have a backlog and couldn't flush it all, return BUSY */
5830 if (test_bit(0, &ctx->sq_check_overflow)) {
5831 if (!list_empty(&ctx->cq_overflow_list) &&
5832 !io_cqring_overflow_flush(ctx, false))
5836 /* make sure SQ entry isn't read before tail */
5837 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5839 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5842 if (nr > IO_PLUG_THRESHOLD) {
5843 io_submit_state_start(&state, nr);
5847 ctx->ring_fd = ring_fd;
5848 ctx->ring_file = ring_file;
5850 for (i = 0; i < nr; i++) {
5851 const struct io_uring_sqe *sqe;
5852 struct io_kiocb *req;
5855 req = io_get_req(ctx, statep);
5856 if (unlikely(!req)) {
5858 submitted = -EAGAIN;
5861 if (!io_get_sqring(ctx, req, &sqe)) {
5862 __io_req_do_free(req);
5866 /* will complete beyond this point, count as submitted */
5869 if (unlikely(req->opcode >= IORING_OP_LAST)) {
5872 io_cqring_add_event(req, err);
5873 io_double_put_req(req);
5877 if (io_op_defs[req->opcode].needs_mm && !*mm) {
5878 mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm);
5879 if (unlikely(mm_fault)) {
5883 use_mm(ctx->sqo_mm);
5887 req->needs_fixed_file = async;
5888 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5890 if (!io_submit_sqe(req, sqe, statep, &link))
5894 if (unlikely(submitted != nr)) {
5895 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5897 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5900 io_queue_link_head(link);
5902 io_submit_state_end(&state);
5904 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5905 io_commit_sqring(ctx);
5910 static int io_sq_thread(void *data)
5912 struct io_ring_ctx *ctx = data;
5913 struct mm_struct *cur_mm = NULL;
5914 const struct cred *old_cred;
5915 mm_segment_t old_fs;
5917 unsigned long timeout;
5920 complete(&ctx->completions[1]);
5924 old_cred = override_creds(ctx->creds);
5926 timeout = jiffies + ctx->sq_thread_idle;
5927 while (!kthread_should_park()) {
5928 unsigned int to_submit;
5930 if (!list_empty(&ctx->poll_list)) {
5931 unsigned nr_events = 0;
5933 mutex_lock(&ctx->uring_lock);
5934 if (!list_empty(&ctx->poll_list))
5935 io_iopoll_getevents(ctx, &nr_events, 0);
5937 timeout = jiffies + ctx->sq_thread_idle;
5938 mutex_unlock(&ctx->uring_lock);
5941 to_submit = io_sqring_entries(ctx);
5944 * If submit got -EBUSY, flag us as needing the application
5945 * to enter the kernel to reap and flush events.
5947 if (!to_submit || ret == -EBUSY) {
5949 * Drop cur_mm before scheduling, we can't hold it for
5950 * long periods (or over schedule()). Do this before
5951 * adding ourselves to the waitqueue, as the unuse/drop
5961 * We're polling. If we're within the defined idle
5962 * period, then let us spin without work before going
5963 * to sleep. The exception is if we got EBUSY doing
5964 * more IO, we should wait for the application to
5965 * reap events and wake us up.
5967 if (!list_empty(&ctx->poll_list) ||
5968 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5969 !percpu_ref_is_dying(&ctx->refs))) {
5970 if (current->task_works)
5976 prepare_to_wait(&ctx->sqo_wait, &wait,
5977 TASK_INTERRUPTIBLE);
5980 * While doing polled IO, before going to sleep, we need
5981 * to check if there are new reqs added to poll_list, it
5982 * is because reqs may have been punted to io worker and
5983 * will be added to poll_list later, hence check the
5986 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
5987 !list_empty_careful(&ctx->poll_list)) {
5988 finish_wait(&ctx->sqo_wait, &wait);
5992 /* Tell userspace we may need a wakeup call */
5993 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
5994 /* make sure to read SQ tail after writing flags */
5997 to_submit = io_sqring_entries(ctx);
5998 if (!to_submit || ret == -EBUSY) {
5999 if (kthread_should_park()) {
6000 finish_wait(&ctx->sqo_wait, &wait);
6003 if (current->task_works) {
6005 finish_wait(&ctx->sqo_wait, &wait);
6008 if (signal_pending(current))
6009 flush_signals(current);
6011 finish_wait(&ctx->sqo_wait, &wait);
6013 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6016 finish_wait(&ctx->sqo_wait, &wait);
6018 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6021 mutex_lock(&ctx->uring_lock);
6022 ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true);
6023 mutex_unlock(&ctx->uring_lock);
6024 timeout = jiffies + ctx->sq_thread_idle;
6027 if (current->task_works)
6035 revert_creds(old_cred);
6042 struct io_wait_queue {
6043 struct wait_queue_entry wq;
6044 struct io_ring_ctx *ctx;
6046 unsigned nr_timeouts;
6049 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6051 struct io_ring_ctx *ctx = iowq->ctx;
6054 * Wake up if we have enough events, or if a timeout occurred since we
6055 * started waiting. For timeouts, we always want to return to userspace,
6056 * regardless of event count.
6058 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6059 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6062 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6063 int wake_flags, void *key)
6065 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6068 /* use noflush == true, as we can't safely rely on locking context */
6069 if (!io_should_wake(iowq, true))
6072 return autoremove_wake_function(curr, mode, wake_flags, key);
6076 * Wait until events become available, if we don't already have some. The
6077 * application must reap them itself, as they reside on the shared cq ring.
