1 // SPDX-License-Identifier: GPL-2.0
3 * Shared application/kernel submission and completion ring pairs, for
4 * supporting fast/efficient IO.
6 * A note on the read/write ordering memory barriers that are matched between
7 * the application and kernel side.
9 * After the application reads the CQ ring tail, it must use an
10 * appropriate smp_rmb() to pair with the smp_wmb() the kernel uses
11 * before writing the tail (using smp_load_acquire to read the tail will
12 * do). It also needs a smp_mb() before updating CQ head (ordering the
13 * entry load(s) with the head store), pairing with an implicit barrier
14 * through a control-dependency in io_get_cqring (smp_store_release to
15 * store head will do). Failure to do so could lead to reading invalid
18 * Likewise, the application must use an appropriate smp_wmb() before
19 * writing the SQ tail (ordering SQ entry stores with the tail store),
20 * which pairs with smp_load_acquire in io_get_sqring (smp_store_release
21 * to store the tail will do). And it needs a barrier ordering the SQ
22 * head load before writing new SQ entries (smp_load_acquire to read
25 * When using the SQ poll thread (IORING_SETUP_SQPOLL), the application
26 * needs to check the SQ flags for IORING_SQ_NEED_WAKEUP *after*
27 * updating the SQ tail; a full memory barrier smp_mb() is needed
30 * Also see the examples in the liburing library:
32 * git://git.kernel.dk/liburing
34 * io_uring also uses READ/WRITE_ONCE() for _any_ store or load that happens
35 * from data shared between the kernel and application. This is done both
36 * for ordering purposes, but also to ensure that once a value is loaded from
37 * data that the application could potentially modify, it remains stable.
39 * Copyright (C) 2018-2019 Jens Axboe
40 * Copyright (c) 2018-2019 Christoph Hellwig
42 #include <linux/kernel.h>
43 #include <linux/init.h>
44 #include <linux/errno.h>
45 #include <linux/syscalls.h>
46 #include <linux/compat.h>
47 #include <net/compat.h>
48 #include <linux/refcount.h>
49 #include <linux/uio.h>
50 #include <linux/bits.h>
52 #include <linux/sched/signal.h>
54 #include <linux/file.h>
55 #include <linux/fdtable.h>
57 #include <linux/mman.h>
58 #include <linux/mmu_context.h>
59 #include <linux/percpu.h>
60 #include <linux/slab.h>
61 #include <linux/kthread.h>
62 #include <linux/blkdev.h>
63 #include <linux/bvec.h>
64 #include <linux/net.h>
66 #include <net/af_unix.h>
68 #include <linux/anon_inodes.h>
69 #include <linux/sched/mm.h>
70 #include <linux/uaccess.h>
71 #include <linux/nospec.h>
72 #include <linux/sizes.h>
73 #include <linux/hugetlb.h>
74 #include <linux/highmem.h>
75 #include <linux/namei.h>
76 #include <linux/fsnotify.h>
77 #include <linux/fadvise.h>
78 #include <linux/eventpoll.h>
79 #include <linux/fs_struct.h>
80 #include <linux/splice.h>
81 #include <linux/task_work.h>
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/io_uring.h>
86 #include <uapi/linux/io_uring.h>
91 #define IORING_MAX_ENTRIES 32768
92 #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES)
95 * Shift of 9 is 512 entries, or exactly one page on 64-bit archs
97 #define IORING_FILE_TABLE_SHIFT 9
98 #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT)
99 #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1)
100 #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE)
103 u32 head ____cacheline_aligned_in_smp;
104 u32 tail ____cacheline_aligned_in_smp;
108 * This data is shared with the application through the mmap at offsets
109 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
111 * The offsets to the member fields are published through struct
112 * io_sqring_offsets when calling io_uring_setup.
116 * Head and tail offsets into the ring; the offsets need to be
117 * masked to get valid indices.
119 * The kernel controls head of the sq ring and the tail of the cq ring,
120 * and the application controls tail of the sq ring and the head of the
123 struct io_uring sq, cq;
125 * Bitmasks to apply to head and tail offsets (constant, equals
128 u32 sq_ring_mask, cq_ring_mask;
129 /* Ring sizes (constant, power of 2) */
130 u32 sq_ring_entries, cq_ring_entries;
132 * Number of invalid entries dropped by the kernel due to
133 * invalid index stored in array
135 * Written by the kernel, shouldn't be modified by the
136 * application (i.e. get number of "new events" by comparing to
139 * After a new SQ head value was read by the application this
140 * counter includes all submissions that were dropped reaching
141 * the new SQ head (and possibly more).
147 * Written by the kernel, shouldn't be modified by the
150 * The application needs a full memory barrier before checking
151 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
155 * Number of completion events lost because the queue was full;
156 * this should be avoided by the application by making sure
157 * there are not more requests pending than there is space in
158 * the completion queue.
160 * Written by the kernel, shouldn't be modified by the
161 * application (i.e. get number of "new events" by comparing to
164 * As completion events come in out of order this counter is not
165 * ordered with any other data.
169 * Ring buffer of completion events.
171 * The kernel writes completion events fresh every time they are
172 * produced, so the application is allowed to modify pending
175 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
178 struct io_mapped_ubuf {
181 struct bio_vec *bvec;
182 unsigned int nr_bvecs;
185 struct fixed_file_table {
189 struct fixed_file_ref_node {
190 struct percpu_ref refs;
191 struct list_head node;
192 struct list_head file_list;
193 struct fixed_file_data *file_data;
194 struct work_struct work;
197 struct fixed_file_data {
198 struct fixed_file_table *table;
199 struct io_ring_ctx *ctx;
201 struct percpu_ref *cur_refs;
202 struct percpu_ref refs;
203 struct completion done;
204 struct list_head ref_list;
209 struct list_head list;
217 struct percpu_ref refs;
218 } ____cacheline_aligned_in_smp;
222 unsigned int compat: 1;
223 unsigned int account_mem: 1;
224 unsigned int cq_overflow_flushed: 1;
225 unsigned int drain_next: 1;
226 unsigned int eventfd_async: 1;
229 * Ring buffer of indices into array of io_uring_sqe, which is
230 * mmapped by the application using the IORING_OFF_SQES offset.
232 * This indirection could e.g. be used to assign fixed
233 * io_uring_sqe entries to operations and only submit them to
234 * the queue when needed.
236 * The kernel modifies neither the indices array nor the entries
240 unsigned cached_sq_head;
243 unsigned sq_thread_idle;
244 unsigned cached_sq_dropped;
245 atomic_t cached_cq_overflow;
246 unsigned long sq_check_overflow;
248 struct list_head defer_list;
249 struct list_head timeout_list;
250 struct list_head cq_overflow_list;
252 wait_queue_head_t inflight_wait;
253 struct io_uring_sqe *sq_sqes;
254 } ____cacheline_aligned_in_smp;
256 struct io_rings *rings;
260 struct task_struct *sqo_thread; /* if using sq thread polling */
261 struct mm_struct *sqo_mm;
262 wait_queue_head_t sqo_wait;
265 * If used, fixed file set. Writers must ensure that ->refs is dead,
266 * readers must ensure that ->refs is alive as long as the file* is
267 * used. Only updated through io_uring_register(2).
269 struct fixed_file_data *file_data;
270 unsigned nr_user_files;
272 struct file *ring_file;
274 /* if used, fixed mapped user buffers */
275 unsigned nr_user_bufs;
276 struct io_mapped_ubuf *user_bufs;
278 struct user_struct *user;
280 const struct cred *creds;
282 /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */
283 struct completion *completions;
285 /* if all else fails... */
286 struct io_kiocb *fallback_req;
288 #if defined(CONFIG_UNIX)
289 struct socket *ring_sock;
292 struct idr io_buffer_idr;
294 struct idr personality_idr;
297 unsigned cached_cq_tail;
300 atomic_t cq_timeouts;
301 unsigned long cq_check_overflow;
302 struct wait_queue_head cq_wait;
303 struct fasync_struct *cq_fasync;
304 struct eventfd_ctx *cq_ev_fd;
305 } ____cacheline_aligned_in_smp;
308 struct mutex uring_lock;
309 wait_queue_head_t wait;
310 } ____cacheline_aligned_in_smp;
313 spinlock_t completion_lock;
316 * ->poll_list is protected by the ctx->uring_lock for
317 * io_uring instances that don't use IORING_SETUP_SQPOLL.
318 * For SQPOLL, only the single threaded io_sq_thread() will
319 * manipulate the list, hence no extra locking is needed there.
321 struct list_head poll_list;
322 struct hlist_head *cancel_hash;
323 unsigned cancel_hash_bits;
324 bool poll_multi_file;
326 spinlock_t inflight_lock;
327 struct list_head inflight_list;
328 } ____cacheline_aligned_in_smp;
330 struct work_struct exit_work;
334 * First field must be the file pointer in all the
335 * iocb unions! See also 'struct kiocb' in <linux/fs.h>
337 struct io_poll_iocb {
340 struct wait_queue_head *head;
346 struct wait_queue_entry wait;
351 struct file *put_file;
355 struct io_timeout_data {
356 struct io_kiocb *req;
357 struct hrtimer timer;
358 struct timespec64 ts;
359 enum hrtimer_mode mode;
364 struct sockaddr __user *addr;
365 int __user *addr_len;
367 unsigned long nofile;
391 /* NOTE: kiocb has the file as the first member, so don't do it here */
399 struct sockaddr __user *addr;
406 struct user_msghdr __user *msg;
412 struct io_buffer *kbuf;
421 struct filename *filename;
422 struct statx __user *buffer;
424 unsigned long nofile;
427 struct io_files_update {
453 struct epoll_event event;
457 struct file *file_out;
458 struct file *file_in;
465 struct io_provide_buf {
474 struct io_async_connect {
475 struct sockaddr_storage address;
478 struct io_async_msghdr {
479 struct iovec fast_iov[UIO_FASTIOV];
481 struct sockaddr __user *uaddr;
483 struct sockaddr_storage addr;
487 struct iovec fast_iov[UIO_FASTIOV];
493 struct io_async_ctx {
495 struct io_async_rw rw;
496 struct io_async_msghdr msg;
497 struct io_async_connect connect;
498 struct io_timeout_data timeout;
503 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
504 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
505 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
506 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
507 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
508 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
516 REQ_F_IOPOLL_COMPLETED_BIT,
517 REQ_F_LINK_TIMEOUT_BIT,
521 REQ_F_TIMEOUT_NOSEQ_BIT,
522 REQ_F_COMP_LOCKED_BIT,
523 REQ_F_NEED_CLEANUP_BIT,
526 REQ_F_BUFFER_SELECTED_BIT,
528 /* not a real bit, just to check we're not overflowing the space */
534 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
535 /* drain existing IO first */
536 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
538 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
539 /* doesn't sever on completion < 0 */
540 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
542 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
543 /* IOSQE_BUFFER_SELECT */
544 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
547 REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT),
548 /* already grabbed next link */
549 REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT),
550 /* fail rest of links */
551 REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT),
552 /* on inflight list */
553 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
554 /* read/write uses file position */
555 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
556 /* must not punt to workers */
557 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
558 /* polled IO has completed */
559 REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT),
560 /* has linked timeout */
561 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
562 /* timeout request */
563 REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT),
565 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
566 /* must be punted even for NONBLOCK */
567 REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT),
568 /* no timeout sequence */
569 REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT),
570 /* completion under lock */
571 REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT),
573 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
574 /* in overflow list */
575 REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT),
576 /* already went through poll handler */
577 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
578 /* buffer already selected */
579 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
583 struct io_poll_iocb poll;
584 struct io_wq_work work;
588 * NOTE! Each of the iocb union members has the file pointer
589 * as the first entry in their struct definition. So you can
590 * access the file pointer through any of the sub-structs,
591 * or directly as just 'ki_filp' in this struct.
597 struct io_poll_iocb poll;
598 struct io_accept accept;
600 struct io_cancel cancel;
601 struct io_timeout timeout;
602 struct io_connect connect;
603 struct io_sr_msg sr_msg;
605 struct io_close close;
606 struct io_files_update files_update;
607 struct io_fadvise fadvise;
608 struct io_madvise madvise;
609 struct io_epoll epoll;
610 struct io_splice splice;
611 struct io_provide_buf pbuf;
614 struct io_async_ctx *io;
616 bool needs_fixed_file;
619 struct io_ring_ctx *ctx;
620 struct list_head list;
623 struct task_struct *task;
629 struct list_head link_list;
631 struct list_head inflight_entry;
633 struct percpu_ref *fixed_file_refs;
637 * Only commands that never go async can use the below fields,
638 * obviously. Right now only IORING_OP_POLL_ADD uses them, and
639 * async armed poll handlers for regular commands. The latter
640 * restore the work, if needed.
643 struct callback_head task_work;
644 struct hlist_node hash_node;
645 struct async_poll *apoll;
647 struct io_wq_work work;
651 #define IO_PLUG_THRESHOLD 2
652 #define IO_IOPOLL_BATCH 8
654 struct io_submit_state {
655 struct blk_plug plug;
658 * io_kiocb alloc cache
660 void *reqs[IO_IOPOLL_BATCH];
661 unsigned int free_reqs;
664 * File reference cache
668 unsigned int has_refs;
669 unsigned int used_refs;
670 unsigned int ios_left;
674 /* needs req->io allocated for deferral/async */
675 unsigned async_ctx : 1;
676 /* needs current->mm setup, does mm access */
677 unsigned needs_mm : 1;
678 /* needs req->file assigned */
679 unsigned needs_file : 1;
680 /* needs req->file assigned IFF fd is >= 0 */
681 unsigned fd_non_neg : 1;
682 /* hash wq insertion if file is a regular file */
683 unsigned hash_reg_file : 1;
684 /* unbound wq insertion if file is a non-regular file */
685 unsigned unbound_nonreg_file : 1;
686 /* opcode is not supported by this kernel */
687 unsigned not_supported : 1;
688 /* needs file table */
689 unsigned file_table : 1;
691 unsigned needs_fs : 1;
692 /* set if opcode supports polled "wait" */
694 unsigned pollout : 1;
695 /* op supports buffer selection */
696 unsigned buffer_select : 1;
699 static const struct io_op_def io_op_defs[] = {
700 [IORING_OP_NOP] = {},
701 [IORING_OP_READV] = {
705 .unbound_nonreg_file = 1,
709 [IORING_OP_WRITEV] = {
714 .unbound_nonreg_file = 1,
717 [IORING_OP_FSYNC] = {
720 [IORING_OP_READ_FIXED] = {
722 .unbound_nonreg_file = 1,
725 [IORING_OP_WRITE_FIXED] = {
728 .unbound_nonreg_file = 1,
731 [IORING_OP_POLL_ADD] = {
733 .unbound_nonreg_file = 1,
735 [IORING_OP_POLL_REMOVE] = {},
736 [IORING_OP_SYNC_FILE_RANGE] = {
739 [IORING_OP_SENDMSG] = {
743 .unbound_nonreg_file = 1,
747 [IORING_OP_RECVMSG] = {
751 .unbound_nonreg_file = 1,
756 [IORING_OP_TIMEOUT] = {
760 [IORING_OP_TIMEOUT_REMOVE] = {},
761 [IORING_OP_ACCEPT] = {
764 .unbound_nonreg_file = 1,
768 [IORING_OP_ASYNC_CANCEL] = {},
769 [IORING_OP_LINK_TIMEOUT] = {
773 [IORING_OP_CONNECT] = {
777 .unbound_nonreg_file = 1,
780 [IORING_OP_FALLOCATE] = {
783 [IORING_OP_OPENAT] = {
789 [IORING_OP_CLOSE] = {
793 [IORING_OP_FILES_UPDATE] = {
797 [IORING_OP_STATX] = {
806 .unbound_nonreg_file = 1,
810 [IORING_OP_WRITE] = {
813 .unbound_nonreg_file = 1,
816 [IORING_OP_FADVISE] = {
819 [IORING_OP_MADVISE] = {
825 .unbound_nonreg_file = 1,
831 .unbound_nonreg_file = 1,
835 [IORING_OP_OPENAT2] = {
841 [IORING_OP_EPOLL_CTL] = {
842 .unbound_nonreg_file = 1,
845 [IORING_OP_SPLICE] = {
848 .unbound_nonreg_file = 1,
850 [IORING_OP_PROVIDE_BUFFERS] = {},
851 [IORING_OP_REMOVE_BUFFERS] = {},
854 static void io_wq_submit_work(struct io_wq_work **workptr);
855 static void io_cqring_fill_event(struct io_kiocb *req, long res);
856 static void io_put_req(struct io_kiocb *req);
857 static void __io_double_put_req(struct io_kiocb *req);
858 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req);
859 static void io_queue_linked_timeout(struct io_kiocb *req);
860 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
861 struct io_uring_files_update *ip,
863 static int io_grab_files(struct io_kiocb *req);
864 static void io_cleanup_req(struct io_kiocb *req);
865 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
866 int fd, struct file **out_file, bool fixed);
867 static void __io_queue_sqe(struct io_kiocb *req,
868 const struct io_uring_sqe *sqe);
870 static struct kmem_cache *req_cachep;
872 static const struct file_operations io_uring_fops;
874 struct sock *io_uring_get_socket(struct file *file)
876 #if defined(CONFIG_UNIX)
877 if (file->f_op == &io_uring_fops) {
878 struct io_ring_ctx *ctx = file->private_data;
880 return ctx->ring_sock->sk;
885 EXPORT_SYMBOL(io_uring_get_socket);
887 static void io_ring_ctx_ref_free(struct percpu_ref *ref)
889 struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
891 complete(&ctx->completions[0]);
894 static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
896 struct io_ring_ctx *ctx;
899 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
903 ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL);
904 if (!ctx->fallback_req)
907 ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL);
908 if (!ctx->completions)
912 * Use 5 bits less than the max cq entries, that should give us around
913 * 32 entries per hash list if totally full and uniformly spread.
915 hash_bits = ilog2(p->cq_entries);
919 ctx->cancel_hash_bits = hash_bits;
920 ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head),
922 if (!ctx->cancel_hash)
924 __hash_init(ctx->cancel_hash, 1U << hash_bits);
926 if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free,
927 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL))
930 ctx->flags = p->flags;
931 init_waitqueue_head(&ctx->cq_wait);
932 INIT_LIST_HEAD(&ctx->cq_overflow_list);
933 init_completion(&ctx->completions[0]);
934 init_completion(&ctx->completions[1]);
935 idr_init(&ctx->io_buffer_idr);
936 idr_init(&ctx->personality_idr);
937 mutex_init(&ctx->uring_lock);
938 init_waitqueue_head(&ctx->wait);
939 spin_lock_init(&ctx->completion_lock);
940 INIT_LIST_HEAD(&ctx->poll_list);
941 INIT_LIST_HEAD(&ctx->defer_list);
942 INIT_LIST_HEAD(&ctx->timeout_list);
943 init_waitqueue_head(&ctx->inflight_wait);
944 spin_lock_init(&ctx->inflight_lock);
945 INIT_LIST_HEAD(&ctx->inflight_list);
948 if (ctx->fallback_req)
949 kmem_cache_free(req_cachep, ctx->fallback_req);
950 kfree(ctx->completions);
951 kfree(ctx->cancel_hash);
956 static inline bool __req_need_defer(struct io_kiocb *req)
958 struct io_ring_ctx *ctx = req->ctx;
960 return req->sequence != ctx->cached_cq_tail
961 + atomic_read(&ctx->cached_cq_overflow);
964 static inline bool req_need_defer(struct io_kiocb *req)
966 if (unlikely(req->flags & REQ_F_IO_DRAIN))
967 return __req_need_defer(req);
972 static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
974 struct io_kiocb *req;
976 req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list);
977 if (req && !req_need_defer(req)) {
978 list_del_init(&req->list);
985 static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
987 struct io_kiocb *req;
989 req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
991 if (req->flags & REQ_F_TIMEOUT_NOSEQ)
993 if (!__req_need_defer(req)) {
994 list_del_init(&req->list);
1002 static void __io_commit_cqring(struct io_ring_ctx *ctx)
1004 struct io_rings *rings = ctx->rings;
1006 /* order cqe stores with ring update */
1007 smp_store_release(&rings->cq.tail, ctx->cached_cq_tail);
1009 if (wq_has_sleeper(&ctx->cq_wait)) {
1010 wake_up_interruptible(&ctx->cq_wait);
1011 kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN);
1015 static inline void io_req_work_grab_env(struct io_kiocb *req,
1016 const struct io_op_def *def)
1018 if (!req->work.mm && def->needs_mm) {
1019 mmgrab(current->mm);
1020 req->work.mm = current->mm;
1022 if (!req->work.creds)
1023 req->work.creds = get_current_cred();
1024 if (!req->work.fs && def->needs_fs) {
1025 spin_lock(¤t->fs->lock);
1026 if (!current->fs->in_exec) {
1027 req->work.fs = current->fs;
1028 req->work.fs->users++;
1030 req->work.flags |= IO_WQ_WORK_CANCEL;
1032 spin_unlock(¤t->fs->lock);
1034 if (!req->work.task_pid)
1035 req->work.task_pid = task_pid_vnr(current);
1038 static inline void io_req_work_drop_env(struct io_kiocb *req)
1041 mmdrop(req->work.mm);
1042 req->work.mm = NULL;
1044 if (req->work.creds) {
1045 put_cred(req->work.creds);
1046 req->work.creds = NULL;
1049 struct fs_struct *fs = req->work.fs;
1051 spin_lock(&req->work.fs->lock);
1054 spin_unlock(&req->work.fs->lock);
1060 static inline void io_prep_async_work(struct io_kiocb *req,
1061 struct io_kiocb **link)
1063 const struct io_op_def *def = &io_op_defs[req->opcode];
1065 if (req->flags & REQ_F_ISREG) {
1066 if (def->hash_reg_file)
1067 io_wq_hash_work(&req->work, file_inode(req->file));
1069 if (def->unbound_nonreg_file)
1070 req->work.flags |= IO_WQ_WORK_UNBOUND;
1073 io_req_work_grab_env(req, def);
1075 *link = io_prep_linked_timeout(req);
1078 static inline void io_queue_async_work(struct io_kiocb *req)
1080 struct io_ring_ctx *ctx = req->ctx;
1081 struct io_kiocb *link;
1083 io_prep_async_work(req, &link);
1085 trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req,
1086 &req->work, req->flags);
1087 io_wq_enqueue(ctx->io_wq, &req->work);
1090 io_queue_linked_timeout(link);
1093 static void io_kill_timeout(struct io_kiocb *req)
1097 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1099 atomic_inc(&req->ctx->cq_timeouts);
1100 list_del_init(&req->list);
1101 req->flags |= REQ_F_COMP_LOCKED;
1102 io_cqring_fill_event(req, 0);
1107 static void io_kill_timeouts(struct io_ring_ctx *ctx)
1109 struct io_kiocb *req, *tmp;
1111 spin_lock_irq(&ctx->completion_lock);
1112 list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list)
1113 io_kill_timeout(req);
1114 spin_unlock_irq(&ctx->completion_lock);
1117 static void io_commit_cqring(struct io_ring_ctx *ctx)
1119 struct io_kiocb *req;
1121 while ((req = io_get_timeout_req(ctx)) != NULL)
1122 io_kill_timeout(req);
1124 __io_commit_cqring(ctx);
1126 while ((req = io_get_deferred_req(ctx)) != NULL)
1127 io_queue_async_work(req);
1130 static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx)
1132 struct io_rings *rings = ctx->rings;
1135 tail = ctx->cached_cq_tail;
1137 * writes to the cq entry need to come after reading head; the
1138 * control dependency is enough as we're using WRITE_ONCE to
1141 if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries)
1144 ctx->cached_cq_tail++;
1145 return &rings->cqes[tail & ctx->cq_mask];
1148 static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx)
1152 if (!ctx->eventfd_async)
1154 return io_wq_current_is_worker();
1157 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
1159 if (waitqueue_active(&ctx->wait))
1160 wake_up(&ctx->wait);
1161 if (waitqueue_active(&ctx->sqo_wait))
1162 wake_up(&ctx->sqo_wait);
1163 if (io_should_trigger_evfd(ctx))
1164 eventfd_signal(ctx->cq_ev_fd, 1);
1167 /* Returns true if there are no backlogged entries after the flush */
1168 static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force)
1170 struct io_rings *rings = ctx->rings;
1171 struct io_uring_cqe *cqe;
1172 struct io_kiocb *req;
1173 unsigned long flags;
1177 if (list_empty_careful(&ctx->cq_overflow_list))
1179 if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) ==
1180 rings->cq_ring_entries))
1184 spin_lock_irqsave(&ctx->completion_lock, flags);
1186 /* if force is set, the ring is going away. always drop after that */
1188 ctx->cq_overflow_flushed = 1;
1191 while (!list_empty(&ctx->cq_overflow_list)) {
1192 cqe = io_get_cqring(ctx);
1196 req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb,
1198 list_move(&req->list, &list);
1199 req->flags &= ~REQ_F_OVERFLOW;
1201 WRITE_ONCE(cqe->user_data, req->user_data);
1202 WRITE_ONCE(cqe->res, req->result);
1203 WRITE_ONCE(cqe->flags, req->cflags);
1205 WRITE_ONCE(ctx->rings->cq_overflow,
1206 atomic_inc_return(&ctx->cached_cq_overflow));
1210 io_commit_cqring(ctx);
1212 clear_bit(0, &ctx->sq_check_overflow);
1213 clear_bit(0, &ctx->cq_check_overflow);
1215 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1216 io_cqring_ev_posted(ctx);
1218 while (!list_empty(&list)) {
1219 req = list_first_entry(&list, struct io_kiocb, list);
1220 list_del(&req->list);
1227 static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
1229 struct io_ring_ctx *ctx = req->ctx;
1230 struct io_uring_cqe *cqe;
1232 trace_io_uring_complete(ctx, req->user_data, res);
1235 * If we can't get a cq entry, userspace overflowed the
1236 * submission (by quite a lot). Increment the overflow count in
1239 cqe = io_get_cqring(ctx);
1241 WRITE_ONCE(cqe->user_data, req->user_data);
1242 WRITE_ONCE(cqe->res, res);
1243 WRITE_ONCE(cqe->flags, cflags);
1244 } else if (ctx->cq_overflow_flushed) {
1245 WRITE_ONCE(ctx->rings->cq_overflow,
1246 atomic_inc_return(&ctx->cached_cq_overflow));
1248 if (list_empty(&ctx->cq_overflow_list)) {
1249 set_bit(0, &ctx->sq_check_overflow);
1250 set_bit(0, &ctx->cq_check_overflow);
1252 req->flags |= REQ_F_OVERFLOW;
1253 refcount_inc(&req->refs);
1255 req->cflags = cflags;
1256 list_add_tail(&req->list, &ctx->cq_overflow_list);
1260 static void io_cqring_fill_event(struct io_kiocb *req, long res)
1262 __io_cqring_fill_event(req, res, 0);
1265 static void __io_cqring_add_event(struct io_kiocb *req, long res, long cflags)
1267 struct io_ring_ctx *ctx = req->ctx;
1268 unsigned long flags;
1270 spin_lock_irqsave(&ctx->completion_lock, flags);
1271 __io_cqring_fill_event(req, res, cflags);
1272 io_commit_cqring(ctx);
1273 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1275 io_cqring_ev_posted(ctx);
1278 static void io_cqring_add_event(struct io_kiocb *req, long res)
1280 __io_cqring_add_event(req, res, 0);
1283 static inline bool io_is_fallback_req(struct io_kiocb *req)
1285 return req == (struct io_kiocb *)
1286 ((unsigned long) req->ctx->fallback_req & ~1UL);
1289 static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx)
1291 struct io_kiocb *req;
1293 req = ctx->fallback_req;
1294 if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req))
1300 static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx,
1301 struct io_submit_state *state)
1303 gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
1304 struct io_kiocb *req;
1307 req = kmem_cache_alloc(req_cachep, gfp);
1310 } else if (!state->free_reqs) {
1314 sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs));
1315 ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs);
1318 * Bulk alloc is all-or-nothing. If we fail to get a batch,
1319 * retry single alloc to be on the safe side.