6079 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6080 const sigset_t __user *sig, size_t sigsz)
6082 struct io_wait_queue iowq = {
6085 .func = io_wake_function,
6086 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6089 .to_wait = min_events,
6091 struct io_rings *rings = ctx->rings;
6095 if (io_cqring_events(ctx, false) >= min_events)
6097 if (!current->task_works)
6103 #ifdef CONFIG_COMPAT
6104 if (in_compat_syscall())
6105 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6109 ret = set_user_sigmask(sig, sigsz);
6115 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6116 trace_io_uring_cqring_wait(ctx, min_events);
6118 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6119 TASK_INTERRUPTIBLE);
6120 if (current->task_works)
6122 if (io_should_wake(&iowq, false))
6125 if (signal_pending(current)) {
6130 finish_wait(&ctx->wait, &iowq.wq);
6132 restore_saved_sigmask_unless(ret == -EINTR);
6134 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6137 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6139 #if defined(CONFIG_UNIX)
6140 if (ctx->ring_sock) {
6141 struct sock *sock = ctx->ring_sock->sk;
6142 struct sk_buff *skb;
6144 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6150 for (i = 0; i < ctx->nr_user_files; i++) {
6153 file = io_file_from_index(ctx, i);
6160 static void io_file_ref_kill(struct percpu_ref *ref)
6162 struct fixed_file_data *data;
6164 data = container_of(ref, struct fixed_file_data, refs);
6165 complete(&data->done);
6168 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6170 struct fixed_file_data *data = ctx->file_data;
6171 struct fixed_file_ref_node *ref_node = NULL;
6172 unsigned nr_tables, i;
6173 unsigned long flags;
6178 spin_lock_irqsave(&data->lock, flags);
6179 if (!list_empty(&data->ref_list))
6180 ref_node = list_first_entry(&data->ref_list,
6181 struct fixed_file_ref_node, node);
6182 spin_unlock_irqrestore(&data->lock, flags);
6184 percpu_ref_kill(&ref_node->refs);
6186 percpu_ref_kill(&data->refs);
6188 /* wait for all refs nodes to complete */
6189 wait_for_completion(&data->done);
6191 __io_sqe_files_unregister(ctx);
6192 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6193 for (i = 0; i < nr_tables; i++)
6194 kfree(data->table[i].files);
6196 percpu_ref_exit(&data->refs);
6198 ctx->file_data = NULL;
6199 ctx->nr_user_files = 0;
6203 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6205 if (ctx->sqo_thread) {
6206 wait_for_completion(&ctx->completions[1]);
6208 * The park is a bit of a work-around, without it we get
6209 * warning spews on shutdown with SQPOLL set and affinity
6210 * set to a single CPU.
6212 kthread_park(ctx->sqo_thread);
6213 kthread_stop(ctx->sqo_thread);
6214 ctx->sqo_thread = NULL;
6218 static void io_finish_async(struct io_ring_ctx *ctx)
6220 io_sq_thread_stop(ctx);
6223 io_wq_destroy(ctx->io_wq);
6228 #if defined(CONFIG_UNIX)
6230 * Ensure the UNIX gc is aware of our file set, so we are certain that
6231 * the io_uring can be safely unregistered on process exit, even if we have
6232 * loops in the file referencing.
6234 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6236 struct sock *sk = ctx->ring_sock->sk;
6237 struct scm_fp_list *fpl;
6238 struct sk_buff *skb;
6241 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6245 skb = alloc_skb(0, GFP_KERNEL);
6254 fpl->user = get_uid(ctx->user);
6255 for (i = 0; i < nr; i++) {
6256 struct file *file = io_file_from_index(ctx, i + offset);
6260 fpl->fp[nr_files] = get_file(file);
6261 unix_inflight(fpl->user, fpl->fp[nr_files]);
6266 fpl->max = SCM_MAX_FD;
6267 fpl->count = nr_files;
6268 UNIXCB(skb).fp = fpl;
6269 skb->destructor = unix_destruct_scm;
6270 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6271 skb_queue_head(&sk->sk_receive_queue, skb);
6273 for (i = 0; i < nr_files; i++)
6284 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6285 * causes regular reference counting to break down. We rely on the UNIX
6286 * garbage collection to take care of this problem for us.
6288 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6290 unsigned left, total;
6294 left = ctx->nr_user_files;
6296 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6298 ret = __io_sqe_files_scm(ctx, this_files, total);
6302 total += this_files;
6308 while (total < ctx->nr_user_files) {
6309 struct file *file = io_file_from_index(ctx, total);
6319 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6325 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6330 for (i = 0; i < nr_tables; i++) {
6331 struct fixed_file_table *table = &ctx->file_data->table[i];
6332 unsigned this_files;
6334 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6335 table->files = kcalloc(this_files, sizeof(struct file *),
6339 nr_files -= this_files;
6345 for (i = 0; i < nr_tables; i++) {
6346 struct fixed_file_table *table = &ctx->file_data->table[i];
6347 kfree(table->files);
6352 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6354 #if defined(CONFIG_UNIX)
6355 struct sock *sock = ctx->ring_sock->sk;
6356 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6357 struct sk_buff *skb;
6360 __skb_queue_head_init(&list);
6363 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6364 * remove this entry and rearrange the file array.
6366 skb = skb_dequeue(head);
6368 struct scm_fp_list *fp;
6370 fp = UNIXCB(skb).fp;
6371 for (i = 0; i < fp->count; i++) {
6374 if (fp->fp[i] != file)
6377 unix_notinflight(fp->user, fp->fp[i]);
6378 left = fp->count - 1 - i;
6380 memmove(&fp->fp[i], &fp->fp[i + 1],
6381 left * sizeof(struct file *));
6388 __skb_queue_tail(&list, skb);
6398 __skb_queue_tail(&list, skb);
6400 skb = skb_dequeue(head);
6403 if (skb_peek(&list)) {
6404 spin_lock_irq(&head->lock);
6405 while ((skb = __skb_dequeue(&list)) != NULL)
6406 __skb_queue_tail(head, skb);
6407 spin_unlock_irq(&head->lock);
6414 struct io_file_put {
6415 struct list_head list;
6419 static void io_file_put_work(struct work_struct *work)
6421 struct fixed_file_ref_node *ref_node;
6422 struct fixed_file_data *file_data;
6423 struct io_ring_ctx *ctx;
6424 struct io_file_put *pfile, *tmp;
6425 unsigned long flags;
6427 ref_node = container_of(work, struct fixed_file_ref_node, work);
6428 file_data = ref_node->file_data;
6429 ctx = file_data->ctx;
6431 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6432 list_del_init(&pfile->list);
6433 io_ring_file_put(ctx, pfile->file);
6437 spin_lock_irqsave(&file_data->lock, flags);
6438 list_del_init(&ref_node->node);
6439 spin_unlock_irqrestore(&file_data->lock, flags);
6441 percpu_ref_exit(&ref_node->refs);
6443 percpu_ref_put(&file_data->refs);
6446 static void io_file_data_ref_zero(struct percpu_ref *ref)
6448 struct fixed_file_ref_node *ref_node;
6450 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6452 queue_work(system_wq, &ref_node->work);
6455 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6456 struct io_ring_ctx *ctx)
6458 struct fixed_file_ref_node *ref_node;
6460 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6462 return ERR_PTR(-ENOMEM);
6464 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6467 return ERR_PTR(-ENOMEM);
6469 INIT_LIST_HEAD(&ref_node->node);
6470 INIT_LIST_HEAD(&ref_node->file_list);
6471 INIT_WORK(&ref_node->work, io_file_put_work);
6472 ref_node->file_data = ctx->file_data;
6477 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6479 percpu_ref_exit(&ref_node->refs);
6483 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6486 __s32 __user *fds = (__s32 __user *) arg;
6491 struct fixed_file_ref_node *ref_node;
6492 unsigned long flags;
6498 if (nr_args > IORING_MAX_FIXED_FILES)
6501 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6502 if (!ctx->file_data)
6504 ctx->file_data->ctx = ctx;
6505 init_completion(&ctx->file_data->done);
6506 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6508 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6509 ctx->file_data->table = kcalloc(nr_tables,
6510 sizeof(struct fixed_file_table),
6512 if (!ctx->file_data->table) {
6513 kfree(ctx->file_data);
6514 ctx->file_data = NULL;
6518 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6519 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6520 kfree(ctx->file_data->table);
6521 kfree(ctx->file_data);
6522 ctx->file_data = NULL;
6526 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6527 percpu_ref_exit(&ctx->file_data->refs);
6528 kfree(ctx->file_data->table);
6529 kfree(ctx->file_data);
6530 ctx->file_data = NULL;
6534 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6535 struct fixed_file_table *table;
6539 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6541 /* allow sparse sets */
6547 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6548 index = i & IORING_FILE_TABLE_MASK;
6556 * Don't allow io_uring instances to be registered. If UNIX
6557 * isn't enabled, then this causes a reference cycle and this
6558 * instance can never get freed. If UNIX is enabled we'll
6559 * handle it just fine, but there's still no point in allowing
6560 * a ring fd as it doesn't support regular read/write anyway.
6562 if (file->f_op == &io_uring_fops) {
6567 table->files[index] = file;
6571 for (i = 0; i < ctx->nr_user_files; i++) {
6572 file = io_file_from_index(ctx, i);
6576 for (i = 0; i < nr_tables; i++)
6577 kfree(ctx->file_data->table[i].files);
6579 kfree(ctx->file_data->table);
6580 kfree(ctx->file_data);
6581 ctx->file_data = NULL;
6582 ctx->nr_user_files = 0;
6586 ret = io_sqe_files_scm(ctx);
6588 io_sqe_files_unregister(ctx);
6592 ref_node = alloc_fixed_file_ref_node(ctx);
6593 if (IS_ERR(ref_node)) {
6594 io_sqe_files_unregister(ctx);
6595 return PTR_ERR(ref_node);
6598 ctx->file_data->cur_refs = &ref_node->refs;
6599 spin_lock_irqsave(&ctx->file_data->lock, flags);
6600 list_add(&ref_node->node, &ctx->file_data->ref_list);
6601 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6602 percpu_ref_get(&ctx->file_data->refs);
6606 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6609 #if defined(CONFIG_UNIX)
6610 struct sock *sock = ctx->ring_sock->sk;
6611 struct sk_buff_head *head = &sock->sk_receive_queue;
6612 struct sk_buff *skb;
6615 * See if we can merge this file into an existing skb SCM_RIGHTS
6616 * file set. If there's no room, fall back to allocating a new skb
6617 * and filling it in.
6619 spin_lock_irq(&head->lock);
6620 skb = skb_peek(head);
6622 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6624 if (fpl->count < SCM_MAX_FD) {
6625 __skb_unlink(skb, head);
6626 spin_unlock_irq(&head->lock);
6627 fpl->fp[fpl->count] = get_file(file);
6628 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6630 spin_lock_irq(&head->lock);
6631 __skb_queue_head(head, skb);
6636 spin_unlock_irq(&head->lock);
6643 return __io_sqe_files_scm(ctx, 1, index);
6649 static int io_queue_file_removal(struct fixed_file_data *data,
6652 struct io_file_put *pfile;
6653 struct percpu_ref *refs = data->cur_refs;
6654 struct fixed_file_ref_node *ref_node;
6656 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6660 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6662 list_add(&pfile->list, &ref_node->file_list);
6667 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6668 struct io_uring_files_update *up,
6671 struct fixed_file_data *data = ctx->file_data;
6672 struct fixed_file_ref_node *ref_node;
6677 unsigned long flags;
6678 bool needs_switch = false;
6680 if (check_add_overflow(up->offset, nr_args, &done))
6682 if (done > ctx->nr_user_files)
6685 ref_node = alloc_fixed_file_ref_node(ctx);
6686 if (IS_ERR(ref_node))
6687 return PTR_ERR(ref_node);
6690 fds = u64_to_user_ptr(up->fds);
6692 struct fixed_file_table *table;
6696 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6700 i = array_index_nospec(up->offset, ctx->nr_user_files);
6701 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6702 index = i & IORING_FILE_TABLE_MASK;
6703 if (table->files[index]) {
6704 file = io_file_from_index(ctx, index);
6705 err = io_queue_file_removal(data, file);
6708 table->files[index] = NULL;
6709 needs_switch = true;
6718 * Don't allow io_uring instances to be registered. If
6719 * UNIX isn't enabled, then this causes a reference
6720 * cycle and this instance can never get freed. If UNIX
6721 * is enabled we'll handle it just fine, but there's
6722 * still no point in allowing a ring fd as it doesn't
6723 * support regular read/write anyway.