1321 if (unlikely(ret <= 0)) {
1322 state->reqs[0] = kmem_cache_alloc(req_cachep, gfp);
1323 if (!state->reqs[0])
1327 state->free_reqs = ret - 1;
1328 req = state->reqs[ret - 1];
1331 req = state->reqs[state->free_reqs];
1336 return io_get_fallback_req(ctx);
1339 static inline void io_put_file(struct io_kiocb *req, struct file *file,
1343 percpu_ref_put(req->fixed_file_refs);
1348 static void __io_req_aux_free(struct io_kiocb *req)
1350 if (req->flags & REQ_F_NEED_CLEANUP)
1351 io_cleanup_req(req);
1355 io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE));
1357 put_task_struct(req->task);
1359 io_req_work_drop_env(req);
1362 static void __io_free_req(struct io_kiocb *req)
1364 __io_req_aux_free(req);
1366 if (req->flags & REQ_F_INFLIGHT) {
1367 struct io_ring_ctx *ctx = req->ctx;
1368 unsigned long flags;
1370 spin_lock_irqsave(&ctx->inflight_lock, flags);
1371 list_del(&req->inflight_entry);
1372 if (waitqueue_active(&ctx->inflight_wait))
1373 wake_up(&ctx->inflight_wait);
1374 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1377 percpu_ref_put(&req->ctx->refs);
1378 if (likely(!io_is_fallback_req(req)))
1379 kmem_cache_free(req_cachep, req);
1381 clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req);
1385 void *reqs[IO_IOPOLL_BATCH];
1390 static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb)
1394 if (rb->need_iter) {
1395 int i, inflight = 0;
1396 unsigned long flags;
1398 for (i = 0; i < rb->to_free; i++) {
1399 struct io_kiocb *req = rb->reqs[i];
1401 if (req->flags & REQ_F_FIXED_FILE) {
1403 percpu_ref_put(req->fixed_file_refs);
1405 if (req->flags & REQ_F_INFLIGHT)
1407 __io_req_aux_free(req);
1412 spin_lock_irqsave(&ctx->inflight_lock, flags);
1413 for (i = 0; i < rb->to_free; i++) {
1414 struct io_kiocb *req = rb->reqs[i];
1416 if (req->flags & REQ_F_INFLIGHT) {
1417 list_del(&req->inflight_entry);
1422 spin_unlock_irqrestore(&ctx->inflight_lock, flags);
1424 if (waitqueue_active(&ctx->inflight_wait))
1425 wake_up(&ctx->inflight_wait);
1428 kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs);
1429 percpu_ref_put_many(&ctx->refs, rb->to_free);
1430 rb->to_free = rb->need_iter = 0;
1433 static bool io_link_cancel_timeout(struct io_kiocb *req)
1435 struct io_ring_ctx *ctx = req->ctx;
1438 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
1440 io_cqring_fill_event(req, -ECANCELED);
1441 io_commit_cqring(ctx);
1442 req->flags &= ~REQ_F_LINK_HEAD;
1450 static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1452 struct io_ring_ctx *ctx = req->ctx;
1453 bool wake_ev = false;
1455 /* Already got next link */
1456 if (req->flags & REQ_F_LINK_NEXT)
1460 * The list should never be empty when we are called here. But could
1461 * potentially happen if the chain is messed up, check to be on the
1464 while (!list_empty(&req->link_list)) {
1465 struct io_kiocb *nxt = list_first_entry(&req->link_list,
1466 struct io_kiocb, link_list);
1468 if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) &&
1469 (nxt->flags & REQ_F_TIMEOUT))) {
1470 list_del_init(&nxt->link_list);
1471 wake_ev |= io_link_cancel_timeout(nxt);
1472 req->flags &= ~REQ_F_LINK_TIMEOUT;
1476 list_del_init(&req->link_list);
1477 if (!list_empty(&nxt->link_list))
1478 nxt->flags |= REQ_F_LINK_HEAD;
1483 req->flags |= REQ_F_LINK_NEXT;
1485 io_cqring_ev_posted(ctx);
1489 * Called if REQ_F_LINK_HEAD is set, and we fail the head request
1491 static void io_fail_links(struct io_kiocb *req)
1493 struct io_ring_ctx *ctx = req->ctx;
1494 unsigned long flags;
1496 spin_lock_irqsave(&ctx->completion_lock, flags);
1498 while (!list_empty(&req->link_list)) {
1499 struct io_kiocb *link = list_first_entry(&req->link_list,
1500 struct io_kiocb, link_list);
1502 list_del_init(&link->link_list);
1503 trace_io_uring_fail_link(req, link);
1505 if ((req->flags & REQ_F_LINK_TIMEOUT) &&
1506 link->opcode == IORING_OP_LINK_TIMEOUT) {
1507 io_link_cancel_timeout(link);
1509 io_cqring_fill_event(link, -ECANCELED);
1510 __io_double_put_req(link);
1512 req->flags &= ~REQ_F_LINK_TIMEOUT;
1515 io_commit_cqring(ctx);
1516 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1517 io_cqring_ev_posted(ctx);
1520 static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt)
1522 if (likely(!(req->flags & REQ_F_LINK_HEAD)))
1526 * If LINK is set, we have dependent requests in this chain. If we
1527 * didn't fail this request, queue the first one up, moving any other
1528 * dependencies to the next request. In case of failure, fail the rest
1531 if (req->flags & REQ_F_FAIL_LINK) {
1533 } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) ==
1534 REQ_F_LINK_TIMEOUT) {
1535 struct io_ring_ctx *ctx = req->ctx;
1536 unsigned long flags;
1539 * If this is a timeout link, we could be racing with the
1540 * timeout timer. Grab the completion lock for this case to
1541 * protect against that.
1543 spin_lock_irqsave(&ctx->completion_lock, flags);
1544 io_req_link_next(req, nxt);
1545 spin_unlock_irqrestore(&ctx->completion_lock, flags);
1547 io_req_link_next(req, nxt);
1551 static void io_free_req(struct io_kiocb *req)
1553 struct io_kiocb *nxt = NULL;
1555 io_req_find_next(req, &nxt);
1559 io_queue_async_work(nxt);
1562 static void io_link_work_cb(struct io_wq_work **workptr)
1564 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
1565 struct io_kiocb *link;
1567 link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
1568 io_queue_linked_timeout(link);
1569 io_wq_submit_work(workptr);
1572 static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt)
1574 struct io_kiocb *link;
1575 const struct io_op_def *def = &io_op_defs[nxt->opcode];
1577 if ((nxt->flags & REQ_F_ISREG) && def->hash_reg_file)
1578 io_wq_hash_work(&nxt->work, file_inode(nxt->file));
1580 *workptr = &nxt->work;
1581 link = io_prep_linked_timeout(nxt);
1583 nxt->work.func = io_link_work_cb;
1587 * Drop reference to request, return next in chain (if there is one) if this
1588 * was the last reference to this request.
1590 __attribute__((nonnull))
1591 static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr)
1593 if (refcount_dec_and_test(&req->refs)) {
1594 io_req_find_next(req, nxtptr);
1599 static void io_put_req(struct io_kiocb *req)
1601 if (refcount_dec_and_test(&req->refs))
1605 static void io_steal_work(struct io_kiocb *req,
1606 struct io_wq_work **workptr)
1609 * It's in an io-wq worker, so there always should be at least
1610 * one reference, which will be dropped in io_put_work() just
1611 * after the current handler returns.
1613 * It also means, that if the counter dropped to 1, then there is
1614 * no asynchronous users left, so it's safe to steal the next work.
1616 if (refcount_read(&req->refs) == 1) {
1617 struct io_kiocb *nxt = NULL;
1619 io_req_find_next(req, &nxt);
1621 io_wq_assign_next(workptr, nxt);
1626 * Must only be used if we don't need to care about links, usually from
1627 * within the completion handling itself.
1629 static void __io_double_put_req(struct io_kiocb *req)
1631 /* drop both submit and complete references */
1632 if (refcount_sub_and_test(2, &req->refs))
1636 static void io_double_put_req(struct io_kiocb *req)
1638 /* drop both submit and complete references */
1639 if (refcount_sub_and_test(2, &req->refs))
1643 static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush)
1645 struct io_rings *rings = ctx->rings;
1647 if (test_bit(0, &ctx->cq_check_overflow)) {
1649 * noflush == true is from the waitqueue handler, just ensure
1650 * we wake up the task, and the next invocation will flush the
1651 * entries. We cannot safely to it from here.
1653 if (noflush && !list_empty(&ctx->cq_overflow_list))
1656 io_cqring_overflow_flush(ctx, false);
1659 /* See comment at the top of this file */
1661 return ctx->cached_cq_tail - READ_ONCE(rings->cq.head);
1664 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
1666 struct io_rings *rings = ctx->rings;
1668 /* make sure SQ entry isn't read before tail */
1669 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
1672 static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req)
1674 if ((req->flags & REQ_F_LINK_HEAD) || io_is_fallback_req(req))
1677 if (!(req->flags & REQ_F_FIXED_FILE) || req->io)
1680 rb->reqs[rb->to_free++] = req;
1681 if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs)))
1682 io_free_req_many(req->ctx, rb);
1686 static int io_put_kbuf(struct io_kiocb *req)
1688 struct io_buffer *kbuf;
1691 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
1692 cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
1693 cflags |= IORING_CQE_F_BUFFER;
1700 * Find and free completed poll iocbs
1702 static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events,
1703 struct list_head *done)
1705 struct req_batch rb;
1706 struct io_kiocb *req;
1708 rb.to_free = rb.need_iter = 0;
1709 while (!list_empty(done)) {
1712 req = list_first_entry(done, struct io_kiocb, list);
1713 list_del(&req->list);
1715 if (req->flags & REQ_F_BUFFER_SELECTED)
1716 cflags = io_put_kbuf(req);
1718 __io_cqring_fill_event(req, req->result, cflags);
1721 if (refcount_dec_and_test(&req->refs) &&
1722 !io_req_multi_free(&rb, req))
1726 io_commit_cqring(ctx);
1727 if (ctx->flags & IORING_SETUP_SQPOLL)
1728 io_cqring_ev_posted(ctx);
1729 io_free_req_many(ctx, &rb);
1732 static void io_iopoll_queue(struct list_head *again)
1734 struct io_kiocb *req;
1737 req = list_first_entry(again, struct io_kiocb, list);
1738 list_del(&req->list);
1739 refcount_inc(&req->refs);
1740 io_queue_async_work(req);
1741 } while (!list_empty(again));
1744 static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events,
1747 struct io_kiocb *req, *tmp;
1754 * Only spin for completions if we don't have multiple devices hanging
1755 * off our complete list, and we're under the requested amount.
1757 spin = !ctx->poll_multi_file && *nr_events < min;
1760 list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) {
1761 struct kiocb *kiocb = &req->rw.kiocb;
1764 * Move completed and retryable entries to our local lists.
1765 * If we find a request that requires polling, break out
1766 * and complete those lists first, if we have entries there.
1768 if (req->flags & REQ_F_IOPOLL_COMPLETED) {
1769 list_move_tail(&req->list, &done);
1772 if (!list_empty(&done))
1775 if (req->result == -EAGAIN) {
1776 list_move_tail(&req->list, &again);
1779 if (!list_empty(&again))
1782 ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
1791 if (!list_empty(&done))
1792 io_iopoll_complete(ctx, nr_events, &done);
1794 if (!list_empty(&again))
1795 io_iopoll_queue(&again);
1801 * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a
1802 * non-spinning poll check - we'll still enter the driver poll loop, but only
1803 * as a non-spinning completion check.
1805 static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
1808 while (!list_empty(&ctx->poll_list) && !need_resched()) {
1811 ret = io_do_iopoll(ctx, nr_events, min);
1814 if (!min || *nr_events >= min)
1822 * We can't just wait for polled events to come to us, we have to actively
1823 * find and complete them.
1825 static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
1827 if (!(ctx->flags & IORING_SETUP_IOPOLL))
1830 mutex_lock(&ctx->uring_lock);
1831 while (!list_empty(&ctx->poll_list)) {
1832 unsigned int nr_events = 0;
1834 io_iopoll_getevents(ctx, &nr_events, 1);
1837 * Ensure we allow local-to-the-cpu processing to take place,
1838 * in this case we need to ensure that we reap all events.
1842 mutex_unlock(&ctx->uring_lock);
1845 static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
1848 int iters = 0, ret = 0;
1851 * We disallow the app entering submit/complete with polling, but we
1852 * still need to lock the ring to prevent racing with polled issue
1853 * that got punted to a workqueue.
1855 mutex_lock(&ctx->uring_lock);
1860 * Don't enter poll loop if we already have events pending.
1861 * If we do, we can potentially be spinning for commands that
1862 * already triggered a CQE (eg in error).
1864 if (io_cqring_events(ctx, false))
1868 * If a submit got punted to a workqueue, we can have the
1869 * application entering polling for a command before it gets
1870 * issued. That app will hold the uring_lock for the duration
1871 * of the poll right here, so we need to take a breather every
1872 * now and then to ensure that the issue has a chance to add
1873 * the poll to the issued list. Otherwise we can spin here
1874 * forever, while the workqueue is stuck trying to acquire the
1877 if (!(++iters & 7)) {
1878 mutex_unlock(&ctx->uring_lock);
1879 mutex_lock(&ctx->uring_lock);
1882 if (*nr_events < min)
1883 tmin = min - *nr_events;
1885 ret = io_iopoll_getevents(ctx, nr_events, tmin);
1889 } while (min && !*nr_events && !need_resched());
1891 mutex_unlock(&ctx->uring_lock);
1895 static void kiocb_end_write(struct io_kiocb *req)
1898 * Tell lockdep we inherited freeze protection from submission
1901 if (req->flags & REQ_F_ISREG) {
1902 struct inode *inode = file_inode(req->file);
1904 __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1906 file_end_write(req->file);
1909 static inline void req_set_fail_links(struct io_kiocb *req)
1911 if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK)
1912 req->flags |= REQ_F_FAIL_LINK;
1915 static void io_complete_rw_common(struct kiocb *kiocb, long res)
1917 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1920 if (kiocb->ki_flags & IOCB_WRITE)
1921 kiocb_end_write(req);
1923 if (res != req->result)
1924 req_set_fail_links(req);
1925 if (req->flags & REQ_F_BUFFER_SELECTED)
1926 cflags = io_put_kbuf(req);
1927 __io_cqring_add_event(req, res, cflags);
1930 static void io_complete_rw(struct kiocb *kiocb, long res, long res2)
1932 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1934 io_complete_rw_common(kiocb, res);
1938 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2)
1940 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
1942 if (kiocb->ki_flags & IOCB_WRITE)
1943 kiocb_end_write(req);
1945 if (res != req->result)
1946 req_set_fail_links(req);
1949 req->flags |= REQ_F_IOPOLL_COMPLETED;
1953 * After the iocb has been issued, it's safe to be found on the poll list.
1954 * Adding the kiocb to the list AFTER submission ensures that we don't
1955 * find it from a io_iopoll_getevents() thread before the issuer is done
1956 * accessing the kiocb cookie.
1958 static void io_iopoll_req_issued(struct io_kiocb *req)
1960 struct io_ring_ctx *ctx = req->ctx;
1963 * Track whether we have multiple files in our lists. This will impact
1964 * how we do polling eventually, not spinning if we're on potentially
1965 * different devices.
1967 if (list_empty(&ctx->poll_list)) {
1968 ctx->poll_multi_file = false;
1969 } else if (!ctx->poll_multi_file) {
1970 struct io_kiocb *list_req;
1972 list_req = list_first_entry(&ctx->poll_list, struct io_kiocb,
1974 if (list_req->file != req->file)
1975 ctx->poll_multi_file = true;
1979 * For fast devices, IO may have already completed. If it has, add
1980 * it to the front so we find it first.
1982 if (req->flags & REQ_F_IOPOLL_COMPLETED)
1983 list_add(&req->list, &ctx->poll_list);
1985 list_add_tail(&req->list, &ctx->poll_list);
1987 if ((ctx->flags & IORING_SETUP_SQPOLL) &&
1988 wq_has_sleeper(&ctx->sqo_wait))
1989 wake_up(&ctx->sqo_wait);
1992 static void io_file_put(struct io_submit_state *state)
1995 int diff = state->has_refs - state->used_refs;
1998 fput_many(state->file, diff);
2004 * Get as many references to a file as we have IOs left in this submission,
2005 * assuming most submissions are for one file, or at least that each file
2006 * has more than one submission.
2008 static struct file *__io_file_get(struct io_submit_state *state, int fd)
2014 if (state->fd == fd) {
2021 state->file = fget_many(fd, state->ios_left);
2026 state->has_refs = state->ios_left;
2027 state->used_refs = 1;
2033 * If we tracked the file through the SCM inflight mechanism, we could support
2034 * any file. For now, just ensure that anything potentially problematic is done
2037 static bool io_file_supports_async(struct file *file)
2039 umode_t mode = file_inode(file)->i_mode;
2041 if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode))
2043 if (S_ISREG(mode) && file->f_op != &io_uring_fops)
2049 static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2050 bool force_nonblock)
2052 struct io_ring_ctx *ctx = req->ctx;
2053 struct kiocb *kiocb = &req->rw.kiocb;
2057 if (S_ISREG(file_inode(req->file)->i_mode))
2058 req->flags |= REQ_F_ISREG;
2060 kiocb->ki_pos = READ_ONCE(sqe->off);
2061 if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) {
2062 req->flags |= REQ_F_CUR_POS;
2063 kiocb->ki_pos = req->file->f_pos;
2065 kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp));
2066 kiocb->ki_flags = iocb_flags(kiocb->ki_filp);
2067 ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags));
2071 ioprio = READ_ONCE(sqe->ioprio);
2073 ret = ioprio_check_cap(ioprio);
2077 kiocb->ki_ioprio = ioprio;
2079 kiocb->ki_ioprio = get_current_ioprio();
2081 /* don't allow async punt if RWF_NOWAIT was requested */
2082 if ((kiocb->ki_flags & IOCB_NOWAIT) ||
2083 (req->file->f_flags & O_NONBLOCK))
2084 req->flags |= REQ_F_NOWAIT;
2087 kiocb->ki_flags |= IOCB_NOWAIT;
2089 if (ctx->flags & IORING_SETUP_IOPOLL) {
2090 if (!(kiocb->ki_flags & IOCB_DIRECT) ||
2091 !kiocb->ki_filp->f_op->iopoll)
2094 kiocb->ki_flags |= IOCB_HIPRI;
2095 kiocb->ki_complete = io_complete_rw_iopoll;
2098 if (kiocb->ki_flags & IOCB_HIPRI)
2100 kiocb->ki_complete = io_complete_rw;
2103 req->rw.addr = READ_ONCE(sqe->addr);
2104 req->rw.len = READ_ONCE(sqe->len);
2105 /* we own ->private, reuse it for the buffer index / buffer ID */
2106 req->rw.kiocb.private = (void *) (unsigned long)
2107 READ_ONCE(sqe->buf_index);
2111 static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
2117 case -ERESTARTNOINTR:
2118 case -ERESTARTNOHAND:
2119 case -ERESTART_RESTARTBLOCK:
2121 * We can't just restart the syscall, since previously
2122 * submitted sqes may already be in progress. Just fail this
2128 kiocb->ki_complete(kiocb, ret, 0);
2132 static void kiocb_done(struct kiocb *kiocb, ssize_t ret)
2134 struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb);
2136 if (req->flags & REQ_F_CUR_POS)
2137 req->file->f_pos = kiocb->ki_pos;
2138 if (ret >= 0 && kiocb->ki_complete == io_complete_rw)
2139 io_complete_rw(kiocb, ret, 0);
2141 io_rw_done(kiocb, ret);
2144 static ssize_t io_import_fixed(struct io_kiocb *req, int rw,
2145 struct iov_iter *iter)
2147 struct io_ring_ctx *ctx = req->ctx;
2148 size_t len = req->rw.len;
2149 struct io_mapped_ubuf *imu;
2150 unsigned index, buf_index;
2154 /* attempt to use fixed buffers without having provided iovecs */
2155 if (unlikely(!ctx->user_bufs))
2158 buf_index = (unsigned long) req->rw.kiocb.private;
2159 if (unlikely(buf_index >= ctx->nr_user_bufs))
2162 index = array_index_nospec(buf_index, ctx->nr_user_bufs);
2163 imu = &ctx->user_bufs[index];
2164 buf_addr = req->rw.addr;
2167 if (buf_addr + len < buf_addr)
2169 /* not inside the mapped region */
2170 if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len)
2174 * May not be a start of buffer, set size appropriately
2175 * and advance us to the beginning.