6725 if (file->f_op == &io_uring_fops) {
6730 table->files[index] = file;
6731 err = io_sqe_file_register(ctx, file, i);
6741 percpu_ref_kill(data->cur_refs);
6742 spin_lock_irqsave(&data->lock, flags);
6743 list_add(&ref_node->node, &data->ref_list);
6744 data->cur_refs = &ref_node->refs;
6745 spin_unlock_irqrestore(&data->lock, flags);
6746 percpu_ref_get(&ctx->file_data->refs);
6748 destroy_fixed_file_ref_node(ref_node);
6750 return done ? done : err;
6753 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6756 struct io_uring_files_update up;
6758 if (!ctx->file_data)
6762 if (copy_from_user(&up, arg, sizeof(up)))
6767 return __io_sqe_files_update(ctx, &up, nr_args);
6770 static void io_free_work(struct io_wq_work *work)
6772 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6774 /* Consider that io_steal_work() relies on this ref */
6778 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6779 struct io_uring_params *p)
6781 struct io_wq_data data;
6783 struct io_ring_ctx *ctx_attach;
6784 unsigned int concurrency;
6787 data.user = ctx->user;
6788 data.free_work = io_free_work;
6790 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6791 /* Do QD, or 4 * CPUS, whatever is smallest */
6792 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6794 ctx->io_wq = io_wq_create(concurrency, &data);
6795 if (IS_ERR(ctx->io_wq)) {
6796 ret = PTR_ERR(ctx->io_wq);
6802 f = fdget(p->wq_fd);
6806 if (f.file->f_op != &io_uring_fops) {
6811 ctx_attach = f.file->private_data;
6812 /* @io_wq is protected by holding the fd */
6813 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6818 ctx->io_wq = ctx_attach->io_wq;
6824 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6825 struct io_uring_params *p)
6829 init_waitqueue_head(&ctx->sqo_wait);
6830 mmgrab(current->mm);
6831 ctx->sqo_mm = current->mm;
6833 if (ctx->flags & IORING_SETUP_SQPOLL) {
6835 if (!capable(CAP_SYS_ADMIN))
6838 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6839 if (!ctx->sq_thread_idle)
6840 ctx->sq_thread_idle = HZ;
6842 if (p->flags & IORING_SETUP_SQ_AFF) {
6843 int cpu = p->sq_thread_cpu;
6846 if (cpu >= nr_cpu_ids)
6848 if (!cpu_online(cpu))
6851 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6855 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6858 if (IS_ERR(ctx->sqo_thread)) {
6859 ret = PTR_ERR(ctx->sqo_thread);
6860 ctx->sqo_thread = NULL;
6863 wake_up_process(ctx->sqo_thread);
6864 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6865 /* Can't have SQ_AFF without SQPOLL */
6870 ret = io_init_wq_offload(ctx, p);
6876 io_finish_async(ctx);
6877 mmdrop(ctx->sqo_mm);
6882 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6884 atomic_long_sub(nr_pages, &user->locked_vm);
6887 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6889 unsigned long page_limit, cur_pages, new_pages;
6891 /* Don't allow more pages than we can safely lock */
6892 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6895 cur_pages = atomic_long_read(&user->locked_vm);
6896 new_pages = cur_pages + nr_pages;
6897 if (new_pages > page_limit)
6899 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6900 new_pages) != cur_pages);
6905 static void io_mem_free(void *ptr)
6912 page = virt_to_head_page(ptr);
6913 if (put_page_testzero(page))
6914 free_compound_page(page);
6917 static void *io_mem_alloc(size_t size)
6919 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6922 return (void *) __get_free_pages(gfp_flags, get_order(size));
6925 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6928 struct io_rings *rings;
6929 size_t off, sq_array_size;
6931 off = struct_size(rings, cqes, cq_entries);
6932 if (off == SIZE_MAX)
6936 off = ALIGN(off, SMP_CACHE_BYTES);
6941 sq_array_size = array_size(sizeof(u32), sq_entries);
6942 if (sq_array_size == SIZE_MAX)
6945 if (check_add_overflow(off, sq_array_size, &off))
6954 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6958 pages = (size_t)1 << get_order(
6959 rings_size(sq_entries, cq_entries, NULL));
6960 pages += (size_t)1 << get_order(
6961 array_size(sizeof(struct io_uring_sqe), sq_entries));
6966 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6970 if (!ctx->user_bufs)
6973 for (i = 0; i < ctx->nr_user_bufs; i++) {
6974 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
6976 for (j = 0; j < imu->nr_bvecs; j++)
6977 unpin_user_page(imu->bvec[j].bv_page);
6979 if (ctx->account_mem)
6980 io_unaccount_mem(ctx->user, imu->nr_bvecs);
6985 kfree(ctx->user_bufs);
6986 ctx->user_bufs = NULL;
6987 ctx->nr_user_bufs = 0;
6991 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
6992 void __user *arg, unsigned index)
6994 struct iovec __user *src;
6996 #ifdef CONFIG_COMPAT
6998 struct compat_iovec __user *ciovs;
6999 struct compat_iovec ciov;
7001 ciovs = (struct compat_iovec __user *) arg;
7002 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7005 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7006 dst->iov_len = ciov.iov_len;
7010 src = (struct iovec __user *) arg;
7011 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7016 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7019 struct vm_area_struct **vmas = NULL;
7020 struct page **pages = NULL;
7021 int i, j, got_pages = 0;
7026 if (!nr_args || nr_args > UIO_MAXIOV)
7029 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7031 if (!ctx->user_bufs)
7034 for (i = 0; i < nr_args; i++) {
7035 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7036 unsigned long off, start, end, ubuf;
7041 ret = io_copy_iov(ctx, &iov, arg, i);
7046 * Don't impose further limits on the size and buffer
7047 * constraints here, we'll -EINVAL later when IO is
7048 * submitted if they are wrong.