2177 offset = buf_addr - imu->ubuf;
2178 iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
2182 * Don't use iov_iter_advance() here, as it's really slow for
2183 * using the latter parts of a big fixed buffer - it iterates
2184 * over each segment manually. We can cheat a bit here, because
2187 * 1) it's a BVEC iter, we set it up
2188 * 2) all bvecs are PAGE_SIZE in size, except potentially the
2189 * first and last bvec
2191 * So just find our index, and adjust the iterator afterwards.
2192 * If the offset is within the first bvec (or the whole first
2193 * bvec, just use iov_iter_advance(). This makes it easier
2194 * since we can just skip the first segment, which may not
2195 * be PAGE_SIZE aligned.
2197 const struct bio_vec *bvec = imu->bvec;
2199 if (offset <= bvec->bv_len) {
2200 iov_iter_advance(iter, offset);
2202 unsigned long seg_skip;
2204 /* skip first vec */
2205 offset -= bvec->bv_len;
2206 seg_skip = 1 + (offset >> PAGE_SHIFT);
2208 iter->bvec = bvec + seg_skip;
2209 iter->nr_segs -= seg_skip;
2210 iter->count -= bvec->bv_len + offset;
2211 iter->iov_offset = offset & ~PAGE_MASK;
2218 static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock)
2221 mutex_unlock(&ctx->uring_lock);
2224 static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock)
2227 * "Normal" inline submissions always hold the uring_lock, since we
2228 * grab it from the system call. Same is true for the SQPOLL offload.
2229 * The only exception is when we've detached the request and issue it
2230 * from an async worker thread, grab the lock for that case.
2233 mutex_lock(&ctx->uring_lock);
2236 static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len,
2237 int bgid, struct io_buffer *kbuf,
2240 struct io_buffer *head;
2242 if (req->flags & REQ_F_BUFFER_SELECTED)
2245 io_ring_submit_lock(req->ctx, needs_lock);
2247 lockdep_assert_held(&req->ctx->uring_lock);
2249 head = idr_find(&req->ctx->io_buffer_idr, bgid);
2251 if (!list_empty(&head->list)) {
2252 kbuf = list_last_entry(&head->list, struct io_buffer,
2254 list_del(&kbuf->list);
2257 idr_remove(&req->ctx->io_buffer_idr, bgid);
2259 if (*len > kbuf->len)
2262 kbuf = ERR_PTR(-ENOBUFS);
2265 io_ring_submit_unlock(req->ctx, needs_lock);
2270 static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len,
2273 struct io_buffer *kbuf;
2276 kbuf = (struct io_buffer *) (unsigned long) req->rw.addr;
2277 bgid = (int) (unsigned long) req->rw.kiocb.private;
2278 kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock);
2281 req->rw.addr = (u64) (unsigned long) kbuf;
2282 req->flags |= REQ_F_BUFFER_SELECTED;
2283 return u64_to_user_ptr(kbuf->addr);
2286 #ifdef CONFIG_COMPAT
2287 static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov,
2290 struct compat_iovec __user *uiov;
2291 compat_ssize_t clen;
2295 uiov = u64_to_user_ptr(req->rw.addr);
2296 if (!access_ok(uiov, sizeof(*uiov)))
2298 if (__get_user(clen, &uiov->iov_len))
2304 buf = io_rw_buffer_select(req, &len, needs_lock);
2306 return PTR_ERR(buf);
2307 iov[0].iov_base = buf;
2308 iov[0].iov_len = (compat_size_t) len;
2313 static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2316 struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr);
2320 if (copy_from_user(iov, uiov, sizeof(*uiov)))
2323 len = iov[0].iov_len;
2326 buf = io_rw_buffer_select(req, &len, needs_lock);
2328 return PTR_ERR(buf);
2329 iov[0].iov_base = buf;
2330 iov[0].iov_len = len;
2334 static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov,
2337 if (req->flags & REQ_F_BUFFER_SELECTED)
2341 else if (req->rw.len > 1)
2344 #ifdef CONFIG_COMPAT
2345 if (req->ctx->compat)
2346 return io_compat_import(req, iov, needs_lock);
2349 return __io_iov_buffer_select(req, iov, needs_lock);
2352 static ssize_t io_import_iovec(int rw, struct io_kiocb *req,
2353 struct iovec **iovec, struct iov_iter *iter,
2356 void __user *buf = u64_to_user_ptr(req->rw.addr);
2357 size_t sqe_len = req->rw.len;
2361 opcode = req->opcode;
2362 if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) {
2364 return io_import_fixed(req, rw, iter);
2367 /* buffer index only valid with fixed read/write, or buffer select */
2368 if (req->rw.kiocb.private && !(req->flags & REQ_F_BUFFER_SELECT))
2371 if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) {
2372 if (req->flags & REQ_F_BUFFER_SELECT) {
2373 buf = io_rw_buffer_select(req, &sqe_len, needs_lock);
2376 return PTR_ERR(buf);
2378 req->rw.len = sqe_len;
2381 ret = import_single_range(rw, buf, sqe_len, *iovec, iter);
2383 return ret < 0 ? ret : sqe_len;
2387 struct io_async_rw *iorw = &req->io->rw;
2390 iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size);
2391 if (iorw->iov == iorw->fast_iov)
2396 if (req->flags & REQ_F_BUFFER_SELECT) {
2397 ret = io_iov_buffer_select(req, *iovec, needs_lock);
2399 ret = (*iovec)->iov_len;
2400 iov_iter_init(iter, rw, *iovec, 1, ret);
2406 #ifdef CONFIG_COMPAT
2407 if (req->ctx->compat)
2408 return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV,
2412 return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter);
2416 * For files that don't have ->read_iter() and ->write_iter(), handle them
2417 * by looping over ->read() or ->write() manually.
2419 static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb,
2420 struct iov_iter *iter)
2425 * Don't support polled IO through this interface, and we can't
2426 * support non-blocking either. For the latter, this just causes
2427 * the kiocb to be handled from an async context.
2429 if (kiocb->ki_flags & IOCB_HIPRI)
2431 if (kiocb->ki_flags & IOCB_NOWAIT)
2434 while (iov_iter_count(iter)) {
2438 if (!iov_iter_is_bvec(iter)) {
2439 iovec = iov_iter_iovec(iter);
2441 /* fixed buffers import bvec */
2442 iovec.iov_base = kmap(iter->bvec->bv_page)
2444 iovec.iov_len = min(iter->count,
2445 iter->bvec->bv_len - iter->iov_offset);
2449 nr = file->f_op->read(file, iovec.iov_base,
2450 iovec.iov_len, &kiocb->ki_pos);
2452 nr = file->f_op->write(file, iovec.iov_base,
2453 iovec.iov_len, &kiocb->ki_pos);
2456 if (iov_iter_is_bvec(iter))
2457 kunmap(iter->bvec->bv_page);
2465 if (nr != iovec.iov_len)
2467 iov_iter_advance(iter, nr);
2473 static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size,
2474 struct iovec *iovec, struct iovec *fast_iov,
2475 struct iov_iter *iter)
2477 req->io->rw.nr_segs = iter->nr_segs;
2478 req->io->rw.size = io_size;
2479 req->io->rw.iov = iovec;
2480 if (!req->io->rw.iov) {
2481 req->io->rw.iov = req->io->rw.fast_iov;
2482 if (req->io->rw.iov != fast_iov)
2483 memcpy(req->io->rw.iov, fast_iov,
2484 sizeof(struct iovec) * iter->nr_segs);
2486 req->flags |= REQ_F_NEED_CLEANUP;
2490 static inline int __io_alloc_async_ctx(struct io_kiocb *req)
2492 req->io = kmalloc(sizeof(*req->io), GFP_KERNEL);
2493 return req->io == NULL;
2496 static int io_alloc_async_ctx(struct io_kiocb *req)
2498 if (!io_op_defs[req->opcode].async_ctx)
2501 return __io_alloc_async_ctx(req);
2504 static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size,
2505 struct iovec *iovec, struct iovec *fast_iov,
2506 struct iov_iter *iter)
2508 if (!io_op_defs[req->opcode].async_ctx)
2511 if (__io_alloc_async_ctx(req))
2514 io_req_map_rw(req, io_size, iovec, fast_iov, iter);
2519 static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2520 bool force_nonblock)
2522 struct io_async_ctx *io;
2523 struct iov_iter iter;
2526 ret = io_prep_rw(req, sqe, force_nonblock);
2530 if (unlikely(!(req->file->f_mode & FMODE_READ)))
2533 /* either don't need iovec imported or already have it */
2534 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2538 io->rw.iov = io->rw.fast_iov;
2540 ret = io_import_iovec(READ, req, &io->rw.iov, &iter, !force_nonblock);
2545 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2549 static int io_read(struct io_kiocb *req, bool force_nonblock)
2551 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2552 struct kiocb *kiocb = &req->rw.kiocb;
2553 struct iov_iter iter;
2555 ssize_t io_size, ret;
2557 ret = io_import_iovec(READ, req, &iovec, &iter, !force_nonblock);
2561 /* Ensure we clear previously set non-block flag */
2562 if (!force_nonblock)
2563 kiocb->ki_flags &= ~IOCB_NOWAIT;
2567 if (req->flags & REQ_F_LINK_HEAD)
2568 req->result = io_size;
2571 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2572 * we know to async punt it even if it was opened O_NONBLOCK
2574 if (force_nonblock && !io_file_supports_async(req->file))
2577 iov_count = iov_iter_count(&iter);
2578 ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count);
2582 if (req->file->f_op->read_iter)
2583 ret2 = call_read_iter(req->file, kiocb, &iter);
2585 ret2 = loop_rw_iter(READ, req->file, kiocb, &iter);
2587 /* Catch -EAGAIN return for forced non-blocking submission */
2588 if (!force_nonblock || ret2 != -EAGAIN) {
2589 kiocb_done(kiocb, ret2);
2592 ret = io_setup_async_rw(req, io_size, iovec,
2593 inline_vecs, &iter);
2596 /* any defer here is final, must blocking retry */
2597 if (!(req->flags & REQ_F_NOWAIT))
2598 req->flags |= REQ_F_MUST_PUNT;
2604 req->flags &= ~REQ_F_NEED_CLEANUP;
2608 static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
2609 bool force_nonblock)
2611 struct io_async_ctx *io;
2612 struct iov_iter iter;
2615 ret = io_prep_rw(req, sqe, force_nonblock);
2619 if (unlikely(!(req->file->f_mode & FMODE_WRITE)))
2622 req->fsize = rlimit(RLIMIT_FSIZE);
2624 /* either don't need iovec imported or already have it */
2625 if (!req->io || req->flags & REQ_F_NEED_CLEANUP)
2629 io->rw.iov = io->rw.fast_iov;
2631 ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter, !force_nonblock);
2636 io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter);
2640 static int io_write(struct io_kiocb *req, bool force_nonblock)
2642 struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
2643 struct kiocb *kiocb = &req->rw.kiocb;
2644 struct iov_iter iter;
2646 ssize_t ret, io_size;
2648 ret = io_import_iovec(WRITE, req, &iovec, &iter, !force_nonblock);
2652 /* Ensure we clear previously set non-block flag */
2653 if (!force_nonblock)
2654 req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT;
2658 if (req->flags & REQ_F_LINK_HEAD)
2659 req->result = io_size;
2662 * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so
2663 * we know to async punt it even if it was opened O_NONBLOCK
2665 if (force_nonblock && !io_file_supports_async(req->file))
2668 /* file path doesn't support NOWAIT for non-direct_IO */
2669 if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) &&
2670 (req->flags & REQ_F_ISREG))
2673 iov_count = iov_iter_count(&iter);
2674 ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count);
2679 * Open-code file_start_write here to grab freeze protection,
2680 * which will be released by another thread in
2681 * io_complete_rw(). Fool lockdep by telling it the lock got
2682 * released so that it doesn't complain about the held lock when
2683 * we return to userspace.
2685 if (req->flags & REQ_F_ISREG) {
2686 __sb_start_write(file_inode(req->file)->i_sb,
2687 SB_FREEZE_WRITE, true);
2688 __sb_writers_release(file_inode(req->file)->i_sb,
2691 kiocb->ki_flags |= IOCB_WRITE;
2693 if (!force_nonblock)
2694 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2696 if (req->file->f_op->write_iter)
2697 ret2 = call_write_iter(req->file, kiocb, &iter);
2699 ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter);
2701 if (!force_nonblock)
2702 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2705 * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just
2706 * retry them without IOCB_NOWAIT.
2708 if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT))
2710 if (!force_nonblock || ret2 != -EAGAIN) {
2711 kiocb_done(kiocb, ret2);
2714 ret = io_setup_async_rw(req, io_size, iovec,
2715 inline_vecs, &iter);
2718 /* any defer here is final, must blocking retry */
2719 req->flags |= REQ_F_MUST_PUNT;
2724 req->flags &= ~REQ_F_NEED_CLEANUP;
2729 static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2731 struct io_splice* sp = &req->splice;
2732 unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL;
2735 if (req->flags & REQ_F_NEED_CLEANUP)
2739 sp->off_in = READ_ONCE(sqe->splice_off_in);
2740 sp->off_out = READ_ONCE(sqe->off);
2741 sp->len = READ_ONCE(sqe->len);
2742 sp->flags = READ_ONCE(sqe->splice_flags);
2744 if (unlikely(sp->flags & ~valid_flags))
2747 ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in,
2748 (sp->flags & SPLICE_F_FD_IN_FIXED));
2751 req->flags |= REQ_F_NEED_CLEANUP;
2753 if (!S_ISREG(file_inode(sp->file_in)->i_mode))
2754 req->work.flags |= IO_WQ_WORK_UNBOUND;
2759 static bool io_splice_punt(struct file *file)
2761 if (get_pipe_info(file))
2763 if (!io_file_supports_async(file))
2765 return !(file->f_flags & O_NONBLOCK);
2768 static int io_splice(struct io_kiocb *req, bool force_nonblock)
2770 struct io_splice *sp = &req->splice;
2771 struct file *in = sp->file_in;
2772 struct file *out = sp->file_out;
2773 unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED;
2774 loff_t *poff_in, *poff_out;
2777 if (force_nonblock) {
2778 if (io_splice_punt(in) || io_splice_punt(out))
2780 flags |= SPLICE_F_NONBLOCK;
2783 poff_in = (sp->off_in == -1) ? NULL : &sp->off_in;
2784 poff_out = (sp->off_out == -1) ? NULL : &sp->off_out;
2785 ret = do_splice(in, poff_in, out, poff_out, sp->len, flags);
2786 if (force_nonblock && ret == -EAGAIN)
2789 io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED));
2790 req->flags &= ~REQ_F_NEED_CLEANUP;
2792 io_cqring_add_event(req, ret);
2794 req_set_fail_links(req);
2800 * IORING_OP_NOP just posts a completion event, nothing else.
2802 static int io_nop(struct io_kiocb *req)
2804 struct io_ring_ctx *ctx = req->ctx;
2806 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2809 io_cqring_add_event(req, 0);
2814 static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2816 struct io_ring_ctx *ctx = req->ctx;
2821 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
2823 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
2826 req->sync.flags = READ_ONCE(sqe->fsync_flags);
2827 if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC))
2830 req->sync.off = READ_ONCE(sqe->off);
2831 req->sync.len = READ_ONCE(sqe->len);
2835 static bool io_req_cancelled(struct io_kiocb *req)
2837 if (req->work.flags & IO_WQ_WORK_CANCEL) {
2838 req_set_fail_links(req);
2839 io_cqring_add_event(req, -ECANCELED);
2847 static void __io_fsync(struct io_kiocb *req)
2849 loff_t end = req->sync.off + req->sync.len;
2852 ret = vfs_fsync_range(req->file, req->sync.off,
2853 end > 0 ? end : LLONG_MAX,
2854 req->sync.flags & IORING_FSYNC_DATASYNC);
2856 req_set_fail_links(req);
2857 io_cqring_add_event(req, ret);
2861 static void io_fsync_finish(struct io_wq_work **workptr)
2863 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2865 if (io_req_cancelled(req))
2868 io_steal_work(req, workptr);
2871 static int io_fsync(struct io_kiocb *req, bool force_nonblock)
2873 /* fsync always requires a blocking context */
2874 if (force_nonblock) {
2875 req->work.func = io_fsync_finish;
2882 static void __io_fallocate(struct io_kiocb *req)
2886 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = req->fsize;
2887 ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off,
2889 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
2891 req_set_fail_links(req);
2892 io_cqring_add_event(req, ret);
2896 static void io_fallocate_finish(struct io_wq_work **workptr)
2898 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
2900 if (io_req_cancelled(req))
2902 __io_fallocate(req);
2903 io_steal_work(req, workptr);
2906 static int io_fallocate_prep(struct io_kiocb *req,
2907 const struct io_uring_sqe *sqe)
2909 if (sqe->ioprio || sqe->buf_index || sqe->rw_flags)
2912 req->sync.off = READ_ONCE(sqe->off);
2913 req->sync.len = READ_ONCE(sqe->addr);
2914 req->sync.mode = READ_ONCE(sqe->len);
2915 req->fsize = rlimit(RLIMIT_FSIZE);
2919 static int io_fallocate(struct io_kiocb *req, bool force_nonblock)
2921 /* fallocate always requiring blocking context */
2922 if (force_nonblock) {
2923 req->work.func = io_fallocate_finish;
2927 __io_fallocate(req);
2931 static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2933 const char __user *fname;
2936 if (sqe->ioprio || sqe->buf_index)
2938 if (req->flags & REQ_F_FIXED_FILE)
2940 if (req->flags & REQ_F_NEED_CLEANUP)
2943 req->open.dfd = READ_ONCE(sqe->fd);
2944 req->open.how.mode = READ_ONCE(sqe->len);
2945 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2946 req->open.how.flags = READ_ONCE(sqe->open_flags);
2947 if (force_o_largefile())
2948 req->open.how.flags |= O_LARGEFILE;
2950 req->open.filename = getname(fname);
2951 if (IS_ERR(req->open.filename)) {
2952 ret = PTR_ERR(req->open.filename);
2953 req->open.filename = NULL;
2957 req->open.nofile = rlimit(RLIMIT_NOFILE);
2958 req->flags |= REQ_F_NEED_CLEANUP;
2962 static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
2964 struct open_how __user *how;
2965 const char __user *fname;
2969 if (sqe->ioprio || sqe->buf_index)
2971 if (req->flags & REQ_F_FIXED_FILE)
2973 if (req->flags & REQ_F_NEED_CLEANUP)
2976 req->open.dfd = READ_ONCE(sqe->fd);
2977 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
2978 how = u64_to_user_ptr(READ_ONCE(sqe->addr2));
2979 len = READ_ONCE(sqe->len);
2981 if (len < OPEN_HOW_SIZE_VER0)
2984 ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how,
2989 if (!(req->open.how.flags & O_PATH) && force_o_largefile())
2990 req->open.how.flags |= O_LARGEFILE;
2992 req->open.filename = getname(fname);
2993 if (IS_ERR(req->open.filename)) {
2994 ret = PTR_ERR(req->open.filename);
2995 req->open.filename = NULL;
2999 req->open.nofile = rlimit(RLIMIT_NOFILE);
3000 req->flags |= REQ_F_NEED_CLEANUP;
3004 static int io_openat2(struct io_kiocb *req, bool force_nonblock)
3006 struct open_flags op;
3013 ret = build_open_flags(&req->open.how, &op);
3017 ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
3021 file = do_filp_open(req->open.dfd, req->open.filename, &op);
3024 ret = PTR_ERR(file);
3026 fsnotify_open(file);
3027 fd_install(ret, file);
3030 putname(req->open.filename);
3031 req->flags &= ~REQ_F_NEED_CLEANUP;
3033 req_set_fail_links(req);
3034 io_cqring_add_event(req, ret);
3039 static int io_openat(struct io_kiocb *req, bool force_nonblock)
3041 req->open.how = build_open_how(req->open.how.flags, req->open.how.mode);
3042 return io_openat2(req, force_nonblock);
3045 static int io_remove_buffers_prep(struct io_kiocb *req,
3046 const struct io_uring_sqe *sqe)
3048 struct io_provide_buf *p = &req->pbuf;
3051 if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off)
3054 tmp = READ_ONCE(sqe->fd);
3055 if (!tmp || tmp > USHRT_MAX)
3058 memset(p, 0, sizeof(*p));
3060 p->bgid = READ_ONCE(sqe->buf_group);
3064 static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf,
3065 int bgid, unsigned nbufs)
3069 /* shouldn't happen */
3073 /* the head kbuf is the list itself */
3074 while (!list_empty(&buf->list)) {
3075 struct io_buffer *nxt;
3077 nxt = list_first_entry(&buf->list, struct io_buffer, list);
3078 list_del(&nxt->list);
3085 idr_remove(&ctx->io_buffer_idr, bgid);
3090 static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock)
3092 struct io_provide_buf *p = &req->pbuf;
3093 struct io_ring_ctx *ctx = req->ctx;
3094 struct io_buffer *head;
3097 io_ring_submit_lock(ctx, !force_nonblock);
3099 lockdep_assert_held(&ctx->uring_lock);
3102 head = idr_find(&ctx->io_buffer_idr, p->bgid);
3104 ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs);
3106 io_ring_submit_lock(ctx, !force_nonblock);
3108 req_set_fail_links(req);
3109 io_cqring_add_event(req, ret);
3114 static int io_provide_buffers_prep(struct io_kiocb *req,
3115 const struct io_uring_sqe *sqe)
3117 struct io_provide_buf *p = &req->pbuf;
3120 if (sqe->ioprio || sqe->rw_flags)
3123 tmp = READ_ONCE(sqe->fd);
3124 if (!tmp || tmp > USHRT_MAX)
3127 p->addr = READ_ONCE(sqe->addr);
3128 p->len = READ_ONCE(sqe->len);
3130 if (!access_ok(u64_to_user_ptr(p->addr), p->len))
3133 p->bgid = READ_ONCE(sqe->buf_group);
3134 tmp = READ_ONCE(sqe->off);
3135 if (tmp > USHRT_MAX)
3141 static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head)
3143 struct io_buffer *buf;
3144 u64 addr = pbuf->addr;
3145 int i, bid = pbuf->bid;
3147 for (i = 0; i < pbuf->nbufs; i++) {
3148 buf = kmalloc(sizeof(*buf), GFP_KERNEL);
3153 buf->len = pbuf->len;
3158 INIT_LIST_HEAD(&buf->list);
3161 list_add_tail(&buf->list, &(*head)->list);
3165 return i ? i : -ENOMEM;
3168 static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock)
3170 struct io_provide_buf *p = &req->pbuf;
3171 struct io_ring_ctx *ctx = req->ctx;
3172 struct io_buffer *head, *list;
3175 io_ring_submit_lock(ctx, !force_nonblock);
3177 lockdep_assert_held(&ctx->uring_lock);
3179 list = head = idr_find(&ctx->io_buffer_idr, p->bgid);
3181 ret = io_add_buffers(p, &head);
3186 ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1,
3189 __io_remove_buffers(ctx, head, p->bgid, -1U);
3194 io_ring_submit_unlock(ctx, !force_nonblock);
3196 req_set_fail_links(req);
3197 io_cqring_add_event(req, ret);
3202 static int io_epoll_ctl_prep(struct io_kiocb *req,
3203 const struct io_uring_sqe *sqe)
3205 #if defined(CONFIG_EPOLL)
3206 if (sqe->ioprio || sqe->buf_index)
3209 req->epoll.epfd = READ_ONCE(sqe->fd);
3210 req->epoll.op = READ_ONCE(sqe->len);
3211 req->epoll.fd = READ_ONCE(sqe->off);
3213 if (ep_op_has_event(req->epoll.op)) {
3214 struct epoll_event __user *ev;
3216 ev = u64_to_user_ptr(READ_ONCE(sqe->addr));
3217 if (copy_from_user(&req->epoll.event, ev, sizeof(*ev)))
3227 static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock)
3229 #if defined(CONFIG_EPOLL)
3230 struct io_epoll *ie = &req->epoll;
3233 ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock);
3234 if (force_nonblock && ret == -EAGAIN)
3238 req_set_fail_links(req);
3239 io_cqring_add_event(req, ret);
3247 static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3249 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3250 if (sqe->ioprio || sqe->buf_index || sqe->off)
3253 req->madvise.addr = READ_ONCE(sqe->addr);
3254 req->madvise.len = READ_ONCE(sqe->len);
3255 req->madvise.advice = READ_ONCE(sqe->fadvise_advice);
3262 static int io_madvise(struct io_kiocb *req, bool force_nonblock)
3264 #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU)
3265 struct io_madvise *ma = &req->madvise;
3271 ret = do_madvise(ma->addr, ma->len, ma->advice);
3273 req_set_fail_links(req);
3274 io_cqring_add_event(req, ret);
3282 static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3284 if (sqe->ioprio || sqe->buf_index || sqe->addr)
3287 req->fadvise.offset = READ_ONCE(sqe->off);
3288 req->fadvise.len = READ_ONCE(sqe->len);
3289 req->fadvise.advice = READ_ONCE(sqe->fadvise_advice);
3293 static int io_fadvise(struct io_kiocb *req, bool force_nonblock)
3295 struct io_fadvise *fa = &req->fadvise;
3298 if (force_nonblock) {
3299 switch (fa->advice) {
3300 case POSIX_FADV_NORMAL:
3301 case POSIX_FADV_RANDOM:
3302 case POSIX_FADV_SEQUENTIAL:
3309 ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice);
3311 req_set_fail_links(req);
3312 io_cqring_add_event(req, ret);
3317 static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3319 const char __user *fname;
3320 unsigned lookup_flags;
3323 if (sqe->ioprio || sqe->buf_index)
3325 if (req->flags & REQ_F_FIXED_FILE)
3327 if (req->flags & REQ_F_NEED_CLEANUP)
3330 req->open.dfd = READ_ONCE(sqe->fd);
3331 req->open.mask = READ_ONCE(sqe->len);
3332 fname = u64_to_user_ptr(READ_ONCE(sqe->addr));
3333 req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3334 req->open.how.flags = READ_ONCE(sqe->statx_flags);
3336 if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags))
3339 req->open.filename = getname_flags(fname, lookup_flags, NULL);
3340 if (IS_ERR(req->open.filename)) {
3341 ret = PTR_ERR(req->open.filename);
3342 req->open.filename = NULL;
3346 req->flags |= REQ_F_NEED_CLEANUP;
3350 static int io_statx(struct io_kiocb *req, bool force_nonblock)
3352 struct io_open *ctx = &req->open;
3353 unsigned lookup_flags;
3361 if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags))
3365 /* filename_lookup() drops it, keep a reference */
3366 ctx->filename->refcnt++;
3368 ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path,
3373 ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags);
3375 if (retry_estale(ret, lookup_flags)) {
3376 lookup_flags |= LOOKUP_REVAL;
3380 ret = cp_statx(&stat, ctx->buffer);
3382 putname(ctx->filename);
3383 req->flags &= ~REQ_F_NEED_CLEANUP;
3385 req_set_fail_links(req);
3386 io_cqring_add_event(req, ret);
3391 static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3394 * If we queue this for async, it must not be cancellable. That would
3395 * leave the 'file' in an undeterminate state.
3397 req->work.flags |= IO_WQ_WORK_NO_CANCEL;
3399 if (sqe->ioprio || sqe->off || sqe->addr || sqe->len ||
3400 sqe->rw_flags || sqe->buf_index)
3402 if (req->flags & REQ_F_FIXED_FILE)
3405 req->close.fd = READ_ONCE(sqe->fd);
3406 if (req->file->f_op == &io_uring_fops ||
3407 req->close.fd == req->ctx->ring_fd)
3413 /* only called when __close_fd_get_file() is done */
3414 static void __io_close_finish(struct io_kiocb *req)
3418 ret = filp_close(req->close.put_file, req->work.files);
3420 req_set_fail_links(req);
3421 io_cqring_add_event(req, ret);
3422 fput(req->close.put_file);
3426 static void io_close_finish(struct io_wq_work **workptr)
3428 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3430 /* not cancellable, don't do io_req_cancelled() */
3431 __io_close_finish(req);
3432 io_steal_work(req, workptr);
3435 static int io_close(struct io_kiocb *req, bool force_nonblock)
3439 req->close.put_file = NULL;
3440 ret = __close_fd_get_file(req->close.fd, &req->close.put_file);
3444 /* if the file has a flush method, be safe and punt to async */
3445 if (req->close.put_file->f_op->flush && force_nonblock) {
3446 /* submission ref will be dropped, take it for async */
3447 refcount_inc(&req->refs);
3449 req->work.func = io_close_finish;
3451 * Do manual async queue here to avoid grabbing files - we don't
3452 * need the files, and it'll cause io_close_finish() to close
3453 * the file again and cause a double CQE entry for this request
3455 io_queue_async_work(req);
3460 * No ->flush(), safely close from here and just punt the
3461 * fput() to async context.
3463 __io_close_finish(req);
3467 static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3469 struct io_ring_ctx *ctx = req->ctx;
3474 if (unlikely(ctx->flags & IORING_SETUP_IOPOLL))
3476 if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index))
3479 req->sync.off = READ_ONCE(sqe->off);
3480 req->sync.len = READ_ONCE(sqe->len);
3481 req->sync.flags = READ_ONCE(sqe->sync_range_flags);
3485 static void __io_sync_file_range(struct io_kiocb *req)
3489 ret = sync_file_range(req->file, req->sync.off, req->sync.len,
3492 req_set_fail_links(req);
3493 io_cqring_add_event(req, ret);
3498 static void io_sync_file_range_finish(struct io_wq_work **workptr)
3500 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3502 if (io_req_cancelled(req))
3504 __io_sync_file_range(req);
3505 io_put_req(req); /* put submission ref */
3508 static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock)
3510 /* sync_file_range always requires a blocking context */
3511 if (force_nonblock) {
3512 req->work.func = io_sync_file_range_finish;
3516 __io_sync_file_range(req);
3520 #if defined(CONFIG_NET)
3521 static int io_setup_async_msg(struct io_kiocb *req,
3522 struct io_async_msghdr *kmsg)
3526 if (io_alloc_async_ctx(req)) {
3527 if (kmsg->iov != kmsg->fast_iov)
3531 req->flags |= REQ_F_NEED_CLEANUP;
3532 memcpy(&req->io->msg, kmsg, sizeof(*kmsg));
3536 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3538 struct io_sr_msg *sr = &req->sr_msg;
3539 struct io_async_ctx *io = req->io;
3542 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3543 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3544 sr->len = READ_ONCE(sqe->len);
3546 #ifdef CONFIG_COMPAT
3547 if (req->ctx->compat)
3548 sr->msg_flags |= MSG_CMSG_COMPAT;
3551 if (!io || req->opcode == IORING_OP_SEND)
3553 /* iovec is already imported */
3554 if (req->flags & REQ_F_NEED_CLEANUP)
3557 io->msg.iov = io->msg.fast_iov;
3558 ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags,
3561 req->flags |= REQ_F_NEED_CLEANUP;
3565 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
3567 struct io_async_msghdr *kmsg = NULL;
3568 struct socket *sock;
3571 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3574 sock = sock_from_file(req->file, &ret);
3576 struct io_async_ctx io;
3580 kmsg = &req->io->msg;
3581 kmsg->msg.msg_name = &req->io->msg.addr;
3582 /* if iov is set, it's allocated already */
3584 kmsg->iov = kmsg->fast_iov;
3585 kmsg->msg.msg_iter.iov = kmsg->iov;
3587 struct io_sr_msg *sr = &req->sr_msg;
3590 kmsg->msg.msg_name = &io.msg.addr;
3592 io.msg.iov = io.msg.fast_iov;
3593 ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg,
3594 sr->msg_flags, &io.msg.iov);
3599 flags = req->sr_msg.msg_flags;
3600 if (flags & MSG_DONTWAIT)
3601 req->flags |= REQ_F_NOWAIT;
3602 else if (force_nonblock)
3603 flags |= MSG_DONTWAIT;
3605 ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags);
3606 if (force_nonblock && ret == -EAGAIN)
3607 return io_setup_async_msg(req, kmsg);
3608 if (ret == -ERESTARTSYS)
3612 if (kmsg && kmsg->iov != kmsg->fast_iov)
3614 req->flags &= ~REQ_F_NEED_CLEANUP;
3615 io_cqring_add_event(req, ret);
3617 req_set_fail_links(req);
3622 static int io_send(struct io_kiocb *req, bool force_nonblock)
3624 struct socket *sock;
3627 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3630 sock = sock_from_file(req->file, &ret);
3632 struct io_sr_msg *sr = &req->sr_msg;
3637 ret = import_single_range(WRITE, sr->buf, sr->len, &iov,
3642 msg.msg_name = NULL;
3643 msg.msg_control = NULL;
3644 msg.msg_controllen = 0;
3645 msg.msg_namelen = 0;
3647 flags = req->sr_msg.msg_flags;
3648 if (flags & MSG_DONTWAIT)
3649 req->flags |= REQ_F_NOWAIT;
3650 else if (force_nonblock)
3651 flags |= MSG_DONTWAIT;
3653 msg.msg_flags = flags;
3654 ret = sock_sendmsg(sock, &msg);
3655 if (force_nonblock && ret == -EAGAIN)
3657 if (ret == -ERESTARTSYS)
3661 io_cqring_add_event(req, ret);
3663 req_set_fail_links(req);
3668 static int __io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3670 struct io_sr_msg *sr = &req->sr_msg;
3671 struct iovec __user *uiov;
3675 ret = __copy_msghdr_from_user(&io->msg.msg, sr->msg, &io->msg.uaddr,
3680 if (req->flags & REQ_F_BUFFER_SELECT) {
3683 if (copy_from_user(io->msg.iov, uiov, sizeof(*uiov)))
3685 sr->len = io->msg.iov[0].iov_len;
3686 iov_iter_init(&io->msg.msg.msg_iter, READ, io->msg.iov, 1,
3690 ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV,
3691 &io->msg.iov, &io->msg.msg.msg_iter);
3699 #ifdef CONFIG_COMPAT
3700 static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req,
3701 struct io_async_ctx *io)
3703 struct compat_msghdr __user *msg_compat;
3704 struct io_sr_msg *sr = &req->sr_msg;
3705 struct compat_iovec __user *uiov;
3710 msg_compat = (struct compat_msghdr __user *) sr->msg;
3711 ret = __get_compat_msghdr(&io->msg.msg, msg_compat, &io->msg.uaddr,
3716 uiov = compat_ptr(ptr);
3717 if (req->flags & REQ_F_BUFFER_SELECT) {
3718 compat_ssize_t clen;
3722 if (!access_ok(uiov, sizeof(*uiov)))
3724 if (__get_user(clen, &uiov->iov_len))
3728 sr->len = io->msg.iov[0].iov_len;
3731 ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV,
3733 &io->msg.msg.msg_iter);
3742 static int io_recvmsg_copy_hdr(struct io_kiocb *req, struct io_async_ctx *io)
3744 io->msg.iov = io->msg.fast_iov;
3746 #ifdef CONFIG_COMPAT
3747 if (req->ctx->compat)
3748 return __io_compat_recvmsg_copy_hdr(req, io);
3751 return __io_recvmsg_copy_hdr(req, io);
3754 static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req,
3755 int *cflags, bool needs_lock)
3757 struct io_sr_msg *sr = &req->sr_msg;
3758 struct io_buffer *kbuf;
3760 if (!(req->flags & REQ_F_BUFFER_SELECT))
3763 kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock);
3768 req->flags |= REQ_F_BUFFER_SELECTED;
3770 *cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT;
3771 *cflags |= IORING_CQE_F_BUFFER;
3775 static int io_recvmsg_prep(struct io_kiocb *req,
3776 const struct io_uring_sqe *sqe)
3778 struct io_sr_msg *sr = &req->sr_msg;
3779 struct io_async_ctx *io = req->io;
3782 sr->msg_flags = READ_ONCE(sqe->msg_flags);
3783 sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr));
3784 sr->len = READ_ONCE(sqe->len);
3785 sr->bgid = READ_ONCE(sqe->buf_group);
3787 #ifdef CONFIG_COMPAT
3788 if (req->ctx->compat)
3789 sr->msg_flags |= MSG_CMSG_COMPAT;
3792 if (!io || req->opcode == IORING_OP_RECV)
3794 /* iovec is already imported */
3795 if (req->flags & REQ_F_NEED_CLEANUP)
3798 ret = io_recvmsg_copy_hdr(req, io);
3800 req->flags |= REQ_F_NEED_CLEANUP;
3804 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
3806 struct io_async_msghdr *kmsg = NULL;
3807 struct socket *sock;
3808 int ret, cflags = 0;
3810 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3813 sock = sock_from_file(req->file, &ret);
3815 struct io_buffer *kbuf;
3816 struct io_async_ctx io;
3820 kmsg = &req->io->msg;
3821 kmsg->msg.msg_name = &req->io->msg.addr;
3822 /* if iov is set, it's allocated already */
3824 kmsg->iov = kmsg->fast_iov;
3825 kmsg->msg.msg_iter.iov = kmsg->iov;
3828 kmsg->msg.msg_name = &io.msg.addr;
3830 ret = io_recvmsg_copy_hdr(req, &io);
3835 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3837 return PTR_ERR(kbuf);
3839 kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr);
3840 iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov,
3841 1, req->sr_msg.len);
3844 flags = req->sr_msg.msg_flags;
3845 if (flags & MSG_DONTWAIT)
3846 req->flags |= REQ_F_NOWAIT;
3847 else if (force_nonblock)
3848 flags |= MSG_DONTWAIT;
3850 ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg,
3851 kmsg->uaddr, flags);
3852 if (force_nonblock && ret == -EAGAIN)
3853 return io_setup_async_msg(req, kmsg);
3854 if (ret == -ERESTARTSYS)
3858 if (kmsg && kmsg->iov != kmsg->fast_iov)
3860 req->flags &= ~REQ_F_NEED_CLEANUP;
3861 __io_cqring_add_event(req, ret, cflags);
3863 req_set_fail_links(req);
3868 static int io_recv(struct io_kiocb *req, bool force_nonblock)
3870 struct io_buffer *kbuf = NULL;
3871 struct socket *sock;
3872 int ret, cflags = 0;
3874 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
3877 sock = sock_from_file(req->file, &ret);
3879 struct io_sr_msg *sr = &req->sr_msg;
3880 void __user *buf = sr->buf;
3885 kbuf = io_recv_buffer_select(req, &cflags, !force_nonblock);
3887 return PTR_ERR(kbuf);
3889 buf = u64_to_user_ptr(kbuf->addr);
3891 ret = import_single_range(READ, buf, sr->len, &iov,
3898 req->flags |= REQ_F_NEED_CLEANUP;
3899 msg.msg_name = NULL;
3900 msg.msg_control = NULL;
3901 msg.msg_controllen = 0;
3902 msg.msg_namelen = 0;
3903 msg.msg_iocb = NULL;
3906 flags = req->sr_msg.msg_flags;
3907 if (flags & MSG_DONTWAIT)
3908 req->flags |= REQ_F_NOWAIT;
3909 else if (force_nonblock)
3910 flags |= MSG_DONTWAIT;
3912 ret = sock_recvmsg(sock, &msg, flags);
3913 if (force_nonblock && ret == -EAGAIN)
3915 if (ret == -ERESTARTSYS)
3920 req->flags &= ~REQ_F_NEED_CLEANUP;
3921 __io_cqring_add_event(req, ret, cflags);
3923 req_set_fail_links(req);
3928 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3930 struct io_accept *accept = &req->accept;
3932 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3934 if (sqe->ioprio || sqe->len || sqe->buf_index)
3937 accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3938 accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
3939 accept->flags = READ_ONCE(sqe->accept_flags);
3940 accept->nofile = rlimit(RLIMIT_NOFILE);
3944 static int __io_accept(struct io_kiocb *req, bool force_nonblock)
3946 struct io_accept *accept = &req->accept;
3947 unsigned file_flags;
3950 file_flags = force_nonblock ? O_NONBLOCK : 0;
3951 ret = __sys_accept4_file(req->file, file_flags, accept->addr,
3952 accept->addr_len, accept->flags,
3954 if (ret == -EAGAIN && force_nonblock)
3956 if (ret == -ERESTARTSYS)
3959 req_set_fail_links(req);
3960 io_cqring_add_event(req, ret);
3965 static void io_accept_finish(struct io_wq_work **workptr)
3967 struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work);
3969 if (io_req_cancelled(req))
3971 __io_accept(req, false);
3972 io_steal_work(req, workptr);
3975 static int io_accept(struct io_kiocb *req, bool force_nonblock)
3979 ret = __io_accept(req, force_nonblock);
3980 if (ret == -EAGAIN && force_nonblock) {
3981 req->work.func = io_accept_finish;
3987 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
3989 struct io_connect *conn = &req->connect;
3990 struct io_async_ctx *io = req->io;
3992 if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL)))
3994 if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags)
3997 conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
3998 conn->addr_len = READ_ONCE(sqe->addr2);
4003 return move_addr_to_kernel(conn->addr, conn->addr_len,
4004 &io->connect.address);
4007 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4009 struct io_async_ctx __io, *io;
4010 unsigned file_flags;
4016 ret = move_addr_to_kernel(req->connect.addr,
4017 req->connect.addr_len,
4018 &__io.connect.address);
4024 file_flags = force_nonblock ? O_NONBLOCK : 0;
4026 ret = __sys_connect_file(req->file, &io->connect.address,
4027 req->connect.addr_len, file_flags);
4028 if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) {
4031 if (io_alloc_async_ctx(req)) {
4035 memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect));
4038 if (ret == -ERESTARTSYS)
4042 req_set_fail_links(req);
4043 io_cqring_add_event(req, ret);
4047 #else /* !CONFIG_NET */
4048 static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4053 static int io_sendmsg(struct io_kiocb *req, bool force_nonblock)
4058 static int io_send(struct io_kiocb *req, bool force_nonblock)
4063 static int io_recvmsg_prep(struct io_kiocb *req,
4064 const struct io_uring_sqe *sqe)
4069 static int io_recvmsg(struct io_kiocb *req, bool force_nonblock)
4074 static int io_recv(struct io_kiocb *req, bool force_nonblock)
4079 static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4084 static int io_accept(struct io_kiocb *req, bool force_nonblock)
4089 static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4094 static int io_connect(struct io_kiocb *req, bool force_nonblock)
4098 #endif /* CONFIG_NET */
4100 struct io_poll_table {
4101 struct poll_table_struct pt;
4102 struct io_kiocb *req;
4106 static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt,
4107 struct wait_queue_head *head)
4109 if (unlikely(poll->head)) {
4110 pt->error = -EINVAL;
4116 add_wait_queue(head, &poll->wait);
4119 static void io_async_queue_proc(struct file *file, struct wait_queue_head *head,
4120 struct poll_table_struct *p)
4122 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4124 __io_queue_proc(&pt->req->apoll->poll, pt, head);
4127 static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll,
4128 __poll_t mask, task_work_func_t func)
4130 struct task_struct *tsk;
4133 /* for instances that support it check for an event match first: */
4134 if (mask && !(mask & poll->events))
4137 trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask);
4139 list_del_init(&poll->wait.entry);
4143 init_task_work(&req->task_work, func);
4145 * If this fails, then the task is exiting. Punt to one of the io-wq
4146 * threads to ensure the work gets run, we can't always rely on exit
4147 * cancelation taking care of this.