7051 if (!iov.iov_base || !iov.iov_len)
7054 /* arbitrary limit, but we need something */
7055 if (iov.iov_len > SZ_1G)
7058 ubuf = (unsigned long) iov.iov_base;
7059 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7060 start = ubuf >> PAGE_SHIFT;
7061 nr_pages = end - start;
7063 if (ctx->account_mem) {
7064 ret = io_account_mem(ctx->user, nr_pages);
7070 if (!pages || nr_pages > got_pages) {
7073 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7075 vmas = kvmalloc_array(nr_pages,
7076 sizeof(struct vm_area_struct *),
7078 if (!pages || !vmas) {
7080 if (ctx->account_mem)
7081 io_unaccount_mem(ctx->user, nr_pages);
7084 got_pages = nr_pages;
7087 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7091 if (ctx->account_mem)
7092 io_unaccount_mem(ctx->user, nr_pages);
7097 down_read(¤t->mm->mmap_sem);
7098 pret = pin_user_pages(ubuf, nr_pages,
7099 FOLL_WRITE | FOLL_LONGTERM,
7101 if (pret == nr_pages) {
7102 /* don't support file backed memory */
7103 for (j = 0; j < nr_pages; j++) {
7104 struct vm_area_struct *vma = vmas[j];
7107 !is_file_hugepages(vma->vm_file)) {
7113 ret = pret < 0 ? pret : -EFAULT;
7115 up_read(¤t->mm->mmap_sem);
7118 * if we did partial map, or found file backed vmas,
7119 * release any pages we did get
7122 unpin_user_pages(pages, pret);
7123 if (ctx->account_mem)
7124 io_unaccount_mem(ctx->user, nr_pages);
7129 off = ubuf & ~PAGE_MASK;
7131 for (j = 0; j < nr_pages; j++) {
7134 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7135 imu->bvec[j].bv_page = pages[j];
7136 imu->bvec[j].bv_len = vec_len;
7137 imu->bvec[j].bv_offset = off;
7141 /* store original address for later verification */
7143 imu->len = iov.iov_len;
7144 imu->nr_bvecs = nr_pages;
7146 ctx->nr_user_bufs++;
7154 io_sqe_buffer_unregister(ctx);
7158 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7160 __s32 __user *fds = arg;
7166 if (copy_from_user(&fd, fds, sizeof(*fds)))
7169 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7170 if (IS_ERR(ctx->cq_ev_fd)) {
7171 int ret = PTR_ERR(ctx->cq_ev_fd);
7172 ctx->cq_ev_fd = NULL;
7179 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7181 if (ctx->cq_ev_fd) {
7182 eventfd_ctx_put(ctx->cq_ev_fd);
7183 ctx->cq_ev_fd = NULL;
7190 static int __io_destroy_buffers(int id, void *p, void *data)
7192 struct io_ring_ctx *ctx = data;
7193 struct io_buffer *buf = p;
7195 __io_remove_buffers(ctx, buf, id, -1U);
7199 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7201 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7202 idr_destroy(&ctx->io_buffer_idr);
7205 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7207 io_finish_async(ctx);
7209 mmdrop(ctx->sqo_mm);
7211 io_iopoll_reap_events(ctx);
7212 io_sqe_buffer_unregister(ctx);
7213 io_sqe_files_unregister(ctx);
7214 io_eventfd_unregister(ctx);
7215 io_destroy_buffers(ctx);
7216 idr_destroy(&ctx->personality_idr);
7218 #if defined(CONFIG_UNIX)
7219 if (ctx->ring_sock) {
7220 ctx->ring_sock->file = NULL; /* so that iput() is called */
7221 sock_release(ctx->ring_sock);
7225 io_mem_free(ctx->rings);
7226 io_mem_free(ctx->sq_sqes);
7228 percpu_ref_exit(&ctx->refs);
7229 if (ctx->account_mem)
7230 io_unaccount_mem(ctx->user,
7231 ring_pages(ctx->sq_entries, ctx->cq_entries));
7232 free_uid(ctx->user);
7233 put_cred(ctx->creds);
7234 kfree(ctx->completions);
7235 kfree(ctx->cancel_hash);
7236 kmem_cache_free(req_cachep, ctx->fallback_req);
7240 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7242 struct io_ring_ctx *ctx = file->private_data;
7245 poll_wait(file, &ctx->cq_wait, wait);
7247 * synchronizes with barrier from wq_has_sleeper call in
7251 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7252 ctx->rings->sq_ring_entries)
7253 mask |= EPOLLOUT | EPOLLWRNORM;
7254 if (io_cqring_events(ctx, false))
7255 mask |= EPOLLIN | EPOLLRDNORM;
7260 static int io_uring_fasync(int fd, struct file *file, int on)
7262 struct io_ring_ctx *ctx = file->private_data;
7264 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7267 static int io_remove_personalities(int id, void *p, void *data)
7269 struct io_ring_ctx *ctx = data;
7270 const struct cred *cred;
7272 cred = idr_remove(&ctx->personality_idr, id);
7278 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7280 mutex_lock(&ctx->uring_lock);
7281 percpu_ref_kill(&ctx->refs);
7282 mutex_unlock(&ctx->uring_lock);
7285 * Wait for sq thread to idle, if we have one. It won't spin on new
7286 * work after we've killed the ctx ref above. This is important to do
7287 * before we cancel existing commands, as the thread could otherwise
7288 * be queueing new work post that. If that's work we need to cancel,
7289 * it could cause shutdown to hang.