4149 ret = task_work_add(tsk, &req->task_work, true);
4150 if (unlikely(ret)) {
4151 tsk = io_wq_get_task(req->ctx->io_wq);
4152 task_work_add(tsk, &req->task_work, true);
4154 wake_up_process(tsk);
4158 static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
4159 __acquires(&req->ctx->completion_lock)
4161 struct io_ring_ctx *ctx = req->ctx;
4163 if (!req->result && !READ_ONCE(poll->canceled)) {
4164 struct poll_table_struct pt = { ._key = poll->events };
4166 req->result = vfs_poll(req->file, &pt) & poll->events;
4169 spin_lock_irq(&ctx->completion_lock);
4170 if (!req->result && !READ_ONCE(poll->canceled)) {
4171 add_wait_queue(poll->head, &poll->wait);
4178 static void io_async_task_func(struct callback_head *cb)
4180 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4181 struct async_poll *apoll = req->apoll;
4182 struct io_ring_ctx *ctx = req->ctx;
4185 trace_io_uring_task_run(req->ctx, req->opcode, req->user_data);
4187 if (io_poll_rewait(req, &apoll->poll)) {
4188 spin_unlock_irq(&ctx->completion_lock);
4192 if (hash_hashed(&req->hash_node))
4193 hash_del(&req->hash_node);
4195 canceled = READ_ONCE(apoll->poll.canceled);
4197 io_cqring_fill_event(req, -ECANCELED);
4198 io_commit_cqring(ctx);
4201 spin_unlock_irq(&ctx->completion_lock);
4205 io_cqring_ev_posted(ctx);
4206 req_set_fail_links(req);
4211 /* restore ->work in case we need to retry again */
4212 memcpy(&req->work, &apoll->work, sizeof(req->work));
4214 __set_current_state(TASK_RUNNING);
4215 mutex_lock(&ctx->uring_lock);
4216 __io_queue_sqe(req, NULL);
4217 mutex_unlock(&ctx->uring_lock);
4222 static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4225 struct io_kiocb *req = wait->private;
4226 struct io_poll_iocb *poll = &req->apoll->poll;
4228 trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data,
4231 return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func);
4234 static void io_poll_req_insert(struct io_kiocb *req)
4236 struct io_ring_ctx *ctx = req->ctx;
4237 struct hlist_head *list;
4239 list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)];
4240 hlist_add_head(&req->hash_node, list);
4243 static __poll_t __io_arm_poll_handler(struct io_kiocb *req,
4244 struct io_poll_iocb *poll,
4245 struct io_poll_table *ipt, __poll_t mask,
4246 wait_queue_func_t wake_func)
4247 __acquires(&ctx->completion_lock)
4249 struct io_ring_ctx *ctx = req->ctx;
4250 bool cancel = false;
4252 poll->file = req->file;
4254 poll->done = poll->canceled = false;
4255 poll->events = mask;
4257 ipt->pt._key = mask;
4259 ipt->error = -EINVAL;
4261 INIT_LIST_HEAD(&poll->wait.entry);
4262 init_waitqueue_func_entry(&poll->wait, wake_func);
4263 poll->wait.private = req;
4265 mask = vfs_poll(req->file, &ipt->pt) & poll->events;
4267 spin_lock_irq(&ctx->completion_lock);
4268 if (likely(poll->head)) {
4269 spin_lock(&poll->head->lock);
4270 if (unlikely(list_empty(&poll->wait.entry))) {
4276 if (mask || ipt->error)
4277 list_del_init(&poll->wait.entry);
4279 WRITE_ONCE(poll->canceled, true);
4280 else if (!poll->done) /* actually waiting for an event */
4281 io_poll_req_insert(req);
4282 spin_unlock(&poll->head->lock);
4288 static bool io_arm_poll_handler(struct io_kiocb *req)
4290 const struct io_op_def *def = &io_op_defs[req->opcode];
4291 struct io_ring_ctx *ctx = req->ctx;
4292 struct async_poll *apoll;
4293 struct io_poll_table ipt;
4296 if (!req->file || !file_can_poll(req->file))
4298 if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED))
4300 if (!def->pollin && !def->pollout)
4303 apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC);
4304 if (unlikely(!apoll))
4307 req->flags |= REQ_F_POLLED;
4308 memcpy(&apoll->work, &req->work, sizeof(req->work));
4310 get_task_struct(current);
4311 req->task = current;
4313 INIT_HLIST_NODE(&req->hash_node);
4317 mask |= POLLIN | POLLRDNORM;
4319 mask |= POLLOUT | POLLWRNORM;
4320 mask |= POLLERR | POLLPRI;
4322 ipt.pt._qproc = io_async_queue_proc;
4324 ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask,
4328 apoll->poll.done = true;
4329 spin_unlock_irq(&ctx->completion_lock);
4330 memcpy(&req->work, &apoll->work, sizeof(req->work));
4334 spin_unlock_irq(&ctx->completion_lock);
4335 trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask,
4336 apoll->poll.events);
4340 static bool __io_poll_remove_one(struct io_kiocb *req,
4341 struct io_poll_iocb *poll)
4343 bool do_complete = false;
4345 spin_lock(&poll->head->lock);
4346 WRITE_ONCE(poll->canceled, true);
4347 if (!list_empty(&poll->wait.entry)) {
4348 list_del_init(&poll->wait.entry);
4351 spin_unlock(&poll->head->lock);
4355 static bool io_poll_remove_one(struct io_kiocb *req)
4357 struct async_poll *apoll = NULL;
4360 if (req->opcode == IORING_OP_POLL_ADD) {
4361 do_complete = __io_poll_remove_one(req, &req->poll);
4364 /* non-poll requests have submit ref still */
4365 do_complete = __io_poll_remove_one(req, &req->apoll->poll);
4370 hash_del(&req->hash_node);
4374 * restore ->work because we need to call io_req_work_drop_env.
4376 memcpy(&req->work, &apoll->work, sizeof(req->work));
4381 io_cqring_fill_event(req, -ECANCELED);
4382 io_commit_cqring(req->ctx);
4383 req->flags |= REQ_F_COMP_LOCKED;
4390 static void io_poll_remove_all(struct io_ring_ctx *ctx)
4392 struct hlist_node *tmp;
4393 struct io_kiocb *req;
4396 spin_lock_irq(&ctx->completion_lock);
4397 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
4398 struct hlist_head *list;
4400 list = &ctx->cancel_hash[i];
4401 hlist_for_each_entry_safe(req, tmp, list, hash_node)
4402 posted += io_poll_remove_one(req);
4404 spin_unlock_irq(&ctx->completion_lock);
4407 io_cqring_ev_posted(ctx);
4410 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
4412 struct hlist_head *list;
4413 struct io_kiocb *req;
4415 list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)];
4416 hlist_for_each_entry(req, list, hash_node) {
4417 if (sqe_addr != req->user_data)
4419 if (io_poll_remove_one(req))
4427 static int io_poll_remove_prep(struct io_kiocb *req,
4428 const struct io_uring_sqe *sqe)
4430 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4432 if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index ||
4436 req->poll.addr = READ_ONCE(sqe->addr);
4441 * Find a running poll command that matches one specified in sqe->addr,
4442 * and remove it if found.
4444 static int io_poll_remove(struct io_kiocb *req)
4446 struct io_ring_ctx *ctx = req->ctx;
4450 addr = req->poll.addr;
4451 spin_lock_irq(&ctx->completion_lock);
4452 ret = io_poll_cancel(ctx, addr);
4453 spin_unlock_irq(&ctx->completion_lock);
4455 io_cqring_add_event(req, ret);
4457 req_set_fail_links(req);
4462 static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error)
4464 struct io_ring_ctx *ctx = req->ctx;
4466 req->poll.done = true;
4467 io_cqring_fill_event(req, error ? error : mangle_poll(mask));
4468 io_commit_cqring(ctx);
4471 static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt)
4473 struct io_ring_ctx *ctx = req->ctx;
4474 struct io_poll_iocb *poll = &req->poll;
4476 if (io_poll_rewait(req, poll)) {
4477 spin_unlock_irq(&ctx->completion_lock);
4481 hash_del(&req->hash_node);
4482 io_poll_complete(req, req->result, 0);
4483 req->flags |= REQ_F_COMP_LOCKED;
4484 io_put_req_find_next(req, nxt);
4485 spin_unlock_irq(&ctx->completion_lock);
4487 io_cqring_ev_posted(ctx);
4490 static void io_poll_task_func(struct callback_head *cb)
4492 struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
4493 struct io_kiocb *nxt = NULL;
4495 io_poll_task_handler(req, &nxt);
4497 struct io_ring_ctx *ctx = nxt->ctx;
4499 mutex_lock(&ctx->uring_lock);
4500 __io_queue_sqe(nxt, NULL);
4501 mutex_unlock(&ctx->uring_lock);
4505 static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
4508 struct io_kiocb *req = wait->private;
4509 struct io_poll_iocb *poll = &req->poll;
4511 return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func);
4514 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
4515 struct poll_table_struct *p)
4517 struct io_poll_table *pt = container_of(p, struct io_poll_table, pt);
4519 __io_queue_proc(&pt->req->poll, pt, head);
4522 static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
4524 struct io_poll_iocb *poll = &req->poll;
4527 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4529 if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index)
4534 events = READ_ONCE(sqe->poll_events);
4535 poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
4537 get_task_struct(current);
4538 req->task = current;
4542 static int io_poll_add(struct io_kiocb *req)
4544 struct io_poll_iocb *poll = &req->poll;
4545 struct io_ring_ctx *ctx = req->ctx;
4546 struct io_poll_table ipt;
4549 INIT_HLIST_NODE(&req->hash_node);
4550 INIT_LIST_HEAD(&req->list);
4551 ipt.pt._qproc = io_poll_queue_proc;
4553 mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events,
4556 if (mask) { /* no async, we'd stolen it */
4558 io_poll_complete(req, mask, 0);
4560 spin_unlock_irq(&ctx->completion_lock);
4563 io_cqring_ev_posted(ctx);
4569 static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
4571 struct io_timeout_data *data = container_of(timer,
4572 struct io_timeout_data, timer);
4573 struct io_kiocb *req = data->req;
4574 struct io_ring_ctx *ctx = req->ctx;
4575 unsigned long flags;
4577 atomic_inc(&ctx->cq_timeouts);
4579 spin_lock_irqsave(&ctx->completion_lock, flags);
4581 * We could be racing with timeout deletion. If the list is empty,
4582 * then timeout lookup already found it and will be handling it.
4584 if (!list_empty(&req->list)) {
4585 struct io_kiocb *prev;
4588 * Adjust the reqs sequence before the current one because it
4589 * will consume a slot in the cq_ring and the cq_tail
4590 * pointer will be increased, otherwise other timeout reqs may
4591 * return in advance without waiting for enough wait_nr.
4594 list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
4596 list_del_init(&req->list);
4599 io_cqring_fill_event(req, -ETIME);
4600 io_commit_cqring(ctx);
4601 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4603 io_cqring_ev_posted(ctx);
4604 req_set_fail_links(req);
4606 return HRTIMER_NORESTART;
4609 static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data)
4611 struct io_kiocb *req;
4614 list_for_each_entry(req, &ctx->timeout_list, list) {
4615 if (user_data == req->user_data) {
4616 list_del_init(&req->list);
4625 ret = hrtimer_try_to_cancel(&req->io->timeout.timer);
4629 req_set_fail_links(req);
4630 io_cqring_fill_event(req, -ECANCELED);
4635 static int io_timeout_remove_prep(struct io_kiocb *req,
4636 const struct io_uring_sqe *sqe)
4638 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4640 if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len)
4643 req->timeout.addr = READ_ONCE(sqe->addr);
4644 req->timeout.flags = READ_ONCE(sqe->timeout_flags);
4645 if (req->timeout.flags)
4652 * Remove or update an existing timeout command
4654 static int io_timeout_remove(struct io_kiocb *req)
4656 struct io_ring_ctx *ctx = req->ctx;
4659 spin_lock_irq(&ctx->completion_lock);
4660 ret = io_timeout_cancel(ctx, req->timeout.addr);
4662 io_cqring_fill_event(req, ret);
4663 io_commit_cqring(ctx);
4664 spin_unlock_irq(&ctx->completion_lock);
4665 io_cqring_ev_posted(ctx);
4667 req_set_fail_links(req);
4672 static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
4673 bool is_timeout_link)
4675 struct io_timeout_data *data;
4678 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4680 if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
4682 if (sqe->off && is_timeout_link)
4684 flags = READ_ONCE(sqe->timeout_flags);
4685 if (flags & ~IORING_TIMEOUT_ABS)
4688 req->timeout.count = READ_ONCE(sqe->off);
4690 if (!req->io && io_alloc_async_ctx(req))
4693 data = &req->io->timeout;
4695 req->flags |= REQ_F_TIMEOUT;
4697 if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr)))
4700 if (flags & IORING_TIMEOUT_ABS)
4701 data->mode = HRTIMER_MODE_ABS;
4703 data->mode = HRTIMER_MODE_REL;
4705 hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode);
4709 static int io_timeout(struct io_kiocb *req)
4711 struct io_ring_ctx *ctx = req->ctx;
4712 struct io_timeout_data *data;
4713 struct list_head *entry;
4715 u32 count = req->timeout.count;
4716 u32 seq = req->sequence;
4718 data = &req->io->timeout;
4721 * sqe->off holds how many events that need to occur for this
4722 * timeout event to be satisfied. If it isn't set, then this is
4723 * a pure timeout request, sequence isn't used.
4726 req->flags |= REQ_F_TIMEOUT_NOSEQ;
4727 spin_lock_irq(&ctx->completion_lock);
4728 entry = ctx->timeout_list.prev;
4732 req->sequence = seq + count;
4735 * Insertion sort, ensuring the first entry in the list is always
4736 * the one we need first.
4738 spin_lock_irq(&ctx->completion_lock);
4739 list_for_each_prev(entry, &ctx->timeout_list) {
4740 struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
4742 long long tmp, tmp_nxt;
4743 u32 nxt_offset = nxt->timeout.count;
4745 if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
4749 * Since seq + count can overflow, use type long
4752 tmp = (long long)seq + count;
4753 nxt_seq = nxt->sequence - nxt_offset;
4754 tmp_nxt = (long long)nxt_seq + nxt_offset;
4757 * cached_sq_head may overflow, and it will never overflow twice
4758 * once there is some timeout req still be valid.
4767 * Sequence of reqs after the insert one and itself should
4768 * be adjusted because each timeout req consumes a slot.
4773 req->sequence -= span;
4775 list_add(&req->list, entry);
4776 data->timer.function = io_timeout_fn;
4777 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode);
4778 spin_unlock_irq(&ctx->completion_lock);
4782 static bool io_cancel_cb(struct io_wq_work *work, void *data)
4784 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
4786 return req->user_data == (unsigned long) data;
4789 static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr)
4791 enum io_wq_cancel cancel_ret;
4794 cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr);
4795 switch (cancel_ret) {
4796 case IO_WQ_CANCEL_OK:
4799 case IO_WQ_CANCEL_RUNNING:
4802 case IO_WQ_CANCEL_NOTFOUND:
4810 static void io_async_find_and_cancel(struct io_ring_ctx *ctx,
4811 struct io_kiocb *req, __u64 sqe_addr,
4814 unsigned long flags;
4817 ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr);
4818 if (ret != -ENOENT) {
4819 spin_lock_irqsave(&ctx->completion_lock, flags);
4823 spin_lock_irqsave(&ctx->completion_lock, flags);
4824 ret = io_timeout_cancel(ctx, sqe_addr);
4827 ret = io_poll_cancel(ctx, sqe_addr);
4831 io_cqring_fill_event(req, ret);
4832 io_commit_cqring(ctx);
4833 spin_unlock_irqrestore(&ctx->completion_lock, flags);
4834 io_cqring_ev_posted(ctx);
4837 req_set_fail_links(req);
4841 static int io_async_cancel_prep(struct io_kiocb *req,
4842 const struct io_uring_sqe *sqe)
4844 if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
4846 if (sqe->flags || sqe->ioprio || sqe->off || sqe->len ||
4850 req->cancel.addr = READ_ONCE(sqe->addr);
4854 static int io_async_cancel(struct io_kiocb *req)
4856 struct io_ring_ctx *ctx = req->ctx;
4858 io_async_find_and_cancel(ctx, req, req->cancel.addr, 0);
4862 static int io_files_update_prep(struct io_kiocb *req,
4863 const struct io_uring_sqe *sqe)
4865 if (sqe->flags || sqe->ioprio || sqe->rw_flags)
4868 req->files_update.offset = READ_ONCE(sqe->off);
4869 req->files_update.nr_args = READ_ONCE(sqe->len);
4870 if (!req->files_update.nr_args)
4872 req->files_update.arg = READ_ONCE(sqe->addr);
4876 static int io_files_update(struct io_kiocb *req, bool force_nonblock)
4878 struct io_ring_ctx *ctx = req->ctx;
4879 struct io_uring_files_update up;
4885 up.offset = req->files_update.offset;
4886 up.fds = req->files_update.arg;
4888 mutex_lock(&ctx->uring_lock);
4889 ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args);
4890 mutex_unlock(&ctx->uring_lock);
4893 req_set_fail_links(req);
4894 io_cqring_add_event(req, ret);
4899 static int io_req_defer_prep(struct io_kiocb *req,
4900 const struct io_uring_sqe *sqe)
4907 if (io_op_defs[req->opcode].file_table) {
4908 ret = io_grab_files(req);
4913 io_req_work_grab_env(req, &io_op_defs[req->opcode]);
4915 switch (req->opcode) {
4918 case IORING_OP_READV:
4919 case IORING_OP_READ_FIXED:
4920 case IORING_OP_READ:
4921 ret = io_read_prep(req, sqe, true);
4923 case IORING_OP_WRITEV:
4924 case IORING_OP_WRITE_FIXED:
4925 case IORING_OP_WRITE:
4926 ret = io_write_prep(req, sqe, true);
4928 case IORING_OP_POLL_ADD:
4929 ret = io_poll_add_prep(req, sqe);
4931 case IORING_OP_POLL_REMOVE:
4932 ret = io_poll_remove_prep(req, sqe);
4934 case IORING_OP_FSYNC:
4935 ret = io_prep_fsync(req, sqe);
4937 case IORING_OP_SYNC_FILE_RANGE:
4938 ret = io_prep_sfr(req, sqe);
4940 case IORING_OP_SENDMSG:
4941 case IORING_OP_SEND:
4942 ret = io_sendmsg_prep(req, sqe);
4944 case IORING_OP_RECVMSG:
4945 case IORING_OP_RECV:
4946 ret = io_recvmsg_prep(req, sqe);
4948 case IORING_OP_CONNECT:
4949 ret = io_connect_prep(req, sqe);
4951 case IORING_OP_TIMEOUT:
4952 ret = io_timeout_prep(req, sqe, false);
4954 case IORING_OP_TIMEOUT_REMOVE:
4955 ret = io_timeout_remove_prep(req, sqe);
4957 case IORING_OP_ASYNC_CANCEL:
4958 ret = io_async_cancel_prep(req, sqe);
4960 case IORING_OP_LINK_TIMEOUT:
4961 ret = io_timeout_prep(req, sqe, true);
4963 case IORING_OP_ACCEPT:
4964 ret = io_accept_prep(req, sqe);
4966 case IORING_OP_FALLOCATE:
4967 ret = io_fallocate_prep(req, sqe);
4969 case IORING_OP_OPENAT:
4970 ret = io_openat_prep(req, sqe);
4972 case IORING_OP_CLOSE:
4973 ret = io_close_prep(req, sqe);
4975 case IORING_OP_FILES_UPDATE:
4976 ret = io_files_update_prep(req, sqe);
4978 case IORING_OP_STATX:
4979 ret = io_statx_prep(req, sqe);
4981 case IORING_OP_FADVISE:
4982 ret = io_fadvise_prep(req, sqe);
4984 case IORING_OP_MADVISE:
4985 ret = io_madvise_prep(req, sqe);
4987 case IORING_OP_OPENAT2:
4988 ret = io_openat2_prep(req, sqe);
4990 case IORING_OP_EPOLL_CTL:
4991 ret = io_epoll_ctl_prep(req, sqe);
4993 case IORING_OP_SPLICE:
4994 ret = io_splice_prep(req, sqe);
4996 case IORING_OP_PROVIDE_BUFFERS:
4997 ret = io_provide_buffers_prep(req, sqe);
4999 case IORING_OP_REMOVE_BUFFERS:
5000 ret = io_remove_buffers_prep(req, sqe);
5003 printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n",
5012 static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5014 struct io_ring_ctx *ctx = req->ctx;
5017 /* Still need defer if there is pending req in defer list. */
5018 if (!req_need_defer(req) && list_empty(&ctx->defer_list))
5021 if (!req->io && io_alloc_async_ctx(req))
5024 ret = io_req_defer_prep(req, sqe);
5028 spin_lock_irq(&ctx->completion_lock);
5029 if (!req_need_defer(req) && list_empty(&ctx->defer_list)) {
5030 spin_unlock_irq(&ctx->completion_lock);
5034 trace_io_uring_defer(ctx, req, req->user_data);
5035 list_add_tail(&req->list, &ctx->defer_list);
5036 spin_unlock_irq(&ctx->completion_lock);
5037 return -EIOCBQUEUED;
5040 static void io_cleanup_req(struct io_kiocb *req)
5042 struct io_async_ctx *io = req->io;
5044 switch (req->opcode) {
5045 case IORING_OP_READV:
5046 case IORING_OP_READ_FIXED:
5047 case IORING_OP_READ:
5048 if (req->flags & REQ_F_BUFFER_SELECTED)
5049 kfree((void *)(unsigned long)req->rw.addr);
5051 case IORING_OP_WRITEV:
5052 case IORING_OP_WRITE_FIXED:
5053 case IORING_OP_WRITE:
5054 if (io->rw.iov != io->rw.fast_iov)
5057 case IORING_OP_RECVMSG:
5058 if (req->flags & REQ_F_BUFFER_SELECTED)
5059 kfree(req->sr_msg.kbuf);
5061 case IORING_OP_SENDMSG:
5062 if (io->msg.iov != io->msg.fast_iov)
5065 case IORING_OP_RECV:
5066 if (req->flags & REQ_F_BUFFER_SELECTED)
5067 kfree(req->sr_msg.kbuf);
5069 case IORING_OP_OPENAT:
5070 case IORING_OP_OPENAT2:
5071 case IORING_OP_STATX:
5072 putname(req->open.filename);
5074 case IORING_OP_SPLICE:
5075 io_put_file(req, req->splice.file_in,
5076 (req->splice.flags & SPLICE_F_FD_IN_FIXED));
5080 req->flags &= ~REQ_F_NEED_CLEANUP;
5083 static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5084 bool force_nonblock)
5086 struct io_ring_ctx *ctx = req->ctx;
5089 switch (req->opcode) {
5093 case IORING_OP_READV:
5094 case IORING_OP_READ_FIXED:
5095 case IORING_OP_READ:
5097 ret = io_read_prep(req, sqe, force_nonblock);
5101 ret = io_read(req, force_nonblock);
5103 case IORING_OP_WRITEV:
5104 case IORING_OP_WRITE_FIXED:
5105 case IORING_OP_WRITE:
5107 ret = io_write_prep(req, sqe, force_nonblock);
5111 ret = io_write(req, force_nonblock);
5113 case IORING_OP_FSYNC:
5115 ret = io_prep_fsync(req, sqe);
5119 ret = io_fsync(req, force_nonblock);
5121 case IORING_OP_POLL_ADD:
5123 ret = io_poll_add_prep(req, sqe);
5127 ret = io_poll_add(req);
5129 case IORING_OP_POLL_REMOVE:
5131 ret = io_poll_remove_prep(req, sqe);
5135 ret = io_poll_remove(req);
5137 case IORING_OP_SYNC_FILE_RANGE:
5139 ret = io_prep_sfr(req, sqe);
5143 ret = io_sync_file_range(req, force_nonblock);
5145 case IORING_OP_SENDMSG:
5146 case IORING_OP_SEND:
5148 ret = io_sendmsg_prep(req, sqe);
5152 if (req->opcode == IORING_OP_SENDMSG)
5153 ret = io_sendmsg(req, force_nonblock);
5155 ret = io_send(req, force_nonblock);
5157 case IORING_OP_RECVMSG:
5158 case IORING_OP_RECV:
5160 ret = io_recvmsg_prep(req, sqe);
5164 if (req->opcode == IORING_OP_RECVMSG)
5165 ret = io_recvmsg(req, force_nonblock);
5167 ret = io_recv(req, force_nonblock);
5169 case IORING_OP_TIMEOUT:
5171 ret = io_timeout_prep(req, sqe, false);
5175 ret = io_timeout(req);
5177 case IORING_OP_TIMEOUT_REMOVE:
5179 ret = io_timeout_remove_prep(req, sqe);
5183 ret = io_timeout_remove(req);
5185 case IORING_OP_ACCEPT:
5187 ret = io_accept_prep(req, sqe);
5191 ret = io_accept(req, force_nonblock);
5193 case IORING_OP_CONNECT:
5195 ret = io_connect_prep(req, sqe);
5199 ret = io_connect(req, force_nonblock);
5201 case IORING_OP_ASYNC_CANCEL:
5203 ret = io_async_cancel_prep(req, sqe);
5207 ret = io_async_cancel(req);
5209 case IORING_OP_FALLOCATE:
5211 ret = io_fallocate_prep(req, sqe);
5215 ret = io_fallocate(req, force_nonblock);
5217 case IORING_OP_OPENAT:
5219 ret = io_openat_prep(req, sqe);
5223 ret = io_openat(req, force_nonblock);
5225 case IORING_OP_CLOSE:
5227 ret = io_close_prep(req, sqe);
5231 ret = io_close(req, force_nonblock);
5233 case IORING_OP_FILES_UPDATE:
5235 ret = io_files_update_prep(req, sqe);
5239 ret = io_files_update(req, force_nonblock);
5241 case IORING_OP_STATX:
5243 ret = io_statx_prep(req, sqe);
5247 ret = io_statx(req, force_nonblock);
5249 case IORING_OP_FADVISE:
5251 ret = io_fadvise_prep(req, sqe);
5255 ret = io_fadvise(req, force_nonblock);
5257 case IORING_OP_MADVISE:
5259 ret = io_madvise_prep(req, sqe);
5263 ret = io_madvise(req, force_nonblock);
5265 case IORING_OP_OPENAT2:
5267 ret = io_openat2_prep(req, sqe);
5271 ret = io_openat2(req, force_nonblock);
5273 case IORING_OP_EPOLL_CTL:
5275 ret = io_epoll_ctl_prep(req, sqe);
5279 ret = io_epoll_ctl(req, force_nonblock);
5281 case IORING_OP_SPLICE:
5283 ret = io_splice_prep(req, sqe);
5287 ret = io_splice(req, force_nonblock);
5289 case IORING_OP_PROVIDE_BUFFERS:
5291 ret = io_provide_buffers_prep(req, sqe);
5295 ret = io_provide_buffers(req, force_nonblock);
5297 case IORING_OP_REMOVE_BUFFERS:
5299 ret = io_remove_buffers_prep(req, sqe);
5303 ret = io_remove_buffers(req, force_nonblock);
5313 if (ctx->flags & IORING_SETUP_IOPOLL) {
5314 const bool in_async = io_wq_current_is_worker();
5316 if (req->result == -EAGAIN)
5319 /* workqueue context doesn't hold uring_lock, grab it now */
5321 mutex_lock(&ctx->uring_lock);
5323 io_iopoll_req_issued(req);
5326 mutex_unlock(&ctx->uring_lock);
5332 static void io_wq_submit_work(struct io_wq_work **workptr)
5334 struct io_wq_work *work = *workptr;
5335 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
5338 /* if NO_CANCEL is set, we must still run the work */
5339 if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) ==
5340 IO_WQ_WORK_CANCEL) {
5346 ret = io_issue_sqe(req, NULL, false);
5348 * We can get EAGAIN for polled IO even though we're
5349 * forcing a sync submission from here, since we can't
5350 * wait for request slots on the block side.