7291 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7294 io_kill_timeouts(ctx);
7295 io_poll_remove_all(ctx);
7298 io_wq_cancel_all(ctx->io_wq);
7300 io_iopoll_reap_events(ctx);
7301 /* if we failed setting up the ctx, we might not have any rings */
7303 io_cqring_overflow_flush(ctx, true);
7304 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7305 wait_for_completion(&ctx->completions[0]);
7306 io_ring_ctx_free(ctx);
7309 static int io_uring_release(struct inode *inode, struct file *file)
7311 struct io_ring_ctx *ctx = file->private_data;
7313 file->private_data = NULL;
7314 io_ring_ctx_wait_and_kill(ctx);
7318 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7319 struct files_struct *files)
7321 struct io_kiocb *req;
7324 while (!list_empty_careful(&ctx->inflight_list)) {
7325 struct io_kiocb *cancel_req = NULL;
7327 spin_lock_irq(&ctx->inflight_lock);
7328 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7329 if (req->work.files != files)
7331 /* req is being completed, ignore */
7332 if (!refcount_inc_not_zero(&req->refs))
7338 prepare_to_wait(&ctx->inflight_wait, &wait,
7339 TASK_UNINTERRUPTIBLE);
7340 spin_unlock_irq(&ctx->inflight_lock);
7342 /* We need to keep going until we don't find a matching req */
7346 if (cancel_req->flags & REQ_F_OVERFLOW) {
7347 spin_lock_irq(&ctx->completion_lock);
7348 list_del(&cancel_req->list);
7349 cancel_req->flags &= ~REQ_F_OVERFLOW;
7350 if (list_empty(&ctx->cq_overflow_list)) {
7351 clear_bit(0, &ctx->sq_check_overflow);
7352 clear_bit(0, &ctx->cq_check_overflow);
7354 spin_unlock_irq(&ctx->completion_lock);
7356 WRITE_ONCE(ctx->rings->cq_overflow,
7357 atomic_inc_return(&ctx->cached_cq_overflow));
7360 * Put inflight ref and overflow ref. If that's
7361 * all we had, then we're done with this request.
7363 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7364 io_put_req(cancel_req);
7369 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7370 io_put_req(cancel_req);
7373 finish_wait(&ctx->inflight_wait, &wait);
7376 static int io_uring_flush(struct file *file, void *data)
7378 struct io_ring_ctx *ctx = file->private_data;
7380 io_uring_cancel_files(ctx, data);
7383 * If the task is going away, cancel work it may have pending
7385 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7386 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7391 static void *io_uring_validate_mmap_request(struct file *file,
7392 loff_t pgoff, size_t sz)
7394 struct io_ring_ctx *ctx = file->private_data;
7395 loff_t offset = pgoff << PAGE_SHIFT;
7400 case IORING_OFF_SQ_RING:
7401 case IORING_OFF_CQ_RING:
7404 case IORING_OFF_SQES:
7408 return ERR_PTR(-EINVAL);
7411 page = virt_to_head_page(ptr);
7412 if (sz > page_size(page))
7413 return ERR_PTR(-EINVAL);
7420 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7422 size_t sz = vma->vm_end - vma->vm_start;
7426 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7428 return PTR_ERR(ptr);
7430 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7431 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7434 #else /* !CONFIG_MMU */
7436 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7438 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7441 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7443 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7446 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7447 unsigned long addr, unsigned long len,
7448 unsigned long pgoff, unsigned long flags)
7452 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7454 return PTR_ERR(ptr);
7456 return (unsigned long) ptr;
7459 #endif /* !CONFIG_MMU */
7461 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7462 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7465 struct io_ring_ctx *ctx;
7470 if (current->task_works)
7473 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7481 if (f.file->f_op != &io_uring_fops)
7485 ctx = f.file->private_data;
7486 if (!percpu_ref_tryget(&ctx->refs))
7490 * For SQ polling, the thread will do all submissions and completions.
7491 * Just return the requested submit count, and wake the thread if
7495 if (ctx->flags & IORING_SETUP_SQPOLL) {
7496 if (!list_empty_careful(&ctx->cq_overflow_list))
7497 io_cqring_overflow_flush(ctx, false);
7498 if (flags & IORING_ENTER_SQ_WAKEUP)
7499 wake_up(&ctx->sqo_wait);
7500 submitted = to_submit;
7501 } else if (to_submit) {
7502 struct mm_struct *cur_mm;
7504 mutex_lock(&ctx->uring_lock);
7505 /* already have mm, so io_submit_sqes() won't try to grab it */
7506 cur_mm = ctx->sqo_mm;
7507 submitted = io_submit_sqes(ctx, to_submit, f.file, fd,
7509 mutex_unlock(&ctx->uring_lock);
7511 if (submitted != to_submit)
7514 if (flags & IORING_ENTER_GETEVENTS) {
7515 unsigned nr_events = 0;
7517 min_complete = min(min_complete, ctx->cq_entries);
7520 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7521 * space applications don't need to do io completion events
7522 * polling again, they can rely on io_sq_thread to do polling
7523 * work, which can reduce cpu usage and uring_lock contention.