5359 req_set_fail_links(req);
5360 io_cqring_add_event(req, ret);
5364 io_steal_work(req, workptr);
5367 static int io_req_needs_file(struct io_kiocb *req, int fd)
5369 if (!io_op_defs[req->opcode].needs_file)
5371 if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg)
5376 static inline struct file *io_file_from_index(struct io_ring_ctx *ctx,
5379 struct fixed_file_table *table;
5381 table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT];
5382 return table->files[index & IORING_FILE_TABLE_MASK];;
5385 static int io_file_get(struct io_submit_state *state, struct io_kiocb *req,
5386 int fd, struct file **out_file, bool fixed)
5388 struct io_ring_ctx *ctx = req->ctx;
5392 if (unlikely(!ctx->file_data ||
5393 (unsigned) fd >= ctx->nr_user_files))
5395 fd = array_index_nospec(fd, ctx->nr_user_files);
5396 file = io_file_from_index(ctx, fd);
5399 req->fixed_file_refs = ctx->file_data->cur_refs;
5400 percpu_ref_get(req->fixed_file_refs);
5402 trace_io_uring_file_get(ctx, fd);
5403 file = __io_file_get(state, fd);
5404 if (unlikely(!file))
5412 static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req,
5413 int fd, unsigned int flags)
5417 if (!io_req_needs_file(req, fd))
5420 fixed = (flags & IOSQE_FIXED_FILE);
5421 if (unlikely(!fixed && req->needs_fixed_file))
5424 return io_file_get(state, req, fd, &req->file, fixed);
5427 static int io_grab_files(struct io_kiocb *req)
5430 struct io_ring_ctx *ctx = req->ctx;
5432 if (req->work.files)
5434 if (!ctx->ring_file)
5438 spin_lock_irq(&ctx->inflight_lock);
5440 * We use the f_ops->flush() handler to ensure that we can flush
5441 * out work accessing these files if the fd is closed. Check if
5442 * the fd has changed since we started down this path, and disallow
5443 * this operation if it has.
5445 if (fcheck(ctx->ring_fd) == ctx->ring_file) {
5446 list_add(&req->inflight_entry, &ctx->inflight_list);
5447 req->flags |= REQ_F_INFLIGHT;
5448 req->work.files = current->files;
5451 spin_unlock_irq(&ctx->inflight_lock);
5457 static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
5459 struct io_timeout_data *data = container_of(timer,
5460 struct io_timeout_data, timer);
5461 struct io_kiocb *req = data->req;
5462 struct io_ring_ctx *ctx = req->ctx;
5463 struct io_kiocb *prev = NULL;
5464 unsigned long flags;
5466 spin_lock_irqsave(&ctx->completion_lock, flags);
5469 * We don't expect the list to be empty, that will only happen if we
5470 * race with the completion of the linked work.
5472 if (!list_empty(&req->link_list)) {
5473 prev = list_entry(req->link_list.prev, struct io_kiocb,
5475 if (refcount_inc_not_zero(&prev->refs)) {
5476 list_del_init(&req->link_list);
5477 prev->flags &= ~REQ_F_LINK_TIMEOUT;
5482 spin_unlock_irqrestore(&ctx->completion_lock, flags);
5485 req_set_fail_links(prev);
5486 io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME);
5489 io_cqring_add_event(req, -ETIME);
5492 return HRTIMER_NORESTART;
5495 static void io_queue_linked_timeout(struct io_kiocb *req)
5497 struct io_ring_ctx *ctx = req->ctx;
5500 * If the list is now empty, then our linked request finished before
5501 * we got a chance to setup the timer
5503 spin_lock_irq(&ctx->completion_lock);
5504 if (!list_empty(&req->link_list)) {
5505 struct io_timeout_data *data = &req->io->timeout;
5507 data->timer.function = io_link_timeout_fn;
5508 hrtimer_start(&data->timer, timespec64_to_ktime(data->ts),
5511 spin_unlock_irq(&ctx->completion_lock);
5513 /* drop submission reference */
5517 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
5519 struct io_kiocb *nxt;
5521 if (!(req->flags & REQ_F_LINK_HEAD))
5523 /* for polled retry, if flag is set, we already went through here */
5524 if (req->flags & REQ_F_POLLED)
5527 nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
5529 if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
5532 req->flags |= REQ_F_LINK_TIMEOUT;
5536 static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5538 struct io_kiocb *linked_timeout;
5539 struct io_kiocb *nxt;
5540 const struct cred *old_creds = NULL;
5544 linked_timeout = io_prep_linked_timeout(req);
5546 if (req->work.creds && req->work.creds != current_cred()) {
5548 revert_creds(old_creds);
5549 if (old_creds == req->work.creds)
5550 old_creds = NULL; /* restored original creds */
5552 old_creds = override_creds(req->work.creds);
5555 ret = io_issue_sqe(req, sqe, true);
5558 * We async punt it if the file wasn't marked NOWAIT, or if the file
5559 * doesn't support non-blocking read/write attempts
5561 if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) ||
5562 (req->flags & REQ_F_MUST_PUNT))) {
5563 if (io_arm_poll_handler(req)) {
5565 io_queue_linked_timeout(linked_timeout);
5569 if (io_op_defs[req->opcode].file_table) {
5570 ret = io_grab_files(req);
5576 * Queued up for async execution, worker will release
5577 * submit reference when the iocb is actually submitted.
5579 io_queue_async_work(req);
5585 /* drop submission reference */
5586 io_put_req_find_next(req, &nxt);
5588 if (linked_timeout) {
5590 io_queue_linked_timeout(linked_timeout);
5592 io_put_req(linked_timeout);
5595 /* and drop final reference, if we failed */
5597 io_cqring_add_event(req, ret);
5598 req_set_fail_links(req);
5604 if (req->flags & REQ_F_FORCE_ASYNC)
5610 revert_creds(old_creds);
5613 static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe)
5617 ret = io_req_defer(req, sqe);
5619 if (ret != -EIOCBQUEUED) {
5621 io_cqring_add_event(req, ret);
5622 req_set_fail_links(req);
5623 io_double_put_req(req);
5625 } else if (req->flags & REQ_F_FORCE_ASYNC) {
5626 ret = io_req_defer_prep(req, sqe);
5627 if (unlikely(ret < 0))
5630 * Never try inline submit of IOSQE_ASYNC is set, go straight
5631 * to async execution.
5633 req->work.flags |= IO_WQ_WORK_CONCURRENT;
5634 io_queue_async_work(req);
5636 __io_queue_sqe(req, sqe);
5640 static inline void io_queue_link_head(struct io_kiocb *req)
5642 if (unlikely(req->flags & REQ_F_FAIL_LINK)) {
5643 io_cqring_add_event(req, -ECANCELED);
5644 io_double_put_req(req);
5646 io_queue_sqe(req, NULL);
5649 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
5650 struct io_submit_state *state, struct io_kiocb **link)
5652 struct io_ring_ctx *ctx = req->ctx;
5656 * If we already have a head request, queue this one for async
5657 * submittal once the head completes. If we don't have a head but
5658 * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be
5659 * submitted sync once the chain is complete. If none of those
5660 * conditions are true (normal request), then just queue it.
5663 struct io_kiocb *head = *link;
5666 * Taking sequential execution of a link, draining both sides
5667 * of the link also fullfils IOSQE_IO_DRAIN semantics for all
5668 * requests in the link. So, it drains the head and the
5669 * next after the link request. The last one is done via
5670 * drain_next flag to persist the effect across calls.
5672 if (req->flags & REQ_F_IO_DRAIN) {
5673 head->flags |= REQ_F_IO_DRAIN;
5674 ctx->drain_next = 1;
5676 if (io_alloc_async_ctx(req))
5679 ret = io_req_defer_prep(req, sqe);
5681 /* fail even hard links since we don't submit */
5682 head->flags |= REQ_F_FAIL_LINK;
5685 trace_io_uring_link(ctx, req, head);
5686 list_add_tail(&req->link_list, &head->link_list);
5688 /* last request of a link, enqueue the link */
5689 if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
5690 io_queue_link_head(head);
5694 if (unlikely(ctx->drain_next)) {
5695 req->flags |= REQ_F_IO_DRAIN;
5696 ctx->drain_next = 0;
5698 if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
5699 req->flags |= REQ_F_LINK_HEAD;
5700 INIT_LIST_HEAD(&req->link_list);
5702 if (io_alloc_async_ctx(req))
5705 ret = io_req_defer_prep(req, sqe);
5707 req->flags |= REQ_F_FAIL_LINK;
5710 io_queue_sqe(req, sqe);
5718 * Batched submission is done, ensure local IO is flushed out.
5720 static void io_submit_state_end(struct io_submit_state *state)
5722 blk_finish_plug(&state->plug);
5724 if (state->free_reqs)
5725 kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs);
5729 * Start submission side cache.
5731 static void io_submit_state_start(struct io_submit_state *state,
5732 unsigned int max_ios)
5734 blk_start_plug(&state->plug);
5735 state->free_reqs = 0;
5737 state->ios_left = max_ios;
5740 static void io_commit_sqring(struct io_ring_ctx *ctx)
5742 struct io_rings *rings = ctx->rings;
5745 * Ensure any loads from the SQEs are done at this point,
5746 * since once we write the new head, the application could
5747 * write new data to them.
5749 smp_store_release(&rings->sq.head, ctx->cached_sq_head);
5753 * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory
5754 * that is mapped by userspace. This means that care needs to be taken to
5755 * ensure that reads are stable, as we cannot rely on userspace always
5756 * being a good citizen. If members of the sqe are validated and then later
5757 * used, it's important that those reads are done through READ_ONCE() to
5758 * prevent a re-load down the line.
5760 static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx)
5762 u32 *sq_array = ctx->sq_array;
5766 * The cached sq head (or cq tail) serves two purposes:
5768 * 1) allows us to batch the cost of updating the user visible
5770 * 2) allows the kernel side to track the head on its own, even
5771 * though the application is the one updating it.
5773 head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
5774 if (likely(head < ctx->sq_entries))
5775 return &ctx->sq_sqes[head];
5777 /* drop invalid entries */
5778 ctx->cached_sq_dropped++;
5779 WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped);
5783 static inline void io_consume_sqe(struct io_ring_ctx *ctx)
5785 ctx->cached_sq_head++;
5788 #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \
5789 IOSQE_IO_HARDLINK | IOSQE_ASYNC | \
5790 IOSQE_BUFFER_SELECT)
5792 static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
5793 const struct io_uring_sqe *sqe,
5794 struct io_submit_state *state, bool async)
5796 unsigned int sqe_flags;
5800 * All io need record the previous position, if LINK vs DARIN,
5801 * it can be used to mark the position of the first IO in the
5804 req->sequence = ctx->cached_sq_head - ctx->cached_sq_dropped;
5805 req->opcode = READ_ONCE(sqe->opcode);
5806 req->user_data = READ_ONCE(sqe->user_data);
5811 /* one is dropped after submission, the other at completion */
5812 refcount_set(&req->refs, 2);
5815 req->needs_fixed_file = async;
5816 INIT_IO_WORK(&req->work, io_wq_submit_work);
5818 if (unlikely(req->opcode >= IORING_OP_LAST))
5821 if (io_op_defs[req->opcode].needs_mm && !current->mm) {
5822 if (unlikely(!mmget_not_zero(ctx->sqo_mm)))
5824 use_mm(ctx->sqo_mm);
5827 sqe_flags = READ_ONCE(sqe->flags);
5828 /* enforce forwards compatibility on users */
5829 if (unlikely(sqe_flags & ~SQE_VALID_FLAGS))
5832 if ((sqe_flags & IOSQE_BUFFER_SELECT) &&
5833 !io_op_defs[req->opcode].buffer_select)
5836 id = READ_ONCE(sqe->personality);
5838 req->work.creds = idr_find(&ctx->personality_idr, id);
5839 if (unlikely(!req->work.creds))
5841 get_cred(req->work.creds);
5844 /* same numerical values with corresponding REQ_F_*, safe to copy */
5845 req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK |
5846 IOSQE_ASYNC | IOSQE_FIXED_FILE |
5847 IOSQE_BUFFER_SELECT | IOSQE_IO_LINK);
5849 fd = READ_ONCE(sqe->fd);
5850 return io_req_set_file(state, req, fd, sqe_flags);
5853 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
5854 struct file *ring_file, int ring_fd, bool async)
5856 struct io_submit_state state, *statep = NULL;
5857 struct io_kiocb *link = NULL;
5858 int i, submitted = 0;
5860 /* if we have a backlog and couldn't flush it all, return BUSY */
5861 if (test_bit(0, &ctx->sq_check_overflow)) {
5862 if (!list_empty(&ctx->cq_overflow_list) &&
5863 !io_cqring_overflow_flush(ctx, false))
5867 /* make sure SQ entry isn't read before tail */
5868 nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx));
5870 if (!percpu_ref_tryget_many(&ctx->refs, nr))
5873 if (nr > IO_PLUG_THRESHOLD) {
5874 io_submit_state_start(&state, nr);
5878 ctx->ring_fd = ring_fd;
5879 ctx->ring_file = ring_file;
5881 for (i = 0; i < nr; i++) {
5882 const struct io_uring_sqe *sqe;
5883 struct io_kiocb *req;
5886 sqe = io_get_sqe(ctx);
5887 if (unlikely(!sqe)) {
5888 io_consume_sqe(ctx);
5891 req = io_alloc_req(ctx, statep);
5892 if (unlikely(!req)) {
5894 submitted = -EAGAIN;
5898 err = io_init_req(ctx, req, sqe, statep, async);
5899 io_consume_sqe(ctx);
5900 /* will complete beyond this point, count as submitted */
5903 if (unlikely(err)) {
5905 io_cqring_add_event(req, err);
5906 io_double_put_req(req);
5910 trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data,
5912 err = io_submit_sqe(req, sqe, statep, &link);
5917 if (unlikely(submitted != nr)) {
5918 int ref_used = (submitted == -EAGAIN) ? 0 : submitted;
5920 percpu_ref_put_many(&ctx->refs, nr - ref_used);
5923 io_queue_link_head(link);
5925 io_submit_state_end(&state);
5927 /* Commit SQ ring head once we've consumed and submitted all SQEs */
5928 io_commit_sqring(ctx);
5933 static inline void io_sq_thread_drop_mm(struct io_ring_ctx *ctx)
5935 struct mm_struct *mm = current->mm;
5943 static int io_sq_thread(void *data)
5945 struct io_ring_ctx *ctx = data;
5946 const struct cred *old_cred;
5947 mm_segment_t old_fs;
5949 unsigned long timeout;
5952 complete(&ctx->completions[1]);
5956 old_cred = override_creds(ctx->creds);
5958 timeout = jiffies + ctx->sq_thread_idle;
5959 while (!kthread_should_park()) {
5960 unsigned int to_submit;
5962 if (!list_empty(&ctx->poll_list)) {
5963 unsigned nr_events = 0;
5965 mutex_lock(&ctx->uring_lock);
5966 if (!list_empty(&ctx->poll_list))
5967 io_iopoll_getevents(ctx, &nr_events, 0);
5969 timeout = jiffies + ctx->sq_thread_idle;
5970 mutex_unlock(&ctx->uring_lock);
5973 to_submit = io_sqring_entries(ctx);
5976 * If submit got -EBUSY, flag us as needing the application
5977 * to enter the kernel to reap and flush events.
5979 if (!to_submit || ret == -EBUSY) {
5981 * Drop cur_mm before scheduling, we can't hold it for
5982 * long periods (or over schedule()). Do this before
5983 * adding ourselves to the waitqueue, as the unuse/drop
5986 io_sq_thread_drop_mm(ctx);
5989 * We're polling. If we're within the defined idle
5990 * period, then let us spin without work before going
5991 * to sleep. The exception is if we got EBUSY doing
5992 * more IO, we should wait for the application to
5993 * reap events and wake us up.
5995 if (!list_empty(&ctx->poll_list) ||
5996 (!time_after(jiffies, timeout) && ret != -EBUSY &&
5997 !percpu_ref_is_dying(&ctx->refs))) {
5998 if (current->task_works)
6004 prepare_to_wait(&ctx->sqo_wait, &wait,
6005 TASK_INTERRUPTIBLE);
6008 * While doing polled IO, before going to sleep, we need
6009 * to check if there are new reqs added to poll_list, it
6010 * is because reqs may have been punted to io worker and
6011 * will be added to poll_list later, hence check the
6014 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
6015 !list_empty_careful(&ctx->poll_list)) {
6016 finish_wait(&ctx->sqo_wait, &wait);
6020 /* Tell userspace we may need a wakeup call */
6021 ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP;
6022 /* make sure to read SQ tail after writing flags */
6025 to_submit = io_sqring_entries(ctx);
6026 if (!to_submit || ret == -EBUSY) {
6027 if (kthread_should_park()) {
6028 finish_wait(&ctx->sqo_wait, &wait);
6031 if (current->task_works) {
6033 finish_wait(&ctx->sqo_wait, &wait);
6036 if (signal_pending(current))
6037 flush_signals(current);
6039 finish_wait(&ctx->sqo_wait, &wait);
6041 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6044 finish_wait(&ctx->sqo_wait, &wait);
6046 ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP;
6049 mutex_lock(&ctx->uring_lock);
6050 ret = io_submit_sqes(ctx, to_submit, NULL, -1, true);
6051 mutex_unlock(&ctx->uring_lock);
6052 timeout = jiffies + ctx->sq_thread_idle;
6055 if (current->task_works)
6059 io_sq_thread_drop_mm(ctx);
6060 revert_creds(old_cred);
6067 struct io_wait_queue {
6068 struct wait_queue_entry wq;
6069 struct io_ring_ctx *ctx;
6071 unsigned nr_timeouts;
6074 static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush)
6076 struct io_ring_ctx *ctx = iowq->ctx;
6079 * Wake up if we have enough events, or if a timeout occurred since we
6080 * started waiting. For timeouts, we always want to return to userspace,
6081 * regardless of event count.
6083 return io_cqring_events(ctx, noflush) >= iowq->to_wait ||
6084 atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts;
6087 static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode,
6088 int wake_flags, void *key)
6090 struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue,
6093 /* use noflush == true, as we can't safely rely on locking context */
6094 if (!io_should_wake(iowq, true))
6097 return autoremove_wake_function(curr, mode, wake_flags, key);
6101 * Wait until events become available, if we don't already have some. The
6102 * application must reap them itself, as they reside on the shared cq ring.