7525 if (ctx->flags & IORING_SETUP_IOPOLL &&
7526 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7527 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7529 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7534 percpu_ref_put(&ctx->refs);
7537 return submitted ? submitted : ret;
7540 #ifdef CONFIG_PROC_FS
7541 static int io_uring_show_cred(int id, void *p, void *data)
7543 const struct cred *cred = p;
7544 struct seq_file *m = data;
7545 struct user_namespace *uns = seq_user_ns(m);
7546 struct group_info *gi;
7551 seq_printf(m, "%5d\n", id);
7552 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7553 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7554 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7555 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7556 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7557 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7558 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7559 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7560 seq_puts(m, "\n\tGroups:\t");
7561 gi = cred->group_info;
7562 for (g = 0; g < gi->ngroups; g++) {
7563 seq_put_decimal_ull(m, g ? " " : "",
7564 from_kgid_munged(uns, gi->gid[g]));
7566 seq_puts(m, "\n\tCapEff:\t");
7567 cap = cred->cap_effective;
7568 CAP_FOR_EACH_U32(__capi)
7569 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7574 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7578 mutex_lock(&ctx->uring_lock);
7579 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7580 for (i = 0; i < ctx->nr_user_files; i++) {
7581 struct fixed_file_table *table;
7584 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7585 f = table->files[i & IORING_FILE_TABLE_MASK];
7587 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7589 seq_printf(m, "%5u: <none>\n", i);
7591 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7592 for (i = 0; i < ctx->nr_user_bufs; i++) {
7593 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7595 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7596 (unsigned int) buf->len);
7598 if (!idr_is_empty(&ctx->personality_idr)) {
7599 seq_printf(m, "Personalities:\n");
7600 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7602 seq_printf(m, "PollList:\n");
7603 spin_lock_irq(&ctx->completion_lock);
7604 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7605 struct hlist_head *list = &ctx->cancel_hash[i];
7606 struct io_kiocb *req;
7608 hlist_for_each_entry(req, list, hash_node)
7609 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7610 req->task->task_works != NULL);
7612 spin_unlock_irq(&ctx->completion_lock);
7613 mutex_unlock(&ctx->uring_lock);
7616 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7618 struct io_ring_ctx *ctx = f->private_data;
7620 if (percpu_ref_tryget(&ctx->refs)) {
7621 __io_uring_show_fdinfo(ctx, m);
7622 percpu_ref_put(&ctx->refs);
7627 static const struct file_operations io_uring_fops = {
7628 .release = io_uring_release,
7629 .flush = io_uring_flush,
7630 .mmap = io_uring_mmap,
7632 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7633 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7635 .poll = io_uring_poll,
7636 .fasync = io_uring_fasync,
7637 #ifdef CONFIG_PROC_FS
7638 .show_fdinfo = io_uring_show_fdinfo,
7642 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7643 struct io_uring_params *p)
7645 struct io_rings *rings;
7646 size_t size, sq_array_offset;
7648 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7649 if (size == SIZE_MAX)
7652 rings = io_mem_alloc(size);
7657 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7658 rings->sq_ring_mask = p->sq_entries - 1;
7659 rings->cq_ring_mask = p->cq_entries - 1;
7660 rings->sq_ring_entries = p->sq_entries;
7661 rings->cq_ring_entries = p->cq_entries;
7662 ctx->sq_mask = rings->sq_ring_mask;
7663 ctx->cq_mask = rings->cq_ring_mask;
7664 ctx->sq_entries = rings->sq_ring_entries;
7665 ctx->cq_entries = rings->cq_ring_entries;
7667 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7668 if (size == SIZE_MAX) {
7669 io_mem_free(ctx->rings);
7674 ctx->sq_sqes = io_mem_alloc(size);
7675 if (!ctx->sq_sqes) {
7676 io_mem_free(ctx->rings);
7685 * Allocate an anonymous fd, this is what constitutes the application
7686 * visible backing of an io_uring instance. The application mmaps this
7687 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7688 * we have to tie this fd to a socket for file garbage collection purposes.
7690 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7695 #if defined(CONFIG_UNIX)
7696 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7702 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7706 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7707 O_RDWR | O_CLOEXEC);
7710 ret = PTR_ERR(file);
7714 #if defined(CONFIG_UNIX)
7715 ctx->ring_sock->file = file;
7717 fd_install(ret, file);
7720 #if defined(CONFIG_UNIX)
7721 sock_release(ctx->ring_sock);
7722 ctx->ring_sock = NULL;
7727 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7729 struct user_struct *user = NULL;
7730 struct io_ring_ctx *ctx;
7736 if (entries > IORING_MAX_ENTRIES) {
7737 if (!(p->flags & IORING_SETUP_CLAMP))
7739 entries = IORING_MAX_ENTRIES;
7743 * Use twice as many entries for the CQ ring. It's possible for the
7744 * application to drive a higher depth than the size of the SQ ring,
7745 * since the sqes are only used at submission time. This allows for
7746 * some flexibility in overcommitting a bit. If the application has
7747 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7748 * of CQ ring entries manually.
7750 p->sq_entries = roundup_pow_of_two(entries);
7751 if (p->flags & IORING_SETUP_CQSIZE) {
7753 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7754 * to a power-of-two, if it isn't already. We do NOT impose
7755 * any cq vs sq ring sizing.
7757 if (p->cq_entries < p->sq_entries)
7759 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7760 if (!(p->flags & IORING_SETUP_CLAMP))
7762 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7764 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7766 p->cq_entries = 2 * p->sq_entries;
7769 user = get_uid(current_user());
7770 account_mem = !capable(CAP_IPC_LOCK);
7773 ret = io_account_mem(user,
7774 ring_pages(p->sq_entries, p->cq_entries));
7781 ctx = io_ring_ctx_alloc(p);
7784 io_unaccount_mem(user, ring_pages(p->sq_entries,
7789 ctx->compat = in_compat_syscall();
7790 ctx->account_mem = account_mem;
7792 ctx->creds = get_current_cred();
7794 ret = io_allocate_scq_urings(ctx, p);
7798 ret = io_sq_offload_start(ctx, p);
7802 memset(&p->sq_off, 0, sizeof(p->sq_off));
7803 p->sq_off.head = offsetof(struct io_rings, sq.head);
7804 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7805 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7806 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7807 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7808 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7809 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7811 memset(&p->cq_off, 0, sizeof(p->cq_off));
7812 p->cq_off.head = offsetof(struct io_rings, cq.head);
7813 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7814 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7815 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7816 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7817 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7820 * Install ring fd as the very last thing, so we don't risk someone
7821 * having closed it before we finish setup
7823 ret = io_uring_get_fd(ctx);
7827 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7828 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7829 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7830 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7833 io_ring_ctx_wait_and_kill(ctx);
7838 * Sets up an aio uring context, and returns the fd. Applications asks for a
7839 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7840 * params structure passed in.