6104 static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
6105 const sigset_t __user *sig, size_t sigsz)
6107 struct io_wait_queue iowq = {
6110 .func = io_wake_function,
6111 .entry = LIST_HEAD_INIT(iowq.wq.entry),
6114 .to_wait = min_events,
6116 struct io_rings *rings = ctx->rings;
6120 if (io_cqring_events(ctx, false) >= min_events)
6122 if (!current->task_works)
6128 #ifdef CONFIG_COMPAT
6129 if (in_compat_syscall())
6130 ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig,
6134 ret = set_user_sigmask(sig, sigsz);
6140 iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts);
6141 trace_io_uring_cqring_wait(ctx, min_events);
6143 prepare_to_wait_exclusive(&ctx->wait, &iowq.wq,
6144 TASK_INTERRUPTIBLE);
6145 if (current->task_works)
6147 if (io_should_wake(&iowq, false))
6150 if (signal_pending(current)) {
6155 finish_wait(&ctx->wait, &iowq.wq);
6157 restore_saved_sigmask_unless(ret == -EINTR);
6159 return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0;
6162 static void __io_sqe_files_unregister(struct io_ring_ctx *ctx)
6164 #if defined(CONFIG_UNIX)
6165 if (ctx->ring_sock) {
6166 struct sock *sock = ctx->ring_sock->sk;
6167 struct sk_buff *skb;
6169 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL)
6175 for (i = 0; i < ctx->nr_user_files; i++) {
6178 file = io_file_from_index(ctx, i);
6185 static void io_file_ref_kill(struct percpu_ref *ref)
6187 struct fixed_file_data *data;
6189 data = container_of(ref, struct fixed_file_data, refs);
6190 complete(&data->done);
6193 static int io_sqe_files_unregister(struct io_ring_ctx *ctx)
6195 struct fixed_file_data *data = ctx->file_data;
6196 struct fixed_file_ref_node *ref_node = NULL;
6197 unsigned nr_tables, i;
6198 unsigned long flags;
6203 spin_lock_irqsave(&data->lock, flags);
6204 if (!list_empty(&data->ref_list))
6205 ref_node = list_first_entry(&data->ref_list,
6206 struct fixed_file_ref_node, node);
6207 spin_unlock_irqrestore(&data->lock, flags);
6209 percpu_ref_kill(&ref_node->refs);
6211 percpu_ref_kill(&data->refs);
6213 /* wait for all refs nodes to complete */
6214 wait_for_completion(&data->done);
6216 __io_sqe_files_unregister(ctx);
6217 nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE);
6218 for (i = 0; i < nr_tables; i++)
6219 kfree(data->table[i].files);
6221 percpu_ref_exit(&data->refs);
6223 ctx->file_data = NULL;
6224 ctx->nr_user_files = 0;
6228 static void io_sq_thread_stop(struct io_ring_ctx *ctx)
6230 if (ctx->sqo_thread) {
6231 wait_for_completion(&ctx->completions[1]);
6233 * The park is a bit of a work-around, without it we get
6234 * warning spews on shutdown with SQPOLL set and affinity
6235 * set to a single CPU.
6237 kthread_park(ctx->sqo_thread);
6238 kthread_stop(ctx->sqo_thread);
6239 ctx->sqo_thread = NULL;
6243 static void io_finish_async(struct io_ring_ctx *ctx)
6245 io_sq_thread_stop(ctx);
6248 io_wq_destroy(ctx->io_wq);
6253 #if defined(CONFIG_UNIX)
6255 * Ensure the UNIX gc is aware of our file set, so we are certain that
6256 * the io_uring can be safely unregistered on process exit, even if we have
6257 * loops in the file referencing.
6259 static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset)
6261 struct sock *sk = ctx->ring_sock->sk;
6262 struct scm_fp_list *fpl;
6263 struct sk_buff *skb;
6266 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL);
6270 skb = alloc_skb(0, GFP_KERNEL);
6279 fpl->user = get_uid(ctx->user);
6280 for (i = 0; i < nr; i++) {
6281 struct file *file = io_file_from_index(ctx, i + offset);
6285 fpl->fp[nr_files] = get_file(file);
6286 unix_inflight(fpl->user, fpl->fp[nr_files]);
6291 fpl->max = SCM_MAX_FD;
6292 fpl->count = nr_files;
6293 UNIXCB(skb).fp = fpl;
6294 skb->destructor = unix_destruct_scm;
6295 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
6296 skb_queue_head(&sk->sk_receive_queue, skb);
6298 for (i = 0; i < nr_files; i++)
6309 * If UNIX sockets are enabled, fd passing can cause a reference cycle which
6310 * causes regular reference counting to break down. We rely on the UNIX
6311 * garbage collection to take care of this problem for us.
6313 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6315 unsigned left, total;
6319 left = ctx->nr_user_files;
6321 unsigned this_files = min_t(unsigned, left, SCM_MAX_FD);
6323 ret = __io_sqe_files_scm(ctx, this_files, total);
6327 total += this_files;
6333 while (total < ctx->nr_user_files) {
6334 struct file *file = io_file_from_index(ctx, total);
6344 static int io_sqe_files_scm(struct io_ring_ctx *ctx)
6350 static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables,
6355 for (i = 0; i < nr_tables; i++) {
6356 struct fixed_file_table *table = &ctx->file_data->table[i];
6357 unsigned this_files;
6359 this_files = min(nr_files, IORING_MAX_FILES_TABLE);
6360 table->files = kcalloc(this_files, sizeof(struct file *),
6364 nr_files -= this_files;
6370 for (i = 0; i < nr_tables; i++) {
6371 struct fixed_file_table *table = &ctx->file_data->table[i];
6372 kfree(table->files);
6377 static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file)
6379 #if defined(CONFIG_UNIX)
6380 struct sock *sock = ctx->ring_sock->sk;
6381 struct sk_buff_head list, *head = &sock->sk_receive_queue;
6382 struct sk_buff *skb;
6385 __skb_queue_head_init(&list);
6388 * Find the skb that holds this file in its SCM_RIGHTS. When found,
6389 * remove this entry and rearrange the file array.
6391 skb = skb_dequeue(head);
6393 struct scm_fp_list *fp;
6395 fp = UNIXCB(skb).fp;
6396 for (i = 0; i < fp->count; i++) {
6399 if (fp->fp[i] != file)
6402 unix_notinflight(fp->user, fp->fp[i]);
6403 left = fp->count - 1 - i;
6405 memmove(&fp->fp[i], &fp->fp[i + 1],
6406 left * sizeof(struct file *));
6413 __skb_queue_tail(&list, skb);
6423 __skb_queue_tail(&list, skb);
6425 skb = skb_dequeue(head);
6428 if (skb_peek(&list)) {
6429 spin_lock_irq(&head->lock);
6430 while ((skb = __skb_dequeue(&list)) != NULL)
6431 __skb_queue_tail(head, skb);
6432 spin_unlock_irq(&head->lock);
6439 struct io_file_put {
6440 struct list_head list;
6444 static void io_file_put_work(struct work_struct *work)
6446 struct fixed_file_ref_node *ref_node;
6447 struct fixed_file_data *file_data;
6448 struct io_ring_ctx *ctx;
6449 struct io_file_put *pfile, *tmp;
6450 unsigned long flags;
6452 ref_node = container_of(work, struct fixed_file_ref_node, work);
6453 file_data = ref_node->file_data;
6454 ctx = file_data->ctx;
6456 list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) {
6457 list_del_init(&pfile->list);
6458 io_ring_file_put(ctx, pfile->file);
6462 spin_lock_irqsave(&file_data->lock, flags);
6463 list_del_init(&ref_node->node);
6464 spin_unlock_irqrestore(&file_data->lock, flags);
6466 percpu_ref_exit(&ref_node->refs);
6468 percpu_ref_put(&file_data->refs);
6471 static void io_file_data_ref_zero(struct percpu_ref *ref)
6473 struct fixed_file_ref_node *ref_node;
6475 ref_node = container_of(ref, struct fixed_file_ref_node, refs);
6477 queue_work(system_wq, &ref_node->work);
6480 static struct fixed_file_ref_node *alloc_fixed_file_ref_node(
6481 struct io_ring_ctx *ctx)
6483 struct fixed_file_ref_node *ref_node;
6485 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL);
6487 return ERR_PTR(-ENOMEM);
6489 if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero,
6492 return ERR_PTR(-ENOMEM);
6494 INIT_LIST_HEAD(&ref_node->node);
6495 INIT_LIST_HEAD(&ref_node->file_list);
6496 INIT_WORK(&ref_node->work, io_file_put_work);
6497 ref_node->file_data = ctx->file_data;
6502 static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node)
6504 percpu_ref_exit(&ref_node->refs);
6508 static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg,
6511 __s32 __user *fds = (__s32 __user *) arg;
6516 struct fixed_file_ref_node *ref_node;
6517 unsigned long flags;
6523 if (nr_args > IORING_MAX_FIXED_FILES)
6526 ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL);
6527 if (!ctx->file_data)
6529 ctx->file_data->ctx = ctx;
6530 init_completion(&ctx->file_data->done);
6531 INIT_LIST_HEAD(&ctx->file_data->ref_list);
6532 spin_lock_init(&ctx->file_data->lock);
6534 nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE);
6535 ctx->file_data->table = kcalloc(nr_tables,
6536 sizeof(struct fixed_file_table),
6538 if (!ctx->file_data->table) {
6539 kfree(ctx->file_data);
6540 ctx->file_data = NULL;
6544 if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill,
6545 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) {
6546 kfree(ctx->file_data->table);
6547 kfree(ctx->file_data);
6548 ctx->file_data = NULL;
6552 if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) {
6553 percpu_ref_exit(&ctx->file_data->refs);
6554 kfree(ctx->file_data->table);
6555 kfree(ctx->file_data);
6556 ctx->file_data = NULL;
6560 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) {
6561 struct fixed_file_table *table;
6565 if (copy_from_user(&fd, &fds[i], sizeof(fd)))
6567 /* allow sparse sets */
6573 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6574 index = i & IORING_FILE_TABLE_MASK;
6582 * Don't allow io_uring instances to be registered. If UNIX
6583 * isn't enabled, then this causes a reference cycle and this
6584 * instance can never get freed. If UNIX is enabled we'll
6585 * handle it just fine, but there's still no point in allowing
6586 * a ring fd as it doesn't support regular read/write anyway.
6588 if (file->f_op == &io_uring_fops) {
6593 table->files[index] = file;
6597 for (i = 0; i < ctx->nr_user_files; i++) {
6598 file = io_file_from_index(ctx, i);
6602 for (i = 0; i < nr_tables; i++)
6603 kfree(ctx->file_data->table[i].files);
6605 kfree(ctx->file_data->table);
6606 kfree(ctx->file_data);
6607 ctx->file_data = NULL;
6608 ctx->nr_user_files = 0;
6612 ret = io_sqe_files_scm(ctx);
6614 io_sqe_files_unregister(ctx);
6618 ref_node = alloc_fixed_file_ref_node(ctx);
6619 if (IS_ERR(ref_node)) {
6620 io_sqe_files_unregister(ctx);
6621 return PTR_ERR(ref_node);
6624 ctx->file_data->cur_refs = &ref_node->refs;
6625 spin_lock_irqsave(&ctx->file_data->lock, flags);
6626 list_add(&ref_node->node, &ctx->file_data->ref_list);
6627 spin_unlock_irqrestore(&ctx->file_data->lock, flags);
6628 percpu_ref_get(&ctx->file_data->refs);
6632 static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file,
6635 #if defined(CONFIG_UNIX)
6636 struct sock *sock = ctx->ring_sock->sk;
6637 struct sk_buff_head *head = &sock->sk_receive_queue;
6638 struct sk_buff *skb;
6641 * See if we can merge this file into an existing skb SCM_RIGHTS
6642 * file set. If there's no room, fall back to allocating a new skb
6643 * and filling it in.
6645 spin_lock_irq(&head->lock);
6646 skb = skb_peek(head);
6648 struct scm_fp_list *fpl = UNIXCB(skb).fp;
6650 if (fpl->count < SCM_MAX_FD) {
6651 __skb_unlink(skb, head);
6652 spin_unlock_irq(&head->lock);
6653 fpl->fp[fpl->count] = get_file(file);
6654 unix_inflight(fpl->user, fpl->fp[fpl->count]);
6656 spin_lock_irq(&head->lock);
6657 __skb_queue_head(head, skb);
6662 spin_unlock_irq(&head->lock);
6669 return __io_sqe_files_scm(ctx, 1, index);
6675 static int io_queue_file_removal(struct fixed_file_data *data,
6678 struct io_file_put *pfile;
6679 struct percpu_ref *refs = data->cur_refs;
6680 struct fixed_file_ref_node *ref_node;
6682 pfile = kzalloc(sizeof(*pfile), GFP_KERNEL);
6686 ref_node = container_of(refs, struct fixed_file_ref_node, refs);
6688 list_add(&pfile->list, &ref_node->file_list);
6693 static int __io_sqe_files_update(struct io_ring_ctx *ctx,
6694 struct io_uring_files_update *up,
6697 struct fixed_file_data *data = ctx->file_data;
6698 struct fixed_file_ref_node *ref_node;
6703 unsigned long flags;
6704 bool needs_switch = false;
6706 if (check_add_overflow(up->offset, nr_args, &done))
6708 if (done > ctx->nr_user_files)
6711 ref_node = alloc_fixed_file_ref_node(ctx);
6712 if (IS_ERR(ref_node))
6713 return PTR_ERR(ref_node);
6716 fds = u64_to_user_ptr(up->fds);
6718 struct fixed_file_table *table;
6722 if (copy_from_user(&fd, &fds[done], sizeof(fd))) {
6726 i = array_index_nospec(up->offset, ctx->nr_user_files);
6727 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
6728 index = i & IORING_FILE_TABLE_MASK;
6729 if (table->files[index]) {
6730 file = io_file_from_index(ctx, index);
6731 err = io_queue_file_removal(data, file);
6734 table->files[index] = NULL;
6735 needs_switch = true;
6744 * Don't allow io_uring instances to be registered. If
6745 * UNIX isn't enabled, then this causes a reference
6746 * cycle and this instance can never get freed. If UNIX
6747 * is enabled we'll handle it just fine, but there's
6748 * still no point in allowing a ring fd as it doesn't
6749 * support regular read/write anyway.
6751 if (file->f_op == &io_uring_fops) {
6756 table->files[index] = file;
6757 err = io_sqe_file_register(ctx, file, i);
6767 percpu_ref_kill(data->cur_refs);
6768 spin_lock_irqsave(&data->lock, flags);
6769 list_add(&ref_node->node, &data->ref_list);
6770 data->cur_refs = &ref_node->refs;
6771 spin_unlock_irqrestore(&data->lock, flags);
6772 percpu_ref_get(&ctx->file_data->refs);
6774 destroy_fixed_file_ref_node(ref_node);
6776 return done ? done : err;
6779 static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg,
6782 struct io_uring_files_update up;
6784 if (!ctx->file_data)
6788 if (copy_from_user(&up, arg, sizeof(up)))
6793 return __io_sqe_files_update(ctx, &up, nr_args);
6796 static void io_free_work(struct io_wq_work *work)
6798 struct io_kiocb *req = container_of(work, struct io_kiocb, work);
6800 /* Consider that io_steal_work() relies on this ref */
6804 static int io_init_wq_offload(struct io_ring_ctx *ctx,
6805 struct io_uring_params *p)
6807 struct io_wq_data data;
6809 struct io_ring_ctx *ctx_attach;
6810 unsigned int concurrency;
6813 data.user = ctx->user;
6814 data.free_work = io_free_work;
6816 if (!(p->flags & IORING_SETUP_ATTACH_WQ)) {
6817 /* Do QD, or 4 * CPUS, whatever is smallest */
6818 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
6820 ctx->io_wq = io_wq_create(concurrency, &data);
6821 if (IS_ERR(ctx->io_wq)) {
6822 ret = PTR_ERR(ctx->io_wq);
6828 f = fdget(p->wq_fd);
6832 if (f.file->f_op != &io_uring_fops) {
6837 ctx_attach = f.file->private_data;
6838 /* @io_wq is protected by holding the fd */
6839 if (!io_wq_get(ctx_attach->io_wq, &data)) {
6844 ctx->io_wq = ctx_attach->io_wq;
6850 static int io_sq_offload_start(struct io_ring_ctx *ctx,
6851 struct io_uring_params *p)
6855 init_waitqueue_head(&ctx->sqo_wait);
6856 mmgrab(current->mm);
6857 ctx->sqo_mm = current->mm;
6859 if (ctx->flags & IORING_SETUP_SQPOLL) {
6861 if (!capable(CAP_SYS_ADMIN))
6864 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
6865 if (!ctx->sq_thread_idle)
6866 ctx->sq_thread_idle = HZ;
6868 if (p->flags & IORING_SETUP_SQ_AFF) {
6869 int cpu = p->sq_thread_cpu;
6872 if (cpu >= nr_cpu_ids)
6874 if (!cpu_online(cpu))
6877 ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread,
6881 ctx->sqo_thread = kthread_create(io_sq_thread, ctx,
6884 if (IS_ERR(ctx->sqo_thread)) {
6885 ret = PTR_ERR(ctx->sqo_thread);
6886 ctx->sqo_thread = NULL;
6889 wake_up_process(ctx->sqo_thread);
6890 } else if (p->flags & IORING_SETUP_SQ_AFF) {
6891 /* Can't have SQ_AFF without SQPOLL */
6896 ret = io_init_wq_offload(ctx, p);
6902 io_finish_async(ctx);
6903 mmdrop(ctx->sqo_mm);
6908 static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages)
6910 atomic_long_sub(nr_pages, &user->locked_vm);
6913 static int io_account_mem(struct user_struct *user, unsigned long nr_pages)
6915 unsigned long page_limit, cur_pages, new_pages;
6917 /* Don't allow more pages than we can safely lock */
6918 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
6921 cur_pages = atomic_long_read(&user->locked_vm);
6922 new_pages = cur_pages + nr_pages;
6923 if (new_pages > page_limit)
6925 } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages,
6926 new_pages) != cur_pages);
6931 static void io_mem_free(void *ptr)
6938 page = virt_to_head_page(ptr);
6939 if (put_page_testzero(page))
6940 free_compound_page(page);
6943 static void *io_mem_alloc(size_t size)
6945 gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
6948 return (void *) __get_free_pages(gfp_flags, get_order(size));
6951 static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries,
6954 struct io_rings *rings;
6955 size_t off, sq_array_size;
6957 off = struct_size(rings, cqes, cq_entries);
6958 if (off == SIZE_MAX)
6962 off = ALIGN(off, SMP_CACHE_BYTES);
6967 sq_array_size = array_size(sizeof(u32), sq_entries);
6968 if (sq_array_size == SIZE_MAX)
6971 if (check_add_overflow(off, sq_array_size, &off))
6980 static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries)
6984 pages = (size_t)1 << get_order(
6985 rings_size(sq_entries, cq_entries, NULL));
6986 pages += (size_t)1 << get_order(
6987 array_size(sizeof(struct io_uring_sqe), sq_entries));
6992 static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx)
6996 if (!ctx->user_bufs)
6999 for (i = 0; i < ctx->nr_user_bufs; i++) {
7000 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7002 for (j = 0; j < imu->nr_bvecs; j++)
7003 unpin_user_page(imu->bvec[j].bv_page);
7005 if (ctx->account_mem)
7006 io_unaccount_mem(ctx->user, imu->nr_bvecs);
7011 kfree(ctx->user_bufs);
7012 ctx->user_bufs = NULL;
7013 ctx->nr_user_bufs = 0;
7017 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst,
7018 void __user *arg, unsigned index)
7020 struct iovec __user *src;
7022 #ifdef CONFIG_COMPAT
7024 struct compat_iovec __user *ciovs;
7025 struct compat_iovec ciov;
7027 ciovs = (struct compat_iovec __user *) arg;
7028 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov)))
7031 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base);
7032 dst->iov_len = ciov.iov_len;
7036 src = (struct iovec __user *) arg;
7037 if (copy_from_user(dst, &src[index], sizeof(*dst)))
7042 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg,
7045 struct vm_area_struct **vmas = NULL;
7046 struct page **pages = NULL;
7047 int i, j, got_pages = 0;
7052 if (!nr_args || nr_args > UIO_MAXIOV)
7055 ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf),
7057 if (!ctx->user_bufs)
7060 for (i = 0; i < nr_args; i++) {
7061 struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
7062 unsigned long off, start, end, ubuf;
7067 ret = io_copy_iov(ctx, &iov, arg, i);
7072 * Don't impose further limits on the size and buffer
7073 * constraints here, we'll -EINVAL later when IO is
7074 * submitted if they are wrong.
7077 if (!iov.iov_base || !iov.iov_len)
7080 /* arbitrary limit, but we need something */
7081 if (iov.iov_len > SZ_1G)
7084 ubuf = (unsigned long) iov.iov_base;
7085 end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
7086 start = ubuf >> PAGE_SHIFT;
7087 nr_pages = end - start;
7089 if (ctx->account_mem) {
7090 ret = io_account_mem(ctx->user, nr_pages);
7096 if (!pages || nr_pages > got_pages) {
7099 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
7101 vmas = kvmalloc_array(nr_pages,
7102 sizeof(struct vm_area_struct *),
7104 if (!pages || !vmas) {
7106 if (ctx->account_mem)
7107 io_unaccount_mem(ctx->user, nr_pages);
7110 got_pages = nr_pages;
7113 imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec),
7117 if (ctx->account_mem)
7118 io_unaccount_mem(ctx->user, nr_pages);
7123 down_read(¤t->mm->mmap_sem);
7124 pret = pin_user_pages(ubuf, nr_pages,
7125 FOLL_WRITE | FOLL_LONGTERM,
7127 if (pret == nr_pages) {
7128 /* don't support file backed memory */
7129 for (j = 0; j < nr_pages; j++) {
7130 struct vm_area_struct *vma = vmas[j];
7133 !is_file_hugepages(vma->vm_file)) {
7139 ret = pret < 0 ? pret : -EFAULT;
7141 up_read(¤t->mm->mmap_sem);
7144 * if we did partial map, or found file backed vmas,
7145 * release any pages we did get
7148 unpin_user_pages(pages, pret);
7149 if (ctx->account_mem)
7150 io_unaccount_mem(ctx->user, nr_pages);
7155 off = ubuf & ~PAGE_MASK;
7157 for (j = 0; j < nr_pages; j++) {
7160 vec_len = min_t(size_t, size, PAGE_SIZE - off);
7161 imu->bvec[j].bv_page = pages[j];
7162 imu->bvec[j].bv_len = vec_len;
7163 imu->bvec[j].bv_offset = off;
7167 /* store original address for later verification */
7169 imu->len = iov.iov_len;
7170 imu->nr_bvecs = nr_pages;
7172 ctx->nr_user_bufs++;
7180 io_sqe_buffer_unregister(ctx);
7184 static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg)
7186 __s32 __user *fds = arg;
7192 if (copy_from_user(&fd, fds, sizeof(*fds)))
7195 ctx->cq_ev_fd = eventfd_ctx_fdget(fd);
7196 if (IS_ERR(ctx->cq_ev_fd)) {
7197 int ret = PTR_ERR(ctx->cq_ev_fd);
7198 ctx->cq_ev_fd = NULL;
7205 static int io_eventfd_unregister(struct io_ring_ctx *ctx)
7207 if (ctx->cq_ev_fd) {
7208 eventfd_ctx_put(ctx->cq_ev_fd);
7209 ctx->cq_ev_fd = NULL;
7216 static int __io_destroy_buffers(int id, void *p, void *data)
7218 struct io_ring_ctx *ctx = data;
7219 struct io_buffer *buf = p;
7221 __io_remove_buffers(ctx, buf, id, -1U);
7225 static void io_destroy_buffers(struct io_ring_ctx *ctx)
7227 idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx);
7228 idr_destroy(&ctx->io_buffer_idr);
7231 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
7233 io_finish_async(ctx);
7235 mmdrop(ctx->sqo_mm);
7237 io_iopoll_reap_events(ctx);
7238 io_sqe_buffer_unregister(ctx);
7239 io_sqe_files_unregister(ctx);
7240 io_eventfd_unregister(ctx);
7241 io_destroy_buffers(ctx);
7242 idr_destroy(&ctx->personality_idr);
7244 #if defined(CONFIG_UNIX)
7245 if (ctx->ring_sock) {
7246 ctx->ring_sock->file = NULL; /* so that iput() is called */
7247 sock_release(ctx->ring_sock);
7251 io_mem_free(ctx->rings);
7252 io_mem_free(ctx->sq_sqes);
7254 percpu_ref_exit(&ctx->refs);
7255 if (ctx->account_mem)
7256 io_unaccount_mem(ctx->user,
7257 ring_pages(ctx->sq_entries, ctx->cq_entries));
7258 free_uid(ctx->user);
7259 put_cred(ctx->creds);
7260 kfree(ctx->completions);
7261 kfree(ctx->cancel_hash);
7262 kmem_cache_free(req_cachep, ctx->fallback_req);
7266 static __poll_t io_uring_poll(struct file *file, poll_table *wait)
7268 struct io_ring_ctx *ctx = file->private_data;
7271 poll_wait(file, &ctx->cq_wait, wait);
7273 * synchronizes with barrier from wq_has_sleeper call in
7277 if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head !=
7278 ctx->rings->sq_ring_entries)
7279 mask |= EPOLLOUT | EPOLLWRNORM;
7280 if (io_cqring_events(ctx, false))
7281 mask |= EPOLLIN | EPOLLRDNORM;
7286 static int io_uring_fasync(int fd, struct file *file, int on)
7288 struct io_ring_ctx *ctx = file->private_data;
7290 return fasync_helper(fd, file, on, &ctx->cq_fasync);
7293 static int io_remove_personalities(int id, void *p, void *data)
7295 struct io_ring_ctx *ctx = data;
7296 const struct cred *cred;
7298 cred = idr_remove(&ctx->personality_idr, id);
7304 static void io_ring_exit_work(struct work_struct *work)
7306 struct io_ring_ctx *ctx;
7308 ctx = container_of(work, struct io_ring_ctx, exit_work);
7310 io_cqring_overflow_flush(ctx, true);
7312 wait_for_completion(&ctx->completions[0]);
7313 io_ring_ctx_free(ctx);
7316 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
7318 mutex_lock(&ctx->uring_lock);
7319 percpu_ref_kill(&ctx->refs);
7320 mutex_unlock(&ctx->uring_lock);
7323 * Wait for sq thread to idle, if we have one. It won't spin on new
7324 * work after we've killed the ctx ref above. This is important to do
7325 * before we cancel existing commands, as the thread could otherwise
7326 * be queueing new work post that. If that's work we need to cancel,
7327 * it could cause shutdown to hang.
7329 while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
7332 io_kill_timeouts(ctx);
7333 io_poll_remove_all(ctx);
7336 io_wq_cancel_all(ctx->io_wq);
7338 io_iopoll_reap_events(ctx);
7339 /* if we failed setting up the ctx, we might not have any rings */
7341 io_cqring_overflow_flush(ctx, true);
7342 idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx);
7343 INIT_WORK(&ctx->exit_work, io_ring_exit_work);
7344 queue_work(system_wq, &ctx->exit_work);
7347 static int io_uring_release(struct inode *inode, struct file *file)
7349 struct io_ring_ctx *ctx = file->private_data;
7351 file->private_data = NULL;
7352 io_ring_ctx_wait_and_kill(ctx);
7356 static void io_uring_cancel_files(struct io_ring_ctx *ctx,
7357 struct files_struct *files)
7359 struct io_kiocb *req;
7362 while (!list_empty_careful(&ctx->inflight_list)) {
7363 struct io_kiocb *cancel_req = NULL;
7365 spin_lock_irq(&ctx->inflight_lock);
7366 list_for_each_entry(req, &ctx->inflight_list, inflight_entry) {
7367 if (req->work.files != files)
7369 /* req is being completed, ignore */
7370 if (!refcount_inc_not_zero(&req->refs))
7376 prepare_to_wait(&ctx->inflight_wait, &wait,
7377 TASK_UNINTERRUPTIBLE);
7378 spin_unlock_irq(&ctx->inflight_lock);
7380 /* We need to keep going until we don't find a matching req */
7384 if (cancel_req->flags & REQ_F_OVERFLOW) {
7385 spin_lock_irq(&ctx->completion_lock);
7386 list_del(&cancel_req->list);
7387 cancel_req->flags &= ~REQ_F_OVERFLOW;
7388 if (list_empty(&ctx->cq_overflow_list)) {
7389 clear_bit(0, &ctx->sq_check_overflow);
7390 clear_bit(0, &ctx->cq_check_overflow);
7392 spin_unlock_irq(&ctx->completion_lock);
7394 WRITE_ONCE(ctx->rings->cq_overflow,
7395 atomic_inc_return(&ctx->cached_cq_overflow));
7398 * Put inflight ref and overflow ref. If that's
7399 * all we had, then we're done with this request.
7401 if (refcount_sub_and_test(2, &cancel_req->refs)) {
7402 io_put_req(cancel_req);
7407 io_wq_cancel_work(ctx->io_wq, &cancel_req->work);
7408 io_put_req(cancel_req);
7411 finish_wait(&ctx->inflight_wait, &wait);
7414 static int io_uring_flush(struct file *file, void *data)
7416 struct io_ring_ctx *ctx = file->private_data;
7418 io_uring_cancel_files(ctx, data);
7421 * If the task is going away, cancel work it may have pending
7423 if (fatal_signal_pending(current) || (current->flags & PF_EXITING))
7424 io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current));
7429 static void *io_uring_validate_mmap_request(struct file *file,
7430 loff_t pgoff, size_t sz)
7432 struct io_ring_ctx *ctx = file->private_data;
7433 loff_t offset = pgoff << PAGE_SHIFT;
7438 case IORING_OFF_SQ_RING:
7439 case IORING_OFF_CQ_RING:
7442 case IORING_OFF_SQES:
7446 return ERR_PTR(-EINVAL);
7449 page = virt_to_head_page(ptr);
7450 if (sz > page_size(page))
7451 return ERR_PTR(-EINVAL);
7458 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7460 size_t sz = vma->vm_end - vma->vm_start;
7464 ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz);
7466 return PTR_ERR(ptr);
7468 pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
7469 return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
7472 #else /* !CONFIG_MMU */
7474 static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
7476 return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL;
7479 static unsigned int io_uring_nommu_mmap_capabilities(struct file *file)
7481 return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE;
7484 static unsigned long io_uring_nommu_get_unmapped_area(struct file *file,
7485 unsigned long addr, unsigned long len,
7486 unsigned long pgoff, unsigned long flags)
7490 ptr = io_uring_validate_mmap_request(file, pgoff, len);
7492 return PTR_ERR(ptr);
7494 return (unsigned long) ptr;
7497 #endif /* !CONFIG_MMU */
7499 SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit,
7500 u32, min_complete, u32, flags, const sigset_t __user *, sig,
7503 struct io_ring_ctx *ctx;
7508 if (current->task_works)
7511 if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP))
7519 if (f.file->f_op != &io_uring_fops)
7523 ctx = f.file->private_data;
7524 if (!percpu_ref_tryget(&ctx->refs))
7528 * For SQ polling, the thread will do all submissions and completions.
7529 * Just return the requested submit count, and wake the thread if
7533 if (ctx->flags & IORING_SETUP_SQPOLL) {
7534 if (!list_empty_careful(&ctx->cq_overflow_list))
7535 io_cqring_overflow_flush(ctx, false);
7536 if (flags & IORING_ENTER_SQ_WAKEUP)
7537 wake_up(&ctx->sqo_wait);
7538 submitted = to_submit;
7539 } else if (to_submit) {
7540 mutex_lock(&ctx->uring_lock);
7541 submitted = io_submit_sqes(ctx, to_submit, f.file, fd, false);
7542 mutex_unlock(&ctx->uring_lock);
7544 if (submitted != to_submit)
7547 if (flags & IORING_ENTER_GETEVENTS) {
7548 unsigned nr_events = 0;
7550 min_complete = min(min_complete, ctx->cq_entries);
7553 * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user
7554 * space applications don't need to do io completion events
7555 * polling again, they can rely on io_sq_thread to do polling
7556 * work, which can reduce cpu usage and uring_lock contention.
7558 if (ctx->flags & IORING_SETUP_IOPOLL &&
7559 !(ctx->flags & IORING_SETUP_SQPOLL)) {
7560 ret = io_iopoll_check(ctx, &nr_events, min_complete);
7562 ret = io_cqring_wait(ctx, min_complete, sig, sigsz);
7567 percpu_ref_put(&ctx->refs);
7570 return submitted ? submitted : ret;
7573 #ifdef CONFIG_PROC_FS
7574 static int io_uring_show_cred(int id, void *p, void *data)
7576 const struct cred *cred = p;
7577 struct seq_file *m = data;
7578 struct user_namespace *uns = seq_user_ns(m);
7579 struct group_info *gi;
7584 seq_printf(m, "%5d\n", id);
7585 seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid));
7586 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid));
7587 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid));
7588 seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid));
7589 seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid));
7590 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid));
7591 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid));
7592 seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid));
7593 seq_puts(m, "\n\tGroups:\t");
7594 gi = cred->group_info;
7595 for (g = 0; g < gi->ngroups; g++) {
7596 seq_put_decimal_ull(m, g ? " " : "",
7597 from_kgid_munged(uns, gi->gid[g]));
7599 seq_puts(m, "\n\tCapEff:\t");
7600 cap = cred->cap_effective;
7601 CAP_FOR_EACH_U32(__capi)
7602 seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8);
7607 static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m)
7611 mutex_lock(&ctx->uring_lock);
7612 seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files);
7613 for (i = 0; i < ctx->nr_user_files; i++) {
7614 struct fixed_file_table *table;
7617 table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT];
7618 f = table->files[i & IORING_FILE_TABLE_MASK];
7620 seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname);
7622 seq_printf(m, "%5u: <none>\n", i);
7624 seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs);
7625 for (i = 0; i < ctx->nr_user_bufs; i++) {
7626 struct io_mapped_ubuf *buf = &ctx->user_bufs[i];
7628 seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf,
7629 (unsigned int) buf->len);
7631 if (!idr_is_empty(&ctx->personality_idr)) {
7632 seq_printf(m, "Personalities:\n");
7633 idr_for_each(&ctx->personality_idr, io_uring_show_cred, m);
7635 seq_printf(m, "PollList:\n");
7636 spin_lock_irq(&ctx->completion_lock);
7637 for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) {
7638 struct hlist_head *list = &ctx->cancel_hash[i];
7639 struct io_kiocb *req;
7641 hlist_for_each_entry(req, list, hash_node)
7642 seq_printf(m, " op=%d, task_works=%d\n", req->opcode,
7643 req->task->task_works != NULL);
7645 spin_unlock_irq(&ctx->completion_lock);
7646 mutex_unlock(&ctx->uring_lock);
7649 static void io_uring_show_fdinfo(struct seq_file *m, struct file *f)
7651 struct io_ring_ctx *ctx = f->private_data;
7653 if (percpu_ref_tryget(&ctx->refs)) {
7654 __io_uring_show_fdinfo(ctx, m);
7655 percpu_ref_put(&ctx->refs);
7660 static const struct file_operations io_uring_fops = {
7661 .release = io_uring_release,
7662 .flush = io_uring_flush,
7663 .mmap = io_uring_mmap,
7665 .get_unmapped_area = io_uring_nommu_get_unmapped_area,
7666 .mmap_capabilities = io_uring_nommu_mmap_capabilities,
7668 .poll = io_uring_poll,
7669 .fasync = io_uring_fasync,
7670 #ifdef CONFIG_PROC_FS
7671 .show_fdinfo = io_uring_show_fdinfo,
7675 static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
7676 struct io_uring_params *p)
7678 struct io_rings *rings;
7679 size_t size, sq_array_offset;
7681 size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset);
7682 if (size == SIZE_MAX)
7685 rings = io_mem_alloc(size);
7690 ctx->sq_array = (u32 *)((char *)rings + sq_array_offset);
7691 rings->sq_ring_mask = p->sq_entries - 1;
7692 rings->cq_ring_mask = p->cq_entries - 1;
7693 rings->sq_ring_entries = p->sq_entries;
7694 rings->cq_ring_entries = p->cq_entries;
7695 ctx->sq_mask = rings->sq_ring_mask;
7696 ctx->cq_mask = rings->cq_ring_mask;
7697 ctx->sq_entries = rings->sq_ring_entries;
7698 ctx->cq_entries = rings->cq_ring_entries;
7700 size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
7701 if (size == SIZE_MAX) {
7702 io_mem_free(ctx->rings);
7707 ctx->sq_sqes = io_mem_alloc(size);
7708 if (!ctx->sq_sqes) {
7709 io_mem_free(ctx->rings);
7718 * Allocate an anonymous fd, this is what constitutes the application
7719 * visible backing of an io_uring instance. The application mmaps this
7720 * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled,
7721 * we have to tie this fd to a socket for file garbage collection purposes.
7723 static int io_uring_get_fd(struct io_ring_ctx *ctx)
7728 #if defined(CONFIG_UNIX)
7729 ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP,
7735 ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC);
7739 file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx,
7740 O_RDWR | O_CLOEXEC);
7743 ret = PTR_ERR(file);
7747 #if defined(CONFIG_UNIX)
7748 ctx->ring_sock->file = file;
7750 fd_install(ret, file);
7753 #if defined(CONFIG_UNIX)
7754 sock_release(ctx->ring_sock);
7755 ctx->ring_sock = NULL;
7760 static int io_uring_create(unsigned entries, struct io_uring_params *p)
7762 struct user_struct *user = NULL;
7763 struct io_ring_ctx *ctx;
7769 if (entries > IORING_MAX_ENTRIES) {
7770 if (!(p->flags & IORING_SETUP_CLAMP))
7772 entries = IORING_MAX_ENTRIES;
7776 * Use twice as many entries for the CQ ring. It's possible for the
7777 * application to drive a higher depth than the size of the SQ ring,
7778 * since the sqes are only used at submission time. This allows for
7779 * some flexibility in overcommitting a bit. If the application has
7780 * set IORING_SETUP_CQSIZE, it will have passed in the desired number
7781 * of CQ ring entries manually.
7783 p->sq_entries = roundup_pow_of_two(entries);
7784 if (p->flags & IORING_SETUP_CQSIZE) {
7786 * If IORING_SETUP_CQSIZE is set, we do the same roundup
7787 * to a power-of-two, if it isn't already. We do NOT impose
7788 * any cq vs sq ring sizing.
7790 if (p->cq_entries < p->sq_entries)
7792 if (p->cq_entries > IORING_MAX_CQ_ENTRIES) {
7793 if (!(p->flags & IORING_SETUP_CLAMP))
7795 p->cq_entries = IORING_MAX_CQ_ENTRIES;
7797 p->cq_entries = roundup_pow_of_two(p->cq_entries);
7799 p->cq_entries = 2 * p->sq_entries;
7802 user = get_uid(current_user());
7803 account_mem = !capable(CAP_IPC_LOCK);
7806 ret = io_account_mem(user,
7807 ring_pages(p->sq_entries, p->cq_entries));
7814 ctx = io_ring_ctx_alloc(p);
7817 io_unaccount_mem(user, ring_pages(p->sq_entries,
7822 ctx->compat = in_compat_syscall();
7823 ctx->account_mem = account_mem;
7825 ctx->creds = get_current_cred();
7827 ret = io_allocate_scq_urings(ctx, p);
7831 ret = io_sq_offload_start(ctx, p);
7835 memset(&p->sq_off, 0, sizeof(p->sq_off));
7836 p->sq_off.head = offsetof(struct io_rings, sq.head);
7837 p->sq_off.tail = offsetof(struct io_rings, sq.tail);
7838 p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask);
7839 p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries);
7840 p->sq_off.flags = offsetof(struct io_rings, sq_flags);
7841 p->sq_off.dropped = offsetof(struct io_rings, sq_dropped);
7842 p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings;
7844 memset(&p->cq_off, 0, sizeof(p->cq_off));
7845 p->cq_off.head = offsetof(struct io_rings, cq.head);
7846 p->cq_off.tail = offsetof(struct io_rings, cq.tail);
7847 p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask);
7848 p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries);
7849 p->cq_off.overflow = offsetof(struct io_rings, cq_overflow);
7850 p->cq_off.cqes = offsetof(struct io_rings, cqes);
7853 * Install ring fd as the very last thing, so we don't risk someone
7854 * having closed it before we finish setup
7856 ret = io_uring_get_fd(ctx);
7860 p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP |
7861 IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS |
7862 IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL;
7863 trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags);
7866 io_ring_ctx_wait_and_kill(ctx);
7871 * Sets up an aio uring context, and returns the fd. Applications asks for a
7872 * ring size, we return the actual sq/cq ring sizes (among other things) in the
7873 * params structure passed in.
7875 static long io_uring_setup(u32 entries, struct io_uring_params __user *params)
7877 struct io_uring_params p;
7881 if (copy_from_user(&p, params, sizeof(p)))
7883 for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
7888 if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL |
7889 IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE |
7890 IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ))
7893 ret = io_uring_create(entries, &p);
7897 if (copy_to_user(params, &p, sizeof(p)))
7903 SYSCALL_DEFINE2(io_uring_setup, u32, entries,
7904 struct io_uring_params __user *, params)
7906 return io_uring_setup(entries, params);
7909 static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args)
7911 struct io_uring_probe *p;
7915 size = struct_size(p, ops, nr_args);
7916 if (size == SIZE_MAX)
7918 p = kzalloc(size, GFP_KERNEL);
7923 if (copy_from_user(p, arg, size))
7926 if (memchr_inv(p, 0, size))
7929 p->last_op = IORING_OP_LAST - 1;
7930 if (nr_args > IORING_OP_LAST)
7931 nr_args = IORING_OP_LAST;
7933 for (i = 0; i < nr_args; i++) {
7935 if (!io_op_defs[i].not_supported)
7936 p->ops[i].flags = IO_URING_OP_SUPPORTED;
7941 if (copy_to_user(arg, p, size))
7948 static int io_register_personality(struct io_ring_ctx *ctx)
7950 const struct cred *creds = get_current_cred();
7953 id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1,
7954 USHRT_MAX, GFP_KERNEL);
7960 static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id)
7962 const struct cred *old_creds;
7964 old_creds = idr_remove(&ctx->personality_idr, id);
7966 put_cred(old_creds);
7973 static bool io_register_op_must_quiesce(int op)
7976 case IORING_UNREGISTER_FILES:
7977 case IORING_REGISTER_FILES_UPDATE:
7978 case IORING_REGISTER_PROBE:
7979 case IORING_REGISTER_PERSONALITY:
7980 case IORING_UNREGISTER_PERSONALITY:
7987 static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode,
7988 void __user *arg, unsigned nr_args)
7989 __releases(ctx->uring_lock)
7990 __acquires(ctx->uring_lock)
7995 * We're inside the ring mutex, if the ref is already dying, then
7996 * someone else killed the ctx or is already going through
7997 * io_uring_register().
7999 if (percpu_ref_is_dying(&ctx->refs))
8002 if (io_register_op_must_quiesce(opcode)) {
8003 percpu_ref_kill(&ctx->refs);
8006 * Drop uring mutex before waiting for references to exit. If
8007 * another thread is currently inside io_uring_enter() it might
8008 * need to grab the uring_lock to make progress. If we hold it
8009 * here across the drain wait, then we can deadlock. It's safe
8010 * to drop the mutex here, since no new references will come in
8011 * after we've killed the percpu ref.
8013 mutex_unlock(&ctx->uring_lock);
8014 ret = wait_for_completion_interruptible(&ctx->completions[0]);
8015 mutex_lock(&ctx->uring_lock);
8017 percpu_ref_resurrect(&ctx->refs);
8024 case IORING_REGISTER_BUFFERS:
8025 ret = io_sqe_buffer_register(ctx, arg, nr_args);
8027 case IORING_UNREGISTER_BUFFERS:
8031 ret = io_sqe_buffer_unregister(ctx);
8033 case IORING_REGISTER_FILES:
8034 ret = io_sqe_files_register(ctx, arg, nr_args);
8036 case IORING_UNREGISTER_FILES:
8040 ret = io_sqe_files_unregister(ctx);
8042 case IORING_REGISTER_FILES_UPDATE:
8043 ret = io_sqe_files_update(ctx, arg, nr_args);
8045 case IORING_REGISTER_EVENTFD:
8046 case IORING_REGISTER_EVENTFD_ASYNC:
8050 ret = io_eventfd_register(ctx, arg);
8053 if (opcode == IORING_REGISTER_EVENTFD_ASYNC)
8054 ctx->eventfd_async = 1;
8056 ctx->eventfd_async = 0;
8058 case IORING_UNREGISTER_EVENTFD:
8062 ret = io_eventfd_unregister(ctx);
8064 case IORING_REGISTER_PROBE:
8066 if (!arg || nr_args > 256)
8068 ret = io_probe(ctx, arg, nr_args);
8070 case IORING_REGISTER_PERSONALITY:
8074 ret = io_register_personality(ctx);
8076 case IORING_UNREGISTER_PERSONALITY:
8080 ret = io_unregister_personality(ctx, nr_args);
8087 if (io_register_op_must_quiesce(opcode)) {
8088 /* bring the ctx back to life */
8089 percpu_ref_reinit(&ctx->refs);
8091 reinit_completion(&ctx->completions[0]);
8096 SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode,
8097 void __user *, arg, unsigned int, nr_args)
8099 struct io_ring_ctx *ctx;
8108 if (f.file->f_op != &io_uring_fops)
8111 ctx = f.file->private_data;
8113 mutex_lock(&ctx->uring_lock);
8114 ret = __io_uring_register(ctx, opcode, arg, nr_args);
8115 mutex_unlock(&ctx->uring_lock);
8116 trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs,
8117 ctx->cq_ev_fd != NULL, ret);
8123 static int __init io_uring_init(void)
8125 #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \
8126 BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \
8127 BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \
8130 #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \
8131 __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename)
8132 BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64);
8133 BUILD_BUG_SQE_ELEM(0, __u8, opcode);
8134 BUILD_BUG_SQE_ELEM(1, __u8, flags);
8135 BUILD_BUG_SQE_ELEM(2, __u16, ioprio);
8136 BUILD_BUG_SQE_ELEM(4, __s32, fd);
8137 BUILD_BUG_SQE_ELEM(8, __u64, off);
8138 BUILD_BUG_SQE_ELEM(8, __u64, addr2);
8139 BUILD_BUG_SQE_ELEM(16, __u64, addr);
8140 BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in);
8141 BUILD_BUG_SQE_ELEM(24, __u32, len);
8142 BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags);
8143 BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags);
8144 BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags);
8145 BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags);
8146 BUILD_BUG_SQE_ELEM(28, __u16, poll_events);
8147 BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags);
8148 BUILD_BUG_SQE_ELEM(28, __u32, msg_flags);
8149 BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags);
8150 BUILD_BUG_SQE_ELEM(28, __u32, accept_flags);
8151 BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags);
8152 BUILD_BUG_SQE_ELEM(28, __u32, open_flags);
8153 BUILD_BUG_SQE_ELEM(28, __u32, statx_flags);
8154 BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice);
8155 BUILD_BUG_SQE_ELEM(28, __u32, splice_flags);
8156 BUILD_BUG_SQE_ELEM(32, __u64, user_data);
8157 BUILD_BUG_SQE_ELEM(40, __u16, buf_index);
8158 BUILD_BUG_SQE_ELEM(42, __u16, personality);
8159 BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in);
8161 BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST);
8162 BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int));
8163 req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
8166 __initcall(io_uring_init);