7842 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7844 struct io_uring_params p;
7848 if (copy_from_user(&p, params, sizeof(p)))
7850 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7855 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7856 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7857 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7860 ret = io_uring_create(entries, &p);
7864 if (copy_to_user(params, &p, sizeof(p)))
7870 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7871 struct io_uring_params __user *, params)
7873 return io_uring_setup(entries, params);
7876 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7878 struct io_uring_probe *p;
7882 size = struct_size(p, ops, nr_args);
7883 if (size == SIZE_MAX)
7885 p = kzalloc(size, GFP_KERNEL);
7890 if (copy_from_user(p, arg, size))
7893 if (memchr_inv(p, 0, size))
7896 p->last_op = IORING_OP_LAST - 1;
7897 if (nr_args > IORING_OP_LAST)
7898 nr_args = IORING_OP_LAST;
7900 for (i = 0; i < nr_args; i++) {
7902 if (!io_op_defs[i].not_supported)
7903 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7908 if (copy_to_user(arg, p, size))
7915 static int io_register_personality(struct io_ring_ctx *ctx)
7917 const struct cred *creds = get_current_cred();
7920 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7921 USHRT_MAX, GFP_KERNEL);
7927 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7929 const struct cred *old_creds;
7931 old_creds = idr_remove(&ctx->personality_idr, id);
7933 put_cred(old_creds);
7940 static bool io_register_op_must_quiesce(int op)
7943 case IORING_UNREGISTER_FILES:
7944 case IORING_REGISTER_FILES_UPDATE:
7945 case IORING_REGISTER_PROBE:
7946 case IORING_REGISTER_PERSONALITY:
7947 case IORING_UNREGISTER_PERSONALITY:
7954 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7955 void __user *arg, unsigned nr_args)
7956 __releases(ctx->uring_lock)
7957 __acquires(ctx->uring_lock)
7962 * We're inside the ring mutex, if the ref is already dying, then
7963 * someone else killed the ctx or is already going through
7964 * io_uring_register().
7966 if (percpu_ref_is_dying(&ctx->refs))
7969 if (io_register_op_must_quiesce(opcode)) {
7970 percpu_ref_kill(&ctx->refs);
7973 * Drop uring mutex before waiting for references to exit. If
7974 * another thread is currently inside io_uring_enter() it might
7975 * need to grab the uring_lock to make progress. If we hold it
7976 * here across the drain wait, then we can deadlock. It's safe
7977 * to drop the mutex here, since no new references will come in
7978 * after we've killed the percpu ref.
7980 mutex_unlock(&ctx->uring_lock);
7981 ret = wait_for_completion_interruptible(&ctx->completions[0]);
7982 mutex_lock(&ctx->uring_lock);
7984 percpu_ref_resurrect(&ctx->refs);
7991 case IORING_REGISTER_BUFFERS:
7992 ret = io_sqe_buffer_register(ctx, arg, nr_args);
7994 case IORING_UNREGISTER_BUFFERS:
7998 ret = io_sqe_buffer_unregister(ctx);
8000 case IORING_REGISTER_FILES:
8001 ret = io_sqe_files_register(ctx, arg, nr_args);
8003 case IORING_UNREGISTER_FILES:
8007 ret = io_sqe_files_unregister(ctx);
8009 case IORING_REGISTER_FILES_UPDATE:
8010 ret = io_sqe_files_update(ctx, arg, nr_args);
8012 case IORING_REGISTER_EVENTFD:
8013 case IORING_REGISTER_EVENTFD_ASYNC:
8017 ret = io_eventfd_register(ctx, arg);
8020 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8021 ctx->eventfd_async = 1;
8023 ctx->eventfd_async = 0;
8025 case IORING_UNREGISTER_EVENTFD:
8029 ret = io_eventfd_unregister(ctx);
8031 case IORING_REGISTER_PROBE:
8033 if (!arg || nr_args > 256)
8035 ret = io_probe(ctx, arg, nr_args);
8037 case IORING_REGISTER_PERSONALITY:
8041 ret = io_register_personality(ctx);
8043 case IORING_UNREGISTER_PERSONALITY:
8047 ret = io_unregister_personality(ctx, nr_args);
8054 if (io_register_op_must_quiesce(opcode)) {
8055 /* bring the ctx back to life */
8056 percpu_ref_reinit(&ctx->refs);
8058 reinit_completion(&ctx->completions[0]);
8063 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8064 void __user *, arg, unsigned int, nr_args)
8066 struct io_ring_ctx *ctx;
8075 if (f.file->f_op != &io_uring_fops)
8078 ctx = f.file->private_data;
8080 mutex_lock(&ctx->uring_lock);
8081 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8082 mutex_unlock(&ctx->uring_lock);
8083 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8084 ctx->cq_ev_fd != NULL, ret);
8090 static int __init io_uring_init(void)
8092 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8093 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8094 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8097 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8098 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8099 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8100 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8101 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8102 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8103 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8104 BUILD_BUG_SQE_ELEM(8, __u64, off);
8105 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8106 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8107 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8108 BUILD_BUG_SQE_ELEM(24, __u32, len);
8109 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8110 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8111 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8112 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8113 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8114 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8115 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8116 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8117 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8118 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8119 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8120 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8121 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8122 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8123 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8124 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8125 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8126 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8128 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8129 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8130 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8133 __initcall(io_uring_init);